Forward Availability v0.2

Forward Availability v0.2 corrects the first-state overconfidence found in v0.1. It remains a separate availability model, not yet an input to the minutes or win-projection pipeline.

Corrected State Contract

The availability panel contains every known rostered player-game, including players who never record an NBA minute. Age comes first from the player-season panel, then from a locally cached roster birth date. A remaining row without an age is retained and receives the pooled availability baseline rather than being dropped from the fit.

For a player's first observed season, v0.1 used the raw observed share as the state. A 0 / 82 rookie season therefore became nearly certain evidence of zero future availability. v0.2 instead starts from the age baseline with a tuned pseudo-game strength \(q_{\mathrm{init}}\):

\[ \widehat p^{\mathrm{post}}_{i,t} = \frac{q_{\mathrm{init}} p^{\mathrm{age}}_{i,t} + A_{i,t}} {q_{\mathrm{init}} + N_{i,t}}. \]

This applies before the player-specific state is carried into the next season. Later observations retain their separately tuned update strength \(q_{\mathrm{update}}\).

Frozen Replay

The corrected panel covers 2015-16 through 2025-26. Hyperparameters were selected from the completed 2020-21 through 2022-23 targets, and the model was then evaluated on frozen 2023-24 through 2025-26. The selected configuration was:

Parameter	Value
State persistence \(\rho\)	0.50
Later-state update strength \(q_{\mathrm{update}}\)	60 pseudo-games
Initial-state strength \(q_{\mathrm{init}}\)	15 pseudo-games
Lagged workload weight	-0.25

Frozen metric, mean of three seasons	Age-only control	Availability v0.2
Exposure-weighted Brier score	0.06133	0.05816
Binomial log loss	0.56145	0.55168
Player-share MAE	0.19984	0.19145
Player-share RMSE	0.25888	0.25338

v0.2 wins every aggregate frozen metric. It improves probability calibration without relying on games played as the target.

Artifact: artifacts/rotation/forward_availability/forward-availability-20260909T233439Z-e341ed5.

Returner Check

Chet Holmgren's missed 2022-23 rookie season is a direct regression check. The raw panel records 0 / 82 availability for that season. v0.1 predicted only 0.3% availability for 2023-24; v0.2 predicts 47.3% before his return and the realized value was 100.0%. The estimate remains deliberately conservative, but it is no longer locked near zero.

Thomas Sorber is also retained despite zero NBA minutes in 2025-26: the panel records 1 / 82 availability at age 20 from the roster birth date. He is no longer silently omitted from the model family.

Integration Boundary

This is the first availability model that clears its age-only benchmark. The next product decision is integration, not a replacement of the minutes model:

\[ E[\mathrm{minutes}_{i,t}] = P(\mathrm{available}_{i,t}) \times E[\mathrm{minutes}_{i,t} \mid \mathrm{available}_{i,t},\ \mathrm{team}]. \]

The probability is player-specific. Conditional minutes remain a joint team allocation, so availability should be wired in only with an explicit reallocation contract.