The Voters and the Model

Every spring about a hundred people who watch more basketball than is healthy fill out an All-NBA ballot. Every spring a model that has watched none of it, in the sense that it has only ever seen the possessions, does the same thing in its own way. I find it reassuring that they mostly agree. Of the fifteen players the voters named to the 2025-26 All-NBA teams, twelve are in the model’s top fifteen.¹ The top of the ballot is the top of the board: Nikola Jokić first by the model, second in the MVP vote; Shai Gilgeous-Alexander, the MVP with 94% of the possible points, second by the model.

That leaves three players the voters liked much more than the model did, and three the model liked much more than the voters did. I’ll take them in turn, and I’ll tell you whose side I’m on, because a columnist who only reports what the model says is just a very expensive printer.

The voters’ three

Jaylen Brown is the big one. Second-team All-NBA, sixth in the MVP vote, and 90th by the model, which rates him +0.03 per 100 possessions: an average player, give or take. This isn’t one strange season. It was +0.14 → +0.08 → +0.03 over the last three.² Brown is exactly the kind of player plus-minus models and voters argue about. He used 35% of Boston’s possessions when he was on the floor, the 2nd-highest share in the league,³ and the model can’t find much of it in the score when he’s on the floor versus when he isn’t. On $53M a year, that’s −$8.8M of surplus. I’m with the model, with the usual caveat that “the model can’t see it” and “it isn’t there” are different sentences.

Jalen Duren made the third team and is 98th by the model. Jalen Johnson made the third team and is 28th. These two are a different kind of disagreement. Johnson’s rating went −0.29 → +0.71 → +1.32; Duren’s went −0.41 → −0.42 → +0.30. Both are breaking out, and a rating built from several seasons of possessions is always a year late to a breakout. It weighs the past because the past is usually right. Usually. Here I’m with the voters, and I’d note that the model’s error bars on both are about a point per 100 wide, which is the model politely agreeing that it isn’t sure.

And here is the thing about Duren. He is on his rookie deal, $6.5M, and even at the model’s modest rating that’s +$35M of surplus. If the voters are right, Detroit has a bargain. If the model is right, Detroit still has a bargain. Uncertainty is expensive when you’re paying for it with a max contract and a gift when it’s on a rookie scale.⁴

The model’s three

Derrick White (12th), Desmond Bane (13th) and Amen Thompson (15th). None made All-NBA. What they have in common is that they make teams better in ways that don’t pile up in the box score. White rated +2.35 per 100 and made All-Defensive first team, so the voters noticed him, just not on the ballot that counts. Bane rated +2.07. And Thompson, whose brother made an All-Defensive team this year,⁵ produced 9.8 wins on a rookie deal: +$90M of surplus, the most of the three.

I’m with the model on all three, and not only out of loyalty. Of all the players in this column, these are the ones a general manager can actually use: the voters have already decided they’re not stars, so nobody is going to charge you a star’s price for them.

Keeping score

Six disagreements. I sided with the model on four and the voters on two, which makes me either admirably independent or unable to commit, depending on how you feel about me. Check back next spring: the ratings will have another season in them, and one of us will look smart.

Awards: the media’s 2025-26 All-NBA and All-Defensive teams and MVP voting, from Basketball-Reference. Everything else: the season board, 538 players ranked by wins above replacement. Wins are on this model’s own scale and dollars at the open-market price of a win; the method explains both.

1. Out of 538 players, ranked by wins above replacement. The model does not know about positions, which the ballot does, so a perfect match was never on the table. ↩
2. Per 100 possessions against an average player, each with a standard error of about 0.74-0.93. A single season could be noise. Three in a row near zero is harder to wave away. ↩
3. Usage rate among the 367 players with 40 or more games, from the season’s box scores. ↩
4. This is also why the model pays for re-signing rights: if a young player turns out great, the club keeps him; if he doesn’t, it doesn’t. That asymmetry makes a wide error bar on a cheap young player worth money, not just worry. ↩
5. Ausar, of Detroit. The model rates the players, not the family, but it’s a good family. ↩