← All findingssource · memory/project_feature_experiments.md
projections-engine prop-model features that were implemented, measured on a held-out backtest, and reverted because they added no accuracy. Don't re-propose these without a materially different approach or new data — they've been tested.
Position-aware DvP — reverted earlier (pre-memory). Splitting defense-vs-position into guard/wing/big buckets gave ~−0.007 aggregate MAE (noise). Team-level DvP is what runs.
Why: Reason — both duplicated signal the model already had; the
project's bar is measured lift on the walk-forward test, not
doc-plausibility. How to apply: before building any doc-suggested
feature, check whether its components are already in STAT_FEATURES
via rolling windows. If yes, expect redundancy. Always A/B with the
env-gated control pattern (one script, toggle the allowlist) before
shipping — see commit 8762d3d for the harness shape. Ship/keep only
on measured improvement; revert cleanly otherwise (no dead feature
code). The game-outcome model from the same arc (game_outcome.py,
elo.py, team_features.py) is a kept new capability — 65.8% hit
rate, separate from this revert. See scope-creep-stay-focused.