Guide · 5 min read
How to read win probabilities
Every game on our picks page comes with a projected margin and a win probability. This guide explains what the probability means.
70% means losing three times in ten
A 70% favorite is expected to lose 30% of the time. Across a weekend with dozens of games at those odds, several favorites should lose. If they never did, the probabilities would be too cautious. Upsets alone do not mean a model is wrong.
Where the numbers come from
A projected margin is the rating gap between two teams, scaled to real scores, plus about 3.6 points for the home team. Real results scatter around those projections, about 19 points either way two times in three. A 7-point favorite therefore wins most of the time but far from always, and the win probability is the share of that scatter that lands on the favorite’s side.
How to judge a model
- Calibration. Of all the games a model called 70%, did the favorite win about 70% of them? This is the key test.
- Out-of-sample grading. A model must be graded on games it had not seen. Ours projects every game using only results from earlier weeks.
- Picking winners. Straight-up accuracy is easy to understand, but it rewards picking obvious favorites. Use it alongside calibration.
This season the Power Board has picked 1,219 winners in 1,699 graded games (72%), with projected margins missing by 15.0 points on average. Every week’s graded picks are public, starting from the weekly reports.
What the model does not know
Injuries, weather, suspensions, motivation, and depth charts are all invisible to a ratings model that reads only scores. These numbers are for information and entertainment, not betting advice.