XGBoost Win-Probability (Phase 74)

Phase 74 is an XGBoost gradient-boosted classifier trained on every closed trade and walk-forward backtest result. It outputs a win probability [0, 1] for every candidate trade the engine considers.

The model lives in our Python research lab on Fly.io (separate from the trading worker). It retrains weekly on Sundays at 06:00 UTC and serves predictions via a bearer-protected /ml/predict endpoint. We never put training compute inside the trading hot path.

XGBoost is registered as the 8th signal in the orchestrator ensemble alongside trend, regime, breaker, RL-policy, IV-skew, news-sentiment and the Manus shadow signals.

Its win-probability is scaled into the ensemble score range [-2, +2]. Default weight: 0 (shadow only). This is the standard shipping pattern — no signal touches sizing until it has 28+ days of green telemetry on the shadow-validation tracker.

Admin controls: toggle, manual retrain, last-24h prediction stats, regime distribution.

Trade win probability (≥0 return over the next 5 days), Platt-calibrated so the raw model output is a real probability you can threshold.

Features include: sleeve identity, regime, recent realised vol, signal stack agreement, time-of-day, exchange, and the historical sleeve-class hit-rate from the past 30 days.