project-conditional-minutes-refuted — Findings

← All findingssource · memory/project_conditional_minutes_refuted.md

DO NOT BUILD. Refuted 2026-08-25 after a full day on it. The idea: for a player whose role changed, project from her own games at a comparable MINUTE LOAD instead of the model's blended rate features.

Why it fails

There is no usage expansion to model. Maya Caldwell's FGA per minute across minute buckets: 0.2468 (<15) / 0.2799 (>=15) / 0.2838 (>=20) / 0.2748 (>=24) — flat. Her FGA triples because her MINUTES triple. The decisive number: pooled career rate x 26 min = 8.07 vs conditional mean 8.22. The conditioning bought +0.15 points on the case that motivated the entire project.

Against actual outcomes it is a null. 862 leak-safe promoted rows with an honest engine pred_min anchor: model q50 MAE 4.5300, cond15 4.5248 (paired z = -0.06). A walk-forward constant q50 * 1.169 scores 4.3919 and beats it (z=+1.75). On 89 real priced rows with real pre-tip lines the full estimator returned +25.7% ROI vs +24.8% for the bare constant — a two-line multiplier matches ~55 lines of stateful logic.

Two methodology errors that produced the false positive

Benchmarked against a strawman. My backtest compared the conditional estimator to a trailing-5 AVERAGE and reported +23.4%. Production does not use a trailing-5 average — it uses LightGBM quantile models. Against the real model q50 the lift vanishes. Always benchmark against what production actually runs.
Projection-vs-projection proves nothing. The "9 of 9 promoted players where conditional > model" figure compares two ESTIMATES. It is worth zero until outcomes are attached.

Also wrong, and corrected

I claimed "her FGA nearly triples so usage scales with role", retracting a CORRECT earlier statement ("she got minutes, not touches") under push-back. Checking FGA per minute takes 30 seconds and settles it. The first answer was right. Do not reverse a position without checking the rate-normalised version.

What DID survive

projected_minutes ingest (PE branch conditional-minutes-estimator, tested)
the >=6-min RotoWire-vs-model disagreement flag, as a suppressor not a re-price
line shopping: +2.84pp ROI, paired t=1.83, block-bootstrap 95% CI [+0.29, +5.33] over slate days. A price fact, independent of any model. 70% of WNBA points props have books >=1 full line apart.
the projection alone measured +0.51% ROI = break-even, independently reproducing the known "projections sit at market parity" result.