What we've found so far

Structure & zones leads the trading styles this month — 0 of 2 models beat the market

average month to date (9 trading days) across 2 models · S&P 500 same window: +2.03% · past week: style +0.79% vs S&P -0.06%

The Structure & zones style — trade around price structure — ranges, breakouts and the levels they fail at — beat the market with 0 of 2 models this month, and 1 of 2 over the past week (+0.79%). Read that as a lead, not a finding. When a style leads on average while only 0 of 2 of its models beat the market, the average is being carried by one or two of them — which is exactly what luck looks like. We print it because it is the top of the board; we will not call it a result until every model under the instruction clears the bar.

Super Bot (Grok and ChatGPT versions)

The Grok version finished +4.95% from Sep 2 against the S&P 500's -0.33% over the same days, with a worst dip of 1.9%. The ChatGPT version, which only started on Sep 9, finished +2.10% against +0.03%. The Grok version beat the Tactical bots (Claude +4.50%, Gemini +3.40%, ChatGPT -5.38%, Grok -7.16%) and the Patience bots from Grok, ChatGPT and Gemini, but not Patience from Claude (+8.66%). In Sep 25-30 it slipped 0.88% while the S&P 500 slipped 0.59%, and on Sep 28 we changed how its rules are judged and how it picks sides, so the days since then (five or fewer) are too few to read: insufficient sample. About 210 closed trades so far, one month, one kind of market. In the first two days of October it gave back about 2%, so this is not a settled result.

The Learner

The version with the learning filter finished +1.50% and its unfiltered twin +1.56%, measured from Sep 3, while the S&P 500 fell 1.36% over the same days. Both beat the market, but the filter itself added nothing. A 0.06-point gap is noise. The filter only switches on once it proves it helps, and it has not shown that yet.

Weather-aware

Both versions beat the S&P 500: Claude +3.81% and Grok +0.85%, against -0.33%, from Sep 2. Their job is smaller bets when the market gets jumpy. Claude's worst dip was 2.8%, smaller than its plain Patience twin's 4.1%, but Grok's was 4.2%, bigger than its twin's 2.8%. Claude's gain trailed plain Patience Claude (+9.20% over the same days), so the safety layer cost it return and protected it in one model only. Two models, one month: too early to call.

Structure & zones

Claude finished +6.25% from Sep 2 against -0.33%. ChatGPT finished +0.22% against -0.58%, and lost 4.07% in the last four sessions of the month. The Grok and Gemini versions were retired on Sep 22 and are not counted. A 6-point gap between two models on the same instruction says the model mattered more than the style. We make no claim about why the style works. This verdict was due around Sep 22 and is late. Both live lanes moved to newer models on Sep 28, so any later read will separate the old and new model.

Tactical (bets on falls as well as rises)

From Sep 2: Claude +4.50%, Gemini +3.40%, ChatGPT -5.38%, Grok -7.16%, against -0.33%. In Sep 25-30 Claude gained 3.96% while Grok lost 4.59%. Same instruction, opposite results. The style did not decide this month; the model did. Worst dips ran from 5.6% to 9.3%.

Moonshot

The ChatGPT version finished +3.36% against -0.58%, with a worst dip of 6.8%. That is the point of an arm built to take concentrated, high-risk bets. The Grok version was retired on Sep 22, so only one model is left and one month proves little.

Patience (the September review)

Patience led August. In September Claude finished +8.66% and Grok +1.77%, but ChatGPT finished -7.23% and Gemini -7.25%, each against -0.58%. The average, -1.01%, was slightly worse than the S&P 500. Two models beat the market and two lost badly: the style did not hold up as a style this month.

System (fixed rules, no learning)

The never-changing rulebook lost 9.45% against the S&P 500's -0.58%, with a worst dip of 10.5%. It lost 6.59% in the last four sessions alone. It led the board in the summer's rising market; a choppy month is where fixed rules give the gains back.

Being good at chess and poker does not make you a better trader

For weeks the models played chess and poker against each other every night (those AI-vs-AI duels are paused now; people can still play the bots free). We tested the obvious hypothesis — that strategic-game skill transfers to markets — by feeding each model its own game record during its strategy rewrites, and running a control arm that never saw it. The control beat both game-trained arms. The experiment was retired. Cross-domain transfer: not detected.

Calendar seasonality has no edge

A widely-sold idea says that if a stock rose in a given calendar window in ≥90% of past years, it will tend to do so again. We tested it properly on our full universe: pick windows using 2015–2022 only, then trade them blind on 2023–2026. Our first pass looked spectacular — and it was our own bug. With a control that differed by seasonality alone, the "highly seasonal" windows lost to ordinary ones in 10 of 10 comparisons. We are not building it.

Holding positions overnight did not pay off as a style

We ran an arm that held positions through the closing bell instead of flattening every evening, across all four models, for four weeks (5–30 August 2026; figures are its final record at retirement). The average lost to buy-and-hold: +1.57% against the market's +2.99% over the same stretch. Two of the four models actually profited from it (Grok +7.2%, Gemini +7.1%) while the other two lost (Claude −0.8%, OpenAI −7.2%) — a split that wide means the style itself was not the driver, the individual model was. The arm is retired; the figures above are its final record.