PARTIAL · filter only
Hypothesis
Gradient-boosted trees on engineered features (vol regime, funding, OI delta, time-of-day) can classify which signals of an existing book are higher-quality, used as a filter rather than a generator.
Math — boosted additive model
The model is an additive ensemble fit greedily to the gradient of the loss:
$$ \hat y = \sum_{m=1}^{M} \nu\, f_m(\mathbf{x}),\qquad f_m \approx -\nabla_{\hat y}\,\mathcal{L} $$
Method
Label baseline signals win/loss, train XGBoost with monotonic constraints, evaluate as a probability gate on out-of-sample signals. Strict temporal split, no leakage.
Results
| AUC (out-of-sample) | 0.58 |
| Top-quartile precision lift | +9pp |
| Standalone generation | no edge |
| Decay TRAIN→TEST | moderate |
Weak-but-real as a filter on an existing edge (AUC 0.58, top-quartile precision +9pp) — never as a signal generator. Most of the lift is the vol-regime and time-of-day features, not the exotic ones. Kept as an optional overlay with monitored decay.
ML earns its keep ranking the quality of signals you already have, not conjuring signals from price alone. And the boring features (regime, time) usually carry the lift.