[warmly] Hey, it’s Duke Sizemore. Quiet August board today, so rather than replay Washington’s late-season slide from yesterday, I want to follow the more interesting thread underneath it: the model has been getting noisy injury information, it has now changed its own settings twice, and that matters when we look at some surprisingly large disagreements with the October betting market.
First, the check-in. Over the latest 25-game sample, COURT VISION is at 60 percent on moneyline picks, with an average margin error of about 14.8 points. Across the broader tracked sample, it’s 66.8 percent over 1,246 games. So the recent window has clearly been weaker.
[thoughtful] The useful part is that we have a pretty specific suspect. Detroit was a 75.2 percent home favorite in the model against Cleveland on May 18 and then lost by 31. That produced a 37.9-point margin error. Tristan Enaruna had a day-to-day designation carrying just 0.1 impact. There were also two New York blowouts, against Philadelphia and Cleveland, where Jalen Brunson was tagged day-to-day with an impact of zero.
In plain English, a label was getting influence even when the accompanying impact estimate basically said, “this shouldn’t matter.” At the same time, the model’s news-correlation pipeline stopped producing player and team tags and was pulled. That pipeline had apparently been helping provide context around exactly this kind of injury noise.
And here’s the fun learning-machine part: the model actually adjusted itself. The day-to-day injury weight went from 0.28 to 0.20, and it has since been cut again, from 0.20 to 0.14. [chuckles] That doesn’t mean the computer has achieved enlightenment. It means recent misses gave us evidence that this particular input was being trusted too much, so its influence got reduced while the news pipeline is being repaired.
There’s another data-quality warning today: no team cleared the power-mover threshold versus August 15, but the model itself says that feed has been frozen at zero movers since August 11 and needs a freshness check. So I’m not going to interpret zero as proof that NBA markets have suddenly entered a state of perfect tranquility.
Instead, Detroit is worth spending some time on because we actually have player-level on/off data there. Elijah Harkless played 26 games and averaged 6.8 points, 2 rebounds and 2.9 assists. He’s currently listed healthy, with no injury report. His on/off net number is minus-0.506.
What does that mean? On/off is comparing how the team performs per possession with a player on the floor versus when he’s off it. Harkless being at minus-0.506 says Detroit’s net performance was about half a point worse by this measure with him on the floor than without him. That’s not the same thing as saying Harkless personally costs Detroit half a point; lineups, opponents and role all get mixed into the result. And with only 26 games, I’d treat it as evidence to investigate, not a verdict.
The Detroit comparison is interesting, though. Gary Harris is at minus-0.385 across 48 sampled games, John Collins minus-0.234 across 69. Wendell Moore Jr. is positive 0.135, although that’s only six games, and Chaz Lanier is positive 0.12 over 33. [curious] So Detroit’s chemistry snapshot isn’t screaming one giant lineup effect. It’s giving us several relatively modest signals with very different sample sizes.
Actually, hold on, let me pull up Sonar real quick. Harkless is listed with Detroit on a two-way contract with the Motor City Cruise, and he was born February 3, 2000. That two-way context is useful when you’re looking at a 26-game NBA sample; role and lineup stability matter when we interpret these numbers.
Now we get to October, because the posted odds give us something much livelier than the August news cycle. Detroit hosts Boston on October 20. The market’s implied Detroit probability is about 54.5 percent; COURT VISION has the Pistons at 65.2 percent. Market spread: Detroit minus 1.5. Model spread: Detroit minus 3.9. That’s a 10.7-percentage-point win-probability disagreement and about 2.4 points on the spread.
But the two biggest gaps on this board are elsewhere. New York hosts Philadelphia on October 20, and the market has the Knicks at roughly 67.2 percent. The model says 90.2 percent. It makes New York minus 13.8 rather than the market’s minus 5.5. [genuinely surprised] An 8.3-point spread gap this far ahead of the game is not subtle. It’s also exactly the kind of number I’d want to keep revisiting as availability information improves, especially given what we just learned about the injury pipeline.
Then Washington hosts Milwaukee on October 21, and this is almost a mirror-image disagreement. The market makes Washington a 70.4 percent implied favorite and has the Wizards minus 6.5. COURT VISION has Washington at only 46 percent and actually makes Milwaukee a one-point favorite. That’s a 24.4-percentage-point gap and seven and a half points of spread disagreement.
Brooklyn-Charlotte is similarly large. The market gives Brooklyn about a 41.3 percent home win probability and Charlotte plus 4.5; the model gives Brooklyn just 16.4 percent and makes Charlotte a 10.1-point favorite. Again, that’s not me telling you an August number is destined to beat an October closing line. It’s where the model and today’s market disagree, and we now know one of the model’s important contextual inputs is under repair.
Around the league, there are at least a couple of real personnel and team-building notes in the data. Dallas has extended Naji Marshall on a three-year, $52.2 million deal. There’s also a report that Victor Wembanyama will host some Spurs teammates for workouts in France. San Antonio is already on that October board too: at home against Oklahoma City on October 21, the model gives the Spurs a 65.9 percent chance compared with the market’s 55.4, and makes them minus 4.1 instead of minus 1.5.
[pause] The next confirmed game in the schedule data is still Miami at Toronto in preseason on October 3, so we’re working well ahead of actual basketball here. That makes uncertainty a feature, not an inconvenience. The numbers will move when information moves, and right now we’ve got unusually good reason to watch exactly how the injury information gets processed.
That’s the thread I’m keeping an eye on: not whether the model missed, but whether we can identify why it missed and make the next probability better. Today, we actually can point to a mechanism and a concrete adjustment from 0.28, to 0.20, and now to 0.14 on that day-to-day injury weight. [warmly] I’m Duke Sizemore. Thanks for hanging out with me, and I’ll catch you on the next COURT VISION Daily Brief.