The Machine
Never Sleeps.

We use large language models to continuously generate new trading strategy hypotheses — entry and exit rules, indicator combinations, and position sizing logic. The AI is constrained by asset class, risk tolerance, and market regime parameters, and is explicitly prohibited from regenerating ideas from our graveyard database.

Strategies are ingested as structured configs and run through a vectorized backtesting engine against years of historical data. We track Sharpe ratio, max drawdown, CAGR, win rate, and profit factor. Anything below threshold is automatically rejected. Speed matters — we test thousands of variants.

The most dangerous failure mode in systematic trading is curve-fitting — a strategy that looks perfect on historical data and fails in live markets. We apply multiple-testing corrections, minimum trade count requirements, and regime decomposition to catch strategies that are exploiting noise rather than signal.

Every failed strategy is permanently stored in our graveyard database with its full parameter set and failure reason. The AI is never allowed to generate an idea that has already been invalidated. This prevents the system from cycling through dead logic and forces genuinely novel hypothesis generation.