L0-MSP v0.0

Season-Opening Minute-Share Persistence is the first preseason baseline in the Rotation Models family. It predicts only a team's first regular-season game, using information that existed before that season's opening date.

Contract

For each team, take the player minute-share allocation in its final completed regular-season game of season \(s-1\). Restrict that allocation to the transaction-derived opening candidate roster \(C_{T,s}\), then renormalize it:

\[ p_i = \frac{y_{i,s-1}^{final}\,\mathbf{1}[i \in C_{T,s}]} {\sum_{j \in C_{T,s}} y_{j,s-1}^{final}}. \]

Players new to the candidate roster begin at zero. L0-MSP reserves uniform cold-start mass \(\epsilon\) across all candidates:

\[ \hat y_i = (1 - \epsilon)p_i + \frac{\epsilon}{|C_{T,s}|}. \]

The source season selects \(\epsilon\) by mean allocation total variation; ties select the smaller value. The following season is then evaluated once with that frozen value.

Initial Evaluation

Using 2024-25 as the source season selected the interior value \(\epsilon = 0.80\). The frozen 2025-26 result did not carry that advantage forward:

Allocation	2024-25 source TV	2024-25 source Brier	2025-26 frozen TV	2025-26 frozen Brier
L0-MSP, frozen \(\epsilon=0.80\)	0.3052	0.0359	0.4409	0.0596
Uniform opening-candidate control, \(\epsilon=1\)	0.3164	0.0370	0.4339	0.0534

The previous final-game allocation has weak source-season signal, but that signal did not generalize to the next opening night. L0-MSP is therefore a useful boundary and control, not a promoted forecast. The next candidate should learn player-specific preseason propensity from prior-season availability and role indicators, rather than increasing reliance on a single final game.

Leakage Boundary

The historical opening-roster mart validates transaction reconstruction against opening-game box scores. Those after-the-fact reconciliation rows are intentionally excluded here: L0-MSP uses only membership_source = transaction_reconstruction. An actual opening player absent from that source state receives zero predicted share and is reported as outside_candidate_actual_share.

This is a harsh but honest preseason baseline. It cannot infer injuries, training-camp cuts, or a newcomer omitted by the transaction feed.

Run

uv run nba-evaluate-l0-opening-minute-share-persistence \
  --tune-season 2024-25 \
  --holdout-season 2025-26

The evaluator writes its epsilon grid, 30 source opening forecasts, 30 frozen holdout forecasts, team-level metrics, and metadata under artifacts/rotation/l0_opening_minute_share_persistence/.