What the Knicks' 2026 Title Teaches Us About Analytics
The Knicks' championship run is a reminder to distinguish player impact, team forecasts and the decisions made inside a series. An unexpected result is a reason to investigate, not proof that a model failed.
Originally published on LinkedIn. Updated 18 October 2026 to correct the playoff figures and distinguish verified results from interpretations of the models. Historical ranking and probability snapshots that I could not substantiate have been removed.
The Knicks’ championship run made me think about a familiar problem in analytics: how easily we turn a useful number into an answer to a different question.
A player impact metric. A team title forecast. An in-game win probability. Each can be useful. They are not interchangeable.
The temptation after an extraordinary run is to say the models got it wrong. That was too strong a conclusion in my original version of this essay. An unlikely outcome can happen under a well-calibrated model. One championship does not settle whether the probabilities were right.
The more interesting question is what the numbers help us understand, and what we still need to investigate.
What actually happened
New York beat San Antonio 4–1 in the Finals, clinching with a 94–90 win on 13 June. Brunson scored 45 and was named unanimous Finals MVP. It was the Knicks’ first title since 1973. NBA Finals game notes
They finished the playoffs 16–3 with a record cumulative point differential of +283. Across those 19 games, that works out to about +14.9 per game. The +19.4 figure in my original essay described an earlier stage of the run, not the completed postseason. NBA postseason record, earlier playoff figures
Game 4 also needs a distinction: their largest deficit was 29, while the halftime deficit was 27. They recovered to win 107–106 on Anunoby’s late tip-in. NBA Game 4 notes
Those results are remarkable without treating them as a verdict against analytics.
Five ways to read the numbers more carefully
1. A player metric is not a Finals MVP ballot
EPM estimates a player’s contribution per 100 possessions, using statistical and adjusted plus-minus components. That is a different question from who had the greatest impact on a particular Finals series.
A disagreement between a leaderboard and an award is a reason to inspect the time window, role and evidence. It does not establish that the metric is blind to the player.
2. Updating a forecast is part of the job
DARKO is a projection system. My original description of it as something that should not be read as a forecast was incorrect.
The useful question is whether a system responds appropriately to new evidence. A short run can contain both a real tactical change and shooting noise. Deciding how much weight to give each is difficult. We need the actual forecast history to judge how well a model handled it.
3. Team context deserves investigation
A series creates repeated encounters with the same opponent. Line-ups, defensive assignments and offensive choices can change as teams learn.
An aggregate rating gives us a starting point. Film, possession data and tactical context help explain what changed. I would want to investigate those changes before attributing a result entirely to a star, a scheme or a model’s supposed blind spot.
4. A comeback does not prove the original probability was wrong
A very small probability is still a probability. To assess calibration, we need many comparable situations and their outcomes.
The original essay asserted that a particular comeback probability should have been higher. Without a reproducible calculation, that was more certainty than the evidence supported. The better question is which information was available at the time and whether it improved predictions across similar games.
5. Agreement is useful, but assumptions still matter
Different forecasts can agree because they capture a strong signal. They can also share data and assumptions. Agreement alone does not tell us which explanation applies.
Equally, a hypothetical 10% title chance is not a prediction that a team cannot win. We should assess forecasts as probabilities, rather than turn them into categorical calls after the result is known.
What I take from it
The gap worth studying is between the question a model answers and the decision someone makes with it.
That applies well beyond basketball. A score can become a verdict. A forecast can become a promise. A useful simplification can become an excuse to stop looking at the underlying system.
The Knicks’ run gives us plenty to investigate. The response should be better questions, clearer comparisons and more careful interpretation. The games are still worth watching, and the numbers are still worth understanding.