How to Judge a Football Prediction Record
Why win rate alone is misleading, what sample size you need, and the warning signs of a prediction record you shouldn't trust.
By GoalPredicts Editorial Team · Reviewed
Win rate depends on what was predicted
A high win rate is easy to achieve by only predicting heavy favourites, and a low win rate can come from a sensible strategy of predicting less likely outcomes. Comparing win rates across different kinds of selections is comparing different things.
A better question is whether outcomes happened about as often as the stated probabilities said they would. A record of selections at around 70% that wins about 70% of the time is behaving as described; one that wins 50% of the time is not.
Sample size
Short runs are dominated by luck. A forecaster with a genuine edge can lose most of its first twenty predictions, and one with no edge at all can win most of them. Meaningful conclusions usually need hundreds of settled predictions, recorded before kickoff.
Measures forecasters use
Log loss and the Brier score reward probabilities that are both confident and correct, and penalise confident mistakes heavily. Lower is better for both. They are more informative than win rate because they use the probability itself, not just the yes-or-no result.
It also helps to compare a forecast with simple baselines, such as always giving each outcome an equal chance, or using how often each outcome has happened historically. A model that can't beat those baselines isn't adding much.
Warning signs
Be cautious of records that only show winners, that start counting from a convenient date, that change past selections, that publish after kickoff, or that promise guaranteed or 'fixed' outcomes. Nobody can guarantee the result of a football match.
Check the timestamps and the denominator
A trustworthy record shows when each prediction was published, and that time should be before kickoff. Predictions added or changed after a match has started can't be judged fairly, because the outcome may already be partly known.
Also check what is being counted. A record that reports '8 winners this week' without saying how many predictions were made in total hides the denominator. Look for the full count of wins, losses and voids over a clearly stated period, and for the same period being used consistently rather than chosen after the fact.
Our rules
GoalPredicts publishes predictions before kickoff with a publication timestamp that can't be changed. Once published, a selection can never be edited; a prediction can only be withdrawn before kickoff, and that is logged. Every settled prediction, including losses and voids, appears on the results page.