Methods & Evidence

Every claim on this site traces to a pre-registered experiment. This page is generated from the same warehouse that audits the pipeline — the cards show what each method changed, what the data said, and what happens next.

The feature menu, measured across domains

Average share of model attribution (SHAP) per engineered feature type — one shared vocabulary, scored in every domain. Red = weak, green = workhorse. Struck-through rows are the FEAT2 retirement set (dead in every domain); the A/B gating that change is running.

Why demand skill ≠ price skill

The models consume search-demand evidence — but across … stock-weeks, its direction agrees with the realized weekly move … of the time — a coin flip. Demand forecasts are the product; weekly price direction is not.