The honest headline
Our main question is simple: is the model better than a basic shortcut, and better than the sportsbook? Here is where that stands today.
Every model, live
Real forecasts on real games. Small samples swing a lot, so each model carries a label for how much its record can tell you yet.
How to read the labels
Too early to judgeFewer than 300 graded forecasts. Luck can explain almost anything at this size.
Early read300 to 999 graded. Patterns start to mean something, but can still flip.
Meaningful sample1,000 or more graded. Enough to take the numbers seriously.
52.4% is the barAt standard -110 odds you need to pick the right side 52.4% of the time just to break even.
Tested on past seasons first
Before a model logs a single live forecast, it has to predict whole past seasons it never saw, one game at a time. These are tests on old games, not live results.
| Area | Model | Result on unseen seasons | Compared with |
|---|
Honest about the limits
It's not a lock machine.Most nights the model and the sportsbook roughly agree. Real edges are rare and small, and any model can lose for weeks.
Game lines are the hardest.In past-season tests, NFL closing spreads were slightly more accurate than our game model. Treat game forecasts as a second opinion.
Late news can beat it.Injury reports are checked in the morning and again before games, but a surprise scratch can slip through.
Some markets are new.NFL props have about two and a half seasons of data behind them. Passing yards are no better than a simple average yet.
Version history
Every change to a model is dated here, including the ones we pulled.