How to read a win probability
A 62% chance is not a prediction that the team will win. It is a claim about a hundred games like this one. Here is how to read the number, and how not to.
A scoreboard tells you what happened. A win probability tells you something stranger: how often this would happen if tonight could be played many times. It is a number about a crowd of games, attached to just one of them. Most of the confusion about probabilities comes from forgetting that.
62% means 38 losses in 100
When a model gives a team a 62% chance, it is not saying the team will win. It is saying that across a hundred games that look like this one, with the same gap between the teams and the same side at home, the favorite takes about 62 and drops about 38. Tonight might be one of the 38. Nothing went wrong if it is.
That is the first habit worth building: read the small number as carefully as the big one. A 75% favorite loses one game in four. Over a baseball season that is a lot of “upsets” that were never really upsets.
Most games are closer than they feel
How far the numbers stretch depends on the sport. In EdgeBoard’s backtest as it stood on October 1, 2026, the model gave a baseball favorite more than 70% in about one game in a hundred, and never reached 80%. In hockey, fewer than one game in ten passed 70%. Football and basketball stretch further, because the better team wins more reliably there: more than a quarter of those games passed 70%.
This surprises people, because pre-game talk is full of certainty. But the results agree with the modest numbers. If baseball favorites really won nine times in ten, the standings would look very different by the end of a season.
A probability is not a margin
60% does not mean “will win by a little”, and 85% does not mean “will win by a lot”. A probability says how often, not by how much. Two evenly matched teams can produce a blowout. A heavy favorite can scrape through in overtime. The number only speaks to who ends up ahead.
What moves the number
In a ratings model like the one EdgeBoard uses, three things set the number for a game:
- The rating gap. Each team carries a rating built from its past results. In this kind of model a gap of 100 points is worth about 64%, and 200 points about 76%.
- The home edge. The home team is treated as a little stronger than its rating. How much depends on the league.
- Short rest. A team that played yesterday, facing one that did not, gives up a few points.
Recent form is not a fourth ingredient. A team on a winning run already has a higher rating because of those wins. Adding form again on top would count the same games twice.
What does not move it
The model has never heard of an injury. It does not know who is pitching, who is in goal, or that a team traded for a star last week. It only sees scores, and it notices a change in a team once the scores start to reflect it. That is a real limit. It is better to know it than to imagine the number is smarter than it is.
One game is not evidence
After a 70% favorite loses, it is tempting to say the model was wrong. After it wins, it is tempting to say the model was right. Neither follows. A model that said 70% and a model that said 55% would both have “called” the win.
The only fair test is to collect many games. Take every game where the model said about 70% and count how often the favorite won. If the answer is close to 70%, the number means what it says. If favorites at “70%” win only 58% of the time, the model is overconfident, and you have learned something useful. This is called calibration, and it is what the chart on the Record page shows.
Three questions to ask of any number
- Was it stated before the game? A number worked out afterward is a different thing from a number put on the record in advance.
- How did the last few hundred turn out? Not the last three. Not the best week.
- Compared with what? Picking the home team every time is right more than half the time in every major league. A model has to beat that to be worth reading.
A probability that passes those three is worth looking at. It still is not a promise about tonight, and it is not betting advice. It is a careful statement of how unsure we ought to be, which is rarer than it should be.