Alpha Signal Lab

What it is

A pre-registered test of one question: does a gradient-boosted model (LightGBM) on twelve standard price/volume features beat plain 12-1 momentum and buy-and-hold on 5-day cross-sectional excess returns, for a 30-name US large-cap universe, under leakage-proof evaluation? The methodology (`specs/methodology.md`) was committed first and never edited after seeing results.

Purged walk-forward cross-validation (5-day purge, ~1% embargo, expanding window) produced 21 out-of-sample folds from 2016 to 2026. Eight automated leakage-audit tests, shuffled-label destruction, split-boundary checks, point-in-time feature recomputation, shift-sensitivity, label-window purity, no-full-series statistics, price sanity, and determinism, all run against the real pipeline code, not mocks, and all pass.

The number that proves the audit is real

• —A deliberately leaky “twin” control, same features, same model, same target, but naive shuffled k-fold instead of purged walk-forward, was run alongside the honest pipeline. Leaky rank IC: 0.1506. Honest rank IC: 0.0100. A fifteen-fold difference from a single methodological mistake, which is exactly what most published “alpha” likely is.
• —A per-fold SHAP diagnostic explains the honest result rather than just reporting it: a volatility ratio dominates the model’s decisions in 21 of 21 folds (not momentum, despite momentum being the only feature with real edge on its own), the model’s picks agree with momentum’s only ~0.05% of the time (so it is not a noisy momentum clone), and it diverges from momentum’s picks on 92% of days, losing money on that swap slightly more than half the time.

Why it matters

The honest model does not beat momentum, and the project says so as the headline, not a caveat. Built via a research/planning pass into an ultracode multi-phase build (scaffold, pipeline, audits/explainability, dashboard, verify), then hardened by three parallel adversarial audits that found and fixed twelve real data-integrity and numeric bugs before publishing. A static GitHub Pages dashboard (dark-mode aware, WCAG-checked contrast) leads with the honest-vs-leaky IC gap as the hero visual, not the model’s own performance.