RAPM Modeling Stints

The first analytical modeling contract combines curated lineup stints with curated possession segments. It is deliberately restricted to regular-season games and retains only stints with positive possession exposure.

Possession exposure

A possession contained within one lineup segment contributes one possession to that segment. When a substitution divides a possession into multiple segments, each segment receives an equal share:

\[ w_{segment} = \frac{1}{\text{segments in possession}} \]

This policy conserves exactly one possession while assigning event points to the lineup actually on the court. Home and away offensive shares are summed separately for every stint. The model weight is their average:

\[ w_{stint} = \frac{\text{home offensive possessions} + \text{away offensive possessions}}{2} \]

The target is home-team point differential per 100 average team possessions:

\[ y_{stint} = 100 \times \frac{\text{home points} - \text{away points}}{w_{stint}} \]

Zero-exposure stints are counted in the manifest and excluded from fitting. They are not silently deleted from canonical lineup data.

Columns

Each row contains:

season, game, period, and stint identifiers;
schedule time and home/away team identities;
exactly five home and five away player IDs;
elapsed-time boundaries and duration;
home, away, and average possession exposure;
points, home margin, and the net-rating target;
game-quality, processing-code, build, and source-hash provenance.

Storage

data/analytical/rapm_stints/{season}/regular/
  _manifest.json
  part-00000.parquet

The manifest records exact curated input manifests, source counts, included and excluded stints, possession totals, player count, modeling code fingerprint, and output file integrity.