AutoHypothesis Strategy Agents

Quant researchers usually test one market idea at a time. AutoHypothesis packages that loop into an agent that writes strategy code, runs backtests, logs results, and tries again inside a fixed research harness. (github.com, github.com) The repository says the agent is allowed to rewrite only three functions in `agent.py`: one for trade signals, one for position sizes, and one for market regime. Everything else — data loading, simulation, metrics, and the command-line runner — stays fixed. (github.com, github.com) That setup turns quant research into a bounded search problem. The repo fixes a development window from 2010 through 2016, a holdback gate for 2017 through 2018, walk-forward validation for 2019 through 2021, and a locked holdout starting in 2022. (github.com, github.com) The scoring rule also pushes the agent away from fragile ideas. AutoHypothesis scores a strategy as Sharpe minus penalties for turnover above 0.3 and drawdowns deeper than 20%, with 10 basis points of transaction cost and a one-day execution lag built into the assumptions. (github.com, github.com) The project’s appeal is reproducibility, not just automation. The README says the user edits `program.md` to change the research directive or target metrics, while the harness writes out `results.csv`, `last_result.json`, and checkpoint files for the best in-sample and holdback-passing agents. (github.com, github.com) That matters in a field where backtests are easy to overfit. Two Sigma’s 2018 technical report on Sharpe ratios says confidence intervals and hypothesis tests matter because a reported Sharpe can look stronger than it is, and the AutoHypothesis README lists survivorship bias and walk-forward leakage among its own limitations. (twosigma.com, github.com) The public materials available through GitHub search confirm the framework, the data splits, and the penalty-based score formula. I could not independently verify the specific Sharpe-0.86 claim from the X post because the accessible repository pages in search did not expose the underlying result file or final metrics page. (github.com, x.com) Even with that gap, the repo shows what the project is trying to standardize: hypothesis generation, controlled testing, and experiment logging in one loop. For anyone building agent-driven quant research, the code is less a finished trading system than a template for running the same research process the same way twice. (github.com, github.com)