Forward Exposure-Gated RAPM
This is the recursive one-number preseason RAPM state. Each completed regular season is fit with prior-centered ridge RAPM. Its resulting player values form the returning-player state for the next season. First-NBA-season players receive the exposure-gated draft/replacement prior only when the necessary components can be estimated from earlier completed seasons.
For target season \(t\), all three cold-start pieces use data through \(t-1\): the draft-rate ridge, the first-year low-exposure logistic gate, and the equal-season mean of tokenized replacement RAPM estimates. No future season is used to form a historical prior.
Current State
The completed run through 2025-26 is
artifacts/models/forward_exposure_gated_rapm/2025-26/forward-exposure-gated-rapm-2025-26-20260806T030613Z-cddcd0ae/.
It publishes 2026-27 returning-player priors from the completed 2025-26 fit.
The direct Draft History and active-roster pipeline supplies an external rookie
profile table before the new class enters the player-season panel.
The 2026-27 Exposure-Gated Player Rankings page publishes the sortable top 100 returning-player table. The direct NBA Draft History ingestion now supports a separate sortable table for the drafted class: 2026-27 Drafted Rookie Rankings. That table is now scored from this completed forward state through 2025-26, not from the earlier standalone cold-start artifact. The frozen 2025-26 rookie branch is evaluated separately in Forward Cold-Start Validation.
Frozen 2025-26 Evaluation
The 2025-26 prior was frozen after 2024-25 and then evaluated on the same regular-season and playoff cohorts as the Frozen Preseason Leaderboard.
| Cohort | Possession RMSE | Possession MAE | Game-margin RMSE | Team NetRtg RMSE | Pythagorean win RMSE |
|---|---|---|---|---|---|
| Regular season | 1.198993 | 1.142055 | 14.8225 | 4.6680 | 10.2778 |
| Playoffs | 1.192821 | 1.136096 | 17.4188 | - | - |
The simpler frozen exposure-gated cold-start model remains better on the main regular-season possession and game-margin metrics. This recursive model is therefore the leading production-state candidate, not a promoted sole gold standard yet. It needs additional target-season holdouts before that choice is settled.
Published Outputs
| File | Contents |
|---|---|
next_season_top_100_returning_rankings.parquet |
Top 100 completed-2025-26 RAPM values used as 2026-27 returning priors |
next_season_returning_rankings.parquet |
Full returning-player state, including 2025-26 possession exposure |
season_player_priors.parquet |
The strictly forward prior vector entering every historical season |
season_cold_start_metadata.parquet |
Per-season gate, draft-rate, and replacement-token settings |
season_replacement_tokens.parquet |
One completed-season token estimate per season |
| Frozen evaluation files | 2025-26 possession, game, team NetRtg, and Pythagorean-win results |