Five-model ML ensemble. 44 tick-level features. AM momentum + PM mean reversion on ES futures, RTH only.
One instrument (ES futures), RTH only (09:30–16:00 ET). Two signal paths fire in parallel. Both inverted — raw ensemble direction flipped before execution, empirically validated by 30 days of backtest. Only signals with confidence ≥ 0.70 reach the broker.
Every tick → 90+ microstructure features (VPIN, Hurst, Kyle's λ, DOM pressure) → four models vote: F1 LSTM, F2 LightGBM, F3 triple-barrier, F3_UPSIDE. Raw direction inverted → conviction gate → IB.
Each closed 5-minute bar → 52 bar-level features (trend strength, session context, range dynamics) → XGBoost trained on triple-barrier labels. Direction inverted. Independent of Path 1 but shares the position lock — first path to fire holds until exit.
Ensemble confidence ≥ 0.70 required. At 0.20: 59% WR, PF 1.07. At 0.70: 77% WR, PF 2.07, 0 losing days over 30-day BT. Quality over quantity — noise signals at 0.2–0.4 destroy edge.
PhD-quality writeup covering feature engineering, model architectures, ablation studies, and statistical results.