Skip to content

Baseline Methodology

Null model

The null prediction is the possession-weighted training mean. Because every target is expressed from the home perspective, its intercept includes average home-court advantage.

Team model

\[ \widehat{y} = \text{HCA} + T_{home} - T_{away} \]

The signed sparse matrix has one column per team. It provides the main schedule-adjusted predictive baseline for RAPM.

One-number RAPM

\[ \widehat{y} = \text{HCA} + \sum_{p \in home} \beta_p - \sum_{p \in away} \beta_p \]

There is one coefficient per player and no offensive/defensive split. The initial model includes no bio features, draft prior, score context, or interaction terms.

Selection and evaluation

Team and player penalties are evaluated on expanding chronological folds. The final test is the last 15% of regular-season games and is not used for lambda selection. After test metrics are recorded, an all-season refit produces the ranking artifact.

Reported metrics include:

  • possession-weighted RMSE and MAE;
  • game-margin RMSE after aggregating stint predictions;
  • skill relative to the mean baseline;
  • skill relative to the team baseline.

Raw on-court net rating is included beside RAPM as a descriptive comparison, not as a predictive model.