Feature Lab

What one Sunday run trains and what each model sees, counted from the artifacts; then the processing walked through on two real signals, a baseline and a baseline-plus-evidence. Experiment cards are on /methods; the strategy lane on /strategy-record.

Models and features, counted
loading feature_census.json…
Baseline processing, walked through: weather · max_temp
One weekly series in. Feature engineering derives 76 own-history features across nine families and 10 seasonal harmonics; XGBoost trains walk-forward on the delta target; importance pruning drops what the model never splits on. Click a family for every feature and what it measures; the SHAP column is each family's share of the latest forecast.

loading…

Baseline → enhanced, walked through: stocks · AAPL
The same 86-feature baseline, then each enhanced variant adds covariate features from the evidence network — 15 / 25 / 35 for Enhanced 1 / 3 / 5 — drawn from the top-ranked sources by the S7b forward objective (search terms, macro, other stocks, the stock's own intraweek structure). The matched search term always has a seat. Pick a variant to see which sources it drew and what each contributed to the latest call.

loading…

Useful evidence: by source, by stock (COVUSE1)
Does any specific evidence source carry accuracy consistently, and which stocks have access to one? FEATREL1's leave-one-out estimator aggregated by source: remove the source's SHAP from the published error and see whether error rises. The null is negative (a useless source costs its own variance), so every number is compared to its own circular-shift null, never to zero, and a flag needs the sign to hold in both halves of the record. Regenerated every Sunday; flag rules were fixed before the first run (AB log COVUSE1, 2026-09-12).

loading covariate_usefulness.json…

Signal-history coverage, week to week
Per domain × model variant: signals, window width, cadence gaps (missing Sundays inside the span), null predictions. Verified by a standing test every week.
domainvariantsignalsweeksgapsnull preds
The cross-domain core
Feature types strong in EVERY domain (ranked by their weakest domain) — the empirically standardized menu. Full heat table on /methods.
Worked example: how a menu change is gated (FEAT2 → FEAT2b)
The template every future feature-engineering version follows.
  1. Map — 1.27M SHAP observations said the long-window family (26–104w trend/slope/strength) carries ~1% attribution in every domain.
  2. Pre-register — menu v2 (drop 11, add 3) with a hard control gate: weather must not regress. Ship criteria fixed before any run.
  3. Result — weather gate held (+2.1%) but pooled z −1.51: no ship. Confound: drops bundled with adds.
  4. Separate — drops-only arm: stocks z −2.11 — the deletions themselves were harmful.
  5. Lesson (now doctrine) — low SHAP share is a MAP of attribution, not a deletion list: low-credit features still work through interactions. Menu v1 stands, confirmed by experiment rather than assumption.
Attribution decomposition: own history vs evidence (COV1)
Three arms on identical windows: own-history only (baseline), covariates only, and both (enhanced). How much of model skill does the evidence network carry?

cov1_attribution.json not loaded.