Ask two tipsters how good they are and both will quote a win rate. The problem is that win rate rewards picking favourites and says nothing about how confident each pick was. A model that calls every match at 51 percent for the favourite can post a respectable win rate while offering no betting value at all. The Brier score fixes this by scoring the probability itself, not just the final pick. It is the standard tool for judging probabilistic forecasts, from weather prediction to sports modelling, and it deserves a place in every bettor's toolkit.

The calculation is simple. For each prediction, take the difference between the predicted probability and the actual outcome, where the outcome is 1 if the event happened and 0 if it did not, then square that difference. Average those squared differences across all predictions and you have the Brier score. Lower is better, with zero meaning perfect prediction and higher values meaning worse. A model that predicts 90 percent for an event that loses scores 0.81 for that bet, a heavy penalty, while a cautious 55 percent call on the same loser scores only 0.30. Confidence is rewarded only when it is right.

Consider two models over a hundred matches. Model A picks winners 58 percent of the time but states every pick at 90 percent confidence. Model B picks winners 55 percent of the time but states honest probabilities around 55 to 65 percent. Win rate crowns Model A, yet Model A is dangerous: its staking, if based on its stated edge, will be far too aggressive. The Brier score exposes this. Model A's confident errors inflate its score, while Model B's modest, honest probabilities keep its score low. For anyone sizing bets from model probabilities, that distinction is worth real money.

The Brier score also lets you compare models across different sports and markets on a common scale, something win rate cannot do, since a 55 percent win rate means something very different on heavy favourites than on long shots. It even gives you a baseline: predicting the base rate every time, for example always saying the home team wins 46 percent of the time if that is the league average, produces a known reference score. Any model that cannot beat that baseline is adding nothing, however impressive its marketing sounds.

A Brier score on its own is hard to interpret without context, so always report it alongside the baseline and the sample size. A score of 0.20 on a balanced two-outcome market is roughly what you would expect from honest but modest predictions, while a score near 0.25 suggests the model is barely better than guessing the base rate. Multi-outcome markets, such as home win, draw, and away win, need a multi-class version where the squared errors across all three outcomes are summed for each match. The principle stays the same: lower is better, and confident errors cost more.

Watch out for small samples. Twenty predictions can produce a flattering or damning Brier score purely by luck, especially if a few long shots land or miss. Treat anything under a hundred predictions as indicative at best, and split results by league or market to check whether one area is carrying the whole score. A model with a good overall Brier score but a terrible score on away underdogs may be systematically mispricing exactly the bets where value usually hides.

When comparing two models, compute the Brier score on the same set of matches, at the same odds snapshot, and prefer the lower score, provided the sample is large enough to trust. Combine it with calibration checks and closing line value for a fuller picture: Brier tells you whether probabilities are accurate, calibration tells you where they bend, and closing line value tells you whether the market agrees. No single metric settles an argument, but a model that wins on Brier, stays calibrated, and beats the closing line has earned a serious look. Anything less is a claim waiting for evidence.

Suppose you track ten bets. Your model gives probabilities of 70 percent to five winners and 60 percent to five losers that actually lost. The Brier score for the winners is the average of the squared gap between 0.70 and 1, which is 0.09 each. For the losers, the gap between 0.60 and 0 is 0.60, squared to 0.36 each. Averaged across all ten, the score is about 0.225. Now imagine a rival model that called the same outcomes at 55 percent and 45 percent respectively. Its score would be far lower, around 0.20, despite looking less confident. The cautious model wins on Brier, and it should, because its probabilities were closer to reality.

Run this calculation on your own picks in a spreadsheet before you trust any new model. It takes minutes, it needs no special software, and it immediately separates honest probability forecasts from confident guesswork. Pair the score with a simple calibration table and you have, in two views, most of what you need to judge a model. Models that refuse to be scored this way are telling you something too.