alpha-signal-lab — pre-registered research

Does gradient boosting on standard price/volume features add predictive power over plain momentum, under leakage-proof evaluation?

Data span: — OOS period: — Last run: — Success criterion: mean rank IC > 0.02 and NW t-stat > 2.0

Leakage audit receipt

Eight pre-registered checks, run against a committed synthetic fixture in CI with no live network calls. Hover or tap a badge for what it verifies.

Honest vs leaky: the headline chart

Identical features, target, model, and hyperparameters. Only the cross-validation changes: PurgedWalkForward (honest) vs naive random 5-fold shuffled CV with no purge or embargo (leaky twin). The gap is pure leakage.

Honest — PurgedWalkForward

purged, embargoed, walk-forward OOS

—

NW t-stat —

Leaky twin — shuffled 5-fold

no purge, no embargo, same everything else

—

NW t-stat —

Δ TBD

Delta between the two mean rank ICs above, computed from the live numbers once both sides are available.

Long-short equity curve

5-day rebalance, equal-weight top-3 long / bottom-3 short, model vs 12-1 momentum rank vs equal-weight buy-and-hold.

Rolling 63-day mean rank IC

Daily cross-sectional Spearman rank IC, smoothed. The dashed line and label mark the pre-registered 0.02 bar.

Per-fold mean rank IC

One LightGBM model per walk-forward fold; bars show that fold's mean OOS rank IC. Gray bars mean no data for that fold, not a measured zero.

Feature importance

SHAP on the final (most recent) fold's model, computed on that fold's OOS test rows only. Stability heatmap shows fold-to-fold rank churn (1 = most important).

Mean |SHAP| (final fold)

Per-fold feature-rank stability

SHAP plots (from explain.py)

Supplementary analyses

Everything below was computed after the primary pre-registered verdict above was sealed. These are diagnostic/robustness checks, not a re-run or reframing of the primary result.

Feature neutralization vs 12-1 momentum

Does the model add anything orthogonal to plain momentum? Rank IC of raw predictions vs predictions cross-sectionally residualized against 12-1 momentum.

Regime and per-ticker breakdown

Rank IC split by realized-volatility regime (median split on trailing 21-day realized volatility), plus per-ticker time-series IC across the universe.

Per-ticker rank IC

Rolling per-fold SHAP importance over time

Mean |SHAP| per feature, one line per feature, across all walk-forward folds, how importance ranks drift over time rather than a single snapshot.

CPCV: a distribution instead of one path

Combinatorial Purged Cross-Validation: the same frozen model refit across many purged/embargoed train/test path combinations, giving a distribution of OOS rank IC instead of the single walk-forward realization above.

Why did the model underperform?

Computed after the primary verdict, from the already-sealed OOS predictions: a diagnostic look at why the model trails the 12-1 momentum baseline, not a re-run or reframing of the primary result.