How reliability is measured

Six checks

Each is scored from 0 to 100%:

• Calibration: in games the model had never seen, did picks given this probability win about that often? The further off, the lower the check.
• Games behind it: how many past games stand behind that measurement. Few games, little to judge by.
• Season: how much of the season the two teams have played. Early on, the model knows less about them.
• Freshness: how old the sports data behind the prediction is.
• Starters: where one player decides a lot — the starting pitcher, the quarterback, who is out in basketball — whether that is known or only assumed.
• Outside estimates: how far the model is from outside estimates of the same game. Those are the prices of prediction markets and the sportsbooks' lines, which iPick Sport reads and does not publish.

Why the last check weighs so much

It is the one that changes most from one game to the next, and it was measured. On 3,580 NFL games from 2013 to 2026, the model's pick won 68% of the time when its probability was within 3 percentage points of the outside estimate, and 56.5% when it was 15 or more away. When the two disagree, it is usually the model that is wrong.

A game with no outside estimate yet is scored as a typical disagreement, never as an agreement: not knowing is not the same as agreeing.

From six checks to one number

Reliability is half the average of the checks and half the weakest of them, so that one bad check cannot be averaged away. High is 70% or more, Medium 55% or more, Low anything under that.

What it changes, and what it does not

Reliability does not move the pick or its probability. It widens the probability range, by half the distance to the outside estimate, and on Picks it is shown beside every pick.

Where to see it

On the Match Center, “Why?” under a game's reliability opens the model's own checks one by one, each with its score. The other numbers of a game are explained in How to read a prediction.