L20-MSP v0.1

Gap-Returner Preseason Minute-Share Persistence extends the static L20 baseline for players who are on the transaction-derived opening roster but had zero NBA minutes in the immediately prior regular season.

Contract

For an incumbent, the score remains total prior-season minutes. For an opening candidate with zero prior-season minutes but an earlier NBA season with minutes, the model substitutes the player's most recent active-season minutes, scaled by \(\gamma\):

\[ S_i = \begin{cases} M_{i,s-1}, & M_{i,s-1}>0 \\ \gamma M_{i,a(i)}, & M_{i,s-1}=0,\ M_{i,a(i)}>0 \\ 0, & \text{otherwise.} \end{cases} \]

Here \(a(i)\) is the latest completed season before \(s-1\) where the player logged regular-season minutes. The scores are roster-normalized and blended with the same uniform cold-start mass \(\epsilon\) as L20-MSP v0.0.

This is deliberately not an injury classifier. It is a strictly pregame proxy for a player returning from any full-season NBA absence, including injury, international play, or time out of the league.

Evaluation

Tune \(\epsilon\) and \(\gamma\) jointly on 2024-25 by mean team allocation total variation, then freeze both values for 2025-26. The target remains each team's cumulative minute allocation over its first 20 games. A lookback of up to three prior seasons supports longer absences without using target-season data.

uv run nba-evaluate-l20-gap-returner-minute-share \
  --tune-season 2024-25 \
  --holdout-season 2025-26 \
  --team-games 20

The result will determine whether this recovery prior supersedes the uniform cold-start treatment for returning players.

Initial Result

The 2024-25 source grid selected \(\epsilon=0.32\) and \(\gamma=0\). Thus, within this one-season tuning slice, carrying forward earlier minutes for a zero-minute returner did not improve mean team allocation total variation. The frozen 2025-26 result is consequently identical to L20-MSP v0.0: mean total variation \(0.252\) and Brier score \(0.024\).

This is a preliminary non-promotion. Only six qualifying gap returners were in the source opening rosters, and players absent for a full season are a mixed population: injury returners, players who left the NBA, and players returning with substantially different roles. The next defensible test, if pursued, is a rolling multi-season selection focused on the gap-returner subset rather than changing the baseline from this sparse result.