[warmly] Hey, it’s Duke Sizemore. Coffee’s here, September is here, and today I want to talk about a team that finished with the strongest rating in our model and then watched more than a hundred Elo points disappear without losing a basketball game. That’s the Knicks. We’ve also got the model changing one of its own injury settings, an interesting Detroit chemistry number, and some very large disagreements with the early October betting lines.
First, the accountability check. Over the latest 25-game sample, COURT VISION picked the moneyline winner 60 percent of the time, with an average margin error of 14.8 points. The broader tracked sample is better: 1,246 games at 66.8 percent moneyline accuracy, with a .211 Brier score. Brier score is basically our probability report card: it punishes you for being confidently wrong, rather than merely asking whether you picked the winner.
And we have been confidently wrong in some interesting places. [thoughtful] Detroit was a 75.2 percent model favorite at home against Cleveland on May 18 and lost by 31. That produced a 37.9-point margin error. The review found a recurring problem around day-to-day injury tags, especially players whose estimated impact was effectively zero. The news-correlation pipeline had also stopped generating player and team tags, which meant the model had lost some of the context it normally uses to interpret those injury labels.
So, yes, the model actually tuned itself. The day-to-day injury weight was cut from .20 to .14 while that noise was being investigated, and the most recent tuning moved it back up to .182 after the miss board suggested genuine availability information was then being discounted too aggressively. That is a nice example of what self-tuning actually means. It’s not the computer waking up and discovering basketball enlightenment. [chuckles] It’s finding an error pattern, changing a parameter, and then checking whether the correction itself went too far.
Now New York. The Knicks finished the modelled schedule at 70-and-32, number one in Elo at 1808, with a 41.5-point recent swing. They won five of the six games in the trajectory we have here. That includes four wins in five against San Antonio and a 130-to-93 win at Cleveland that alone added 21.6 Elo points.
For four games where we have the advanced sample, New York scored 128.8 points per 100 possessions and allowed 102.6. That’s a plus-26.2 net rating. Tiny sample, enormous number. You absolutely should not project plus-26 basketball over a season, but it tells you why the model loved the finish.
Then preseason regression comes along and says: okay, prove it again. New York’s new prior is 1699.8 Elo, third rather than first. The carry component is 1700.2 and the market component is 1699.5, so those two inputs are almost perfectly aligned. In plain English, Elo is our running estimate of team strength: beat good teams, especially unexpectedly, and it rises; lose unexpectedly and it falls. A preseason prior deliberately compresses the extremes because rosters, health and circumstances change between seasons. So the drop from 1808 to about 1700 is not a claim that the Knicks suddenly got 108 points worse. It is uncertainty being put back into the forecast.
One important uncertainty is Jalen Brunson. He’s listed day-to-day and is expected to undergo surgery on the right wrist and forearm area, with at least two months of rehab expected. Given everything I just told you about the model’s problems with day-to-day labels, that is exactly the sort of case where the label and the actual medical context need to be separated carefully.
There isn’t a meaningful fresh market move today, either. The power-movers board has nobody crossing its threshold versus August 31. And the title-race snapshot is marked stale, so I’m not going to turn old futures prices into a new story.
Let me move over to Detroit, because the chemistry data gives us something worth unpacking. Elijah Harkless has a minus-.506 on/off net figure across 26 sampled games, the weakest figure in Detroit’s group here. Harkless averaged 6.8 points, 2 rebounds and 2.9 assists in 26 games, and he currently has no health report.
What does minus-.506 mean? It’s comparing Detroit’s net performance with Harkless on the court against its performance with him off. The team-level swing was about half a point worse with him on. That does not mean Harkless personally costs Detroit half a point every night. Lineups, opponents, roles and teammates all live inside an on/off statistic. It’s a clue about how his minutes functioned, not an individual verdict. Gary Harris is also negative in this sample at minus-.385 over 48 games, while Chaz Lanier is plus-.12 over 33.
[amused] Actually, hold on, let me pull up Sonar real quick, because there’s a Harkless detail sitting here that fits nicely. He played for Oklahoma in 2020-21 and became one of the key defensive backcourt pieces on a Sooners team that reached the NCAA tournament. So there’s some useful context for the player behind that on/off number: his résumé includes defense as a real part of his profile, even though Detroit’s team results in his minutes last season were slightly worse by this measure.
There are a couple of actual personnel stories around the league as well. Dallas extended Naji Marshall for three years and $52.2 million. Victor Wembanyama is set to host some Spurs teammates for workouts in France. And on the Knicks’ organizational side, executive Gersson Rosas is reported to be leaving after four years with the team.
Now the early betting board, because this is where things get spicy without us needing to pretend October is tomorrow. [curious] On October 20, Detroit hosts Boston. The market has Detroit at about a 54.5 percent home probability and minus-1.5; our model has the Pistons at 65.2 percent and roughly minus-3.9. That's about an 11-point probability disagreement.
But New York against Philadelphia later that day is much larger. The market makes the Knicks about 67.7 percent to win at home and has them minus-5.5. COURT VISION currently says 90.2 percent and minus-13.8. That’s a 22.5 percentage-point probability edge and an 8.3-point spread gap toward New York. With Brunson’s rehab situation and weeks still remaining before the game, treat that as an early model-market disagreement, not a number carved into granite.
And the board contains two disagreements in the other direction on October 21. Washington is priced as about a 69.7 percent home favorite against Milwaukee, while our model has Washington at only 46 percent and actually makes Milwaukee a one-point favorite. Brooklyn is getting 4.5 at home against Charlotte, but our model has Brooklyn at only 16.4 percent to win and makes the Nets roughly 10-point underdogs. Those are exactly the games I want to revisit as health, rotations and preseason information arrive, because large September edges can be real information—or a neon sign showing you where one side of the model is stale.
The next confirmed game in the feed is still Miami at Toronto on October 3, so we have some runway before actual basketball starts testing these preseason assumptions. Until then, the interesting work is figuring out which signals deserve to survive the summer and which ones need to be discounted.
[warmly] I’ll keep watching that Knicks reset, Brunson’s health context, and whether those big early market disagreements survive as October gets closer. Thanks for hanging out with me. I’m Duke Sizemore, and this is COURT VISION: The Daily Brief. Catch you next time.