why
A personal project to learn how to automate investing properly: not a trading bot bolted onto a hunch, but a research system that is honest about what works and what's just noise.
what i built
- a research pipeline where every artifact is a file in git. hypotheses are written in plain language, turned into strategy code, backtested, and versioned so every result is diffable and auditable.
- guardrails against fooling yourself: a hard no-lookahead rule enforced in the engine, held-out validation data the optimizer never sees, and a correctness suite that runs before any research does.
- exhaustive parameter sweeps with robustness filters: a configuration only survives if it holds up across multiple non-overlapping market periods.
- regime detection: a hidden markov model classifies market conditions so strategies can adapt to the environment they're in.
- AI in the loop, on a leash: an LLM proposes structural strategy ideas and explains results in plain english, while the numbers are left to systematic search.
- ML research on features, carry and strategy complementarity.
- live execution: intraday futures strategies running unattended on a dedicated machine against a brokerage API.
what it shows
How to build ML systems that don't lie to you: reproducibility, validation discipline and automation that keeps running without babysitting.