Stop guessing. Start building.

Before you set an alert for COT extremes, know their rhythm. In our 4,889-backtest sweep of Larry Williams' COT index on the E-mini S&P 500 (2006-2025), one practical detail rarely gets mentioned: how often the signal everyone talks about actually occurs. We counted every extreme crossing, a trader group's COT index reaching the classic Williams levels (75 and above for bull, 25 and below for bear) at a 24-week lookback, with each report counted from its release day: - 411 extreme crossings, across all three groups and both sides. - Each group hit an extreme 3.0 to 4.4 times per year, per side. - Once entered, an extreme run lasted 3.6 to 5.2 weekly reports on average. So a COT extreme behaves less like a signal firing and more like a regime arriving: it shows up a few times a year, then holds for about a month. That cadence is one reason for the central trade-off. Layer a slow, once-a-week gate onto a fast 2-to-3 day RSI(2) swing strategy and the gate sits closed most days. Across the filtered variants, the median trade count was 95 against the unfiltered baseline's 406. The median CAR/MaxDD, a separate median, was 0.07 against 0.51. None of the 4,888 filtered variants beat the baseline on that measure. Slow signals make good context. Poor triggers. The COT report is a calendar, not a stopwatch. Full study: https://statoasis.com/overfit/research/larry-williams-cot-filter

87 of 165. That is how often the "fixed" rainbow moving average beat the classic one. The classic rainbow stacks five moving averages evenly: 10, 20, 30, 40, 50. The obvious fix is to space them by one constant ratio instead, 10, 15, 22, 33, 50, so every step is the same +49.5%. I tested both on 11 futures from 2005 to 2026, with 15 matched pairs of stacks that share the same first and last line. Each pair was compared market by market, 165 comparisons for every signal and side. On the pullback signal, traded long and short, the ratio stack had the better return to drawdown in 87 of 165. Across all six signal and side cells it won between 63 and 88. Near half, every time. So neither spacing wins consistently. What did move them was where the stack starts: every ratio stack starting at 8 to 12 bars, with a step ratio of 1.3 to 1.8, had a positive median return to drawdown on the pullback signal. 18 of 18. If you run a rainbow, set your first line before you touch the spacing. Frictionless, one exit, no stop. A trend filter to build on, not a finished system. The full study, the charts and the video walkthrough: https://statoasis.com/overfit/research/rainbow-moving-average

Zero of 2,068 eligible optimizer variants beat buying and holding out-of-sample. That is the result of the optimizer sweep I ran alongside this prompting workflow: 2,304 variants swept on SPY, 2,068 of them with enough trades to score, and not one that cleared the baseline out-of-sample. The same study tested 5 AI-written rules exactly as given. 0 of the 5 beat buy-and-hold either, and the AI that wrote them was trained on text published during that out-of-sample decade. So I split the two things apart. This page is the workflow. Four prompts, in order, to get a specific and testable strategy out of a model instead of a paragraph of hedging. Experience and criteria first. Then the market edge, named. Then exits. Then filters that remove trades. It is deliberately not the evidence that the output works. A guide that teaches a workflow and also grades it tends to grade it generously. Use the prompts. Then go and validate what comes out, because the prompts are the easy half. The four prompts, in order: https://statoasis.com/overfit/research/use-ai-to-develop-profitable-trading-strategies-here-is-how