Impact Rate · public validation

Does Impact Rate predict what happens next?

Yes, with an important caveat. On 295 player holdouts, the live bounded model beat a box-score-only baseline on all four published metrics. The first fully frozen prospective test remains pending.

Model bounded-hybrid-v1-2026-07-21Target next-season RAPMProspective test pending

Flagship scoreboard

2025–26 pseudo-holdout

Development seasons: 2021–22 through 2024–25 · 295 qualified player outcomes.

MetricImpact RateBox onlyResult

Prediction error (RMSE)Typical error, with larger misses penalized more. Lower is better.2.0692.1604.2% lower

Absolute error (MAE)Average absolute miss in RAPM points per 100 possessions. Lower is better.1.7041.7723.9% lower

Rank agreement (Spearman)Agreement between predicted and observed player order. Higher is better.0.4190.338+0.081

Explained variance (R²)Share of player-to-player outcome variation explained. Higher is better.0.2660.199+0.066

The test that matters next

2026–27 prospective scoreboard

Formula, predictions, hashes, and pass/fail gates were frozen before the outcome season.

StatuspendingAwaiting completed-season RAPM

FrozenJuly 21, 2026387 player predictions

Minimum evidence250 players70.0% target join rate

Season gate1,230 gamesExact-score and processed

01
data complete
2026–27 RAPM quality is complete with at least 1,230 exact-score and processed games.
Pending
02
coverage
At least 250 frozen predictions join to at least 70% of qualified target players.
Pending
03
rmse improvement vs raw
Bounded candidate RMSE improves on Raw-only by at least 1.0%.
Pending
04
spearman gain vs raw
Bounded candidate Spearman exceeds Raw-only by at least 0.025.
Pending
05
beats published bpm
Bounded candidate RMSE is lower than published-BPM-only RMSE.
Pending
06
beats rapm only
Bounded candidate RMSE is lower than multi-season-RAPM-only RMSE.
Pending
07
near uncapped rmse
Bounded candidate RMSE is no more than 0.020 worse than uncapped hybrid RMSE.
Pending
08
near uncapped spearman
Bounded candidate Spearman is no more than 0.015 below uncapped hybrid Spearman.
Pending
09
frozen integrity
Prediction-file hash, candidate id, cap, and model coefficients exactly match this manifest.
Pending

Reproducibility

What was predicted

The target is a player's next-season, complete-stint total RAPM. The comparison asks whether a bounded blend of stabilized box production and multi-season lineup evidence predicts that outcome better than the same calibrated box prior by itself.

Observation weights are based on target-season RAPM possessions. Named players were not selection gates. The 2026–27 prediction file covers 387 players and is locked by SHA-256: bde57792d14fc126194a037070255e37ae9ab6140ba51587380176f32d83a640.

Source files: research/modern_impact_validation/outputs/bounded_hybrid_results.json, research/modern_impact_validation/prospective/bounded_hybrid_v1_manifest.json, and research/modern_impact_validation/prospective/bounded_hybrid_v1_gate_2027.json.

Read the model methodology →