Peter, a systematic trader, is experimenting with a different approach to AI application. Instead of having AI trade directly, he has built an automated research pipeline that allows Claude Code to autonomously develop futures strategies while he sleeps.

The entire process is based on a Python pipeline, with Claude running structured research cycles in headless mode. Each cycle consists of four stages: a senior trader persona proposes a structural hypothesis; Claude implements the strategy code and runs backtests; a senior analyst persona evaluates it with hard data; and finally, a structured review takes place—accepting or rejecting, updating the research status, and proposing the next direction.

To prevent pure data mining, safety measures are built into every cycle. Hypotheses must be structural, not parameter scans; every accepted condition must be tested against a random entry baseline; data must pass half-split stability tests; and a termination threshold is set to stop research if there is no improvement over multiple cycles.

Some netizens commented that overnight autonomous research is where agents truly shine. The gap between a demo version and a "run while I sleep" version lies mainly in error handling and state recovery.

Peter emphasizes that this does not replace trading intuition. He still needs to define the underlying strategy logic, the market structures he believes in, and the research rules. What AI replaces are hours of coding, backtesting, and spreadsheet work that previously had to be done manually.


Another key decision was integrating a production-grade framework. Peter tested NautilusTrader, so the AI only needs to write strategy logic rather than building backtesting infrastructure. The framework reduces the surface area for errors and avoids reinventing everything in homebrewed scripts.

Feature testing is the highlight of this pipeline. Strategies collect features for every trade—volatility regime, time of day, directional context, impulse size, etc. After each backtest, the analysis step performs quintile analysis, winner/loser forensics, and stability checks.

The most valuable output isn't what it found, but what it eliminated. One cycle discovered a filter that seemingly improved the result of every trade by $50. However, a random baseline test showed that random entries using the same filter earned more. This signal was actually destroying value and was ultimately rejected.

A trading researcher shared a similar experience, testing 23,520 parameter combinations, with 96.7% failing the termination threshold. The 3.3% of surviving strategies all had structural hypotheses, passed half-split stability tests and cross-market validation, and achieved a Sharpe ratio of 4.29 over 5.5 years.

The true value of this method lies in consistency and speed. It's the same checklist every time; steps aren't skipped due to fatigue or excitement over a certain curve. Originally, only one idea could be tested in a weekend; now, multiple can be tested overnight, with every hypothesis undergoing the same rigorous scrutiny.