CONTRACT SURPLUS MODEL

What it measures

A hierarchical regression of log cap share on last season's output, where each stat has a league price that roles can bend toward their own, and where a maximum salary counts as "at least this much".

How to read it

fair salary (point estimate and 80% interval); surplus (point estimate and 80% interval); probability that surplus is positive; verdict, max-ceiling and limited-availability flags; dollar waterfall by stat group, age and availability; similar producers and similar contracts; market summary (league and role prices, age curve)

Why it exists

The model is Bayesian because the question needs an honest range, not a single number: fair salary comes with a posterior predictive interval, and verdicts are withheld when that interval straddles zero. Role-level prices are partially pooled, so a big man's rebounds and a guard's rebounds can be priced differently without small groups overfitting. Maximum contracts are censored: the league cannot pay a star more than his tier, so his salary is evidence of a floor, not a price. Structure that failed on real fits is recorded in the code and not repeated: free and learned-scale group effects did not converge, so season and role effects are sum-to-zero with fixed scales. The model publishes only after beating the method it replaces on a season it never saw.

Limits

• Box-score features miss most defensive and off-ball value.
• Position family is inferred from the box score for players with no listed position (76% accuracy).