[warmly] Hey, it’s Duke Sizemore. The Knicks are giving us a nice little offseason modeling lesson today, because two things can be true at once: New York finished the modelled run as the hottest team on the board, and our preseason process immediately knocked them from number one to number three. I want to explain why that isn’t actually a contradiction. We’ve also got a very large Knicks-Sixers disagreement with the sportsbook sitting out there for October, and the model has been tinkering with its own injury settings again.
First, the accountability check. Over the latest 25-game sample, we’re at 60 percent moneyline accuracy, with an average margin error of 14.8 points. That’s not a particularly pretty margin number. One of the clearest misses was Detroit-Cleveland on May 18: Detroit had a 75.2 percent home win probability and lost by 31. The resulting margin error was 37.9 points.
The model’s latest diagnosis keeps coming back to injury noise. Low-impact day-to-day labels have been influencing projections even when they haven’t represented meaningful absences. At the same time, the news-correlation pipeline that was supposed to add context produced zero player and team tags over its latest seven-day window and was pulled. So the system was relying more heavily on injury labels without one of the signals meant to help interpret those labels.
[curious] And yes, the self-tuning process actually changed a setting. The day-to-day injury weight was cut from point-two to point-one-four, then most recently moved back up to point-one-eight-two after another miss suggested the model had gone too far in the other direction. That little back-and-forth is important. Learning doesn’t mean every adjustment marches perfectly toward the answer. Sometimes you turn the knob down, discover you turned it down too far, and bring it partway back.
That matters today because Jalen Brunson is currently listed day-to-day, with the supplied report saying he’s set for surgery on his right wrist and forearm area and expected to need at least two months of rehab. So rather than blindly treating the words “day-to-day” as one universal quantity, this is exactly the kind of case where the underlying medical context matters.
Now, New York. Their closing run really was strong. The Knicks finished at 1808 Elo, number one in the model, with a 70-and-32 record. They won four of the five listed games against San Antonio, including 94-90 on June 14. But the biggest Elo jolt in this sequence actually came against Cleveland on May 26: a 130-93 Knicks win took them from 1767 to 1788, a 21.6-point jump.
Elo, in plain English, is our rolling estimate of team strength. Beat somebody the model thinks you should beat and you gain a little. Produce a result that’s much stronger than expected and you gain more. That Cleveland result was strong enough to move the estimate substantially. Across the recent stretch, New York’s swing was 41.5 points.
[thoughtful] Then preseason blending comes along and says: don’t assume the last team we saw is exactly the team we’ll get months later. New York’s carry Elo is 1700.2 and its market Elo is 1704.4. Blend those and the preseason number lands at 1702.3, third in the league rather than first. Oklahoma City is at 1746 and San Antonio at 1744. New York’s supplied title probability is 8.23 percent at plus-900 odds, although the title-race feed is flagged stale, so I would not treat that as fresh market information today. And there were no qualifying day-over-day power movers from August 21 to August 22 either.
There are a couple of actual offseason changes worth keeping on the radar. A supplied headline says Knicks executive Gersson Rosas is leaving after four years. Dallas, meanwhile, extended Naji Marshall for three years and 52.2 million dollars. And on the Spurs side, Victor Wembanyama is set to host some teammates for workouts in France. Those aren’t Elo moves by themselves; they’re pieces of context we’ll eventually get to test against the basketball.
For today’s chemistry microscope, Detroit gives us something unusual. Elijah Harkless averaged 6.8 points, 2 rebounds and 2.9 assists over 26 games, and he currently has no health report. His on/off net number in this data is minus point-five-zero-six across those 26 games.
What does that mean? An on/off net-rating figure compares how the team performs with a player on the court against how it performs when he’s off. Harkless sitting around minus half a point means Detroit’s net performance was modestly worse in his minutes relative to the minutes without him. It is not the same thing as saying Harkless personally costs Detroit half a point every game. Lineups, opponents, roles and teammates all get tangled into these numbers. Gary Harris is at minus point-three-eight-five over 48 games and John Collins at minus point-two-three-four over 69, while Chaz Lanier is plus point-one-two over 33. These are useful clues, not isolated player grades.
Actually, hold on, let me pull up Sonar real quick. [chuckles] Harkless had a pretty winding route before this point: he started his college basketball at UNLV and later played at Oklahoma. That fits nicely with why I’m reluctant to flatten a player into one on/off number. Careers and roles are rarely that tidy.
Now for the October board, because this is where things get spicy even this far ahead. On October 20, Philadelphia visits New York. The market has the Knicks at about a 67.7 percent home win probability and minus five-and-a-half. Our model has New York at 90.2 percent and minus 13.8. That’s a 22.46 percentage-point probability disagreement and an 8.3-point spread disagreement. [genuinely surprised] That is enormous.
It’s also exactly where I’d be cautious given what we just discussed about Brunson and injury processing. The model likes New York much more than the current posted price, but October availability is a real unknown in August, and our own recent evaluation says injury handling has been a weak spot. So this is a model-market disagreement worth tracking, not a reason to pretend uncertainty disappeared.
There are two other big discrepancies on October 21. Washington is priced by the market as roughly a 69.7 percent home favorite against Milwaukee, with Washington minus six-and-a-half. The model actually has Washington at only 46.0 percent and makes Milwaukee the slight favorite, equivalent to Washington plus one. Brooklyn-Charlotte goes the other way too: the market gives Brooklyn about a 41.3 percent home chance and plus four-and-a-half, while the model is way down at 16.4 percent and has Brooklyn plus 10.1.
Those early lines are going to be fascinating to revisit once rotations and health become clearer. Before that, the next confirmed game in this feed is Miami at Toronto on October 3, so right now we’re measuring expectations rather than reacting to fresh results.
[pause] And that’s really today’s story. The Knicks can finish at 1808 and still open the preseason at 1702 without the model changing its mind. One number describes where they finished; the other asks how much of that ending should survive an offseason reset. Meanwhile, the model itself is still figuring out how aggressively to trust those little day-to-day injury labels. I kind of love that these two ideas arrive together: basketball changes, and the uncertainty around basketball changes too.
[warmly] I’ll keep an eye on that huge Knicks-Philly gap as October gets closer and see whether the market, the model, or both start moving. Thanks for hanging out with me. I’m Duke Sizemore, and this is COURT VISION: The Daily Brief. Catch you next time.