How to read a prediction
Every game on iPick Sport carries the same few numbers. This page says what each of them means, and what it does not.
The pick and its probability
The pick — iPick — and its percentage are the model's probability weighed against outside estimates of the same game; where there is no outside estimate yet, they are the model's alone. 57% means that, out of 100 games like this one, the pick is expected to win about 57 — and to lose about 43.
What the model says by itself is under Details, as “Model's own probability”.
The two teams' percentages add up to 100. A game at 51% and 49% has a pick too, and it is close to a coin flip.
Pick strength: Toss-up, Lean, Strong
Every list shows a pick in one of three levels of its own probability — a name for the number, not a second opinion:
- Toss-up: under 55%. Close to a coin flip; it loses about half the time.
- Lean: 55% to 65%.
- Strong: 65% or more.
How picks of each level have really done — how many won, next to how many they said would — is on the Track record page and beside each badge. A game's own view names the probability more finely (coin flip, slight, moderate, strong, very strong).
Reliability
A game's own view also shows a reliability: how far the model's own number deserves to be believed for this game, mostly from how far it is from outside estimates of the same game. Measured on the published record in October 2026, it did not separate winners — picks of high, medium and low reliability won about equally often — so the lists show the pick's strength instead.
What it is made of is on its own page: How reliability is measured.
The model's range
The model's range — 52% to 62%, say — is where the model, by itself, puts the pick's real chance once its own past misses are counted. It is wider when few past games stand behind that probability, when the model has been off at that level, and when outside estimates of the same game are far from the model's. The pick's percentage can fall outside it: that is a game where the outside estimates and the model are far apart.
Expected margin and expected total
A second model estimates by how much the winner wins, and how many runs, points or goals the two teams score between them. “ATL by 0.8 ± 4.5” reads: on average Atlanta wins by 0.8, and that estimate has typically been off by about 4.5 in past games.
In baseball and hockey a margin under one run or goal is usual. It is an average over many games like this one, not a score.
The pick and the margin come from two different estimates, so now and then the margin leans to the team that is not the pick. It is then almost always a small fraction of its own ±, and the page says “Close game” instead of naming a team.
While a game is played, and after
- Live: while a game is being played, the percentage shown is ESPN's live win probability. The pick's percentage from before the game is kept in the game's own view, as “Before the game”.
- Final: the score, and a green tick when the pick won or a red cross when it lost.
- Retro: a prediction computed after the game from what was known before it. It is marked, and counted apart from the ones made before the game.
The Picks page
Picks lists the day's picks most likely to be right: the same pick and the same percentage as on every other page, in order.
What none of it is
A probability is an estimate, not a promise, and none of this is advice: see the Terms of Use. How the numbers are made is in What each model knows.