Articles
Findings from the experiments behind HoopModel: what makes NBA games predictable, what doesn't, and how we decide whether a model change is real.
Minutes explain only a sixth of a player's scoring miss
When a points projection misses, playing time gets the blame. We handed the model the real minutes after the fact and only about a sixth of the miss went away. What that says about reading projections.
Why we predict steals and blocks as chances, not numbers
Nobody gets 0.7 steals. We turn each projection into the chance of 0, 1 or 2+, and three seasons of results show it works for typical players, slips for the busiest shot-blockers, and resists an easy fix.
The simple baseline that wasn't: testing our player projections
Before launch, our player model had to beat two simple methods. The famous one, Marcel, lost to a plain 10-game average on most stats. Our model beat both almost everywhere, with two honest exceptions.
What a 70% win probability actually means, and the fix that broke our picks
A win probability is a promise about frequencies. We added a layer to make ours more honest, and for two months it made one model pick the home team in every game. What we learned, and what we do now.
Early-season picks are weaker, and so is the betting market's
October is the hardest month of the season for our models and for the betting market. Three seasons of results show why, and why feeding the models more old data wouldn't fix it.
Signal or noise? How we decide whether a model change is real
A new feature that 'adds 2.6 points of accuracy' on a month of games usually adds nothing. Here is the testing routine that caught it, and the one change it let through.
Why nobody predicts NBA margins within five points
Our models miss the final margin by about 11.6 points on average, and the betting market by about 10.8. That isn't a bug. Here's how much of an NBA margin is predictable at all.
Last season's impact ratings don't help predict this season
All-in-one player impact metrics look like an obvious input. We tested them twice, in the player model and the team model, and both times they added nothing the models didn't already know.
Playoff games are harder to predict, and it's not the model's fault
Trained on the regular season, our models lose about 7 points of winner accuracy in the playoffs, yet their margin ranges stay just as honest. What changes when the postseason starts.