Our All-Stars and award winners

Our major award selections

Point selections rank expected contribution. The frequencies show how often a player leads in the model’s uncertain paths; they are merit probabilities, not actual voting odds. Small samples can produce surprising and fragile selections.

Our representative teams

Our All-Stars are positionless: 12 from each conference, with the first five labelled starters. These are our merit rosters. Official voting, roster formats and games-played eligibility exceptions are outside this definition.

Every observed player

Contribution is estimated points added above league average; defensive contribution is estimated points prevented. The range contains the middle 80% of model paths. Teammates’ effects are correlated and can be difficult to separate.

What this model measures

When a full game passes strict lineup, score and minutes checks, the model uses rotation stints. Otherwise it uses the game’s verified team scores with actual player minute shares. A game enters once at one resolution. Aggregate-game rows add extra model noise and assume possession shares follow minute shares. They cannot distinguish teammates as well as rotations.

Stan fits partially pooled offensive and defensive player effects, a league scoring level, home offence and a shared game scoring environment. Generic zero-centred player priors are identical for everybody. Scoring errors use a robust Student-t distribution and estimated possessions. These are statistical attributions within shared lineups, rather than isolated causal player effects.

Current-season regular games supply all evidence for the current player ratings. No preseason player ratings, prior-season performance, salaries, draft slots, market prices or voting history enter this fit. Reported current roster membership and rookie status supply identity and selection metadata only. An unrecognised rookie status is withheld from rookie selections. Most Improved independently fits last season as a comparison only, using coarser team-game observations. The change includes uncertainty and differences in evidence resolution; it does not prove individual development.

Season-end paths hold measured impact and current minutes per appearance fixed, then simulate remaining appearances from this season’s participation and starts. They assume an 82-game team season, current team membership and continuing roles. Departed players keep earned contribution and receive no future minutes. This does not forecast injuries, trades, improvement or future coaching decisions; October evidence makes the extrapolation particularly uncertain.

MVP and All-NBA use total net contribution; DPOY and All-Defence use defensive contribution. Rookie selections use net contribution among reported rookies. Sixth Man uses net contribution among players with a bench majority; a projected point selection needs at least 50% probability of finishing with more bench appearances than starts. Coach and clutch awards require other definitions and are not assessed here.