Backtest & Model Evaluation
Ablation study · Phase 2 ensemble results · On-demand training · Thesis Section 8.2
Phase 2 AUC 0.8325 · F1 0.5164 · +2.42% vs baseline
Ablation Study — Section 8.2
Click row to view equity curve| Experiment | Features | Test AUC | Test F1 | Sharpe | Max DD |
|---|---|---|---|---|---|
Baseline OHLCV only | 7 | 0.6120 | 0.4800 | 0.31 | -28.4% |
+ Order Book OHLCV + LOB (15 features) | 15 | 0.7230 | 0.5200 | 0.68 | -22.1% |
+ Derivatives OHLCV + LOB + Futures (24) | 24 | 0.7800 | 0.5500 | 0.92 | -18.6% |
Phase 2 Ensemble GRU + LGBM + XGB → Meta (24 selected) | 24 | 0.8325 | 0.5164 | 1.14 | -14.2% |
Equity Curve
Phase 2 Ensemble
AUC: 0.8325Sharpe: 1.14Return: +-37.5%
Drawdown
Max: -14.2%Thesis Contribution — Section 8.2
Ablation study isolating statistical contribution of each data layer. Phase 2 Stacked Ensemble achieves AUC 0.8325 and F1 0.5164, a +2.42% improvement over the Phase 1 XGBoost baseline (AUC 0.8083). Target: liq_cascade_target — binary label for liquidation cascade events in next 15-min candle. Answers: "Does stacking GRU + ensemble trees via meta-learner improve cascade prediction?" — Yes.