Tested strategies

Rules. Version A: buy at the next open when the 2-day RSI closes below 10; sell when the close is above its 5-day simple moving average. Version B: the same, but buy only when the close is above its 200-day simple moving average.

Market. SPY, daily bars, long only, one trade at a time.

Data. US stocks and ETFs on daily bars, split-adjusted, with cash dividends credited on each ex-date (total returns). Run through the Agent API by Claude Code, on a research account used only for this series.

Costs. The defaults: commission 0.005% per round trip and slippage 0.01% per fill, with next-bar fills.

Versions tried. Two rules, each on two windows: four development runs. Version B on 1993–2024 was graded against the 2 distinct versions this account had tested on SPY in the last 30 days.

Beat buy-and-hold over the same dates? No, on every run.

Held-out test. Run once, on Version B, over 2 January 2025 – 23 October 2026: 20 trades, 80% won, grade C (51/100), below buy-and-hold. Twenty trades is a small sample. This account's held-out test on SPY is now used.

Grade C (64/100) is the most a result can score before its held-out test. A high win rate is not an edge: neither version beat holding SPY over the same dates.

Rules. Buy at the next open when the 14-day RSI closes below 30; sell when it closes above 70. No stop.

Market. SPY, daily bars, long only, one trade at a time.

Data. US stocks and ETFs on daily bars, split-adjusted, with cash dividends credited on each ex-date (total returns). Run in the app, by its own agent, in one chat on the same research account as episode 1.

Costs. The defaults: commission 0.005% per round trip and slippage 0.01% per fill, with next-bar fills.

Versions tried. One rule, one development run. The app counted 5 distinct versions tested on SPY on this account in the last 30 days, episode 1's included.

Beat buy-and-hold over the same dates? No, on both runs.

Held-out test. Run once on 30 October 2026, the same code unchanged, over 29 January 1993 – 8 November 2018 (25.7 years): 20 trades, 85.0% won, grade C (50/100), below buy-and-hold. The 20th trade was still open on 8 November 2018 and is marked there. The development run warmed up its indicator on bars before November 2018 but took no trades there. Episode 1 used the Agent API's held-out test, which counts once per account and market; in the app it counts once per chat and market, so this chat's is now used.

A high win rate on few trades tells you little. The first run's own warning: "Only 8 trades — too few for the statistics to be reliable. Aim for 30+ before trusting the verdict." And on the held-out years: "Grade C: treat this as a first look, not proof."