Models
The modeling program starts with transparent predictive baselines before adding player priors, separate offensive and defensive effects, or nonlinear lineup interactions.
Baseline ladder
| Model | Information available |
|---|---|
| Mean | Training-window average home net rating |
| Team | Signed home and away team identities |
| RAPM | Signed identities of all ten players |
| Bayesian RAPM | RAPM posterior uncertainty under an explicit Gaussian model |
| RAPM aging model | Forward player-season priors from prior RAPM and age |
| Age-informed prior RAPM | Lineup RAPM centered on the frozen aging forecast |
| Forward RAPM calibration | Frozen lagged RAPM mapped to team wins using realized lineup seconds |
| Frozen lagged RAPM | Preseason 2024-25 player values scored on all 2025-26 oracle lineups |
| Frozen aging prior | Preseason age, experience, prior RAPM, draft, and physical-profile values scored on the same oracle lineups |
| Frozen O/D RAPM | Forward offense and defense player values scored on the same oracle lineups |
| Box-score RAPM prior | Returning-player forecast from lagged RAPM and possession-native box profile |
| Additive neural RAPM | Possession-level signed scalar player embeddings |
| Deep Sets | Nonlinear permutation-invariant lineup aggregation |
| CatBoost | Categorical player states and boosted-tree interactions |
| RAPM + Transformer | Frozen ridge prediction plus contextual player-player attention residual |
The first four use the same regular-season stint target, chronological game splits, possession weights, and test window. The first three compare predictive information sets. Bayesian RAPM deliberately retains the ridge point estimate and adds coefficient, rank, and predictive uncertainty. Neural models move to single-lineup possession rows while retaining chronological game boundaries. RAPM + Transformer reconnects the two samples by converting a leakage-safe, stint-weighted ridge prediction into the possession target and learning only its nonlinear residual.
See Baseline methodology for the model contract, the 2025-26 RAPM case study for a worked diagnostic review, the Bayesian RAPM methodology and 2025-26 Bayesian case study for the probabilistic baseline, and Promoted rankings for reviewed public releases. Neural Networks defines the staged additive, Deep Sets, and RAPM + Transformer program. Tree Models defines the orthogonal CatBoost baseline. Leaderboard defines the shared regular-holdout and playoff metrics and maintains the cross-model scoreboard. Forward RAPM Calibration maps frozen player priors to regular-season team wins. The Frozen Preseason Leaderboard holds out the complete target regular season and playoffs without a player refit. RAPM Aging Model defines the first temporal player prior, while the Modeling Roadmap records the approved joint dynamic RAPM extension.