<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:itunes="http://www.itunes.com/dtds/podcast-1.0.dtd" xmlns:googleplay="http://www.google.com/schemas/play-podcasts/1.0"><channel><title><![CDATA[Verbatim]]></title><description><![CDATA[A one-person quant lab building custom neural networks to trade short-term stock moves, in public. Every experiment posted here before it runs. Machine-graded, published unedited, losses included. Publisher of The Graded Tape. https://gradedtape.com]]></description><link>https://letters.verbatimsigma.com</link><image><url>https://substackcdn.com/image/fetch/$s_!8yF1!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fab70c646-1706-46b2-9028-bfdbea86a2a7_1024x1024.png</url><title>Verbatim</title><link>https://letters.verbatimsigma.com</link></image><generator>Substack</generator><lastBuildDate>Sat, 21 Nov 2026 03:54:08 GMT</lastBuildDate><atom:link href="https://letters.verbatimsigma.com/feed" rel="self" type="application/rss+xml"/><copyright><![CDATA[Verbatim]]></copyright><language><![CDATA[en]]></language><webMaster><![CDATA[verbatimsigma@substack.com]]></webMaster><itunes:owner><itunes:email><![CDATA[verbatimsigma@substack.com]]></itunes:email><itunes:name><![CDATA[Verbatim]]></itunes:name></itunes:owner><itunes:author><![CDATA[Verbatim]]></itunes:author><googleplay:owner><![CDATA[verbatimsigma@substack.com]]></googleplay:owner><googleplay:email><![CDATA[verbatimsigma@substack.com]]></googleplay:email><googleplay:author><![CDATA[Verbatim]]></googleplay:author><itunes:block><![CDATA[Yes]]></itunes:block><item><title><![CDATA[I built a model to say when to trust my trading model]]></title><description><![CDATA[A small model built to say when the big one could be trusted, and a FAIL I published.]]></description><link>https://letters.verbatimsigma.com/p/i-built-a-model-to-say-when-to-trust</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/i-built-a-model-to-say-when-to-trust</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Thu, 05 Nov 2026 14:02:34 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!hxIC!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faab21ee2-9e50-4045-8230-1c3b77fee1c9_2600x1520.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Written in July 2026. Anything that has changed since then is marked with an <strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>.</em></p><p>I liked the idea in this story, and it failed its test. Nobody required me to publish the result.</p><h2>The idea</h2><p><strong><span data-color="#3c78d8" style="color: rgb(60, 120, 216);">Lookout</span></strong>, the main model in my algorithmic trading program, does one job. It ranks stocks by the probability of an imminent large move. Like any model, it has good days and bad days, and it occurred to me that the good and bad days might themselves be predictable. So the idea was a second, much smaller model whose only task was to grade how far the first model could be trusted, one day at a time, sorting each day into one of three classes (LOW, MID, or HIGH) by its confidence that the big model&#8217;s rankings would hold up that day.</p><p>The appeal was practical. If the grades were real, position sizes could follow them, larger on the days the layer graded HIGH and smaller on the days it graded LOW. </p><h4><span data-color="#9184d9" style="color: rgb(145, 132, 217);">A model that knows when to distrust itself would be valuable</span></h4><h2>The rule</h2><p>Following the program&#8217;s standing practice, I pre-registered the trust layer before any code was written. The design doc stated the hypothesis, and it stated an adoption rule. For the layer to be used in trading, the grades had to mean what they claimed on out-of-sample days. Days stamped HIGH had to be better days for the big model, and days stamped LOW had to be worse ones. A grade that works in only one direction is worse than no grade at all, because sizing decisions would rely on it in both directions.</p><p>The rule was quantitative. It set four gates, every one required, on at least 200 out-of-sample predictions. A prediction stamped LOW had to be followed by a below-median result for the big model at least 60 percent of the time. The LOW class&#8217;s average had to sit at least 0.02 below the HIGH class&#8217;s. Both had to hold again in a leak control that graded one window further ahead, so no test day was shared. And weighting by trust had to beat not weighting at all. The evaluation then ran the same way everything in this program runs, on walk-forward windows the layer had never seen. I wrote down sizes in advance because a prediction that only states a direction can count a tiny effect as a win.</p><p>I signed the adoption rule while I was still impartial, before LOW and HIGH were results I had feelings about.</p><h2>The verdict</h2><p>The layer was built, trained, and evaluated on walk-forward windows like everything else in the program.</p><p>The HIGH class worked, somewhat. Days the layer flagged as trustworthy tended to be better days for the big model, and 75.5 percent of those predictions landed above the model&#8217;s trailing median, against about 61 percent of predictions in general. It was a small effect, pointing the right way. Had the experiment ended there, I would have been writing a very different and considerably smugger post.</p><p>The LOW class failed. On the measure the rule used, the effect pointed the wrong way. Days the layer flagged as low-trust landed below the big model&#8217;s trailing median only 31.2 percent of the time, less often than an ordinary prediction did (about 39 percent), when the rule needed at least 60 percent. Their average score did come in below the other classes, by enough to clear the 0.02 gap gate, but on a class of only sixteen predictions, which settles very little. So the layer could partly recognize the big model&#8217;s good days and was unreliable about the bad ones, which were the half that risk management needed.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!hxIC!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faab21ee2-9e50-4045-8230-1c3b77fee1c9_2600x1520.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!hxIC!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faab21ee2-9e50-4045-8230-1c3b77fee1c9_2600x1520.png 424w, https://substackcdn.com/image/fetch/$s_!hxIC!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faab21ee2-9e50-4045-8230-1c3b77fee1c9_2600x1520.png 848w, https://substackcdn.com/image/fetch/$s_!hxIC!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faab21ee2-9e50-4045-8230-1c3b77fee1c9_2600x1520.png 1272w, https://substackcdn.com/image/fetch/$s_!hxIC!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faab21ee2-9e50-4045-8230-1c3b77fee1c9_2600x1520.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!hxIC!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faab21ee2-9e50-4045-8230-1c3b77fee1c9_2600x1520.png" width="1456" height="851" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/aab21ee2-9e50-4045-8230-1c3b77fee1c9_2600x1520.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:851,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:219858,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221691791?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faab21ee2-9e50-4045-8230-1c3b77fee1c9_2600x1520.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!hxIC!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faab21ee2-9e50-4045-8230-1c3b77fee1c9_2600x1520.png 424w, https://substackcdn.com/image/fetch/$s_!hxIC!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faab21ee2-9e50-4045-8230-1c3b77fee1c9_2600x1520.png 848w, https://substackcdn.com/image/fetch/$s_!hxIC!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faab21ee2-9e50-4045-8230-1c3b77fee1c9_2600x1520.png 1272w, https://substackcdn.com/image/fetch/$s_!hxIC!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faab21ee2-9e50-4045-8230-1c3b77fee1c9_2600x1520.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>The pre-registered rule required every gate, and the LOW gate failed both here and in the leak control, so the verdict was FAIL, and the ledger records it that way. I shelved the layer before it was wired into the nightly run, so no position size ever depended on its grades.</p><p><em><strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>: its code was archived in an October cleanup of closed studies, and its FAIL stays on the record.</em></p><p>I like this result now. The test asked whether <strong><span data-color="#3c78d8" style="color: rgb(60, 120, 216);">Lookout</span></strong>&#8217;s bad days could be predicted, and the answer was that its good days could be spotted and its bad days could not. That is something I know about the system I trade, and I would not know it if the test had been designed to be easy to pass.</p><p>Then I was tempted to adopt just the HIGH half, since it works. The rule I signed asked about the whole layer, and rewriting a rule after seeing the results is what the rule exists to prevent. If the HIGH class alone is worth having, that is a new hypothesis, and it can have its own design doc, its own bar, and its own entry in the trial ledger like every other idea.</p><h2>Why publish this</h2><p>Most public track records are filtered, and you usually cannot see what was filtered out. A notebook that records only wins is essentially just advertising, because the reader has no way to know how many attempts produced the wins on display. The multiple-comparisons arithmetic from the last letter applies to publications the same way it applies to backtests, because the problem is the count of attempts left out.</p><p>Publishing this failure shows you part of that count. If a success appears in this notebook, you will be able to judge it against entries like this one, where the test said no and I published it anyway.</p><p>This notebook publishes its FAIL verdicts. Subscribe for the whole ledger, failed attempts included.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><p>Writing each loss down keeps the tests hard enough to fail, which is the only reason a pass means anything. When the live diary starts, with real money behind its percentages, every win it reports will sit next to entries like this one. The failed trust layer showed me where my model&#8217;s knowledge of its own reliability stops.</p><p><em><strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>: the live diary has not started; apart from a few days of hand trading at half size in August, nothing has traded real money. The public record today is <a href="https://gradedtape.com/">The Graded Tape</a>, at fixed paper sizes. Through Oct 21, 2026, 63 trading days and 11,888 signals graded, it reads -$1,348.45.</em></p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included.</em></p><p>Paid subscribers get the build notes below, covering six minutes and forty seconds from rule to verdict, and the rule I forgot to write.</p><h2>How I built it</h2><p>The trust layer&#8217;s design doc went into git at 2:04 p.m. on July 12, 2026, with 174 lines, the four gates, and a line saying the first version stays display-only whatever the results. The code, 12 tests and the FAIL verdict went in six minutes and forty seconds later. The agent wrote about 700 lines in plain NumPy and wrote the report card; no GPU was involved.</p>
      <p>
          <a href="https://letters.verbatimsigma.com/p/i-built-a-model-to-say-when-to-trust">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[Pre-registration for traders]]></title><description><![CDATA[The one-page design doc that decides a verdict before the results exist, field by field, free to copy.]]></description><link>https://letters.verbatimsigma.com/p/pre-registration-for-traders</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/pre-registration-for-traders</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Tue, 03 Nov 2026 14:02:45 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!7cNE!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c330cb7-df3a-4d84-933d-b2afacbb7e0a_1800x3320.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Written in July 2026. Anything that has changed since then is marked with an <strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>.</em></p><p>Every verdict in this notebook, the wins and the losses, was decided by a document written before the results existed. I call these documents <span data-color="#a64d79" style="color: rgb(166, 77, 121);">pre-registrations</span>, a practice from clinical trials and from psychology after its replication crisis. The figure below is a real design doc from my program, with the proprietary details removed. The structure is the useful part, and you are free to copy it.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!7cNE!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c330cb7-df3a-4d84-933d-b2afacbb7e0a_1800x3320.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!7cNE!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c330cb7-df3a-4d84-933d-b2afacbb7e0a_1800x3320.png 424w, https://substackcdn.com/image/fetch/$s_!7cNE!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c330cb7-df3a-4d84-933d-b2afacbb7e0a_1800x3320.png 848w, https://substackcdn.com/image/fetch/$s_!7cNE!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c330cb7-df3a-4d84-933d-b2afacbb7e0a_1800x3320.png 1272w, https://substackcdn.com/image/fetch/$s_!7cNE!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c330cb7-df3a-4d84-933d-b2afacbb7e0a_1800x3320.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!7cNE!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c330cb7-df3a-4d84-933d-b2afacbb7e0a_1800x3320.png" width="1456" height="2686" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/4c330cb7-df3a-4d84-933d-b2afacbb7e0a_1800x3320.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:2686,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:416309,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221691005?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c330cb7-df3a-4d84-933d-b2afacbb7e0a_1800x3320.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!7cNE!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c330cb7-df3a-4d84-933d-b2afacbb7e0a_1800x3320.png 424w, https://substackcdn.com/image/fetch/$s_!7cNE!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c330cb7-df3a-4d84-933d-b2afacbb7e0a_1800x3320.png 848w, https://substackcdn.com/image/fetch/$s_!7cNE!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c330cb7-df3a-4d84-933d-b2afacbb7e0a_1800x3320.png 1272w, https://substackcdn.com/image/fetch/$s_!7cNE!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c330cb7-df3a-4d84-933d-b2afacbb7e0a_1800x3320.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p></p><h2>Prove me wrong, bro</h2><p>A design doc opens with the hypothesis, and it has to be a single sentence that can be proven wrong. For example, &#8220;this specific change, evaluated on these specific windows, will beat the current model by at least this much.&#8221; If I cannot state an idea in a form that a number could contradict, it is still a hope, and no result can disprove it. I drop plenty of ideas while writing that sentence, hours before any compute gets rented.</p><h2>The bar, chosen while you are still impartial</h2><p>Next comes the <strong>success bar</strong>, and it gets chosen before any results exist. In my program the standing bar is that a challenger must beat the incumbent by more than +0.02 on the evaluation metric, measured across matched walk-forward windows, with a bootstrap confidence interval clear of zero. The margin exists because a new idea adds complexity, and a challenger that ties the incumbent while being more complex is the worse system. The confidence interval exists because without it, luck alone can clear any bar I set.</p><p>The interval comes from a paired bootstrap. Both models are graded on the identical walk-forward windows, so each window yields one number, the challenger&#8217;s score minus the incumbent&#8217;s on that window. Pairing this way removes the effect of the window itself. Whatever market conditions that test period contained, both models faced them, so they cancel in the difference. The bootstrap then resamples that list of per-window differences with replacement a thousand times, computing the mean difference each time, and the middle 95 percent of those thousand means becomes the interval. So the challenger must beat the incumbent across a thousand reshuffles of the same test history before the verdict counts. The resampler&#8217;s random seed is fixed, so anyone replaying the experiment, including a suspicious future me, gets the identical interval to the last digit.</p><p>I choose the bar early because once results arrive, I am no longer impartial. I start arguing for whichever outcome I was hoping to see. Choosing the bar first means I set it before I know which side I&#8217;m on.</p><h2>The cost</h2><p>Every doc states what the experiment will cost to run, in dollars of rented compute. fWith the cost set in advance, a failed experiment is money I already agreed to spend for the answer. It also forces me to rank which questions to ask first, because the budget is finite.</p><h2>Ties go to the incumbent</h2><p>This line is standing policy and never negotiated per experiment. When a challenger and the incumbent are statistically indistinguishable, the incumbent wins. New ideas carry the burden of proof, and the older, simpler model stays until something measurably better shows up. This rule has settled more of my close calls than any other clause in the template.</p><h2>The trial ledger</h2><p>Every experiment, the moment its design doc is signed, adds one to a count in a ledger, and that count appears in this notebook. Nine experiments in arc 1 means nine ledger entries, including the seven or so I would have preferred to forget.</p><p>The ledger exists because of the multiple-comparisons trap, the way backtests mislead even careful people. Run twenty independent tests of worthless ideas at a conventional significance threshold and you should expect roughly one to pass by pure chance. If you then report only the one that passed, you have reported a false discovery without changing a single number. The problem is the count of attempts you left out, the denominator. The ledger keeps that count where I cannot leave it out, so any win I report can be read against the full number of attempts behind it.</p><p>The trial ledger is public, so every win I report comes with its count of attempts. Subscribe and hold me to it.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><h2>The sweep rule</h2><p>This is the rule I would have broken first. When an experiment fails, the temptation is to adjust a parameter and run it again, and again, until something clears the bar. So the template carries an explicit rule that a parameter sweep following a failure is a new experiment, requiring its own design doc, its own pre-registered bar, and its own entry in the trial ledger. Retries are allowed, and every one gets recorded, because twenty hidden retries carry the same odds as twenty hidden experiments.</p><p>The template fits on one page and needs no mathematics beyond what your backtesting already uses. The hard part is following rules you wrote before you saw the results, which is harder on results day than it sounds in July. Every idea in this program now goes through the same page. The next letter is about a results day when the document said no.</p><p><em><strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>: for the second model, the one that trades once a day and is in research now, the document comes at the end. Its exploratory studies are logged and counted as they run, and any configuration that becomes a candidate gets one sealed, pre-registered test on data it has never seen, with the count of configurations searched to find it written into the document.</em></p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included.</em></p><p>Paid subscribers get the build notes below, covering where the template came from and the plain text file that counts every try.</p><h2>How I built it</h2><p>The template grew out of the first few design docs. A typical one from July is 61 lines under four headings (the hypothesis, the fixed inputs, the design, and what kills it). Its header carries the price (about $45 to $55 for a 10 to 12 hour walk) and a status line that read AWAITING SIGN-OFF until I signed it. The agent drafts, I sign, then it runs. The sweep rule started as a single parenthetical in a July 9 doc.</p>
      <p>
          <a href="https://letters.verbatimsigma.com/p/pre-registration-for-traders">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[Weekly note 2: four questions to ask any backtest]]></title><description><![CDATA[What I check before I trust a backtest, mine included. Paid subscriptions now start with a 14-day free trial.]]></description><link>https://letters.verbatimsigma.com/p/weekly-note-2-four-questions-to-ask</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/weekly-note-2-four-questions-to-ask</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Fri, 30 Oct 2026 13:01:59 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!k-Y7!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F402aa17d-5119-402d-a8cb-1006c4573dfe_2600x2000.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Friday, October 30, 2026.</em></p><p>This week&#8217;s letters covered walk-forward testing and the three months of data I keep off-limits to myself. Both exist because backtests are easy to fool. These are the four questions I now ask of any backtest before I believe it, my own included.</p><ol><li><p><strong>Was every number graded on days the strategy never saw?</strong> If the rules were tuned on the same days they were scored on, the score measures the tuning. Walk-forward testing (Tuesday&#8217;s letter) is the fix.</p></li><li><p><strong>How many versions were tried before this one?</strong> Test twenty worthless ideas and about one will pass by chance. A result means little without the count of attempts behind it. In July I tested nine ideas and one produced a clean win, and I print the nine.</p></li><li><p><strong>What does it cost to trade?</strong> In September I measured my own paper bot. A round trip cost about 142 basis points of the money traded, against an edge of about 16 per signal. A backtest that ignores costs can show an edge that trading never delivers.</p></li><li><p><strong>What does something simple score on the same days?</strong> If a plain baseline gets most of the way there, the complicated strategy has to earn its complexity. A plain logistic regression got 96 percent of the way to my deep model&#8217;s score.</p></li></ol><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!k-Y7!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F402aa17d-5119-402d-a8cb-1006c4573dfe_2600x2000.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!k-Y7!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F402aa17d-5119-402d-a8cb-1006c4573dfe_2600x2000.png 424w, https://substackcdn.com/image/fetch/$s_!k-Y7!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F402aa17d-5119-402d-a8cb-1006c4573dfe_2600x2000.png 848w, https://substackcdn.com/image/fetch/$s_!k-Y7!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F402aa17d-5119-402d-a8cb-1006c4573dfe_2600x2000.png 1272w, https://substackcdn.com/image/fetch/$s_!k-Y7!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F402aa17d-5119-402d-a8cb-1006c4573dfe_2600x2000.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!k-Y7!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F402aa17d-5119-402d-a8cb-1006c4573dfe_2600x2000.png" width="1456" height="1120" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/402aa17d-5119-402d-a8cb-1006c4573dfe_2600x2000.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1120,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:162626,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221689917?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F402aa17d-5119-402d-a8cb-1006c4573dfe_2600x2000.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!k-Y7!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F402aa17d-5119-402d-a8cb-1006c4573dfe_2600x2000.png 424w, https://substackcdn.com/image/fetch/$s_!k-Y7!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F402aa17d-5119-402d-a8cb-1006c4573dfe_2600x2000.png 848w, https://substackcdn.com/image/fetch/$s_!k-Y7!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F402aa17d-5119-402d-a8cb-1006c4573dfe_2600x2000.png 1272w, https://substackcdn.com/image/fetch/$s_!k-Y7!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F402aa17d-5119-402d-a8cb-1006c4573dfe_2600x2000.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p></p><p>None of these needs special software. They need the discipline to ask them before you see a result you like.</p><h2>The Tape</h2><p>The Graded Tape grades every signal my model fires, every morning, wins and misses, at <a href="https://gradedtape.com/">gradedtape.com</a>. Its numbers change nightly, so I link to it here instead of printing a figure that would be out of date by the time you read this.</p><h2>From the archive</h2><p>Tuesday&#8217;s letter was &#8220;Walk-forward testing, explained for people who lost money after a good backtest&#8221;, and Thursday&#8217;s was &#8220;Why I froze three months of data I&#8217;m not allowed to look at&#8221;. The September research log is free to read in full and shows these questions applied to a month of real tests.</p><p>The note goes out free every Friday.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><p>Starting today, a paid subscription comes with a 14-day free trial. Paid subscribers get the build notes at the end of each letter, the working papers that close each arc and a monthly research log with each study&#8217;s full tables.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?coupon=28b03538&amp;utm_content=221689917&quot;,&quot;text&quot;:&quot;Get 14 day free trial&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?coupon=28b03538&amp;utm_content=221689917"><span>Get 14 day free trial</span></a></p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included.</em></p>]]></content:encoded></item><item><title><![CDATA[Why I froze three months of data I'm not allowed to look at]]></title><description><![CDATA[The winner's curse, a holdout used exactly once, and the Sunday evening the sealed months were finally read.]]></description><link>https://letters.verbatimsigma.com/p/why-i-froze-three-months-of-data</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/why-i-froze-three-months-of-data</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Thu, 29 Oct 2026 13:01:49 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!QiMf!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ac6dec9-d5c8-4df2-8f17-520ecb88f2a1_2600x1800.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Written in July 2026. Anything that has changed since then is marked with an <strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>.</em></p><p>Last letter I said that even after you adopt walk-forward evaluation and every grade is earned on unseen days, one problem remains. I ran the program with it for a long time before I noticed it.</p><h2>The winner&#8217;s curse</h2><p>Suppose you test nine ideas, each one graded correctly on out-of-sample windows. Each measured score is the idea&#8217;s true quality plus some noise, because finite test windows always carry noise, and the noise can go in either direction. Then you do the natural thing and pick the highest scorer.</p><p>By picking the maximum of nine noisy measurements, you have selected partly for quality and partly for luck. The winner is disproportionately likely to be an idea whose noise happened to go in its favor, which means the winner&#8217;s measured score is biased upward even though every individual measurement was correct. The statistics here are old. The expected value of the maximum of several noisy draws sits above the expected value of any single draw, and the gap grows with both the noise in each measurement and the number of ideas compared. Test more ideas or shorten the test windows and you get more of this false edge, without a single wrong number anywhere. Auction theorists named it the winner&#8217;s curse, after the finding that the company that wins the bidding for an oilfield of uncertain value is usually the one that most overestimated it. The same holds for the idea that wins a comparison.</p><p>Every test in this example was clean. The bias comes in at the moment of choosing, which feels like nothing more than reading the top line of a results table. The arc 1 campaign chose its winner from nine ideas, so its winner&#8217;s measured edge is, by construction, a little optimistic, and walk-forward testing cannot fix that.</p><h2>The holdout</h2><p>The fix is old and a little painful. Keep some data out of every selection decision, and use it exactly once. Clinical trials have worked this way for decades. My part was to sit down one evening, design the trading-program version, write the protocol into the ledger, and then follow it. I enjoyed the design work. The hard part is writing a rule strict enough that I can&#8217;t argue my way around it later.</p><p>In my program this took the form of a fixed date. I chose April 2026 as the boundary, and everything after it was excluded from every adjudication decision the program makes, down to the tie-breaks. Design docs may not cite the frozen months, a &#8220;quick sanity look&#8221; is not permitted, and the trial ledger records that nothing has read them for the purpose of choosing anything.</p><p><strong>Lookout</strong>, the intraday ensemble this notebook is mostly about, keeps walking forward through the frozen period along with its challengers. The daily walk-forward process continues to train, test, and write grades into the ledgers for those months, so the pipelines keep running, the data stays current, and the eventual reading will reflect the system operating normally. The scores exist on disk. The rule is that no decision is allowed to depend on them, so the freeze is mostly a rule about my own behavior. Knowing the numbers are sitting there, computed and unread, is hard to ignore. Those files hold the answer to whether the whole campaign was real, and I had scheduled the reading for two days later.</p><h2>One reading, no second attempt</h2><p>The frozen period gets read at exactly one scheduled event, a final confirmation. The finished, chosen package (the incumbent model plus every adopted change from the campaign, exactly as it would trade) is graded on the frozen months once, with no second attempt.</p><p>If a disappointing first reading could be followed by a tweak and a retest, the frozen data would start being used for selection, and the winner&#8217;s curse would be back. So I fixed the terms in advance. One reading, of one pre-specified package, and the result publishes either way. If the frozen months come back consistent with the walk-forward results, that is meaningful evidence the campaign&#8217;s verdicts were real, because this data played no part in producing them. If they come back badly, then I have learned that my selection process manufactured some of its apparent edge, which is painful and much cheaper to learn here than in the market. Either way the program ends up with a confirmed package or a known flaw, and I found I was looking forward to the reading.</p><p>This holdout&#8217;s verdict is a few paragraphs down. Subscribe to see the next one when it is read.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><p>The freeze cost three months of the newest, most relevant data, held out of every decision during the stretch when I was making the most decisions. I set the date on the morning of July 10, before any of the round&#8217;s verdicts, and read the sealed months on the evening of July</p><ol><li><p>Two days sounds short for a holdout, but those two days held every</p></li></ol><p>decision of the campaign (the nine verdicts, the tie-break and the ensemble&#8217;s adoption), all made without the newest three months. The same freeze procedure now applies to every later campaign.</p><h2>The reading</h2><p>I read it on a Sunday evening in July. I was nervous, because three months of held-out data could easily have shown the campaign was wrong. The results held.</p><p>Each of the four Lookout variants scored <em>higher</em> on the frozen quarter than it had over the sixty windows before the boundary, with every confidence interval sitting entirely above zero. The deltas were +0.033, +0.025, +0.035, and +0.027. The frozen months were evidently an easier market for the model, so I didn&#8217;t read much into the levels. The more useful number is the ensemble&#8217;s premium over its best single member, which came out to +0.0288 on the frozen data. The value adopted before the freeze, measured on data the selection process had already used, was +0.0287. That frozen figure comes from only the five frozen windows with saved checkpoints, and it has no interval, so the match to the fourth decimal is partly luck. It didn&#8217;t shrink. For scale, a premium of that size works out to roughly three extra correct calls in every hundred of the model&#8217;s most confident picks.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!QiMf!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ac6dec9-d5c8-4df2-8f17-520ecb88f2a1_2600x1800.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!QiMf!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ac6dec9-d5c8-4df2-8f17-520ecb88f2a1_2600x1800.png 424w, https://substackcdn.com/image/fetch/$s_!QiMf!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ac6dec9-d5c8-4df2-8f17-520ecb88f2a1_2600x1800.png 848w, https://substackcdn.com/image/fetch/$s_!QiMf!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ac6dec9-d5c8-4df2-8f17-520ecb88f2a1_2600x1800.png 1272w, https://substackcdn.com/image/fetch/$s_!QiMf!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ac6dec9-d5c8-4df2-8f17-520ecb88f2a1_2600x1800.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!QiMf!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ac6dec9-d5c8-4df2-8f17-520ecb88f2a1_2600x1800.png" width="1456" height="1008" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/4ac6dec9-d5c8-4df2-8f17-520ecb88f2a1_2600x1800.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1008,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:265329,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221665662?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ac6dec9-d5c8-4df2-8f17-520ecb88f2a1_2600x1800.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!QiMf!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ac6dec9-d5c8-4df2-8f17-520ecb88f2a1_2600x1800.png 424w, https://substackcdn.com/image/fetch/$s_!QiMf!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ac6dec9-d5c8-4df2-8f17-520ecb88f2a1_2600x1800.png 848w, https://substackcdn.com/image/fetch/$s_!QiMf!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ac6dec9-d5c8-4df2-8f17-520ecb88f2a1_2600x1800.png 1272w, https://substackcdn.com/image/fetch/$s_!QiMf!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ac6dec9-d5c8-4df2-8f17-520ecb88f2a1_2600x1800.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p></p><p>This holdout is now used up, since it only works once, and a new one starts accumulating for whatever package needs judging next. The model starts the paper diary tomorrow with that offline evidence behind it.</p><p><em><strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>: the offline evidence held, but the model has not made money. The paper record that model started is <a href="https://gradedtape.com/">The Graded Tape</a>, which grades its signals every morning at fixed paper sizes; through Oct 21, 2026, 63 trading days and 11,888 signals graded, it shows a 34.0% hit rate and -$1,348.45.</em></p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included.</em></p><p>Paid subscribers get the build notes below, covering the ten lines that set the date, the script that read it once, and the one detail that didn&#8217;t line up.</p><h2>How I built it</h2><p>The freeze is a few lines of code and one paragraph. On the morning of Friday, July 10, 2026, I set the date at April 10, and the agent committed a ten-line paragraph to the design doc saying every window tested on or after that date is out of every decision, for every variant. The adjudication script carries the date as a constant and drops any window whose test days touch it. It ran at 10:48 that night.</p>
      <p>
          <a href="https://letters.verbatimsigma.com/p/why-i-froze-three-months-of-data">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[Walk-forward testing, explained for people who lost money after a good backtest]]></title><description><![CDATA[How a model gets graded only on days it has never seen, what that costs, and the one failure it still cannot catch.]]></description><link>https://letters.verbatimsigma.com/p/walk-forward-testing-explained-for</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/walk-forward-testing-explained-for</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Tue, 27 Oct 2026 13:02:40 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!_iq5!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffee9cf78-946c-4c53-83c2-ee78b3a8edbb_2600x1600.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Written in July 2026. Anything that has changed since then is marked with an <strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>.</em></p><p>If you have ever had a strategy look great in a backtest and then lose money in its very first live month, so have I. It took me a long time to see why. The backtest had answered a different question from the one I was asking.</p><p>New letters arrive by email, free. Subscribe and the next one comes to you.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><h2>What an ordinary backtest measures</h2><p>An ordinary backtest measures how well a strategy fits the history I used while building it. The question I care about is how the strategy will do on days that have not happened yet. The two stop matching as soon as you start tuning. Every time you adjust a parameter, look at the resulting equity curve, and keep the change when the curve improves, you have used the test days to build the strategy. Do that a few dozen times, which takes one slow afternoon with a laptop, and the strategy has been fitted to the exact days it is graded on. The final backtest score then measures how well you fitted those days and says nothing about how the strategy will do next.</p><p>No single adjustment feels like cheating, because each one seems like careful work.</p><h2>Sliding the window</h2><p>Walk-forward evaluation is the standard fix. You pick a cutoff date and train the model only on data from before it. The tuning happens inside that same boundary. A short validation period is taken from the end of the training span, the most recent data the model is allowed to see, and every decision that shapes the model (when to stop training, which configuration to keep) uses that period and nothing newer. Most homemade test setups let future data in at this step, and a model tuned with a look at the future has seen part of the test in advance.</p><p>Then the model is graded on the days immediately after the cutoff, days it has never seen in any form. The grade is the same strict statistic from arc 1, precision among the model&#8217;s most confident names with false alarms held to a one percent budget, computed on those unseen days alone. The grade is written to the ledger, and then the whole process moves forward one day and repeats. It trains on a slightly newer past, tunes on a slightly newer validation period, is graded on a slightly newer future, and records the grade. Because the step is a single day and each window spans much longer, consecutive windows overlap heavily, by design. The same model family is retested under gradually changing conditions, and because the step is one day, no shift between market regimes can fall between two tests.</p><p>In my program this procedure produces roughly 869 overlapping windows covering January 2023 through the middle of 2026, and every grade in that walk was earned by <strong>Lookout</strong>, the intraday ensemble this notebook is mostly about, on days it had never seen. I looked for any way the model could see its test days before being graded and did not find one. The overlap means neighboring grades are correlated, because they share most of their training data, so the program never treats the 869 as independent results. Comparisons between models happen pairwise on matched windows, which the pre-registration letter covers.</p><p>Three and a half years covers several kinds of market, including the choppy, directionless stretches of 2023, the steadier market of 2024, and the rougher conditions of 2025 and early 2026. A model that only fits one kind of market shows it, because the windows run through all of them and the ledger keeps each period&#8217;s grades where you can see them. Watching those 869 grades come in, each one earned on days the model never saw, is one of the parts of this project I enjoy most.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!_iq5!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffee9cf78-946c-4c53-83c2-ee78b3a8edbb_2600x1600.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!_iq5!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffee9cf78-946c-4c53-83c2-ee78b3a8edbb_2600x1600.png 424w, https://substackcdn.com/image/fetch/$s_!_iq5!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffee9cf78-946c-4c53-83c2-ee78b3a8edbb_2600x1600.png 848w, https://substackcdn.com/image/fetch/$s_!_iq5!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffee9cf78-946c-4c53-83c2-ee78b3a8edbb_2600x1600.png 1272w, https://substackcdn.com/image/fetch/$s_!_iq5!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffee9cf78-946c-4c53-83c2-ee78b3a8edbb_2600x1600.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!_iq5!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffee9cf78-946c-4c53-83c2-ee78b3a8edbb_2600x1600.png" width="1456" height="896" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/fee9cf78-946c-4c53-83c2-ee78b3a8edbb_2600x1600.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:896,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:195494,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221663200?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffee9cf78-946c-4c53-83c2-ee78b3a8edbb_2600x1600.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!_iq5!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffee9cf78-946c-4c53-83c2-ee78b3a8edbb_2600x1600.png 424w, https://substackcdn.com/image/fetch/$s_!_iq5!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffee9cf78-946c-4c53-83c2-ee78b3a8edbb_2600x1600.png 848w, https://substackcdn.com/image/fetch/$s_!_iq5!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffee9cf78-946c-4c53-83c2-ee78b3a8edbb_2600x1600.png 1272w, https://substackcdn.com/image/fetch/$s_!_iq5!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffee9cf78-946c-4c53-83c2-ee78b3a8edbb_2600x1600.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h2>What walk-forward costs</h2><p>The first cost is compute. Instead of training a model once and grading it once, you are training and grading it hundreds of times, one fresh fit per window. Every experiment costs that many times more, and a large share of the $423 I spent on the arc 1 campaign went to this procedure. When people ask why I rent data-center GPUs to evaluate ideas for a small account, this is most of the reason. The harness, at least, was a one-time cost. It was written once, reworked at length, and every future idea now runs through it, so each new hypothesis pays for its own compute but not for the engineering.</p><p>The second cost is that the score comes out lower. A conventional backtest, tuned on the days it grades, reports flattering numbers; walk-forward reports what the model earns on unseen days, which is almost always lower. The lower number is the one to trust, and the gap between the two shows how much the ordinary backtest overstated the result. When a change to my testing procedure makes the score go up, I check the change before I believe the score.</p><p>I don&#8217;t mind the lower number, because a score earned this way is one I can rely on. When Lookout&#8217;s grade held up across 869 windows it never saw in advance, through three different kinds of market, I could treat it as a property of the system and build plans on it.</p><p>Subscribe to get each new grade as it comes in, good or bad.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><h2>What it still cannot fix</h2><p>Walk-forward stops a model from being graded on the days it was tuned on. It does nothing about the person running the experiments, who tests many ideas, watches the grades come in, and then picks the luckiest one. That problem is called the winner&#8217;s curse. Clean evaluation does not remove it, so it needs a separate check. The next letter covers that check, which keeps three months of my own data off-limits until one scheduled reading.</p><p><em><strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>: a September audit of the data pipeline found three small look-ahead leaks in the model&#8217;s feature inputs; removing them moved the grade by at most 0.0015, so the grades stand, but my claim above that I found no way for the model to see its test days was too strong. The plans built on those grades have not made money, and an October test of trading the signals through the day lost money in its main configuration even before costs. One step of that walk also skipped fifteen sessions of June 2025 that the data had silently lost; the last letter of this arc covers it.</em></p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included.</em></p><p>Paid subscribers get the build notes below, covering the 800-line runner, the one-line fix that kept it running, and what a full walk costs.</p><h2>How I built it</h2><p>I built the harness in one day, July 8, 2026, with Claude Code doing the typing. Before writing anything new, the agent reran my old training as a control, and my earlier results didn&#8217;t reproduce. Those bugs got fixed first.</p><p>The new runner is about 800 lines plus 50 tests, and it replaced about 2,100 lines of older scripts. Each window writes one line to a log and flushes it to disk before the next window starts, so a crash costs one window. Every window also grades four dumb baselines beside the model, including a plain logistic regression.</p><p>On the first rented H100 run, every job died after about 27 windows with &#8220;too many open files&#8221;. The fix was one line in the launch command. A full walk takes about 26 hours, roughly $40 to $60 per variant.</p><p>Try this: ask your assistant to rerun your best old backtest from a clean checkout before it builds anything new.</p>]]></content:encoded></item><item><title><![CDATA[Weekly note 1: a positive week on the Tape, and two ideas that only won in hindsight]]></title><description><![CDATA[Three sessions up on the Tape at paper sizes before costs, and two older ideas that beat the index only when chosen with hindsight.]]></description><link>https://letters.verbatimsigma.com/p/weekly-note-1-a-positive-week-on</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/weekly-note-1-a-positive-week-on</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Fri, 23 Oct 2026 13:03:33 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!VIXi!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F576960f0-3b59-46ea-be49-0f244942c399_2600x1440.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Friday, October 23, 2026. Tape data through Wednesday, Oct 21 (gradedtape.com, posted Oct 22 at 07:00 EDT); lab notes through Thursday, Oct 22.</em></p><p>This is the first weekly note on the Tape and the lab, sent while my historical archive continues to roll out the story so far.</p><h2>The Tape this week</h2><p>This week&#8217;s first three sessions, Oct 19-21, gave the Tape 572 graded signals, of which 205 reached their +5% target inside the two-hour window, a hit rate of 35.8%. At fixed paper sizes the three sessions gained $528.68, and all three were positive. The record since July 14 stands at 63 trading days, 11,888 signals graded, a 34.0% hit rate and -$1,348.45, data through Oct 21, 2026, on <a href="https://gradedtape.com/">gradedtape.com</a>.</p><p>The hit rate ran a little above the record&#8217;s average, and this week the hits paid more than the misses cost. The Tape grades the signal at paper sizes before trading costs. The September research log measured those costs at a round trip several times the edge per flag, so a positive paper week is not money. The program&#8217;s bar, beating an index fund after costs, is tested in the lab.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!VIXi!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F576960f0-3b59-46ea-be49-0f244942c399_2600x1440.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!VIXi!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F576960f0-3b59-46ea-be49-0f244942c399_2600x1440.png 424w, https://substackcdn.com/image/fetch/$s_!VIXi!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F576960f0-3b59-46ea-be49-0f244942c399_2600x1440.png 848w, https://substackcdn.com/image/fetch/$s_!VIXi!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F576960f0-3b59-46ea-be49-0f244942c399_2600x1440.png 1272w, https://substackcdn.com/image/fetch/$s_!VIXi!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F576960f0-3b59-46ea-be49-0f244942c399_2600x1440.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!VIXi!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F576960f0-3b59-46ea-be49-0f244942c399_2600x1440.png" width="1456" height="806" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/576960f0-3b59-46ea-be49-0f244942c399_2600x1440.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:806,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:215575,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221660827?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F576960f0-3b59-46ea-be49-0f244942c399_2600x1440.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!VIXi!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F576960f0-3b59-46ea-be49-0f244942c399_2600x1440.png 424w, https://substackcdn.com/image/fetch/$s_!VIXi!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F576960f0-3b59-46ea-be49-0f244942c399_2600x1440.png 848w, https://substackcdn.com/image/fetch/$s_!VIXi!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F576960f0-3b59-46ea-be49-0f244942c399_2600x1440.png 1272w, https://substackcdn.com/image/fetch/$s_!VIXi!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F576960f0-3b59-46ea-be49-0f244942c399_2600x1440.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h2>In the lab</h2><p><strong><span data-color="#6aa84f" style="color: rgb(106, 168, 79);">Swell</span></strong> is the daily model meant to beat an index fund after costs. Two older ideas were replayed first, each quarter&#8217;s version chosen only from what was known before it, from January 2023 into October 2026.</p><ul><li><p><strong>A rule book that follows the stock purchases members of Congress disclose. </strong>At three of the four account sizes I test it made less than the index after costs, and at all four it had deeper drops. <strong>Verdict: did not clear.</strong></p></li><li><p><strong>An overnight sleeve beside the index,</strong> buying <strong><span data-color="#3c78d8" style="color: rgb(60, 120, 216);">Lookout</span></strong>&#8217;s top-ranked names at the close and selling them at the next open. It did no better than the same rule on random picks. <strong>Verdict: lost to the index at every account size tested.</strong></p></li></ul><p>Then every neural network got a standing test. Each one retrains each morning on what it knew the day before and trades the next session, walked forward from January 2023, with all of 2026 sealed until a design is frozen.</p><ul><li><p><strong>A stock-picking network.</strong> Its rankings had no measurable relation to what the stocks then did (an average daily correlation of -0.001, 90 percent interval -0.018 to +0.015), and none of its five independently trained copies came out ahead at any account size. Adding <strong><span data-color="#3c78d8" style="color: rgb(60, 120, 216);">Lookout</span></strong>&#8217;s scores as inputs changed nothing; a linear model and a tree model did no better. <strong>Verdict: lost to the index at every account size tested.</strong></p></li><li><p><strong>A market-state network,</strong> which picks no stocks and decides each day how much of the account sits in the index fund and how much in Treasury bills, reading the index&#8217;s recent path and how much big moves are triggering one another (the subject of arc three). It was built and read on Sunday. <strong>Verdict: lost to the index at every account size tested</strong> (a Sharpe of 1.08 against the index&#8217;s 1.49, deeper drops, and none of its five copies ahead).</p></li></ul><p>This week had no new verdicts.</p><ul><li><p><strong>A transformer</strong> that reads each stock&#8217;s recent trading days alongside the whole day&#8217;s market was declared Sunday evening, built and piloted. Its full walk needs a rented GPU for a few hours and has not run yet.</p></li><li><p><strong>A learning-curve read</strong> asked whether more history makes the stock-picking network better. Copies frozen at four dates in 2023 to 2025 scored no differently from the live, daily-updated version on the same later periods; every interval included zero.</p></li><li><p><strong>The paper-only overnight sleeve</strong> has now shadowed 10 sessions and is behind on both of its tracking lines, so its buys stay on hold.</p></li></ul><h2>One verdict</h2><p>The congress book and the overnight sleeve failed in the same way. When I looked back over the whole stretch and picked the best-scoring version of each, both came out ahead of the index after costs at every account size I test. That reading is in-sample, because the version was chosen by someone who already knew how every one of those days turned out, which no trader knows on the morning of a trade. Most of the congress version&#8217;s gain, and at the smallest size more than all of it, came from 2025 alone.</p><p>The fair form of the question replays the calendar. At the start of each quarter, choose using only the data before it, trade the quarter, pay for any switch, and repeat. Asked that way, neither idea cleared the bar at any account size.</p><p>Tuesday&#8217;s letter, &#8220;Walk-forward testing, explained for people who lost money after a good backtest,&#8221; is the long version of that replayed calendar, told about the first model in July.</p><p>From next week the letters go out on Tuesdays and Thursdays, and this note goes out free every Friday, the losing verdicts included.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><p>Paid subscribers also get the working papers that close each arc and a monthly research log with each study&#8217;s full tables. The first one, <a href="https://letters.verbatimsigma.com/p/the-research-log-september-2026">September&#8217;s</a>, is free to read in full.</p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included.</em></p>]]></content:encoded></item><item><title><![CDATA[The research log, September 2026]]></title><description><![CDATA[The allocator failed its full-history test, Lookout's paper losses traced to the fills, and a faster data feed would not have helped. This first issue is free in full.]]></description><link>https://letters.verbatimsigma.com/p/the-research-log-september-2026</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/the-research-log-september-2026</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Thu, 22 Oct 2026 16:20:26 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!hx79!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fff9b171c-55e0-4615-b406-bbdf0ae5d8fa_2600x1520.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Research log, September 2026. It covers every test from 1 to 30 September. Nothing ran from 1 to 8 September, because the nightly data chain was down (incidents, at the end).</em></p><p>This log lists every test I ran in September and how each one came out. This first issue of the Research Log is free in full, so you can see what a paid subscription gets each month. From the October issue on, the verdicts stay free and the designs, tables and intervals behind them are for paid subscribers.</p><p>September was the month the program stopped asking whether <strong><span data-color="#3c78d8" style="color: rgb(60, 120, 216);">Lookout</span></strong> can rank stocks (it can) and started asking where the money goes when it trades them. The answer was the fills. The same month, Ballast, the daily allocator, failed the test that decided its future.</p><ol><li><p><strong>An audit of my August claims.</strong> Of eight claims re-derived from the raw evidence, one held as worded. The numbers reproduced in seven of eight; the wording and the paper trail did not.</p></li><li><p><strong>A risk control for the earlier daily network.</strong> It failed its volatility rule, passed after two fixes, and paid for both with a lower Sharpe, mostly by holding more cash.</p></li><li><p><strong>Corrupted training days.</strong> 38 of 427 stored days of feature data had been damaged by a nightly automation job. Retraining on the repaired days moved Lookout&#8217;s precision by less than the bar I set in advance, so the served networks stayed. (Note, I found it interesting that you don&#8217;t need flawless data for precision. There&#8217;s some resilience to my model)</p></li><li><p><strong>Ballast&#8217;s full-history test.</strong> FAIL. Over 780 trading days its Sharpe after costs was 0.16, against 0.95 for an equal-weight basket and 1.26 for an S&amp;P 500 index fund. It held 42 percent of the account in cash on an average day.</p></li><li><p><strong>Where the paper bot&#8217;s money went.</strong> The leak is the price paid to get in. A round trip cost about 142 basis points of the money traded, against an edge of about 16 basis points per graded flag.</p></li><li><p><strong>Price bands on the Tape.</strong> Through 25 September the Tape was positive only on stocks under $1, and the paper bot lost money in every band. You can see my daily progress at <a href="http://www.gradedtape.com">The Graded Tape</a>, which is always open and free.</p></li><li><p><strong>The re-aim of <span data-color="#3c78d8" style="color: rgb(60, 120, 216);">Lookout</span>.</strong> Adopted 27 September: retrain Lookout on a cost-aware target and test one trading protocol over the full history. </p></li><li><p><strong>Would a real-time data feed help?</strong> No. None of 16 tested combinations made money, and a faster feed made the result slightly worse. Limit orders beat market orders by about 74 basis points per flag.</p></li><li><p><strong>Three look-ahead leaks.</strong> Confirmed in the data, then knocked out of the live model one at a time. None of them moved the ensemble&#8217;s precision by more than 0.0015. Again, good demonstration of model resilience.</p></li><li><p><strong>An outside review of my plan.</strong> Five Opus 5.5 reviewers scored my 4.8 out of 10: strong on methodology, weak on any route to money.</p></li></ol><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!hx79!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fff9b171c-55e0-4615-b406-bbdf0ae5d8fa_2600x1520.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!hx79!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fff9b171c-55e0-4615-b406-bbdf0ae5d8fa_2600x1520.png 424w, https://substackcdn.com/image/fetch/$s_!hx79!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fff9b171c-55e0-4615-b406-bbdf0ae5d8fa_2600x1520.png 848w, https://substackcdn.com/image/fetch/$s_!hx79!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fff9b171c-55e0-4615-b406-bbdf0ae5d8fa_2600x1520.png 1272w, https://substackcdn.com/image/fetch/$s_!hx79!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fff9b171c-55e0-4615-b406-bbdf0ae5d8fa_2600x1520.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!hx79!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fff9b171c-55e0-4615-b406-bbdf0ae5d8fa_2600x1520.png" width="1456" height="851" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/ff9b171c-55e0-4615-b406-bbdf0ae5d8fa_2600x1520.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:851,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:167143,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221655458?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fff9b171c-55e0-4615-b406-bbdf0ae5d8fa_2600x1520.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!hx79!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fff9b171c-55e0-4615-b406-bbdf0ae5d8fa_2600x1520.png 424w, https://substackcdn.com/image/fetch/$s_!hx79!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fff9b171c-55e0-4615-b406-bbdf0ae5d8fa_2600x1520.png 848w, https://substackcdn.com/image/fetch/$s_!hx79!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fff9b171c-55e0-4615-b406-bbdf0ae5d8fa_2600x1520.png 1272w, https://substackcdn.com/image/fetch/$s_!hx79!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fff9b171c-55e0-4615-b406-bbdf0ae5d8fa_2600x1520.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>What September changed is the question I was pursuing. Every result above says Lookout&#8217;s ranking is real, <strong>but my way of trading it is lacking</strong>. Below, each entry gets its design, its tables and intervals, and what it led to.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><h1>The research log, September 2026</h1><h2>How to read this log</h2><p>Every test here was judged by a rule written down before its result existed. </p><p>Let me start with a vocab glossary:</p><ul><li><p><strong>Sharpe</strong> is return per unit of volatility.</p></li><li><p><strong>Maximum drawdown</strong> is the worst fall from a peak.</p></li><li><p><strong>Precision at a 1 percent false-positive budget</strong> is Lookout&#8217;s score, the share of its most confident flags that make the move.</p></li><li><p><strong>A basis point</strong> is a hundredth of a percent.</p></li><li><p><strong>A 95 percent interval</strong> comes from resampling the results. A result counts only when its interval clears zero.</p></li><li><p><strong>Walk-forward</strong> means every decision uses only data from before it.</p></li><li><p><strong>The deflated Sharpe ratio</strong> charges a Sharpe for every trial run before it.</p></li></ul><p>In September the reads still used Sharpe after costs as their main bar. The program&#8217;s current criterion, &#8220;active dollars against an index fund at four test account sizes,&#8221; was set on 13 October, so the October log uses it and this research log post mostly does not.</p><h2>1. The audit of my August claims (read 9 September)</h2><p><strong>The question.</strong> By the end of August I had eight claims the program leaned on, from &#8220;the allocator sweep found no skill&#8221; to &#8220;the Tape is graded with no manual edits&#8221;. Do they hold when someone else re-derives each one from the raw evidence?</p><p><strong>How it was tested.</strong></p><ul><li><p>One reviewer per claim assembled the evidence.</p></li><li><p>Three more tried to refute it from three angles: the arithmetic, whether the data was available at the time it was used, and whether production matched the claim.</p></li><li><p>They were told to rule &#8220;refuted&#8221; whenever a claim could not be checked.</p></li><li><p>A vote to uphold counted only if it cited evidence beyond what the first reviewer found.</p></li></ul><p><strong>The verdict.</strong> One claim confirmed as worded, seven refuted as worded.</p><p><strong>The numbers.</strong></p><ul><li><p>The confirmed claim was the August allocator sweep&#8217;s no-skill reading, re-run at SPA p-values of 0.957, 0.943 and 1.000.</p></li><li><p>In seven of the eight claims the headline numbers reproduced from raw data. What failed was the paper trail, the match to production, or the wording.</p></li><li><p>The earlier daily network&#8217;s August &#8220;pass&#8221; rested on a preregistration that had never been frozen. Scored strictly, its volatility rule failed, with one seed over the bar on 23 percent of days against an allowed 10.</p></li><li><p>The Tape&#8217;s page said &#8220;manual edits none, ever&#8221;, while sessions had been regraded offline after a grading incident. On 10 September the page was corrected to &#8220;27 sessions regraded offline on Aug 21, 2026; no manual edits since&#8221;.</p></li><li><p>That day the Tape read 33 sessions, 6,336 graded, 33.3 percent and -$1,159.21 at fixed paper sizes, checked to the cent against 49 morning reports.</p></li></ul><p><strong>The lesson.</strong> A number can reproduce while the claim built on it is still wrong. I now check every claim three ways: the number, where it came from, and the words around it. This is mostly semantics, but if I&#8217;m building in public and highlighting my transparency as a perk of this Substack, I may as well share.</p><h2>2. A risk control for the earlier daily network (9 to 27 September)</h2><p><strong>The question.</strong> Before Swell (another one of my model designs) took its current form, the daily model was a network that wrote a portfolio directly, and its training kept drifting toward an all-cash book. Could a redesigned training objective keep it invested while holding its volatility inside a set band? This was a condition for renting a GPU day for a larger sweep. It was not a money test.</p><p><strong>How it was tested.</strong></p><ul><li><p>One configuration, four monthly retrains in mid-2024, 87 trading days served, three seeds.</p></li><li><p>A served day counted as over the band if its volatility ran more than 5 percent above the target.</p></li><li><p>More than 10 percent of days over failed the read.</p></li><li><p>Sharpe here is after costs, on the portfolio&#8217;s weights, over those 87 days.</p></li></ul><p><strong>The verdicts, in order.</strong></p><ul><li><p><strong>9 September, a clean re-run from committed code.</strong> FAIL. One seed ran over the band on 20 of 87 days.</p></li><li><p><strong>9 to 10 September, a 314-day check of the larger sweep&#8217;s exact command. </strong>Realized volatility ran at more than twice the target, with 197 of 314 days over. The problem was in the design, not in one seed.</p></li><li><p><strong>10 September, fix one.</strong> A hard cap scales the book down whenever its volatility exceeds the target. PASS, with no seed over on any day. The median Sharpe fell from 1.02 to 0.80.</p></li><li><p><strong>27 September, fix two.</strong> Training that is aware of the volatility band, scored on runs made 10 September. PASS, with the hard cap needed on only 2 or 3 of 87 days per seed. The median Sharpe fell again, to 0.47. Its own verdict says volatility fell &#8220;mainly&#8221; through holding more cash.</p></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!4ZE6!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F49fe5856-1d83-4085-b736-61e4b9db611c_2600x844.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!4ZE6!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F49fe5856-1d83-4085-b736-61e4b9db611c_2600x844.png 424w, https://substackcdn.com/image/fetch/$s_!4ZE6!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F49fe5856-1d83-4085-b736-61e4b9db611c_2600x844.png 848w, https://substackcdn.com/image/fetch/$s_!4ZE6!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F49fe5856-1d83-4085-b736-61e4b9db611c_2600x844.png 1272w, https://substackcdn.com/image/fetch/$s_!4ZE6!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F49fe5856-1d83-4085-b736-61e4b9db611c_2600x844.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!4ZE6!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F49fe5856-1d83-4085-b736-61e4b9db611c_2600x844.png" width="1456" height="473" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/49fe5856-1d83-4085-b736-61e4b9db611c_2600x844.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:473,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:71908,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221655458?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F49fe5856-1d83-4085-b736-61e4b9db611c_2600x844.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!4ZE6!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F49fe5856-1d83-4085-b736-61e4b9db611c_2600x844.png 424w, https://substackcdn.com/image/fetch/$s_!4ZE6!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F49fe5856-1d83-4085-b736-61e4b9db611c_2600x844.png 848w, https://substackcdn.com/image/fetch/$s_!4ZE6!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F49fe5856-1d83-4085-b736-61e4b9db611c_2600x844.png 1272w, https://substackcdn.com/image/fetch/$s_!4ZE6!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F49fe5856-1d83-4085-b736-61e4b9db611c_2600x844.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p><strong>What followed.</strong> The larger sweep was amended and staged, and never ran. The line folded into Swell&#8217;s October work.</p><p><strong>The lesson.</strong> A volatility target that lives only in the training objective is not enforced. Once it was enforced, the network met it by holding cash, the same habit that sank Ballast in entry 4.</p><h2>3. Corrupted training days (found 17 September, replayed 18 to 19 September, corrected 28 September)</h2><p><strong>The question.</strong> An audit of the feature pipeline on 17 September found that the nightly job had been damaging one group of overnight and prior-day features. Lookout&#8217;s networks had trained on the damaged days. Did that make them worse?</p><p><strong>What went wrong.</strong></p><ul><li><p>Every night the job rebuilds its twelve most recent dates. It zero-filled that feature group for the oldest of the twelve and wrote the damaged day over the stored one.</p></li><li><p>38 of 427 stored days were affected, across late March, July to mid-August and late August.</p></li><li><p>The fix went in that evening and was proven on data the same night. The damaged dates were rebuilt.</p></li></ul><p><strong>How it was tested.</strong></p><ul><li><p>Each of Lookout&#8217;s four networks was retrained on the repaired data over 66 walk-forward windows and compared with the served version, window by window.</p></li><li><p>The rule was set before the replay ran. A mean change in precision under 0.01 means no material effect. Over 0.02 means swapping in the retrained networks.</p></li></ul><p><strong>The verdict.</strong> No material effect, so no swap. Model design is resilient.</p><p><strong>The numbers.</strong></p><ul><li><p>On 28 September I found that the replay&#8217;s starting point had already trained on the test days of its first 15 windows, so the read uses the 51 windows after them.</p></li><li><p>On those 51 windows the retrained networks scored lower than the served ones by 0.0062 [0.0021 to 0.0102], 0.0063 [0.0029 to 0.0097], 0.0091 [0.0056 to 0.0127] and 0.0050 [0.0014 to 0.0085].</p></li><li><p>Every difference sits under the 0.01 line, so the verdict held.</p></li><li><p>The replay is not a pure test of the corruption, because the retrained networks also differed from the served ones in other ways.</p></li></ul><p><strong>The lesson.</strong> A job that rewrites stored history needs a check that what it wrote is plausible. The damaged days showed one feature value on every row, against about 0.05 percent of rows on a healthy day, and nothing was checking for it.</p><h2>4. Ballast&#8217;s full-history test (run 10 September, read 27 September)</h2><p><strong>The question.</strong> <strong><span data-color="#6aa84f" style="color: rgb(106, 168, 79);">Ballast</span></strong><span data-color="#6aa84f" style="color: rgb(106, 168, 79);"> is a daily allocator, a forecaster feeding an optimizer that sets position sizes. </span>It had passed its earlier gates on the period it was chosen on. Over every day it could be tested on, did it beat holding an S&amp;P 500 index fund and an equal-weight basket of its own stocks, after costs? On 10 September I replaced a planned wait for live results with this walk, because the walk could answer the question now.</p><p><strong>How it was tested.</strong></p><ul><li><p>The frozen candidate was walked over 780 trading days, 1 August 2023 to 9 September 2026.</p></li><li><p>38 monthly retrains, with a seven-day gap between training and trading.</p></li><li><p>Costs came from spreads. Whole shares in a small account, long only plus cash.</p></li><li><p>The bar, written before the walk, was a net Sharpe at 1x costs strictly above both baselines.</p></li><li><p>The deflated Sharpe ratio was reported but not part of the bar.</p></li><li><p>A third arm, the same optimizer fed trailing-average returns in place of the forecaster, is shown for reference.</p></li></ul><p><strong>The verdict.</strong> FAIL. <strong><span data-color="#6aa84f" style="color: rgb(106, 168, 79);">Ballast</span></strong> was parked as an allocator model. It still runs every night on paper as a check, and an extended read runs once December&#8217;s data exists, or if I feel like restarting it sooner.</p><p><strong>The numbers.</strong></p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!qZK7!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f422097-a543-4bab-994f-743ed74f22c5_2454x844.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!qZK7!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f422097-a543-4bab-994f-743ed74f22c5_2454x844.png 424w, https://substackcdn.com/image/fetch/$s_!qZK7!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f422097-a543-4bab-994f-743ed74f22c5_2454x844.png 848w, https://substackcdn.com/image/fetch/$s_!qZK7!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f422097-a543-4bab-994f-743ed74f22c5_2454x844.png 1272w, https://substackcdn.com/image/fetch/$s_!qZK7!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f422097-a543-4bab-994f-743ed74f22c5_2454x844.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!qZK7!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f422097-a543-4bab-994f-743ed74f22c5_2454x844.png" width="1456" height="501" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/3f422097-a543-4bab-994f-743ed74f22c5_2454x844.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:501,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:100747,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221655458?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f422097-a543-4bab-994f-743ed74f22c5_2454x844.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!qZK7!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f422097-a543-4bab-994f-743ed74f22c5_2454x844.png 424w, https://substackcdn.com/image/fetch/$s_!qZK7!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f422097-a543-4bab-994f-743ed74f22c5_2454x844.png 848w, https://substackcdn.com/image/fetch/$s_!qZK7!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f422097-a543-4bab-994f-743ed74f22c5_2454x844.png 1272w, https://substackcdn.com/image/fetch/$s_!qZK7!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f422097-a543-4bab-994f-743ed74f22c5_2454x844.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><ul><li><p>Sharpe margin against the equal-weight basket: -0.79, with a 95 percent interval of -1.74 to +0.12.</p></li><li><p>Sharpe margin against the index fund: -1.10, interval -2.18 to -0.04.</p></li><li><p>Deflated Sharpe ratio: 0.13.</p></li><li><p>Year by year, Ballast trailed the index fund in every year of the walk.</p></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!7u1V!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fad79faa1-0a97-47a4-8412-d8db382677d3_2382x844.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!7u1V!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fad79faa1-0a97-47a4-8412-d8db382677d3_2382x844.png 424w, https://substackcdn.com/image/fetch/$s_!7u1V!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fad79faa1-0a97-47a4-8412-d8db382677d3_2382x844.png 848w, https://substackcdn.com/image/fetch/$s_!7u1V!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fad79faa1-0a97-47a4-8412-d8db382677d3_2382x844.png 1272w, https://substackcdn.com/image/fetch/$s_!7u1V!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fad79faa1-0a97-47a4-8412-d8db382677d3_2382x844.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!7u1V!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fad79faa1-0a97-47a4-8412-d8db382677d3_2382x844.png" width="1456" height="516" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/ad79faa1-0a97-47a4-8412-d8db382677d3_2382x844.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:516,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:85769,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221655458?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fad79faa1-0a97-47a4-8412-d8db382677d3_2382x844.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!7u1V!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fad79faa1-0a97-47a4-8412-d8db382677d3_2382x844.png 424w, https://substackcdn.com/image/fetch/$s_!7u1V!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fad79faa1-0a97-47a4-8412-d8db382677d3_2382x844.png 848w, https://substackcdn.com/image/fetch/$s_!7u1V!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fad79faa1-0a97-47a4-8412-d8db382677d3_2382x844.png 1272w, https://substackcdn.com/image/fetch/$s_!7u1V!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fad79faa1-0a97-47a4-8412-d8db382677d3_2382x844.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><ul><li><p>On the span its earlier gates used, July 2024 to June 2026, Ballast read 1.10, against 1.07 for the basket and 1.11 for the index fund. Its earlier pass came from the period it was chosen on.</p></li></ul><p><strong>Why it lost.</strong> A diagnosis split the gap into parts, with its reading rule frozen before any number was computed.</p><ul><li><p><strong>Cash was the largest part.</strong> Ballast was 57.6 percent invested on an average day, so 42.4 percent of the account sat in cash earning nothing.</p></li><li><p>Cash cost it 11.09 percent a year against the basket (interval -21.32 to -0.93) and 11.97 percent a year against the index fund (interval -21.11 to -3.37).</p></li><li><p><strong>Costs came next</strong>, about a quarter of the Sharpe gap to the basket.</p></li><li><p>Cash and costs were the only parts whose intervals excluded zero.</p></li><li><p><strong>The forecasts showed no measurable skill.</strong> Their daily rank correlation with what happened was +0.0104, interval -0.0039 to +0.0252.</p></li></ul><p><strong>The lesson.</strong> A portfolio that holds 42 percent cash cannot keep up with an index fund that is always fully invested, and a forecast with no measurable skill gives the optimizer no reason to hold more stock. Both findings went into the criterion the program uses now, which judges every book against an index fund held in the same account.</p><h2>5. Where the paper bot&#8217;s money went (frozen and read 27 September)</h2><p><strong>The question.</strong> Lookout&#8217;s paper bot trades its flags on the broker&#8217;s paper simulator. Over its first eleven sessions after it resumed (11 to 25 September, 535 filled lots), the Tape&#8217;s grade of those flags was positive but within noise, and the bot lost money. Where does the money go between the flag and the fill?</p><p><strong>How it was tested.</strong></p><ul><li><p>Each lot&#8217;s result was split exactly into four parts: which flags it picked and when it entered, the exit rule, the price paid to enter, and the price received on exit.</p></li><li><p>Intervals came from resampling whole sessions 10,000 times.</p></li><li><p>The reading rule, frozen before computing, called a part a leak only if it was negative and its 95 percent interval excluded zero.</p></li><li><p>It was a reading of the evidence, not a pass or fail, and it changed no trading rule.</p></li></ul><p><strong>The verdict.</strong><span data-color="#674ea7" style="color: rgb(103, 78, 167);"> &#8220;Most of the money leaks in the price the bot pays to get in and out, not in which flags it picks, when it enters, or how it exits.&#8221; </span>The entry fill was the only part that met the leak rule.</p><p><strong>The numbers,</strong> in basis points of the money traded.</p><ul><li><p>Entering cost about 102 and exiting about 40, a round trip of about 142.</p></li><li><p>The Tape&#8217;s edge was about 16 basis points per graded flag, so the round trip cost six to nine times the edge.</p></li><li><p>A check added after the fact found 351 of 1,070 fills priced outside the range printed in their minute, or in a minute with no print at all. Moving every such fill inside the printed range still left a round trip of about 90.</p></li><li><p>Entry timing and the exit rule both came out positive and unresolved.</p></li><li><p>An independent check recomputed all 535 lots from the raw inputs with its own code and found no disagreements.</p></li><li><p>The bot made money in 2 of its 14 September sessions.</p></li></ul><p><strong>The lesson.</strong> An edge of 16 basis points per flag does not survive a round trip of 90 to 142. No exit rule changes that, because the exits were not the problem.</p><h2>6. Price bands on the Tape (read 27 September)</h2><p><strong>The question.</strong> Lookout&#8217;s precision is much higher on low-priced stocks (arc one covers why). Does the Tape&#8217;s money follow?</p><p><strong>How it was tested.</strong></p><ul><li><p>The Tape&#8217;s own grading, grouped by share price, over every graded session from 14 July to 25 September.</p></li><li><p>The Tape grades at fixed paper sizes, before trading costs.</p></li><li><p>I reproduced the table exactly from the archived morning reports with the Tape&#8217;s own code.</p></li></ul><p><strong>The numbers.</strong> 45 graded sessions, 8,506 graded signals, -$822.78 in all.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!UdiB!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6df581bb-2521-48cd-8a5f-6de5587a5252_1800x844.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!UdiB!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6df581bb-2521-48cd-8a5f-6de5587a5252_1800x844.png 424w, https://substackcdn.com/image/fetch/$s_!UdiB!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6df581bb-2521-48cd-8a5f-6de5587a5252_1800x844.png 848w, https://substackcdn.com/image/fetch/$s_!UdiB!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6df581bb-2521-48cd-8a5f-6de5587a5252_1800x844.png 1272w, https://substackcdn.com/image/fetch/$s_!UdiB!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6df581bb-2521-48cd-8a5f-6de5587a5252_1800x844.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!UdiB!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6df581bb-2521-48cd-8a5f-6de5587a5252_1800x844.png" width="1456" height="683" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/6df581bb-2521-48cd-8a5f-6de5587a5252_1800x844.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:683,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:73769,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221655458?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6df581bb-2521-48cd-8a5f-6de5587a5252_1800x844.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!UdiB!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6df581bb-2521-48cd-8a5f-6de5587a5252_1800x844.png 424w, https://substackcdn.com/image/fetch/$s_!UdiB!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6df581bb-2521-48cd-8a5f-6de5587a5252_1800x844.png 848w, https://substackcdn.com/image/fetch/$s_!UdiB!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6df581bb-2521-48cd-8a5f-6de5587a5252_1800x844.png 1272w, https://substackcdn.com/image/fetch/$s_!UdiB!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6df581bb-2521-48cd-8a5f-6de5587a5252_1800x844.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><ul><li><p>The paper bot lost money in every one of its own price bands over the same eleven sessions as entry 5.</p></li></ul><p><strong>What followed.</strong></p><ul><li><p>I kept the bot&#8217;s universe as it was. Price and liquidity floors became tested arms of the re-aim walk rather than live limits.</p></li><li><p>On the live Tape through 16 October, the under-$1 band stood at +$57.33 on 3,964 graded, so most of its September gain had gone.</p></li></ul><p><strong>The lesson.</strong> The one positive band is positive before costs, and entry 5 shows what costs do to an edge this size. A band that was ahead in September and nearly flat a month later is not a result to trade on.</p><h2>7. The re-aim of <span data-color="#3c78d8" style="color: rgb(60, 120, 216);">Lookout</span> (proposed 18 September, adopted 27 September, amended 30 September)</h2><p>This entry is a change of plan, not a test.</p><p><strong>Where things stood.</strong> Lookout&#8217;s ranking skill was well measured, with precision near 36 percent at a 1 percent false-positive budget against a base rate of about 1.3 percent. No trading protocol had made money on the vendor&#8217;s bars.</p><p><strong>The plan, in three working steps.</strong></p><ol><li><p>Make the live model match the tested one.</p></li><li><p>Retrain Lookout on a target that subtracts the cost of getting in, judged against the current target.</p></li><li><p>Walk one preregistered trading protocol over the full history. It passes only if it is positive after costs with its interval clear of zero. Otherwise Lookout is a ranking instrument and not yet a trading one.</p></li></ol><p><strong>Set on 30 September, after the outside review in entry 10.</strong></p><ul><li><p>The trial count the read would be charged for: twelve.</p></li><li><p>The walk&#8217;s start: 3 January 2023.</p></li><li><p>The consequence of a fail, written down in advance. Lookout&#8217;s paper bot stays a measurement, and Lookout&#8217;s scores become a candidate input for the daily model.</p></li><li><p>The same night I ruled that the public record would stay private until a Lookout protocol made money. I reversed that on 18 October and published everything, losses included.</p></li></ul><p><strong>What happened.</strong> The re-aim was frozen on 7 October and read on 10 October. Both steps failed. The October log has the full result.</p><h2>8. Would a real-time data feed help? (frozen and read 28 September)</h2><p><strong>The question.</strong> The live chain works from data delayed by a median of 17 minutes. If Lookout&#8217;s flags are right but arrive too late, how much would a real-time feed recover?</p><p><strong>How it was tested.</strong></p><ul><li><p>45 sessions from 14 July to 25 September, 14,002 flags, 8,460 of them graded in common.</p></li><li><p>Each flag was re-traded at delays of 17, 5, 2 and 1 minutes.</p></li><li><p>Two order types, a market order or a limit order just below the prior close, at the costs measured in entry 5.</p></li><li><p>The rule, frozen before any number existed, said a real-time plan was worth pricing only if at least one of 16 decision cells (four price bands, two short delays, two order types) came out positive by the interval rule.</p></li></ul><p><strong>The verdict.</strong> The rule did not fire. 0 of 16 cells positive, 12 negative, 4 unresolved.</p><p><strong>The numbers.</strong></p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!M3gA!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F759e9709-5e56-4fe5-9fb6-7205e6d318a7_2036x1006.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!M3gA!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F759e9709-5e56-4fe5-9fb6-7205e6d318a7_2036x1006.png 424w, https://substackcdn.com/image/fetch/$s_!M3gA!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F759e9709-5e56-4fe5-9fb6-7205e6d318a7_2036x1006.png 848w, https://substackcdn.com/image/fetch/$s_!M3gA!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F759e9709-5e56-4fe5-9fb6-7205e6d318a7_2036x1006.png 1272w, https://substackcdn.com/image/fetch/$s_!M3gA!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F759e9709-5e56-4fe5-9fb6-7205e6d318a7_2036x1006.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!M3gA!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F759e9709-5e56-4fe5-9fb6-7205e6d318a7_2036x1006.png" width="1456" height="719" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/759e9709-5e56-4fe5-9fb6-7205e6d318a7_2036x1006.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:719,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:115007,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221655458?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F759e9709-5e56-4fe5-9fb6-7205e6d318a7_2036x1006.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!M3gA!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F759e9709-5e56-4fe5-9fb6-7205e6d318a7_2036x1006.png 424w, https://substackcdn.com/image/fetch/$s_!M3gA!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F759e9709-5e56-4fe5-9fb6-7205e6d318a7_2036x1006.png 848w, https://substackcdn.com/image/fetch/$s_!M3gA!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F759e9709-5e56-4fe5-9fb6-7205e6d318a7_2036x1006.png 1272w, https://substackcdn.com/image/fetch/$s_!M3gA!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F759e9709-5e56-4fe5-9fb6-7205e6d318a7_2036x1006.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><ul><li><p>A faster feed made the result slightly worse before costs.</p></li><li><p>The order type was worth about six times any change in speed.</p></li><li><p>The limit-order fill model is optimistic, so the limit numbers are a best case.</p></li><li><p>The Tape&#8217;s published rows were reproduced, 3,667 of 3,667 on 12 sessions, and a second reviewer hand-checked 80 random flags.</p></li></ul><p><strong>What followed.</strong> No real-time feed was bought, and the re-aim walk added limit-entry arms.</p><p><strong>The lesson.</strong> Arriving earlier does not help when the cost of getting in is the problem. The cheapest improvement in this test was the order type, and even the best cell did not clear zero.</p><h2>9. Three look-ahead leaks (confirmed 18 September, knocked out 28 September, read 30 September)</h2><p><strong>The question.</strong> The pipeline audit found three ways future information reached Lookout&#8217;s inputs, and two reviewers confirmed all three on the data. How much does the served ensemble lean on them?</p><p><strong>The three leaks.</strong></p><ul><li><p>Two company-description inputs were taken as of the run date and copied onto every past date.</p></li><li><p>Some inputs were computed from the whole trading day, including bars after the one being scored.</p></li><li><p>One input was built from filings dated after the bar.</p></li></ul><p><strong>How it was tested.</strong></p><ul><li><p>On the frozen networks, each group&#8217;s values were overwritten with a neutral value at scoring time.</p></li><li><p>37 windows, with paired resampling 10,000 times.</p></li><li><p>The rule, set before the read, called a group free under a 0.01 change in the ensemble&#8217;s precision, and ordered a retrain over 0.02.</p></li></ul><p><strong>The verdict.</strong> All three free.</p><p><strong>The numbers,</strong> change in the ensemble&#8217;s precision.</p><ul><li><p>Company-description pair: +0.0001 (interval -0.0012 to +0.0012).</p></li><li><p>Whole-day inputs: +0.0015 (-0.0004 to +0.0037).</p></li><li><p>Filing-date input: +0.0001 (-0.0004 to +0.0006).</p></li><li><p>Single networks moved more than the ensemble did.</p></li></ul><p><strong>What followed.</strong> All three groups were dropped from the re-aim&#8217;s training anyway, since there was no reason to keep inputs that cannot exist at the time of the trade.</p><p><strong>The lesson.</strong> A leak can be real and harmless at the same time. The knockout measured the harm before I paid for a retrain.</p><h2>10. An outside review of the plan (30 September)</h2><p><strong>The question.</strong> Is the plan the right path to the goal, and is the change I proposed that night better?</p><p><strong>How it was tested.</strong> Five reviewers took one lens each, market microstructure, signal research, validation, live retail trading and a red team. Two refuters cross-examined each reviewer.</p><p><strong>The verdict.</strong></p><ul><li><p>The plan scored 4.8 out of 10 (spread 4 to 5), and the proposed change 4.2.</p></li><li><p>The summary called the program &#8220;on the right track in method and direction, and off the shortest path in tempo and odds&#8221;.</p></li><li><p>Measurement discipline scored near 9. The route to money scored 3 to 5.</p></li><li><p>All twenty of the reviewers&#8217; recommendations were refuted, mostly because they assumed inputs the program does not have.</p></li><li><p>The point no reviewer disputed was that the program&#8217;s checks had removed false edge without creating real edge.</p></li></ul><p><strong>What followed.</strong> The full-history start date and the trial count in entry 7, and the advance consequence of a fail.</p><h2>Incidents</h2><p><strong>The data chain outage, 30 August to 9 September.</strong></p><ul><li><p>A firmware update on the storage server locked the data folders, and every nightly step after it failed.</p></li><li><p>The alarm named only the steps that failed downstream, so the cause went unreported for ten days.</p></li><li><p>The fix made the first step write its own failure flag.</p></li><li><p>Six live sessions, 31 August, 1 to 4 September and 8 September, have no record on the Tape.</p></li><li><p>Nothing in this log ran before 9 September.</p></li></ul><p><strong>The corrupted feature days.</strong> Entry 3.</p><h2>What this log does not contain</h2><p>Feature definitions and counts, score thresholds and universe screen parameters stay private for now. Dollar account sizes and position sizes are never published, so the paper bot&#8217;s results are given in basis points of the money traded. Both rules apply to every tier.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included.</em></p>]]></content:encoded></item><item><title><![CDATA[Arc two: The Machinery]]></title><description><![CDATA[Five checks I built to keep from fooling myself that my neural network is good at trading.]]></description><link>https://letters.verbatimsigma.com/p/arc-two-the-machinery</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/arc-two-the-machinery</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Tue, 20 Oct 2026 13:04:11 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!8AQS!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faa99a818-3e4e-430a-affe-56d66ac60140_2600x2320.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Written in July 2026. Anything that has changed since then is marked with an <strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>.</em></p><p>If you are just arriving, the first arc of this notebook tested nine pre-registered ideas about <strong><span data-color="#3c78d8" style="color: rgb(60, 120, 216);">Lookout</span></strong>, the intraday neural network model that this notebook is mostly about, so far. The tests ran on rented GPU time, cost me $423, and the verdicts were mostly unflattering: several noes, a couple of ties, one marginal call, and one clean win. The founding thesis of the whole project, the architectural belief I had organized two years of work around, was tested and lost, and I retired it.</p><p>Arc 1, &#8220;<strong><span data-color="#a64d79" style="color: rgb(166, 77, 121);">The Campaign</span>,</strong>&#8221; covered what the rules decided. Arc 2, <strong><span data-color="#3d85c6" style="color: rgb(61, 133, 198);">The Machinery</span></strong>,&#8221; covers how the rules work. Most of the ideas in arc 1 were retired. The rules that retired persist, and they are the reason the verdicts are believable.</p><p>I built each rule after I caught myself about to cheat the Scientific Method in a specific way. I tested it until it worked, then made it a standing part of the program, and it now applies to every result whether I like the result or not. </p><p>The next 5 letters describe the machinery, one piece each, in order:</p><ol><li><p>Walk-forward evaluation, which grades a model only on days it has never seen, and why its score is almost always the lower one.</p></li><li><p>The freeze, three months of my own data I&#8217;m not allowed to consult, until one scheduled reading.</p></li><li><p>The parts of a pre-registration document, which you&#8217;re welcome to copy.</p></li><li><p>A failure I published on purpose.</p></li><li><p>Fifteen trading days my dataset lost, and the check that now looks for missing days every night.</p></li></ol><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!8AQS!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faa99a818-3e4e-430a-affe-56d66ac60140_2600x2320.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!8AQS!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faa99a818-3e4e-430a-affe-56d66ac60140_2600x2320.png 424w, https://substackcdn.com/image/fetch/$s_!8AQS!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faa99a818-3e4e-430a-affe-56d66ac60140_2600x2320.png 848w, https://substackcdn.com/image/fetch/$s_!8AQS!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faa99a818-3e4e-430a-affe-56d66ac60140_2600x2320.png 1272w, https://substackcdn.com/image/fetch/$s_!8AQS!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faa99a818-3e4e-430a-affe-56d66ac60140_2600x2320.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!8AQS!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faa99a818-3e4e-430a-affe-56d66ac60140_2600x2320.png" width="1456" height="1299" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/aa99a818-3e4e-430a-affe-56d66ac60140_2600x2320.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1299,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:223442,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221220687?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faa99a818-3e4e-430a-affe-56d66ac60140_2600x2320.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!8AQS!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faa99a818-3e4e-430a-affe-56d66ac60140_2600x2320.png 424w, https://substackcdn.com/image/fetch/$s_!8AQS!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faa99a818-3e4e-430a-affe-56d66ac60140_2600x2320.png 848w, https://substackcdn.com/image/fetch/$s_!8AQS!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faa99a818-3e4e-430a-affe-56d66ac60140_2600x2320.png 1272w, https://substackcdn.com/image/fetch/$s_!8AQS!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faa99a818-3e4e-430a-affe-56d66ac60140_2600x2320.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>None of this is my invention, but I&#8217;m applying it, stubbornly, to a small trading account.</p><p>Five letters on the checks that decide what I trade. Subscribe to read them as they come out.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><p>The arc after this one covers the market behavior <strong><span data-color="#3d85c6" style="color: rgb(61, 133, 198);">Lookout</span></strong> tries to catch. Later comes a live diary, where the program trades in public and the notebook gets a P&amp;L column. Every number in that diary will depend on the checks in this arc, which is why they come first.</p><p><em><strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>: the live diary has not started. I traded real money by hand, at half size, for a few days in August and then stopped; every bot since trades on a broker&#8217;s paper simulator. The P&amp;L column that exists today is <a href="https://gradedtape.com/">The Graded Tape</a>, which grades every signal at fixed paper sizes. Through Oct 16, 2026, 60 trading days and 11,316 signals graded, it reads <strong><span data-color="#980000" style="color: rgb(152, 0, 0);">-$1,877.13</span></strong>.</em></p><h2>How I built it</h2><p>In the rest of <strong><span data-color="#3d85c6" style="color: rgb(61, 133, 198);">The Machinery</span></strong> arc (the next 5 letters), this &#8220;How I built it,&#8221; section is for paid subscribers. Here it&#8217;s free, so you can see what it holds.</p><p>Every check in this arc came out of one Claude Code session that ran from July 7 to July 13, 2026, six days and 191 commits across five repositories. I didn&#8217;t write the code. I picked the bars, signed each design doc before anything ran, set the freeze date, and approved every dollar of rented GPU time.</p><p>When I went to bed, I left standing orders written as if-then rules. One of them told the agent to retry a failed nightly training job once, then skip it and put a banner on the morning report. The agent kept working overnight without making up policy at 3 a.m., and every rule that could flatter a result was in git before the result existed.</p><p>Try this: before your next overnight agent run, write five if-then orders for the failures you expect, and commit them first.</p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included.</em></p>]]></content:encoded></item><item><title><![CDATA[My model is 6x more precise on low-priced stocks]]></title><description><![CDATA[Same model, same test, six times the precision. Then I checked the base rates. The arc one working papers are below, free.]]></description><link>https://letters.verbatimsigma.com/p/my-model-is-6x-more-precise-on-low</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/my-model-is-6x-more-precise-on-low</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Mon, 19 Oct 2026 20:43:10 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!tvgS!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5f730493-cc76-41a4-a51d-546964636c01_2600x1480.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Written in July 2026. Anything that has changed since then is marked with an <strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>.</em></p><p>The best finding of the campaign was where my model already works best.</p><p>When my founding thesis lost, working out <em>why</em> meant re-scoring the single model, <strong><span data-color="#3c78d8" style="color: rgb(60, 120, 216);">Lookout</span>, </strong>started from on different slices of the market. One of those slices scored far higher than I expected.</p><h2>The split</h2><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!tvgS!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5f730493-cc76-41a4-a51d-546964636c01_2600x1480.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!tvgS!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5f730493-cc76-41a4-a51d-546964636c01_2600x1480.png 424w, https://substackcdn.com/image/fetch/$s_!tvgS!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5f730493-cc76-41a4-a51d-546964636c01_2600x1480.png 848w, https://substackcdn.com/image/fetch/$s_!tvgS!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5f730493-cc76-41a4-a51d-546964636c01_2600x1480.png 1272w, https://substackcdn.com/image/fetch/$s_!tvgS!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5f730493-cc76-41a4-a51d-546964636c01_2600x1480.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!tvgS!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5f730493-cc76-41a4-a51d-546964636c01_2600x1480.png" width="1456" height="829" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/5f730493-cc76-41a4-a51d-546964636c01_2600x1480.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:829,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:191469,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221213094?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5f730493-cc76-41a4-a51d-546964636c01_2600x1480.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!tvgS!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5f730493-cc76-41a4-a51d-546964636c01_2600x1480.png 424w, https://substackcdn.com/image/fetch/$s_!tvgS!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5f730493-cc76-41a4-a51d-546964636c01_2600x1480.png 848w, https://substackcdn.com/image/fetch/$s_!tvgS!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5f730493-cc76-41a4-a51d-546964636c01_2600x1480.png 1272w, https://substackcdn.com/image/fetch/$s_!tvgS!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5f730493-cc76-41a4-a51d-546964636c01_2600x1480.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>Same model, graded the same way, and it&#8217;s roughly six times more precise among low-priced names than among the largest names, where most systematic funds trade. About one low-priced flag in five goes on to make the move.</p><h2>Base rates first</h2><p>Low-priced names make large intraday moves far more often than the largest names do. In these same windows, at any given minute, about one low-priced name in a hundred was within an hour of making the move, against fewer than one of the largest names in a thousand, which is roughly thirteen times as many targets. The score holds the false-alarm budget fixed within each slice, so a slice can&#8217;t inflate its score just by being noisier, but having more genuine movers to find still helps it a great deal.</p><p>Measured against those base rates, the comparison turns around. Relative to how rarely the largest names move, the model&#8217;s lift there is the larger one, about forty-five times the base rate, against about twenty-one times among low-priced names. So most of the six-fold gap comes from having more movers to find rather than from a smarter model. Still, more of the low-priced flags you could act on go on to make the move.</p><h2>Why the opportunity persists</h2><p>Visible opportunities usually get traded away, so this one needs a reason. My first guess in July was capacity, since big funds can&#8217;t trade small names without moving them. That guess doesn&#8217;t fit this slice. It&#8217;s cut by share price and trading volume, so it holds some of the most heavily traded products in the market, leveraged index funds included. The plainer reason is arithmetic. A five percent move is an ordinary day for a three-dollar stock or a fund built to move three times the index, and a rare one for a giant.</p><p>A finding like this still has to survive real costs. Subscribe and watch that measurement happen in public.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><h2>What changed</h2><p>My daily watchlist now defaults to the low-priced end of the market, and I&#8217;ve mostly stopped looking at the largest names for now, for this model. I spent most of this campaign trying to make the model smarter. Re-scoring the same model on different slices of the market cost nothing, and it changed my trading more than any of the architecture experiments did.</p><h2>The caveats</h2><p>Low-priced names come with wider spreads as a share of the price, uglier fills, and exits that can take patience at the worst possible moments. Some of the paper edge will go to paying those costs. How much is a measurable question, and the live diary later in this series exists to measure it with logged fills. I wrote that into the record on purpose, so you can hold me to it.</p><p><em><strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>: no published diary of real fills exists yet, but <a href="https://gradedtape.com/">The Graded Tape</a> has been taking the measurement in public, at fixed paper trade sizes, and so far the low-priced precision has not shown up as money. Through Oct 16, 2026, 60 trading days and 11,316 graded signals, the record stands at <strong><span data-color="#980000" style="color: rgb(152, 0, 0);">-$1,877.13</span></strong>, with 34.0% of signals reaching the Tape&#8217;s target.</em></p><p>From arc two onward, the working papers sit below the line for paid subscribers, twenty dollars a month or two hundred a year. The founding tier, four hundred a year, is for readers who want to fund the lab harder, with the rate locked and first claim on anything that opens later, if and when a model passes its sealed test.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><p>If someone you know builds models for their own trading, this is the arc to send them.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/p/my-model-is-6x-more-precise-on-low?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/p/my-model-is-6x-more-precise-on-low?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p><h2>The working papers</h2><p>Below is the paperwork the six posts were written from, so you can check any number in them. Nothing in it revises a verdict.</p><p>This appendix is free, and it stays free. From here on every arc closes this way, with the working papers below the line for subscribers, while the verdicts, the hit rates, the misses, and anything else that would change your read on whether this works stay above it permanently.</p><h2>Appendix: working papers, arc one</h2><p>Every number below is traced to the document or data file it came out of, and each source is named in plain words below the table it feeds.</p><h3>1. The decision rule, as pre-registered</h3><p><em>As signed, before any result existed.</em></p><ul><li><p>Dethrone bar: improve mean P@1%FPR by more than +0.02 over the incumbent</p></li><li><p>Interval: paired bootstrap over per-window differences, 1,000 resamples, fixed seed, 95% interval clear of zero</p></li><li><p>Tie-break: ties go to the simpler and older model</p></li><li><p>Evidence freeze: 2026-04-10; windows past it excluded from adjudication</p></li><li><p>Scored: 793 matched windows, 60 frozen windows excluded</p></li><li><p>Outcome: the record&#8217;s dethrone flag reads false. Nothing cleared the bar</p></li></ul><p>Source: the round-one adjudication verdict record. The cross-sectional arm has its own pre-registration, the cross-sectional round-two design document, drafted 2026-07-09 and signed 2026-07-10. The excitation arm&#8217;s, pre-registered 2026-07-11 before any training run, is the excitation-features experiment document. Both are preserved with their original hypothesis sections intact and an adjudication header added on top, so the original argument is still readable under the verdict.</p><h3>2. The four-way tie, to four decimals</h3><p><em>Mean P@1%FPR, 793 matched windows.</em></p><ul><li><p>Variant D, protection mechanism removed: 0.3632</p></li><li><p>Variant C, pruned inputs: 0.3628</p></li><li><p>Variant A, incumbent: 0.3624</p></li><li><p>Variant B, two-head: 0.3603</p></li><li><p>Per-window logistic control: 0.3506</p></li></ul><p>Source: the round-one adjudication verdict record, its per-arm means block. All four deep arms sit inside the 0.02 tie band; the record carries the tie set explicitly and names Variant D as the winner under the simpler-wins tie-break. The logistic reaches 96.5 percent of the winner&#8217;s score.</p><h3>3. The intervals</h3><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!CThW!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03240323-25a1-44be-91a1-357f8dbcde21_2600x1280.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!CThW!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03240323-25a1-44be-91a1-357f8dbcde21_2600x1280.png 424w, https://substackcdn.com/image/fetch/$s_!CThW!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03240323-25a1-44be-91a1-357f8dbcde21_2600x1280.png 848w, https://substackcdn.com/image/fetch/$s_!CThW!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03240323-25a1-44be-91a1-357f8dbcde21_2600x1280.png 1272w, https://substackcdn.com/image/fetch/$s_!CThW!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03240323-25a1-44be-91a1-357f8dbcde21_2600x1280.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!CThW!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03240323-25a1-44be-91a1-357f8dbcde21_2600x1280.png" width="1456" height="717" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/03240323-25a1-44be-91a1-357f8dbcde21_2600x1280.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:717,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:179951,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221213094?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03240323-25a1-44be-91a1-357f8dbcde21_2600x1280.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!CThW!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03240323-25a1-44be-91a1-357f8dbcde21_2600x1280.png 424w, https://substackcdn.com/image/fetch/$s_!CThW!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03240323-25a1-44be-91a1-357f8dbcde21_2600x1280.png 848w, https://substackcdn.com/image/fetch/$s_!CThW!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03240323-25a1-44be-91a1-357f8dbcde21_2600x1280.png 1272w, https://substackcdn.com/image/fetch/$s_!CThW!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03240323-25a1-44be-91a1-357f8dbcde21_2600x1280.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!Aomr!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b16620c-6d0b-4bc5-b65a-01e0d163da7f_2244x742.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!Aomr!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b16620c-6d0b-4bc5-b65a-01e0d163da7f_2244x742.png 424w, https://substackcdn.com/image/fetch/$s_!Aomr!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b16620c-6d0b-4bc5-b65a-01e0d163da7f_2244x742.png 848w, https://substackcdn.com/image/fetch/$s_!Aomr!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b16620c-6d0b-4bc5-b65a-01e0d163da7f_2244x742.png 1272w, https://substackcdn.com/image/fetch/$s_!Aomr!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b16620c-6d0b-4bc5-b65a-01e0d163da7f_2244x742.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!Aomr!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b16620c-6d0b-4bc5-b65a-01e0d163da7f_2244x742.png" width="1456" height="481" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/4b16620c-6d0b-4bc5-b65a-01e0d163da7f_2244x742.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:481,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:73146,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221213094?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b16620c-6d0b-4bc5-b65a-01e0d163da7f_2244x742.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!Aomr!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b16620c-6d0b-4bc5-b65a-01e0d163da7f_2244x742.png 424w, https://substackcdn.com/image/fetch/$s_!Aomr!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b16620c-6d0b-4bc5-b65a-01e0d163da7f_2244x742.png 848w, https://substackcdn.com/image/fetch/$s_!Aomr!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b16620c-6d0b-4bc5-b65a-01e0d163da7f_2244x742.png 1272w, https://substackcdn.com/image/fetch/$s_!Aomr!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b16620c-6d0b-4bc5-b65a-01e0d163da7f_2244x742.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>Source: same file, its bootstrap block against Variant A. Look at the first row. The two-head arm&#8217;s whole interval sits below zero, so it is reliably a little worse than the model it was meant to beat, and it still belongs in the tie set because the gap is a tenth of the bar.</p><p>The one line in that file that cleared its bar is the ensemble. The record&#8217;s ensemble line shows +0.0287 with a 95% interval of [+0.0231, +0.0350] against the best single arm, measured on the retained checkpoint grid before the freeze date, with its cost noted as the nightly chains of every member.</p><h3>4. The worst stretch</h3><p><em>Worst rolling ten-window mean.</em></p><ul><li><p>Variant A, incumbent: 0.2333</p></li><li><p>Variant D: 0.2612</p></li><li><p>Variant B: 0.2889</p></li><li><p>Variant C: 0.2946</p></li></ul><p>Source: same file, its worst-rolling-stretch block. The champion&#8217;s worst ten consecutive windows average 0.2333, roughly 36 percent below its full-period 0.3624. This is the closest thing a ranking model has to a drawdown, and it&#8217;s the first number I&#8217;d look for if someone handed me this table.</p><h3>5. Premium over the logistic, year by year</h3><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!HNvj!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45458cc8-6346-43a6-9777-6b7878ac02e6_2600x1480.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!HNvj!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45458cc8-6346-43a6-9777-6b7878ac02e6_2600x1480.png 424w, https://substackcdn.com/image/fetch/$s_!HNvj!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45458cc8-6346-43a6-9777-6b7878ac02e6_2600x1480.png 848w, https://substackcdn.com/image/fetch/$s_!HNvj!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45458cc8-6346-43a6-9777-6b7878ac02e6_2600x1480.png 1272w, https://substackcdn.com/image/fetch/$s_!HNvj!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45458cc8-6346-43a6-9777-6b7878ac02e6_2600x1480.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!HNvj!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45458cc8-6346-43a6-9777-6b7878ac02e6_2600x1480.png" width="1456" height="829" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/45458cc8-6346-43a6-9777-6b7878ac02e6_2600x1480.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:829,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:278330,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221213094?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45458cc8-6346-43a6-9777-6b7878ac02e6_2600x1480.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!HNvj!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45458cc8-6346-43a6-9777-6b7878ac02e6_2600x1480.png 424w, https://substackcdn.com/image/fetch/$s_!HNvj!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45458cc8-6346-43a6-9777-6b7878ac02e6_2600x1480.png 848w, https://substackcdn.com/image/fetch/$s_!HNvj!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45458cc8-6346-43a6-9777-6b7878ac02e6_2600x1480.png 1272w, https://substackcdn.com/image/fetch/$s_!HNvj!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F45458cc8-6346-43a6-9777-6b7878ac02e6_2600x1480.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!_Jg8!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9a58a3d-f056-4fe2-8985-302f6d8dee47_1800x844.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!_Jg8!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9a58a3d-f056-4fe2-8985-302f6d8dee47_1800x844.png 424w, https://substackcdn.com/image/fetch/$s_!_Jg8!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9a58a3d-f056-4fe2-8985-302f6d8dee47_1800x844.png 848w, https://substackcdn.com/image/fetch/$s_!_Jg8!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9a58a3d-f056-4fe2-8985-302f6d8dee47_1800x844.png 1272w, https://substackcdn.com/image/fetch/$s_!_Jg8!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9a58a3d-f056-4fe2-8985-302f6d8dee47_1800x844.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!_Jg8!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9a58a3d-f056-4fe2-8985-302f6d8dee47_1800x844.png" width="1456" height="683" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/c9a58a3d-f056-4fe2-8985-302f6d8dee47_1800x844.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:683,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:69288,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221213094?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9a58a3d-f056-4fe2-8985-302f6d8dee47_1800x844.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!_Jg8!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9a58a3d-f056-4fe2-8985-302f6d8dee47_1800x844.png 424w, https://substackcdn.com/image/fetch/$s_!_Jg8!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9a58a3d-f056-4fe2-8985-302f6d8dee47_1800x844.png 848w, https://substackcdn.com/image/fetch/$s_!_Jg8!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9a58a3d-f056-4fe2-8985-302f6d8dee47_1800x844.png 1272w, https://substackcdn.com/image/fetch/$s_!_Jg8!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9a58a3d-f056-4fe2-8985-302f6d8dee47_1800x844.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>Source: same file, the per-year premium block and the windows-per-year count beside it. The incumbent&#8217;s advantage over a logistic regression falls in every year of the record, from +0.0188 in 2023 to +0.0040 in 2026, and 2026 is a partial year of 66 windows, so the last row is thin. The other three arms hold up better in this table, and I wouldn&#8217;t push the interpretation further than the table supports (the recomputation caveat at the end of this appendix applies to this row). It is still the row that most argues against paying for a deep model, and it&#8217;s why the logistic control retrains forever.</p><h3>6. The thesis that lost, as filed</h3><p>The cross-sectional round-two design document carries a status header marking it adjudicated and superseded as a live design, with the sections below it preserved as the pre-registration record. The verdict line says it loses to the champion on matched population, -0.0059 with an interval of [-0.0086, -0.0032] over 71 windows, and gets no further funding. That header sits on top of the document that argued for the idea in the first place, so the argument and the result live on the same page.</p><h3>7. The near miss that stayed out</h3><p>The excitation-arm adjudication verdict record covers the excitation-state features.</p><ul><li><p>Challenger mean: 0.3636</p></li><li><p>Incumbent mean: 0.3624</p></li><li><p>Mean difference: +0.0012</p></li><li><p>95% interval: [0.0000, +0.0024]</p></li><li><p>Windows: 792 matched, 61 frozen excluded</p></li><li><p>Bar: more than +0.02</p></li><li><p>Adopted: false</p></li></ul><p>The same file sets its uniform-edge flag to true and carries a reading protocol committed in advance, which says a uniform edge over every member is a possible artifact and should be read skeptically. Its per-year block shows the largest effect in 2026 at +0.0054 over 66 windows, which is the shape a small artifact would take, and why the skeptical reading was committed in advance. One row is marked unavailable. The logistic comparison could not be made for this replay, because the replay was run without baselines.</p><h3>8. What it cost</h3><p><em>Amount.</em></p><ul><li><p>Capacity and ablation variants (four questions, one shared walk): $60</p></li><li><p>Rank-aware loss: $50</p></li><li><p>Cross-sectional attention: $180</p></li><li><p>Annual normalization refresh: $40</p></li><li><p>Ensemble, inference only: $15</p></li><li><p>Excitation features: $50</p></li><li><p>Itemised subtotal (per-question splits are estimates): $395</p></li><li><p>Shared machine time and failed runs, belonging to no single question: $28.12</p></li><li><p><strong>Total, verified against invoices</strong>: <strong>$423.12</strong></p></li></ul><p>Rented on two single-GPU machines, at $6.14 an hour on the first, then $4.29 an hour after I moved to Lambda Labs, a more cost-effective host for the same kind of card. Storage and platform subscription fees are excluded, since they carry the whole lab rather than this campaign. The $180 bought the news that my founding thesis was wrong; the $15 bought the campaign&#8217;s only win.</p><h3>9. The crash logs</h3><p>Seven preserved crash evidence logs are kept in the final box archive.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!f_KF!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff17e846c-e3b7-4748-86b4-533c3e8191df_2600x1389.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!f_KF!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff17e846c-e3b7-4748-86b4-533c3e8191df_2600x1389.png 424w, https://substackcdn.com/image/fetch/$s_!f_KF!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff17e846c-e3b7-4748-86b4-533c3e8191df_2600x1389.png 848w, https://substackcdn.com/image/fetch/$s_!f_KF!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff17e846c-e3b7-4748-86b4-533c3e8191df_2600x1389.png 1272w, https://substackcdn.com/image/fetch/$s_!f_KF!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff17e846c-e3b7-4748-86b4-533c3e8191df_2600x1389.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!f_KF!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff17e846c-e3b7-4748-86b4-533c3e8191df_2600x1389.png" width="1456" height="778" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/f17e846c-e3b7-4748-86b4-533c3e8191df_2600x1389.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:778,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:190262,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221213094?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff17e846c-e3b7-4748-86b4-533c3e8191df_2600x1389.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!f_KF!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff17e846c-e3b7-4748-86b4-533c3e8191df_2600x1389.png 424w, https://substackcdn.com/image/fetch/$s_!f_KF!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff17e846c-e3b7-4748-86b4-533c3e8191df_2600x1389.png 848w, https://substackcdn.com/image/fetch/$s_!f_KF!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff17e846c-e3b7-4748-86b4-533c3e8191df_2600x1389.png 1272w, https://substackcdn.com/image/fetch/$s_!f_KF!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff17e846c-e3b7-4748-86b4-533c3e8191df_2600x1389.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>All seven are the same failure, a CUDA launch failure, surfacing partway through a walk and taking the run with it. Five separate training arms were hit. The numbering carries its own footnote. The archive holds a second and a third crash for one arm and a second for another, with no first crash log kept for either, so at least one further failure happened whose log did not make it into the archive. Rented GPUs fail more often than the pricing page suggests, and the campaign included several hours spent restarting from checkpoints.</p><h3>10. The raw ledgers</h3><p>The final box archive holds one JSON line per walk-forward window, per arm.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!OvZ1!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff4495d27-3e27-48ec-8a96-bd61d322d1b5_1800x844.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!OvZ1!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff4495d27-3e27-48ec-8a96-bd61d322d1b5_1800x844.png 424w, https://substackcdn.com/image/fetch/$s_!OvZ1!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff4495d27-3e27-48ec-8a96-bd61d322d1b5_1800x844.png 848w, https://substackcdn.com/image/fetch/$s_!OvZ1!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff4495d27-3e27-48ec-8a96-bd61d322d1b5_1800x844.png 1272w, https://substackcdn.com/image/fetch/$s_!OvZ1!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff4495d27-3e27-48ec-8a96-bd61d322d1b5_1800x844.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!OvZ1!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff4495d27-3e27-48ec-8a96-bd61d322d1b5_1800x844.png" width="1456" height="683" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/f4495d27-3e27-48ec-8a96-bd61d322d1b5_1800x844.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:683,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:58558,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221213094?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff4495d27-3e27-48ec-8a96-bd61d322d1b5_1800x844.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!OvZ1!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff4495d27-3e27-48ec-8a96-bd61d322d1b5_1800x844.png 424w, https://substackcdn.com/image/fetch/$s_!OvZ1!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff4495d27-3e27-48ec-8a96-bd61d322d1b5_1800x844.png 848w, https://substackcdn.com/image/fetch/$s_!OvZ1!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff4495d27-3e27-48ec-8a96-bd61d322d1b5_1800x844.png 1272w, https://substackcdn.com/image/fetch/$s_!OvZ1!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff4495d27-3e27-48ec-8a96-bd61d322d1b5_1800x844.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>Matched by test date, the four complete ledgers share 838 windows; 45 of those touch the freeze, which leaves the 793 adjudication reports. Each line carries the window&#8217;s train, validation and test dates, the run&#8217;s configuration and the scores at every horizon. The champion&#8217;s ledger also carries the baselines scored on the same window; the other three arms ran without them.</p><p>One caveat for anyone who does recompute. The window filter moves the fourth decimal. Score every pre-freeze row in the champion&#8217;s ledger instead of the matched 793 and it comes out at 0.3621 rather than 0.3624. The worst-ten column and the 2023 and 2024 premium rows fall straight out of the archived files. The challengers&#8217; 2025 premium cells land within about a thousandth of the printed values. The 2026 row comes from the original adjudication ledgers, which this archive does not fully preserve. Only 9 of its 66 windows here carry a baseline, and on those nine all four premiums come out negative. If you find a number here that you cannot reproduce, tell me, and either the appendix or the verdict gets a correction with your name on it.</p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included.</em></p><p>Free subscribers get the build notes below, covering why this breakdown cost almost nothing to make, and the three numbers to print for any slice.</p>
      <p>
          <a href="https://letters.verbatimsigma.com/p/my-model-is-6x-more-precise-on-low">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[A logistic regression scores 0.35. My deep model scores 0.36. Is that worth it?]]></title><description><![CDATA[A plain logistic regression got 96 percent of the way to my deep learning score. What the trees scored told me where the edge comes from.]]></description><link>https://letters.verbatimsigma.com/p/a-logistic-regression-scores-035</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/a-logistic-regression-scores-035</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Mon, 19 Oct 2026 20:30:07 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!WuiF!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6c089a2f-9761-4940-ac8c-29d4435da92c_2600x1639.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Written in July 2026. Anything that has changed since then is marked with an <strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>.</em></p><p>Before I trust any score my model gets, I check what the dumbest credible alternative scores on the same test.</p><p>I hold <strong><span data-color="#3c78d8" style="color: rgb(60, 120, 216);">Lookout</span></strong>, my intraday neural network ensemble, against two controls, each retrained inside the same walk-forward windows as the real thing. One is a plain logistic regression, which retrains in every window and never gets turned off, and the other is a set of gradient-boosted trees, run once as a full replay of 2023. Each control sees exactly the training history the deep model saw and gets graded on the same unseen days, so every comparison is like-for-like.</p><p>Every score below is P@1%FPR, which works like this:</p><ol><li><p>The model ranks every stock it watches by its confidence that a large move is imminent.</p></li><li><p>Walk down that ranking, flagging names, until the false alarms among your flags add up to one percent of all the names that stayed quiet.</p></li><li><p>Stop there and ask what fraction of the flags were correct. That fraction is the score.</p></li></ol><p>It&#8217;s a deliberately harsh grader, because it only looks at the top of the ranking under a tight false-alarm budget, which is where the trades come from.</p><h2>The controls</h2><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!WuiF!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6c089a2f-9761-4940-ac8c-29d4435da92c_2600x1639.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!WuiF!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6c089a2f-9761-4940-ac8c-29d4435da92c_2600x1639.png 424w, https://substackcdn.com/image/fetch/$s_!WuiF!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6c089a2f-9761-4940-ac8c-29d4435da92c_2600x1639.png 848w, https://substackcdn.com/image/fetch/$s_!WuiF!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6c089a2f-9761-4940-ac8c-29d4435da92c_2600x1639.png 1272w, https://substackcdn.com/image/fetch/$s_!WuiF!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6c089a2f-9761-4940-ac8c-29d4435da92c_2600x1639.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!WuiF!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6c089a2f-9761-4940-ac8c-29d4435da92c_2600x1639.png" width="1456" height="918" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/6c089a2f-9761-4940-ac8c-29d4435da92c_2600x1639.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:918,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:226865,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221211855?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6c089a2f-9761-4940-ac8c-29d4435da92c_2600x1639.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!WuiF!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6c089a2f-9761-4940-ac8c-29d4435da92c_2600x1639.png 424w, https://substackcdn.com/image/fetch/$s_!WuiF!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6c089a2f-9761-4940-ac8c-29d4435da92c_2600x1639.png 848w, https://substackcdn.com/image/fetch/$s_!WuiF!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6c089a2f-9761-4940-ac8c-29d4435da92c_2600x1639.png 1272w, https://substackcdn.com/image/fetch/$s_!WuiF!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F6c089a2f-9761-4940-ac8c-29d4435da92c_2600x1639.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>The logistic gets <em>embarrassingly</em> close to the deep model. And the gradient-boosted trees, the tool that wins most tabular-data competitions, land on the logistic&#8217;s score, well below my model&#8217;s.</p><h2>What the trees told me</h2><p>Trees and logistic regressions both look at a stock&#8217;s condition at one moment. My model looks at the sequence: how the last few hours unfolded, rather than only where things stand now. So when the trees (run here off the shelf, with fixed settings) land exactly on the logistic, and the sequence model clears them both, that narrows down where the edge comes from. Two different methods got the same score from that one-moment view, so there probably isn&#8217;t much more in it. Whatever extra the deep model knows most likely comes from the <em>shape</em> of the path.</p><p>That sorts future experiment ideas into two piles. Ideas that add information about the path have a chance, and ideas that only re-weight the one-moment view probably won&#8217;t help.</p><h2>A point or two of precision</h2><p>At the same one-percent false-alarm budget, the gap between <strong><span data-color="#3c78d8" style="color: rgb(60, 120, 216);">Lookout</span></strong> and the logistic buys roughly one to two extra correct catches per hundred flags. That sounds small, and spread evenly across the market it would be. They aren&#8217;t. Positions get taken from the very top of the ranking, where those extra catches concentrate, and a percentage point of precision at the top compounds across hundreds of trading days into a difference that shows up in an account. For now, that&#8217;s worth running a much more complicated model.</p><p><em><strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>: that compounding has not shown up yet. In October 2026 a pre-registered test of trading <strong><span data-color="#3c78d8" style="color: rgb(60, 120, 216);">Lookout</span></strong>&#8217;s signals intraday, after costs, failed, and every variant tested lost money. The rankings still carry real signal (they rank what moves), but the signal has not yet turned into money at intraday horizons.</em></p><p>Every model here answers to a baseline, forever. Subscribe to watch what that does to a trading program.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><h2>A cheap side experiment</h2><p>One afternoon went to a side experiment. Using a classical mathematical summary of price paths, no neural network anywhere at decision time, I recovered roughly 65 percent of the deep model&#8217;s edge over the logistic, running comfortably on a laptop CPU. For now it&#8217;s recorded and set aside, with a pre-registered path to promotion if the economics of my setup ever change.</p><p><em><strong><span data-color="#8071d3" style="color: rgb(128, 113, 211);">October update</span></strong>: it took that path within weeks. The path-summary model was later added as the ensemble&#8217;s sixth member, under the same rules as every other idea.</em></p><h2>The controls stay on</h2><p>Every answer here cost almost nothing, because the controls retrain inside the same harness as the real model, window by window, and the logistic does it automatically and forever. If you run backtests, score a dumb baseline on the same days as your model. It&#8217;s cheap, and it tells you what your score is worth.</p><p>The logistic keeps running as the program trades in public, and I hope it beats the deep model someday. If it does, the harness catches it and I trade a simpler, cheaper machine. If it never does, my trust in the complicated one grows a little with each window.</p><p><em><strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>: the logistic still votes in the ensemble every day, but since the nightly retrains began in July its score on its own is no longer logged, so the nightly comparison this paragraph describes is not running. The July replays above are where that comparison was made.</em></p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included.</em></p><p>Free subscribers get the build notes below, covering how the controls run alongside every training run, and the crash that kept the trees in their own process.</p><h2>How I built it</h2>
      <p>
          <a href="https://letters.verbatimsigma.com/p/a-logistic-regression-scores-035">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[My only clean win cost fifteen dollars]]></title><description><![CDATA[Nine ideas and $423 produced one clean win: a fifteen-dollar average of four tied models and a plain logistic regression.]]></description><link>https://letters.verbatimsigma.com/p/my-only-clean-win-cost-fifteen-dollars</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/my-only-clean-win-cost-fifteen-dollars</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Mon, 19 Oct 2026 20:23:04 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!QV5i!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F35d8acda-ba52-49a6-9b26-35f22bd62e8a_2600x1160.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Written in July 2026. Anything that has changed since then is marked with an <strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>.</em></p><p>Nine ideas and $423 produced exactly one clean win. The winner cost fifteen dollars, contained no new ideas, and took an afternoon.</p><h2>How it works</h2><ol><li><p>Take the four statistically tied variants from the last post.</p></li><li><p>Add the plain logistic control as a fifth member.</p></li><li><p>Average their probability estimates.</p></li></ol><p>Each day, for every name under consideration, all five emit a probability that the name is about to make its move, and the ensemble&#8217;s score is the plain average of the five, with nothing to tune. That average is what <strong><span data-color="#3c78d8" style="color: rgb(60, 120, 216);">Lookout</span></strong>, my intraday neural network ensemble, runs on today (a sixth member has since been added under the same rules).</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!QV5i!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F35d8acda-ba52-49a6-9b26-35f22bd62e8a_2600x1160.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!QV5i!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F35d8acda-ba52-49a6-9b26-35f22bd62e8a_2600x1160.png 424w, https://substackcdn.com/image/fetch/$s_!QV5i!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F35d8acda-ba52-49a6-9b26-35f22bd62e8a_2600x1160.png 848w, https://substackcdn.com/image/fetch/$s_!QV5i!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F35d8acda-ba52-49a6-9b26-35f22bd62e8a_2600x1160.png 1272w, https://substackcdn.com/image/fetch/$s_!QV5i!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F35d8acda-ba52-49a6-9b26-35f22bd62e8a_2600x1160.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!QV5i!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F35d8acda-ba52-49a6-9b26-35f22bd62e8a_2600x1160.png" width="1456" height="650" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/35d8acda-ba52-49a6-9b26-35f22bd62e8a_2600x1160.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:650,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:150264,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221211139?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F35d8acda-ba52-49a6-9b26-35f22bd62e8a_2600x1160.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!QV5i!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F35d8acda-ba52-49a6-9b26-35f22bd62e8a_2600x1160.png 424w, https://substackcdn.com/image/fetch/$s_!QV5i!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F35d8acda-ba52-49a6-9b26-35f22bd62e8a_2600x1160.png 848w, https://substackcdn.com/image/fetch/$s_!QV5i!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F35d8acda-ba52-49a6-9b26-35f22bd62e8a_2600x1160.png 1272w, https://substackcdn.com/image/fetch/$s_!QV5i!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F35d8acda-ba52-49a6-9b26-35f22bd62e8a_2600x1160.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>The clever ideas had to beat the incumbent by +0.02 with the confidence interval clear of zero, judged by the same paired bootstrap that settles everything in this notebook, and none of them did. The average only had to beat its best member with an interval clear of zero. It cleared the clever ideas&#8217; bar too, with room to spare, at roughly an eight percent relative improvement on the precision score, or close to three more correct catches in every hundred flags.</p><h2>Same average, different days</h2><p>This came from looking at daily rankings instead of summary scores. The four variants land on the same average, but they look at different things, are built differently and trained under different constraints, so on any given day they disagree, sometimes sharply, about which names belong at the top of the list. When variant A is right and variant C is wrong, and the next day it flips, an average keeps the signal they share and lets most of the disagreement (mostly noise) cancel out.</p><p>The math is old: the shared signal survives averaging untouched, while errors cancel roughly in proportion to how independent they are. Models this close in skill and this different in construction are the ones worth averaging.</p><h2>Why a weaker member helps</h2><p>The odd member is the fifth one. Adding a logistic regression that scores visibly below the other four sounds like it should make the average worse, and at first I resisted it for that reason. It helps anyway, because its errors are the most different in kind: it fails where the deep variants don&#8217;t, and gets a few right where all four are wrong together. An average gains from mistakes that land in different places, which is why the weakest member still earns its seat. (Probably true of most committees, though people take the point better when it&#8217;s about models.)</p><h2>What &#8220;free&#8221; means here</h2><p>Free here means the production cost is computing a mean, with nothing new to train and no new data. It also has no settings, and settings are where I fool myself, because anything with settings tempts me to adjust them until the number looks better. Something free could fairly be held to a lower bar; this one cleared the full bar anyway, so adopting it was the easiest call of the campaign.</p><p>Cheap wins and expensive lessons both go on the record here. Subscribe and get the ledger as it grows.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><h2>If you train models</h2><p>If you train models on market data and keep getting annoyingly similar scores from different configurations, you may already have the win. Those similar configurations are different, comparably skilled models, and their average is very likely better than the best of them.</p><p>That average went into production. It retrains every night on a small computer at home while I sleep, and its signals are the ones graded in public every morning on <a href="https://gradedtape.com/">The Graded Tape</a>, misses and all. Of everything in the program, it&#8217;s the part I understand best.</p><p><em>October update: the average still retrains every night, and <a href="https://gradedtape.com/">The Graded Tape</a> still grades its signals at fixed paper sizes. Through Oct 16, 2026 it reads -$1,877.13 over 60 trading days. Better ranking hasn&#8217;t turned into money yet.</em></p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included.</em></p><p>Free subscribers get the build notes below, covering the 200-line averager, the bug in its first run, and the habit that would have saved a later test.</p><h2>How I built it</h2><p>The averager trained nothing. It is about 200 lines and ran in roughly 45 minutes of GPU time, which is most of the fifteen dollars. During the long runs each variant saved a checkpoint every 25th window. The script reloaded those, scored 30 windows from before the held-back months, fit a fresh logistic on each window&#8217;s training days, and averaged the five probabilities. Averaging ranks instead gave the same answer to the third decimal.</p>
      <p>
          <a href="https://letters.verbatimsigma.com/p/my-only-clean-win-cost-fifteen-dollars">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[Four models tied. That was the right answer.]]></title><description><![CDATA[Written in July 2026.]]></description><link>https://letters.verbatimsigma.com/p/four-models-tied-that-was-the-right</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/four-models-tied-that-was-the-right</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Mon, 19 Oct 2026 20:16:51 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!L7ru!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe8569361-39e8-4cfc-b96f-aeeba086e3a0_2600x1480.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Written in July 2026. Anything that has changed since then is marked with an <strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>.</em></p><p>I lined up my four best variants of <strong><span data-color="#3c78d8" style="color: rgb(60, 120, 216);">Lookout</span></strong>, my intraday neural network ensemble, for a final scored comparison across more than three years of walk-forward windows, and I expected a clear winner. They tied, and it took me about a day to see what that was telling me.</p><p>A walk-forward window is a small test the model has never seen: train on the days just before a date, score on the unseen stretch that follows, then roll the date forward and do it again. Do that across more than three years and you get roughly 793 separate tests per contender, all on days the model never trained on.</p><h2>The result</h2><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!L7ru!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe8569361-39e8-4cfc-b96f-aeeba086e3a0_2600x1480.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!L7ru!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe8569361-39e8-4cfc-b96f-aeeba086e3a0_2600x1480.png 424w, https://substackcdn.com/image/fetch/$s_!L7ru!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe8569361-39e8-4cfc-b96f-aeeba086e3a0_2600x1480.png 848w, https://substackcdn.com/image/fetch/$s_!L7ru!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe8569361-39e8-4cfc-b96f-aeeba086e3a0_2600x1480.png 1272w, https://substackcdn.com/image/fetch/$s_!L7ru!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe8569361-39e8-4cfc-b96f-aeeba086e3a0_2600x1480.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!L7ru!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe8569361-39e8-4cfc-b96f-aeeba086e3a0_2600x1480.png" width="1456" height="829" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/e8569361-39e8-4cfc-b96f-aeeba086e3a0_2600x1480.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:829,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:211935,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221210026?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe8569361-39e8-4cfc-b96f-aeeba086e3a0_2600x1480.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!L7ru!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe8569361-39e8-4cfc-b96f-aeeba086e3a0_2600x1480.png 424w, https://substackcdn.com/image/fetch/$s_!L7ru!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe8569361-39e8-4cfc-b96f-aeeba086e3a0_2600x1480.png 848w, https://substackcdn.com/image/fetch/$s_!L7ru!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe8569361-39e8-4cfc-b96f-aeeba086e3a0_2600x1480.png 1272w, https://substackcdn.com/image/fetch/$s_!L7ru!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe8569361-39e8-4cfc-b96f-aeeba086e3a0_2600x1480.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>The top four are separated by three thousandths of a point on the precision statistic this program uses (the fraction of the model&#8217;s most confident flags that pay off, with false alarms held to a one percent budget). My pre-registered minimum-detectable-effect bar was 0.02, chosen in advance as roughly the smallest improvement I could reliably tell apart from window-to-window luck, and it sits nearly seven times above the largest gap among the four. Statistically, the four are the same model.</p><h2>The plateau</h2><p>My first reaction to the tie was disappointment. More capacity, fewer feature inputs and a more elaborate architecture all landed on the same score.</p><p>When several different attempts to be smarter all arrive at the same place, the data is telling you the bottleneck is no longer the model. Most of the signal in these feature inputs has already been found, and the useful question changes from &#8220;which architecture is best&#8221; to &#8220;what new information could I feed any of them.&#8221; That question steered the next round of research.</p><p>A lot of the good results in this notebook start out looking like failures. Subscribe to catch the next one.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><h2>Ties go to the simpler model</h2><p>Under the tie-break rule that I set in advance, the win goes to Variant D, the one with an entire training-time protection mechanism removed. Fewer components at the same score is strictly better, since a deleted component can no longer fail, drift or need tuning.</p><p>The deleted mechanism protects against so-called &#8220;catastrophic forgetting&#8221; as a model retrains through time. It&#8217;s a celebrated technique in the academic literature, and I had assumed it mattered. On my data it was a no-op. Plenty of published techniques work on the benchmarks they were invented for and do nothing on a different problem, and the only way to find out is to run the boring experiment.</p><h2>The logistic regression</h2><p>Notice the fifth row. A plain logistic regression, roughly the simplest statistical model that could plausibly apply, gets you about 96 percent of the way to any single deep variant&#8217;s score. The remaining gap is real, and a later post in this arc asks whether it&#8217;s worth the considerable trouble. Still, if you&#8217;re a retail quant about to commit to deep learning, look at that number first. I keep the logistic regression in every comparison <em>because</em> it&#8217;s embarrassing.</p><h2>Next</h2><p>Four models that tie on average still disagree on individual days, and that disagreement is useful. The next post shows what happened when I stopped trying to break the tie and started using it. It&#8217;s the only clean win of the campaign.</p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included.</em></p><p>Free subscribers get the build notes below, covering eight runs on one rented GPU, and everything that broke along the way.</p><h2>How I built it</h2><p>All the variants ran at once on one rented H100. My model is small enough that a single card could feed eight runs, each in its own terminal session and each a flag on the same runner. A smoke test of every variant before launch caught a bug in the two-head version.</p>
      <p>
          <a href="https://letters.verbatimsigma.com/p/four-models-tied-that-was-the-right">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[My founding thesis lost]]></title><description><![CDATA[The idea I built the whole program around lost.]]></description><link>https://letters.verbatimsigma.com/p/my-founding-thesis-lost</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/my-founding-thesis-lost</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Mon, 19 Oct 2026 20:05:31 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!KYmI!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fed238951-f0a8-4653-8f3f-ad1ba1115826_2600x1160.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Written in July 2026. Anything that has changed since then is marked with an <strong><span data-color="#8071d3" style="color: rgb(128, 113, 211);">October update</span></strong>.</em></p><p>One core belief I had going into this project was that <span data-color="#93c47d" style="color: rgb(147, 196, 125);">any model looking at one stock at a time misses too much.</span></p><p>Markets move together. Sectors surge, money rotates, panic spreads from name to name, and it seemed obvious to me that a model able to see the whole cross-section of the market at once would beat one that studies each name in isolation.</p><p>I believed it enough to design the whole program around proving it. <strong><span data-color="#3c78d8" style="color: rgb(60, 120, 216);">Lookout</span></strong>, the per-stock neural network ensemble I built first, was in my mind just the baseline my real idea would someday beat.</p><p>Last week my real idea was tested, and it failed.</p><h2>The test I could not argue with</h2><p>There&#8217;s an easy way to lie to yourself when testing your favorite idea. Build it, run it, and when it loses, decide the <em>test</em> was unfair. Wrong data, wrong period, needed more tuning, surely worth one more attempt&#8230;right?</p><p>Every retry is another chance to win by luck, though. Retry enough and one wins, and you&#8217;ll be tempted to call it a discovery.</p><p>I needed to keep myself from fooling myself, so I locked the following controls before the result existed, in four clauses:</p><ul><li><p><strong>Same windows.</strong> Both models evaluated on the identical walk-forward test windows, 71 matched periods in all.</p></li><li><p><strong>Same population.</strong> The cross-sectional model trades a particular slice of the market, so the incumbent was re-scored on exactly that slice.</p></li><li><p><strong>Same bar as every other idea.</strong> Beat the incumbent by +0.02 with a bootstrap confidence interval clear of zero.</p></li><li><p><strong>Ties go to the incumbent.</strong> New ideas carry the burden of proof.</p></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!KYmI!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fed238951-f0a8-4653-8f3f-ad1ba1115826_2600x1160.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!KYmI!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fed238951-f0a8-4653-8f3f-ad1ba1115826_2600x1160.png 424w, https://substackcdn.com/image/fetch/$s_!KYmI!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fed238951-f0a8-4653-8f3f-ad1ba1115826_2600x1160.png 848w, https://substackcdn.com/image/fetch/$s_!KYmI!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fed238951-f0a8-4653-8f3f-ad1ba1115826_2600x1160.png 1272w, https://substackcdn.com/image/fetch/$s_!KYmI!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fed238951-f0a8-4653-8f3f-ad1ba1115826_2600x1160.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!KYmI!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fed238951-f0a8-4653-8f3f-ad1ba1115826_2600x1160.png" width="1456" height="650" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/ed238951-f0a8-4653-8f3f-ad1ba1115826_2600x1160.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:650,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:138893,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221208290?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fed238951-f0a8-4653-8f3f-ad1ba1115826_2600x1160.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!KYmI!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fed238951-f0a8-4653-8f3f-ad1ba1115826_2600x1160.png 424w, https://substackcdn.com/image/fetch/$s_!KYmI!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fed238951-f0a8-4653-8f3f-ad1ba1115826_2600x1160.png 848w, https://substackcdn.com/image/fetch/$s_!KYmI!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fed238951-f0a8-4653-8f3f-ad1ba1115826_2600x1160.png 1272w, https://substackcdn.com/image/fetch/$s_!KYmI!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fed238951-f0a8-4653-8f3f-ad1ba1115826_2600x1160.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>Look at the confidence interval: the whole of it sits below zero. If it had straddled zero, I could have told myself a story about noise and bad luck on a few windows. Instead the market-aware model was reliably a little worse than the model it was supposed to beat, across 71 test windows from October 2023 into January 2024.</p><h2>Why it lost</h2><p>I only partly know why. Two observations held up when I checked them.</p><p>First, <strong><span data-color="#3c78d8" style="color: rgb(60, 120, 216);">Lookout</span></strong> already saw more of the market than I&#8217;d assumed. Several of its feature inputs already carry market context, so the cross-sectional architecture was partly re-learning things the incumbent already knew, while being much harder to train.</p><p>Second, sharing information across thousands of names also shares their noise, and minute-by-minute stock data carries a lot more noise than signal.</p><h2>The exit</h2><p>The controls I set were pre-registered rules that left nothing for me to negotiate. The idea was retired in writing, in the design doc, with its numbers attached, to prevent myself from quietly bringing it back after a bad month and give it one more try.</p><p>Losing was also useful. Working out <em>why</em> it lost showed where my incumbent is most precise, and that was worth more than the thesis it replaced.</p><p>Retiring the expensive idea freed the budget for a modest one, averaging my four nearly-tied model variants, which delivered the campaign&#8217;s only clean win at almost no production cost. The cross-sectional machinery didn&#8217;t go to waste either. The code, the market-slice tooling and the evaluation harness are all still in use.</p><p>Without the rule, I&#8217;d have kept tuning my favorite idea until it looked good, then traded it. Instead I retired it and moved my watchlist toward the stocks where the incumbent is most precise.</p><p><em><strong><span data-color="#8071d3" style="color: rgb(128, 113, 211);">October update</span></strong>: that finding held up as a ranking result, but no edge has yet survived trading costs, and live trading has been dormant since August. The arc&#8217;s closer and <a href="http://gradedtape.com">The Graded Tape</a> have the numbers.</em></p><p>Losses get the same detail here as wins. Subscribe and watch the next ones come in.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><p>The next letter is a four-way tie, and the rule that picked the winner.</p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included.</em></p><p>Free subscribers get the build notes below, on what it took to build the model that lost and the tool that made the comparison fair.</p><h2>How I built it</h2><p>The cross-sectional model took about four hours of agent time to build, on the evening of July 9, in six commits and roughly 1,100 lines including tests. Most of the work was the data path. My loader served one stock&#8217;s history at a time, and the new model needed every stock at the same minute in one batch. The agent added an index that groups sequences by the minute they end, and a sampler that hands the model one market snapshot per batch.</p>
      <p>
          <a href="https://letters.verbatimsigma.com/p/my-founding-thesis-lost">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[I spent $423 asking nine questions about my trading model. Most of the answers were no.]]></title><description><![CDATA[Nine pre-registered experiments on a rented datacenter GPU. One clean win, two ties, and my founding thesis retired.]]></description><link>https://letters.verbatimsigma.com/p/i-spent-423-asking-nine-questions</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/i-spent-423-asking-nine-questions</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Mon, 19 Oct 2026 19:37:40 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!9bPr!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9b077471-0176-45c4-910f-e3e857f58284_2560x1720.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Written in July 2026. Anything that has changed since then is marked with an <strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>.</em></p><p>Last week I rented a datacenter GPU and spent several days using it to test nine ideas against my own trading model.</p><p>I trade a small account. A couple of years of discretionary trading taught me that the biggest risk in my account was <em>me</em>, believing a backtest because I wanted it to be true. So when I started building Lookout, my neural network ensemble for finding intraday opportunities, I adopted one rule before writing any code.</p><p>Every idea gets a written hypothesis, a success bar and a price tag, all committed before I&#8217;m allowed to see any results. Scientists call this pre-registration.</p><p>New letters arrive by email, free. Subscribe and the next one comes to you.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><h2>The setup</h2><p>All through the trading day, <strong>Lookout</strong> ranks thousands of stocks by the probability that they&#8217;re<span data-color="#8071d3" style="color: rgb(128, 113, 211);"> </span><span data-color="#9184d9" style="color: rgb(145, 132, 217);">about to make a large move within the next hour </span>or so. I grade it with a deliberately harsh statistic. Among the names the model is most confident about, holding false positive rates to a budget of one percent (P@1%FPR), what fraction actually go on to make the move? Random guessing scores about 0.01 on this test, since large intraday moves are rare. My baseline model scores around 0.34 to 0.36, depending on the period. That gap is why I trusted it.</p><p>Then I tested it. Each of the nine ideas below tried either to beat that baseline or to show it wasn&#8217;t real, and each was written down first, with what was being tested, the bar it had to clear (for most, beat the incumbent by more than +0.02 on matched test windows, with a bootstrap confidence interval clear of zero), and what it would cost to find out.</p><h2>The scoreboard</h2><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!9bPr!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9b077471-0176-45c4-910f-e3e857f58284_2560x1720.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!9bPr!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9b077471-0176-45c4-910f-e3e857f58284_2560x1720.png 424w, https://substackcdn.com/image/fetch/$s_!9bPr!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9b077471-0176-45c4-910f-e3e857f58284_2560x1720.png 848w, https://substackcdn.com/image/fetch/$s_!9bPr!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9b077471-0176-45c4-910f-e3e857f58284_2560x1720.png 1272w, https://substackcdn.com/image/fetch/$s_!9bPr!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9b077471-0176-45c4-910f-e3e857f58284_2560x1720.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!9bPr!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9b077471-0176-45c4-910f-e3e857f58284_2560x1720.png" width="1456" height="978" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/9b077471-0176-45c4-910f-e3e857f58284_2560x1720.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:978,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:249490,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221203963?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9b077471-0176-45c4-910f-e3e857f58284_2560x1720.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!9bPr!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9b077471-0176-45c4-910f-e3e857f58284_2560x1720.png 424w, https://substackcdn.com/image/fetch/$s_!9bPr!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9b077471-0176-45c4-910f-e3e857f58284_2560x1720.png 848w, https://substackcdn.com/image/fetch/$s_!9bPr!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9b077471-0176-45c4-910f-e3e857f58284_2560x1720.png 1272w, https://substackcdn.com/image/fetch/$s_!9bPr!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9b077471-0176-45c4-910f-e3e857f58284_2560x1720.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>(The itemized rows sum to about $395; the rest of the $423 was shared machine time and failed runs that belong to no single question. Storage and platform subscription fees are not counted, since they carry the whole lab rather than this one campaign.)</p><p>That table is mostly no&#8217;s, with two ties, a couple of near-misses and one clean win. Each no kept me from trading an idea that didn&#8217;t work.</p><p>The lab publishes either way. Subscribe and watch the next campaign run.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><h2>What the no&#8217;s bought</h2><p>The failed experiments were the valuable ones. Take row 6. That was my founding thesis, the belief this whole project was organized around, and it lost clearly on a matched comparison that was designed before the result existed.</p><p>I know what would have happened without the rule, because I&#8217;ve lived it. I would have tweaked and re-run, tweaked and re-run, and eventually &#8220;discovered&#8221; a configuration that beat the baseline by chance. Then I&#8217;d have put it into production and lost real money on it.</p><p>So the $423 mostly bought the knowledge that<span data-color="#9184d9" style="color: rgb(145, 132, 217);"> the model I already had was </span><em><span data-color="#9184d9" style="color: rgb(145, 132, 217);">better</span></em><span data-color="#9184d9" style="color: rgb(145, 132, 217);"> </span>than eight of my nine ideas, and that the ninth, a plain average of the four tied variants with a simple logistic regression added as a fifth member, was worth +0.029. That&#8217;s roughly an eight percent relative improvement, about three more correct calls in every hundred signals, and its production cost is computing a mean.</p><h2>The rule that made it work</h2><p>One number per idea, written down first. A bar the idea must clear, a confidence interval so a lucky result can&#8217;t clear the bar by chance, and a standing tie-break: when in doubt, the simpler and older thing wins.</p><p>(For the technically inclined, the interval comes from a paired bootstrap. I resample the per-window score differences a thousand times with a fixed seed and read off the middle 95 percent, so &#8220;beats the incumbent&#8221; has to survive a thousand alternate versions of the same test period before it counts.)</p><p>The rule decided every verdict in that table. I set every rule before the results exist, so there&#8217;s nothing to argue about when they arrive.</p><h2>Where this goes</h2><p>The campaign ended with a model I trust, because eight attempts to beat it failed under (what I think are) fair rules. Working out why the thesis lost showed where the model is most precise, which changed my watchlist. The aftershock measurement was too weak to pick stocks, so it went into my risk management instead. And the ensemble, the boring average, is what <strong>Lookout v1.0</strong> runs on today (with a sixth member added since). It retrains itself every night while I sleep, and its signals are graded in public every morning on <a href="https://gradedtape.com/">The Graded Tape</a>, misses and all.</p><p><em><strong><span data-color="#9184d9" style="color: rgb(145, 132, 217);">October update</span></strong>: the ensemble still retrains every night, and the Tape still grades its signals every morning. Its precision held up. It has not made money so far. Through Oct 16, 2026 the Tape reads -$1,877.13 at fixed paper sizes over 60 trading days.</em></p><p>In the next letter my founding thesis loses, and I share why I&#8217;m glad it did.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/p/i-spent-423-asking-nine-questions?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/p/i-spent-423-asking-nine-questions?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included.</em></p><p>For <span data-color="#a64d79" style="color: rgb(166, 77, 121);">The Campaign </span>arc, Free subscribers will continue to get access to the build notes below, covering the 122-line script that judged all nine questions, and how to check one like it by hand.</p><h2>How I built it</h2><p>The script that judged all nine questions is 122 lines of Python and needs nothing beyond the standard library. An agent wrote it in one sitting on the evening of the verdict.</p><p>It reads the file every training run writes, which holds one line of JSON per walk-forward window, with the dates, the settings, the code version and every score. Runs append as they go, so a crash costs one window and the run picks up from the last line.</p><p>The script does five things, in order:</p>
      <p>
          <a href="https://letters.verbatimsigma.com/p/i-spent-423-asking-nine-questions">
              Read more
          </a>
      </p>
   ]]></content:encoded></item><item><title><![CDATA[Arc one: the campaign]]></title><description><![CDATA[Five days of rented GPU time, nine of my best ideas. Most of them lost.]]></description><link>https://letters.verbatimsigma.com/p/arc-one-the-campaign</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/arc-one-the-campaign</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Mon, 19 Oct 2026 19:25:10 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!30vu!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F94883f20-1040-47a2-85cb-e57181911c4f_2600x2000.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Written in July 2026. Anything that has changed since then is marked with an <span data-color="#8071d3" style="color: rgb(128, 113, 211);">October update</span>.</em></p><p>Over the course of five days in July I rented some powerful computing hardware and put nine of my best ideas for <strong><span data-color="#0b5394" style="color: rgb(11, 83, 148);">Lookout</span></strong>, my intraday neural network ensemble, through pre-registered trials. Each one got a hypothesis, a success bar and a budget, all written down before I was allowed to see a single result.</p><p>For most of the nine the bar was the same, to beat the incumbent model by at least +0.02 on a deliberately harsh precision statistic, on identical walk-forward test windows, with a bootstrapped confidence interval entirely clear of zero.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!30vu!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F94883f20-1040-47a2-85cb-e57181911c4f_2600x2000.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!30vu!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F94883f20-1040-47a2-85cb-e57181911c4f_2600x2000.png 424w, https://substackcdn.com/image/fetch/$s_!30vu!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F94883f20-1040-47a2-85cb-e57181911c4f_2600x2000.png 848w, https://substackcdn.com/image/fetch/$s_!30vu!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F94883f20-1040-47a2-85cb-e57181911c4f_2600x2000.png 1272w, https://substackcdn.com/image/fetch/$s_!30vu!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F94883f20-1040-47a2-85cb-e57181911c4f_2600x2000.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!30vu!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F94883f20-1040-47a2-85cb-e57181911c4f_2600x2000.png" width="1456" height="1120" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/94883f20-1040-47a2-85cb-e57181911c4f_2600x2000.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1120,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:162626,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221196722?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F94883f20-1040-47a2-85cb-e57181911c4f_2600x2000.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!30vu!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F94883f20-1040-47a2-85cb-e57181911c4f_2600x2000.png 424w, https://substackcdn.com/image/fetch/$s_!30vu!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F94883f20-1040-47a2-85cb-e57181911c4f_2600x2000.png 848w, https://substackcdn.com/image/fetch/$s_!30vu!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F94883f20-1040-47a2-85cb-e57181911c4f_2600x2000.png 1272w, https://substackcdn.com/image/fetch/$s_!30vu!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F94883f20-1040-47a2-85cb-e57181911c4f_2600x2000.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p></p><p>The next 6 letters describe the story in order:</p><ol><li><p>The scoreboard.</p></li><li><p>My founding thesis loses.</p></li><li><p>A four-way tie, and what it told me.</p></li><li><p>The fifteen-dollar idea that beat the five-hundred-dollar ones.</p></li><li><p>The two simple baselines every model is checked against.</p></li><li><p>The stocks where my model is most precise.</p></li></ol><p>Nine pre-registered verdicts, every number with its interval. Subscribe and follow the <span data-color="#a64d79" style="color: rgb(166, 77, 121);">Campaign</span> in the order it happened.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><p>After this<span data-color="#3c78d8" style="color: rgb(60, 120, 216);"> </span><span data-color="#a64d79" style="color: rgb(166, 77, 121);">Campaign</span><span data-color="#3c78d8" style="color: rgb(60, 120, 216);"> </span>arc I&#8217;ll share the<span data-color="#6aa84f" style="color: rgb(106, 168, 79);"> Machinery </span>arc (why you can trust these results), the physics of how market moves cluster, the portfolio layer, and the live diary, where the whole program trades every day in public. The goal is trading with real money, and this campaign is why I felt confident about getting there.</p><p><em><span data-color="#8071d3" style="color: rgb(128, 113, 211);">October update:</span> <strong><span data-color="#1155cc" style="color: rgb(17, 85, 204);">Lookout</span></strong>&#8217;s rankings held up, but in October a pre-registered test of trading them through the day lost money in its main configuration even before costs, and every variant lost after them. Later letters show how the program found that out and what I did next.</em></p><h2>How I built it</h2><h5><em>Note: In the rest of <span data-color="#a64d79" style="color: rgb(166, 77, 121);">The Campaign</span><span data-color="#38761d" style="color: rgb(56, 118, 29);"> </span>arc (the next 6 posts), the <strong>&#8220;How I built it&#8221;</strong> section sits below a <span data-color="#674ea7" style="color: rgb(103, 78, 167);">Freewall</span> line for subscribers who share their email. Here it&#8217;s free, so you can see what it holds. In The Machinery arc, the &#8220;How I built it&#8221; section will be reserved for paid subscribers. </em></h5><p></p><p>The campaign started on my laptop, the day before I rented anything. I gave a Claude Code session a written plan and a free hand with the model&#8217;s code, with one condition. No powerful GPU machines get rented until this 32GB RAM laptop shows the model is worth renting for.</p><p>The agent wrote tests first, and the tests showed my model couldn&#8217;t reproduce its own April numbers. Four bugs did it:</p><ol><li><p><span data-color="#b45f06" style="color: rgb(180, 95, 6);">The label looked one bar too early.</span></p></li><li><p><span data-color="#b45f06" style="color: rgb(180, 95, 6);">Short sequences were scored from padding.</span></p></li><li><p><span data-color="#b45f06" style="color: rgb(180, 95, 6);">An input clamp flattened some inputs into constants.</span></p></li><li><p><span data-color="#b45f06" style="color: rgb(180, 95, 6);">A misnamed key logged the headline score as zero.</span></p></li></ol><p>With those fixed, I wrote down four <span data-color="#a64d79" style="color: rgb(166, 77, 121);">go/no-go</span><span data-color="#1155cc" style="color: rgb(17, 85, 204);"> </span>gates and ran 113 walk-forward windows on the laptop, which took 2.7 hours. It passed all four. My first rental plan was eight GPUs for about $300. The model turned out small enough that one card would sit mostly idle, so I rented one.</p><p>Over the six days of the campaign that one session made 108 commits and about 18,000 lines of code, roughly 5,000 of them tests.</p><p>Try this: before you rent anything, ask your assistant for tests that pin your metric on a case you can work out by hand, and run them against the code you already have.</p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included.</em></p>]]></content:encoded></item><item><title><![CDATA[The front-row seat (What is Verbatim?)]]></title><description><![CDATA[What Verbatim is, who it's for, where the record stands, and what it costs.]]></description><link>https://letters.verbatimsigma.com/p/the-front-row-seat-what-is-verbatim</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/the-front-row-seat-what-is-verbatim</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Sun, 18 Oct 2026 23:53:45 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!YnUT!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F234f54a0-9881-4722-9a0a-9db70a3c7ef7_2600x1440.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>Verbatim is the newsletter you&#8217;re reading. It&#8217;s a working lab notebook kept by one retail trader who builds custom neural networks to trade, and it publishes every verdict, win or lose.</p><p><a href="https://gradedtape.com/">The Graded Tape</a> is its public record. Every morning a machine grades the signals from <em>Lookout</em>, my custom intraday neural network ensemble, against rules I wrote down in advance, and posts the results, misses included.</p><p>The name means what it says. You get what the machine said, exactly as it said it, including the mornings I wish it had said something else.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!YnUT!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F234f54a0-9881-4722-9a0a-9db70a3c7ef7_2600x1440.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!YnUT!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F234f54a0-9881-4722-9a0a-9db70a3c7ef7_2600x1440.png 424w, https://substackcdn.com/image/fetch/$s_!YnUT!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F234f54a0-9881-4722-9a0a-9db70a3c7ef7_2600x1440.png 848w, https://substackcdn.com/image/fetch/$s_!YnUT!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F234f54a0-9881-4722-9a0a-9db70a3c7ef7_2600x1440.png 1272w, https://substackcdn.com/image/fetch/$s_!YnUT!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F234f54a0-9881-4722-9a0a-9db70a3c7ef7_2600x1440.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!YnUT!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F234f54a0-9881-4722-9a0a-9db70a3c7ef7_2600x1440.png" width="1456" height="806" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/234f54a0-9881-4722-9a0a-9db70a3c7ef7_2600x1440.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:806,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:214897,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221065325?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F234f54a0-9881-4722-9a0a-9db70a3c7ef7_2600x1440.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!YnUT!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F234f54a0-9881-4722-9a0a-9db70a3c7ef7_2600x1440.png 424w, https://substackcdn.com/image/fetch/$s_!YnUT!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F234f54a0-9881-4722-9a0a-9db70a3c7ef7_2600x1440.png 848w, https://substackcdn.com/image/fetch/$s_!YnUT!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F234f54a0-9881-4722-9a0a-9db70a3c7ef7_2600x1440.png 1272w, https://substackcdn.com/image/fetch/$s_!YnUT!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F234f54a0-9881-4722-9a0a-9db70a3c7ef7_2600x1440.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h2>The goal</h2><p>I want to grow my own trading account faster than an index fund would, after every trading cost, consistently, at the size my account actually is.</p><h4>It sounds easy, but most trading systems fail this test.</h4><p>Plenty of systems beat the market in a backtest, before costs, at an account size that isn&#8217;t real / liquid. Very few beat just buying the index and sitting on it once the spreads, the fees and the bad fills are paid, with drawdowns no worse than the index&#8217;s. That&#8217;s the only scoreboard I keep here.</p><h2>Who this is for</h2><p>Me, about four years ago.</p><p>You know markets well enough to have real ideas, and you have some money you&#8217;d put behind an idea you thought might work. What you could never do was build anything that reliably tested a quant thesis, because of the technical barrier. Getting from a market hunch to a validated model took engineering, and engineering was a whole profession you didn&#8217;t have.</p><p>That&#8217;s changed. The tools that let a curious non-specialist build, test and pick apart a real model showed up more or less all at once, over the last 4 years. AI coding assistants accelerated the process, and you&#8217;ve started to suspect you could learn to do this yourself.</p><p>I know the feeling. I started this project without ever having set up a proper development environment and built every piece of it alongside those assistants as they grew up. <a href="https://letters.verbatimsigma.com/p/how-i-built-it-from-copy-pasting">How I built it</a> tells that story.</p><p>Meanwhile, the newsletters that sell trading signals give you almost nothing you can verify. The track records are filtered, the hot streaks start exactly when the marketing does, and nobody tells you what the results were compared against, on what data, or whether the rules were set before or after the fact. You want to understand a pick well enough to make your own.</p><p>If that&#8217;s you, I am writing this newsletter for you.</p><h2>What you get</h2><p>A front-row seat to a one-person quant lab building custom deep learning and neural network trading models in the open. Every verdict comes with the bar it had to clear, written down before I saw the answer, plus what the test cost in real dollars ($USD) and how it came out, whether I like the result or not.</p><p>If a real edge ever shows up in these pages, the reason to believe it will be the long, visible pile of entries where the machinery said &#8220;you&#8217;re wrong,&#8221; and I wrote it down here anyway. You&#8217;ll also be able to see exactly how it was built, so you can build your own if you want to.</p><h2>Where things stand</h2><p>I started writing this newsletter archive in July 2026, and each letter is dated, so you can read the story in the order it happened.</p><h2>As of October, one of my custom ensemble neural models, <em>Lookout,</em> <span data-color="#6aa84f" style="color: rgb(106, 168, 79);">ranks the market well.</span></h2><p>Its signals separate the names <em>about to move</em> from the ones that aren&#8217;t, and the first arcs show how I know, with receipts.</p><p>What <em>Lookout</em> hasn&#8217;t done is make me money. In October 2026, I ran a pre-registered test of trading those signals through the day. The main version lost money even before costs, and after costs every variant lost. <a href="https://gradedtape.com/">The Graded Tape</a> tells the same story from a different angle. Hover or tap any day on its charts and you&#8217;ll see a record that is mostly red.</p><p>I traded the signals with real money for a few days in August 2026, by hand and at half size, and then stopped. Every bot I run now trades on a broker&#8217;s paper simulator, under rules set in advance, and every model is judged against the goal above, on days it has never seen.</p><p>So the program turned. Everything I built for the intraday work now feeds a <span data-color="#3c78d8" style="color: rgb(60, 120, 216);">second model</span> designed to trade once a day and aimed straight at beating the index. It&#8217;s in research now, in the open, and its failures get logged the same way the first model&#8217;s did. You&#8217;ll read about it as it happens.</p><p>What I&#8217;m offering right now is the build toward an edge. There&#8217;s no proven edge here you could get rich on this afternoon, and pretending otherwise would be easier to sell&#8230;<em>and a lie</em>.</p><p>I&#8217;d rather sell the climb and price the rest when we get there.</p><h2>What a seat costs</h2><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!HTP5!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa118e38a-df6a-4f00-b0f6-78ad7335648d_2600x1800.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!HTP5!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa118e38a-df6a-4f00-b0f6-78ad7335648d_2600x1800.png 424w, https://substackcdn.com/image/fetch/$s_!HTP5!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa118e38a-df6a-4f00-b0f6-78ad7335648d_2600x1800.png 848w, https://substackcdn.com/image/fetch/$s_!HTP5!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa118e38a-df6a-4f00-b0f6-78ad7335648d_2600x1800.png 1272w, https://substackcdn.com/image/fetch/$s_!HTP5!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa118e38a-df6a-4f00-b0f6-78ad7335648d_2600x1800.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!HTP5!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa118e38a-df6a-4f00-b0f6-78ad7335648d_2600x1800.png" width="1456" height="1008" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/a118e38a-df6a-4f00-b0f6-78ad7335648d_2600x1800.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1008,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:225023,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221065325?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa118e38a-df6a-4f00-b0f6-78ad7335648d_2600x1800.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!HTP5!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa118e38a-df6a-4f00-b0f6-78ad7335648d_2600x1800.png 424w, https://substackcdn.com/image/fetch/$s_!HTP5!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa118e38a-df6a-4f00-b0f6-78ad7335648d_2600x1800.png 848w, https://substackcdn.com/image/fetch/$s_!HTP5!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa118e38a-df6a-4f00-b0f6-78ad7335648d_2600x1800.png 1272w, https://substackcdn.com/image/fetch/$s_!HTP5!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa118e38a-df6a-4f00-b0f6-78ad7335648d_2600x1800.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p><strong><span data-color="#674ea7" style="color: rgb(103, 78, 167);">Free: </span></strong>Every letter, every verdict, and The Graded Tape. That&#8217;s everything you need to judge whether this is the real thing, and it costs nothing. </p><p><strong><span data-color="#674ea7" style="color: rgb(103, 78, 167);">Letters: </span></strong>twenty dollars a month or two hundred a year. Everything above, plus three things. A &#8220;How I built it&#8221; section at the end of many of my posts: the tools, the build decisions, what broke and one step you can copy. The working papers that close each arc from arc 2 onward: the signed decision rules, the full result tables, every confidence interval, and the ugliest numbers in the pack. And the monthly research log. The letters tell you what happened, and the paid sections show you how to check it and how to build it. This tier pays for the data, the rented compute, the nightly machinery and the hours that run the lab.</p><p>If the record earns it, the price may go up for new subscribers. Anyone already in keeps the rate they joined at.</p><p><strong><span data-color="#674ea7" style="color: rgb(103, 78, 167);">Founding:</span> </strong>four hundred dollars a year, for readers who want to back the lab harder. Your rate is locked for as long as you subscribe, and you get first claim on anything that opens later, if and when a model passes its sealed test.</p><p><strong><span data-color="#674ea7" style="color: rgb(103, 78, 167);">Later: </span></strong>and only if it&#8217;s earned. If a model passes its sealed test, a single final reading on data it has never touched, I&#8217;ll lay out what a seat that follows its signals would look like and price it then, in public, against the evidence. Same for the dashboards and daily model outputs I use myself. I&#8217;m naming them now so you can see where this could go, and I&#8217;m not putting a date on any of it.</p><p>To follow along, subscribe free and each new letter lands in your inbox, the Friday note included. The paid tiers are on the same page whenever you want them.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><p>If you&#8217;re new, start with <a href="https://letters.verbatimsigma.com/p/what-this-is-and-why-it-exists">What this is, and why it exists</a>, and keep <a href="https://gradedtape.com/">The Graded Tape</a> open to check my work.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/p/the-front-row-seat-what-is-verbatim?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/p/the-front-row-seat-what-is-verbatim?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included. The Graded Tape is a paper, counterfactual record at notional sizes, not realized trading, and not audited.</em></p>]]></content:encoded></item><item><title><![CDATA[How I built it: from copy-pasting ChatGPT code to a team of agents]]></title><description><![CDATA[From pasting ChatGPT code into files on a new iMac in 2022 to nine repositories and a team of AI agents, and why the rules still matter more than the tools.]]></description><link>https://letters.verbatimsigma.com/p/how-i-built-it-from-copy-pasting</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/how-i-built-it-from-copy-pasting</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Sun, 18 Oct 2026 23:24:53 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!1hWy!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff3ebda84-845d-45b3-86ab-b4de83cca13f_2600x1680.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<h3><span data-color="#674ea7" style="color: rgb(103, 78, 167);">I&#8217;m not an engineer. A single retail trader can run a lab like this </span><em><span data-color="#674ea7" style="color: rgb(103, 78, 167);">because</span></em><span data-color="#674ea7" style="color: rgb(103, 78, 167);"> of AI coding tools, and I suspect some of you are here for that part at least as much as for the markets.</span></h3><h2>November 2022</h2><p>It started in November 2022, with a brand new M1 iMac and the first public release of ChatGPT. I knew enough to be dangerous up to that point. Product development work and years of hobby projects left me comfortable with HTML, CSS and JavaScript, some Python and Java, and just enough Processing and Arduino to have some fun. What I had never done was build software the way engineers do. I had never opened VS Code, didn&#8217;t know what a virtual environment was, had never heard of a package manager, and had never made a git commit.</p><p>So my first workflow was an obvious one. I described what I wanted, copied the code ChatGPT gave me into a file, ran it, pasted the error back, and went around again&#8230;and again&#8230;and again&#8230;and again. It worked often enough to keep me going. I could see where generative AI was heading, and I rode the wave early.</p><h2>The long apprenticeship</h2><p>The next couple of years were an apprenticeship in everything around code and software engineering. I learned the software development lifecycle, serverless functions on AWS (and later on DigitalOcean), S3 storage, fetching data on a schedule, automations that sent me emails, data transformation, data science workflows, thesis writing, and the quant finance that all of it was supposed to serve. <span data-color="#674ea7" style="color: rgb(103, 78, 167);">I started over many times. </span>About two years ago I started over once more, this time with the rules from the second letter, and those rules still govern everything here.</p><p>By the first half of 2024 the project had real plumbing. I had bought API subscriptions to Polygon.io (now Massive) for market data, including its bulk flat files, and to QuiverQuant for alternative data, and I had fetchers pulling both automatically, with deploy and scheduling scripts for AWS Lambda and a plan to move to DigitalOcean and Paperspace.</p><p>The assistants of that era, ChatGPT first and then Cody from Sourcegraph, could only take in one file, or about a hundred lines, at a time. So I assembled context by hand. A few of those prompt files survive in the old project: a list of file names, a sentence reading <em>&#8220;Today&#8217;s work will solely be related to the files... within this directory,&#8221;</em> and ten rows of data pasted inline so the assistant could see what a table looked like. </p><p><span data-color="#3d85c6" style="color: rgb(61, 133, 198);">In effect&#8230;I was the context window.</span></p><h2>June 2025: the shift</h2><p>By late 2024, I had moved my work into a dev container, a development environment defined in a configuration file so it can be rebuilt identically anywhere, running in GitHub Codespaces. In June 2025 I installed Claude Code, Anthropic&#8217;s command-line coding agent, and it changed how I worked. Cursor and Copilot had been around for a while, and I used both, but neither fit the way I worked across a whole codebase.</p><p>Instead of me feeding it a hundred lines at a time, the agent could read the whole project, run the code, see the error itself and try again.</p><p>In early July I finished its standing instructions file, and on July 20 the first commit landed with Claude listed as a co-author.</p><p>On July 26, 2025 I committed my first agent &#8220;team,&#8221; a set of personas the agent could step into. There was a data architect, a quant ML code reviewer, and personas for UX, product and monetization. </p><p>The next day&#8217;s commit message begins &#8220;ran out of tokens&#8221; and goes on, &#8220;Next time I startup claude, I may need a full review by the team.&#8221; Within a day I&#8217;d found out what it could do and what it cost. By August, most commits carried Claude as a co-author.</p><h2>Rented and owned</h2><p>By September 2025 I was renting H100s, the datacenter GPUs, by the hour on Paperspace; a commit that month recorded a 24x speedup from tuning for them. At the end of October a Mac mini joined my desk as an always-on home server, and the nightly automation moved onto it.</p><p>From November 2025 into February 2026 I stepped away.</p><p>In March 2026 I came back and split my one big repository into separate ones for the<span data-color="#674ea7" style="color: rgb(103, 78, 167);"> data fetchers</span>, the <span data-color="#45818e" style="color: rgb(69, 129, 142);">feature pipeline</span> and <span data-color="#a64d79" style="color: rgb(166, 77, 121);">the model</span>, dropped the DigitalOcean dependencies, and moved to self-hosting. Over the following weeks the fetchers got containers, data manifests and a nightly schedule.</p><h2>The tall bar</h2><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!1hWy!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff3ebda84-845d-45b3-86ab-b4de83cca13f_2600x1680.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!1hWy!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff3ebda84-845d-45b3-86ab-b4de83cca13f_2600x1680.png 424w, https://substackcdn.com/image/fetch/$s_!1hWy!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff3ebda84-845d-45b3-86ab-b4de83cca13f_2600x1680.png 848w, https://substackcdn.com/image/fetch/$s_!1hWy!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff3ebda84-845d-45b3-86ab-b4de83cca13f_2600x1680.png 1272w, https://substackcdn.com/image/fetch/$s_!1hWy!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff3ebda84-845d-45b3-86ab-b4de83cca13f_2600x1680.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!1hWy!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff3ebda84-845d-45b3-86ab-b4de83cca13f_2600x1680.png" width="1456" height="941" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/f3ebda84-845d-45b3-86ab-b4de83cca13f_2600x1680.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:941,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:227115,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221062340?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff3ebda84-845d-45b3-86ab-b4de83cca13f_2600x1680.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!1hWy!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff3ebda84-845d-45b3-86ab-b4de83cca13f_2600x1680.png 424w, https://substackcdn.com/image/fetch/$s_!1hWy!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff3ebda84-845d-45b3-86ab-b4de83cca13f_2600x1680.png 848w, https://substackcdn.com/image/fetch/$s_!1hWy!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff3ebda84-845d-45b3-86ab-b4de83cca13f_2600x1680.png 1272w, https://substackcdn.com/image/fetch/$s_!1hWy!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff3ebda84-845d-45b3-86ab-b4de83cca13f_2600x1680.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>July 2026 is the tall bar I the bar cart, about 690 commits in a single month, more than the whole year before it combined (sort of, because I switch from Github and Codespaces to self-hosted git). That month I built a dashboard, started these letters, and wrote the code behind the public track-record sites. It is also when the agents began working in teams, with several at once on one job and separate agents set loose on the finished work to find its mistakes.</p><p>By August 2026 an audit could put fifteen reviewer agents over the code at the same time and come back with seventy findings. I don&#8217;t recommend it; you&#8217;ll eat your tokens real quick.</p><p>Today the program spans nine repositories and nearly 2,200 commits, more than 1,900 of them co-authored by Claude. Agents write the code and run the studies, a second and stronger model reviews what they produced, and I decide what gets built and which results count.</p><p>If you want to watch one person and a team of agents build this in the open, with every verdict published, subscribe and the next one arrives in your inbox.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><h2>What the tools closed</h2><p>In 2022, the thing standing between my market ideas and a tested model was engineering I couldn&#8217;t do. The tools closed that gap, and the chart above shows how.</p><p>They only closed it after I was willing to start over many times and to learn the tooling for real: what a virtual environment is for, why version control matters, the benefits of containerization, and how to read the error the agent has just read. An agent that can write a thousand lines in a minute multiplies whatever understanding you bring to it, and I have watched it multiply a misunderstanding at the same speed. Most of the guardrails in this program, the audits, the locked-down home server, the reviewer model, are there to catch the second kind.</p><h2>Where the money stands</h2><p>None of this machinery has produced money yet. Through October 16, 2026, The Graded Tape reads <span data-color="#980000" style="color: rgb(152, 0, 0);">-$1,877.13</span> at fixed paper sizes (thank goodness), and in October a pre-registered test of trading Lookout&#8217;s signals through the day lost money. The tools make it cheap to ask a question and just as cheap to fool yourself with the answer. The rules from the second letter are what stop me.</p><p>If you know someone building things with AI coding tools, or wondering whether they could, this one might be worth passing along.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/p/how-i-built-it-from-copy-pasting?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/p/how-i-built-it-from-copy-pasting?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included.</em></p>]]></content:encoded></item><item><title><![CDATA[The rules of this notebook]]></title><description><![CDATA[Hold me to &#8216;em, Chat]]></description><link>https://letters.verbatimsigma.com/p/the-rules-of-this-notebook</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/the-rules-of-this-notebook</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Sun, 18 Oct 2026 23:02:45 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!QdaL!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2601e1bd-5b72-4968-9bfc-77deeb86a4da_2600x1320.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>Before my first real post, I want to share my operating rules where everyone can see them.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!QdaL!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2601e1bd-5b72-4968-9bfc-77deeb86a4da_2600x1320.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!QdaL!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2601e1bd-5b72-4968-9bfc-77deeb86a4da_2600x1320.png 424w, https://substackcdn.com/image/fetch/$s_!QdaL!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2601e1bd-5b72-4968-9bfc-77deeb86a4da_2600x1320.png 848w, https://substackcdn.com/image/fetch/$s_!QdaL!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2601e1bd-5b72-4968-9bfc-77deeb86a4da_2600x1320.png 1272w, https://substackcdn.com/image/fetch/$s_!QdaL!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2601e1bd-5b72-4968-9bfc-77deeb86a4da_2600x1320.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!QdaL!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2601e1bd-5b72-4968-9bfc-77deeb86a4da_2600x1320.png" width="1456" height="739" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/2601e1bd-5b72-4968-9bfc-77deeb86a4da_2600x1320.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:739,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:207119,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221062060?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2601e1bd-5b72-4968-9bfc-77deeb86a4da_2600x1320.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!QdaL!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2601e1bd-5b72-4968-9bfc-77deeb86a4da_2600x1320.png 424w, https://substackcdn.com/image/fetch/$s_!QdaL!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2601e1bd-5b72-4968-9bfc-77deeb86a4da_2600x1320.png 848w, https://substackcdn.com/image/fetch/$s_!QdaL!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2601e1bd-5b72-4968-9bfc-77deeb86a4da_2600x1320.png 1272w, https://substackcdn.com/image/fetch/$s_!QdaL!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2601e1bd-5b72-4968-9bfc-77deeb86a4da_2600x1320.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h2>What I publish</h2><p>Verdicts, either way. Every verdict appears with its success bar, written down before the test ran, its cost, and its result, whether the result is good or bad.</p><p>Confidence intervals, always. When I say a model improved by +0.029, I will show you the range the improvement lives in, so you can see whether the claim clears the noise or got lucky. The range comes from a paired bootstrap: resample the per-window score differences a thousand times, with a fixed random seed so the interval itself is reproducible, and read off the middle 95 percent. A claim has to survive a thousand alternate versions of its own test period before it appears here as a win.</p><p>Real costs in real dollars. GPU rentals, data, the price of finding things out. Account performance appears in percentages; the account is small and its size is not the interesting part.</p><p>Incidents. When the pipeline breaks, when I mess up my own data with a careless command, when fifteen days go missing from a dataset, you get the forensic report. The failures are where I learn the most.</p><p>Fills. Every completed round trip on the paper simulator is logged against what the model expected, and once live trading begins, real fills get the same treatment. The gap between paper and practice (traders call it slippage) is a number that almost never gets published. I&#8217;m going to publish mine, to be part of the solution and all that.</p><p>Rules first, receipts forever on <a href="https://gradedtape.com/">The Graded Tape</a>. Subscribe to Verbatim and hold me to every one of them as the program heads for the live market.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><h2>What stays private, for now</h2><p>The Lookout model&#8217;s exact inputs, thresholds, and screens. Any edge I find is what pays for the lab, so those stay private. When something stays private, I&#8217;ll openly share why, instead of writing around it.</p><h2>The rules the program itself runs under</h2><p>Each of these will get its own post in time:</p><ol><li><p>Every verdict is decided by a bar written down before the test runs, with a budget.</p></li><li><p>All scoring is walk-forward: a model trains only on data that existed before each test window, is graded on the unseen days that follow, and then the window rolls ahead and the exercise repeats through the years. No verdict rests on a question the model had already seen answered.</p></li><li><p>Every model has to outrun boring baselines tested on the same windows.</p></li><li><p>A slice of recent history stays sealed until one scheduled final reading.</p></li><li><p>Ties go to the simpler method, and new ideas carry the burden of proof.</p></li><li><p>No result counts as evidence until the process that produced it can be replayed from scratch.</p></li></ol><p>None of this is exotic; it&#8217;s roughly what any decent research lab does. I just do it on a small trading account, in public.</p><h2>Where we are headed</h2><p>The arcs ahead, roughly in order:</p><ol><li><p>The campaign. Nine of my best ideas meet the rules, and the rules win.</p></li><li><p>The machinery. The checks that keep the experiments honest: walk-forward testing, a sealed quarter of data, the pre-registration template, and a failure published on purpose.</p></li><li><p>The physics. Big market moves cluster like earthquake aftershocks, measurably, and the mathematics seismologists use for aftershocks describes them well. As a way to pick stocks it failed; it found its job sizing my positions instead, shrinking them when the market runs hot.</p></li><li><p>The turn. What happened when Lookout&#8217;s signals met real trading costs, and why the program re-aimed at a plainer goal: grow the account faster than an index fund would, after every cost.</p></li><li><p>The models after Lookout. Ballast, a portfolio allocator that decides what to hold and at what size, failed its own full-history test and now runs on paper as a check. Swell, a network that trains and trades once a day, is being put through its gates now, and has earned nothing yet.</p></li><li><p>The diary. Whatever earns its way there trades live and writes down what happened, every morning, indefinitely.</p></li></ol><p>Somewhere along the way, the instruments this program runs on may become available to subscribers as research tools. As a reminder, this is a notebook for education and research, never advice. Hold me to all of it.</p><p>From arc two on, the working papers sit below the line for paid subscribers, twenty dollars a month or two hundred a year. The founding tier, four hundred a year, is for readers who want to fund the lab harder, with the rate locked and first claim on anything that opens later, if and when a model passes its sealed test.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included.</em></p>]]></content:encoded></item><item><title><![CDATA[How I got here]]></title><description><![CDATA[A short story of the afternoon I stopped accepting evidence from myself that I would never accept from a junior data analyst.]]></description><link>https://letters.verbatimsigma.com/p/how-i-got-here</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/how-i-got-here</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Sun, 18 Oct 2026 22:46:51 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!VFWt!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa23f2236-20fb-4cb7-af60-0413379a99d9_2600x1680.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>Like a lot of systematic traders, I started as a retail trader with a theory. I had several theories, one per season, and the (expensive and ugly) brokerage statements to show for it.</p><p>It took me a long time to see the pattern.</p><ol><li><p>There was the stretch where I only took setups from a book whose author seemed trustworthy&#8230; on account of his expensive vest and collared shirt.</p></li><li><p>There was my momentum phase, which taught me that I could correctly identify momentum&#8230;shortly after it ended.</p></li><li><p>There was a long period of what I would have called instinct, which in retrospect was a disciplined process for buying things that had just finished going up.</p></li></ol><p>None of this was overly risky. I traded small, which I&#8217;d love to credit to wisdom, though the account balance enforced it. But the lessons added up, and somewhere in there I noticed the common factor in every losing theory: the person holding it. My selective memory made me attribute the wins to anything except luck, and my backtests had a way of confirming whatever I had already decided.</p><p>Over the years, I learned more about good data science practice, and it eventually occurred to me that I would never accept from a junior data analyst the standard of evidence I routinely accepted from myself about trades. If an analyst told me a model was good, I would ask, &#8220;Good compared to what baseline, on what held-out data, decided by what threshold, chosen before or after you saw the results?&#8221; I had never once asked myself those questions with my own money on the line.</p><p>So I started over, with rules. Every idea written down before testing, with a bar it must clear and a price for finding out the answers.</p><p>&#8220;So what?&#8221; and &#8220;Compared to what?&#8221; and &#8220;Why did that work?&#8221; became my favorite questions.</p><p>Two of the rules got stricter than the rest:</p><ol><li><p>All scoring would now be done &#8220;walk-forward,&#8221; meaning the model trains only on data from before each test window, gets graded on the unseen days that follow, and then the window rolls ahead and the whole exercise repeats, year after year, so no score ever comes from a question whose answer the model had already been shown.</p></li><li><p>I moved to testing against &#8220;held out&#8221; data that the process is forbidden to touch until a single, final, pre-scheduled reading. No improvement counts unless its confidence interval, built by resampling the results, sits clear of zero. Failures get recorded where future-me can see them, because future-me has a documented history of creative forgetting.</p></li></ol><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!VFWt!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa23f2236-20fb-4cb7-af60-0413379a99d9_2600x1680.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!VFWt!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa23f2236-20fb-4cb7-af60-0413379a99d9_2600x1680.png 424w, https://substackcdn.com/image/fetch/$s_!VFWt!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa23f2236-20fb-4cb7-af60-0413379a99d9_2600x1680.png 848w, https://substackcdn.com/image/fetch/$s_!VFWt!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa23f2236-20fb-4cb7-af60-0413379a99d9_2600x1680.png 1272w, https://substackcdn.com/image/fetch/$s_!VFWt!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa23f2236-20fb-4cb7-af60-0413379a99d9_2600x1680.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!VFWt!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa23f2236-20fb-4cb7-af60-0413379a99d9_2600x1680.png" width="1456" height="941" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/a23f2236-20fb-4cb7-af60-0413379a99d9_2600x1680.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:941,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:200064,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221059112?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa23f2236-20fb-4cb7-af60-0413379a99d9_2600x1680.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!VFWt!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa23f2236-20fb-4cb7-af60-0413379a99d9_2600x1680.png 424w, https://substackcdn.com/image/fetch/$s_!VFWt!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa23f2236-20fb-4cb7-af60-0413379a99d9_2600x1680.png 848w, https://substackcdn.com/image/fetch/$s_!VFWt!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa23f2236-20fb-4cb7-af60-0413379a99d9_2600x1680.png 1272w, https://substackcdn.com/image/fetch/$s_!VFWt!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa23f2236-20fb-4cb7-af60-0413379a99d9_2600x1680.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>The surprise was how much of what I thought made my good trades didn&#8217;t survive the rules, and how quickly the process started producing things worth keeping anyway.</p><p>This letter is the running audit trail of those rules at work. Subscribe and get every entry, verdicts included.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><p></p><p>The first arc is the campaign where nine of my best ideas met the rules and mostly lost. Lookout, the neural ensemble that came out of it, earned my trust as a forecaster long before it had a chance to earn a dollar, and it hasn&#8217;t earned one yet. After that come the machinery, some physics borrowed from earthquake research, and the live tape, where the trust gets tested in dollars.</p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included.</em></p>]]></content:encoded></item><item><title><![CDATA[What this is, and why it exists]]></title><description><![CDATA[A one-person quant lab building custom neural networks to trade short-term stock moves, in public. Every experiment posted here before it runs. Machine-graded, published unedited, losses included.]]></description><link>https://letters.verbatimsigma.com/p/what-this-is-and-why-it-exists</link><guid isPermaLink="false">https://letters.verbatimsigma.com/p/what-this-is-and-why-it-exists</guid><dc:creator><![CDATA[Verbatim]]></dc:creator><pubDate>Sun, 18 Oct 2026 22:17:01 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!3RQS!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fede08242-ea7e-4edf-8094-1fb2c3d4a30c_2600x1040.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>This is a lab notebook kept by one retail trader who builds custom neural networks to find short-horizon stock opportunities. Every verdict is published, including the ideas that fail.</p><p>If you have spent any time around trading newsletters, you already know the gimmicks of the genre. A confident voice + a chart with lots of green and arrows pointing up + a historical track record that begins when the good streak did, and a sense that the author has never once been wrong.</p><p>I have read a bunch of these, and in my earlier trading years I subscribed fervently. One day, it clicked with me that none of them could be validated from the outside, by the audience.</p><p>So, I stopped subscribing and <span data-color="#6aa84f" style="color: rgb(106, 168, 79);">decided to build, instead</span>.</p><p>Verbatim tries the opposite of these stereotypical newsletters. Every claim here comes out of a process built to stop me from fooling myself that I&#8217;m good at trading, and I run that process in public, where you can check it.</p><h2>How it works</h2><p>Before any idea gets a verdict (or any of my cash), I write down my hypothesis, the exact success bar it must clear, and what the test will cost. Scientists call this pre-registration. It exists because anyone allowed to peek at the answer first will reliably find whatever they were hoping to find - aka, they fooled themselves.</p><p>I&#8217;d rather keep my bar consistent, and high.</p><p>Two things have to happen before a win counts. The experiment has to clear a minimum effect size, and its confidence interval has to sit clear of zero.</p><p>I build that interval by resampling the test results with a fixed random seed, so even the interval is reproducible.</p><p>When an idea fails, I publish it here with the same detail as a success, with the numbers and confidence intervals attached.</p><p>When something breaks, which happens all the time, I&#8217;ll post about it too.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!3RQS!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fede08242-ea7e-4edf-8094-1fb2c3d4a30c_2600x1040.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!3RQS!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fede08242-ea7e-4edf-8094-1fb2c3d4a30c_2600x1040.png 424w, https://substackcdn.com/image/fetch/$s_!3RQS!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fede08242-ea7e-4edf-8094-1fb2c3d4a30c_2600x1040.png 848w, https://substackcdn.com/image/fetch/$s_!3RQS!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fede08242-ea7e-4edf-8094-1fb2c3d4a30c_2600x1040.png 1272w, https://substackcdn.com/image/fetch/$s_!3RQS!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fede08242-ea7e-4edf-8094-1fb2c3d4a30c_2600x1040.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!3RQS!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fede08242-ea7e-4edf-8094-1fb2c3d4a30c_2600x1040.png" width="1456" height="582" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/ede08242-ea7e-4edf-8094-1fb2c3d4a30c_2600x1040.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:582,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:131866,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://letters.verbatimsigma.com/i/221056527?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fede08242-ea7e-4edf-8094-1fb2c3d4a30c_2600x1040.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!3RQS!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fede08242-ea7e-4edf-8094-1fb2c3d4a30c_2600x1040.png 424w, https://substackcdn.com/image/fetch/$s_!3RQS!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fede08242-ea7e-4edf-8094-1fb2c3d4a30c_2600x1040.png 848w, https://substackcdn.com/image/fetch/$s_!3RQS!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fede08242-ea7e-4edf-8094-1fb2c3d4a30c_2600x1040.png 1272w, https://substackcdn.com/image/fetch/$s_!3RQS!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fede08242-ea7e-4edf-8094-1fb2c3d4a30c_2600x1040.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h2>Approach</h2><p>I pretty much follow the scientific method:</p><ol><li><p>Become curious about an idea</p></li><li><p>Design an experiment to explore it</p></li><li><p>Ask the hardest question at a reasonably solvable bar</p></li><li><p>Run the cheapest test that I can afford<span data-color="#a64d79" style="color: rgb(166, 77, 121);">*</span> to prove myself wrong</p></li></ol><p><em><span data-color="#a64d79" style="color: rgb(166, 77, 121);">*Note: I&#8217;ll never compete in HFT with the big guys, and I don&#8217;t think I need to in order to make money.</span></em></p><p>I don&#8217;t do speculative work. I&#8217;m definitely hands-on, and the first thing I built is &#8220;Lookout&#8221;, an intraday neural network and autonomous trading system described in this first batch of letters. It paper trades daily under set rules, and every fill gets recorded against what was expected.</p><p>The gap between backtest and reality is itself an interesting thing to measure (and now publish). So far, the edge has not survived in execution, but you&#8217;ll see exactly how that goes along the way.</p><h2>Who this is for</h2><p>You will probably enjoy this letter if you are:</p><ul><li><p>a retail trader who has noticed how rarely anything sold to retail traders comes with verifiable evidence</p></li><li><p>a programmer or data person who has wondered what would happen if someone applied real experimental discipline to markets at a small scale</p></li><li><p>someone building with AI coding tools who wants to see how far one person can take them</p></li><li><p>a quantitatively curious investor who wants to see how the sausage is made</p></li><li><p>someone who enjoys watching a person argue with his own optimism in public, using statistics and deep learning</p></li></ul><p>There are no picks here, just the process and its results, misses included.</p><p>Disclaimers:</p><ul><li><p>Yes, I&#8217;m building <strong><span data-color="#674ea7" style="color: rgb(103, 78, 167);">multiple models</span></strong> and an app that I&#8217;ll release only if it&#8217;s worthwhile</p></li><li><p>Training the Lookout neural network ensemble cost me real money, itemized down to the dollar in the first arc. The training happened before Verbatim launched. I had to prove to myself I could post something interesting.</p></li></ul><h2>What you get</h2><p>I&#8217;ll share regularly, in arcs, and each arc is a complete story.</p><p>The first arc covers a research campaign in which I spent $423 of rented supercomputer (H100) time trying to prove my own model wrong in nine different ways.</p><p>Spoiler, I mostly succeeded. Learn from me so you spend less money.</p><p>Later arcs cover the machinery of not fooling yourself, some strange physics that describes how market moves cluster, and what happened when the signals met real trading costs, which is where the program turned.</p><p>Paid subscriptions fund the lab behind Verbatim, and every verdict stays free either way.</p><p>There&#8217;s also a founding tier, four hundred dollars a year, for anyone who wants to fund the lab harder. Your rate locks for as long as you subscribe, and founding subscribers get first claim on anything that opens later, if and when a model passes its sealed test.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/subscribe?"><span>Subscribe now</span></a></p><p>Posts arrive a few times a week while the archive builds. Down the road, the tools this program runs on, the dashboards and daily model outputs I use myself, may become available to subscribers.</p><h2>The fine print</h2><p>Everything published here exists for education and research purposes. It is not investment advice, not a recommendation to buy or sell any security, and not a solicitation of anything. I am not an investment adviser, and nothing here creates an advisory relationship. I write about my own small account and my own process, I will be wrong regularly and in writing, and any tools that eventually accompany this letter will be research tools, offered to study a process. Remember that markets take money from people, every day, and nothing in this notebook changes that.</p><p>If that appeals to you, welcome. The first arc starts with a scoreboard. Nine ideas, one rule, $423. Most of the answers were &#8220;no,&#8221; but eventually the Lookout neural network ensemble survived. Its signals are graded by a machine every morning and published on <a href="https://gradedtape.com/">The Graded Tape</a>, which stays a paper record at stated sizes, with every signal that could be entered under its rules, misses included.</p><p>Those signals have not made live money yet. The Tape is mostly red, a pre-registered test of trading them through the day lost money in October 2026 (its main configuration even before costs), and the program has turned toward a second model being built to beat the index after costs. These letters run on a delay and each one is dated, so read the Tape for where things stand today. I&#8217;m looking forward to showing you the whole climb, the parts that failed included.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://letters.verbatimsigma.com/p/what-this-is-and-why-it-exists?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://letters.verbatimsigma.com/p/what-this-is-and-why-it-exists?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p><div><hr></div><p><em>Lab notebook, education and research purposes only. Not investment advice, not a solicitation; no advisory relationship is created by reading this. I trade a small account and describe my own process, errors included.</em></p>]]></content:encoded></item></channel></rss>