Skip to content

Frozen Model Tournament

The tournament is the promotion gate for successive HIPSTER PM feature releases. It tests each challenger against the incumbent that survived every earlier round, rather than interpreting small differences in a global ranking table as automatic wins.

Contract

All models use the same frozen forecasts for 2023-24 through 2025-26. Games are resampled with replacement within each season, preserving season balance and all within-game possession dependence. Each draw evaluates both models on the identical sampled games.

The promotion metric is full-game margin RMSE. For each match we compute

\[ \Delta = \operatorname{RMSE}_{\mathrm{challenger}} - \operatorname{RMSE}_{\mathrm{incumbent}}. \]

The challenger normally advances only when the upper endpoint of its paired 95% bootstrap interval is below zero. Possession RMSE, possession MAE, and winner accuracy are retained as supporting diagnostics.

For a coherent strict simplification of the active feature contract, the published reference may instead be changed under a parsimony non-inferiority decision: the full-game interval must not indicate a practically material degradation (upper endpoint no greater than +0.015 RMSE), and the challenger must remove an interpretable feature family rather than add capacity. This is a model-selection preference, not evidence that the simpler model is statistically superior. Future simplification decisions use this same margin.

Current Tournament

10,000 paired game-block bootstrap draws, seed 20260814.

Round Incumbent Challenger Incumbent RMSE Challenger RMSE Delta Paired 95% CI P(challenger better) Result
1 Complete player-prior RAPM Original HPM v1 14.4756 14.3802 -0.0954 [-0.1763, -0.0153] 99.03% Promoted
2 Original HPM v1 HPM v2 shooting composition 14.3802 14.3908 +0.0105 [-0.0072, +0.0284] 12.45% Retained incumbent
3 Original HPM v1 HPM v2.1 rebound capacity 14.3802 14.3869 +0.0067 [-0.0155, +0.0288] 27.73% Retained incumbent
4 Original HPM v1 HPM v2.2 usage allocation 14.3802 14.4050 +0.0247 [-0.0106, +0.0595] 8.49% Retained incumbent
5 Original HPM v1 HPM v2.3 shot portfolio 14.3802 14.4288 +0.0486 [+0.0132, +0.0831] 0.40% Retained incumbent
6 Original HPM v1 HPM x1 ORB claim context 14.3802 14.4665 +0.0862 [+0.0083, +0.1637] 1.45% Retained incumbent
7 Original HPM v1 HPM x3 ORB claim rebound replacement 14.3802 14.3774 -0.0029 [-0.0168, +0.0113] 65.17% Promoted on parsimony non-inferiority
8 HPM x3 ORB claim rebound replacement HPM x4 ORB claims + blocks 14.3774 14.4036 +0.0262 [-0.0102, +0.0632] 8.05% Retained incumbent
9 HPM x3 ORB claim rebound replacement HPM x5 interaction-only creation 14.3774 14.3800 +0.0027 [-0.0214, +0.0267] 42.29% Retained incumbent

The published regular-season reference is therefore HPM x3 ORB claim rebound replacement. Its full-game interval does not establish superiority, but it removes four rebound response functions and the rebound-by-usage interaction without evidence of a practically material full-game degradation. Original HPM v1 remains the strict-superiority benchmark.

Run

uv run nba-run-frozen-model-tournament --draws 10000

Each immutable run writes paired metric intervals and one primary-metric row per tournament round under artifacts/models/frozen_model_tournament/.