The Brier Score, Explained

The Brier score is the standard way to grade a probability forecast: it takes the probability you assigned to an outcome, compares it with what actually happened, squares the difference, and averages that across every forecast you made. Lower is better, and it rewards being both accurate and honest about your confidence.

Saying an event is “likely” is not something you can grade. Saying it is “70% likely” is. Once a forecast is a number, you can check it against reality and put a score on it. The Brier score is the most widely used way to do exactly that, and it is what turns a pile of predictions into a track record.

Where the score comes from

The score was introduced in 1950 by the meteorologist Glenn W. Brier, in a short paper on how to grade weather forecasts that are stated as probabilities. His problem was practical: a forecaster who says “70% chance of rain” should be judged not on a single day but on how well those numbers hold up over a long run of days. You can read the original Verification of Forecasts Expressed in Terms of Probability from Monthly Weather Review. The same idea now scores forecasters far beyond the weather, from elections to sports to markets.

The formula in plain words

For a single yes-or-no forecast, code the outcome as 1 if the event happened and 0 if it did not. Take the probability you gave (a number between 0 and 1), subtract the outcome, and square the result. That squared gap is the score for that one forecast.

Your overall Brier score is just the average of those squared gaps across all your forecasts. That is the whole thing: the mean squared difference between the probability you gave and the outcome coded 1 or 0. Nothing more exotic than an average of squares.

If you want the exact notation and the standard variants for questions with more than two outcomes, the Brier score overview lays them out. For everyday forecasting the two-outcome version above is all you need.

A worked example

Suppose you make two forecasts. First, you say there is a 90% chance your team wins tonight, and they win. Second, you say there is a 30% chance it rains tomorrow, and it does rain.

1. Forecast one: you gave 0.9, the outcome was 1 (they won). The gap is 0.9 minus 1, which is minus 0.1. Squared, that is 0.01. A small penalty, because you were confident and right.
2. Forecast two: you gave 0.3, the outcome was 1 (it rained). The gap is 0.3 minus 1, which is minus 0.7. Squared, that is 0.49. A large penalty, because you leaned the wrong way.
3. Average the two: (0.01 plus 0.49) divided by 2 equals 0.25. That is your Brier score across these two forecasts.

Notice how one confident miss dominates. The correct, confident call earned almost nothing in penalty, while the single wrong lean cost nearly half a point on its own. That asymmetry is the heart of how the score behaves.

Reading the 0 to 1 range

Because it is an average of squared gaps, the Brier score always sits between 0 and 1. Both ends are easy to picture.

• 0 is perfect: every time you said 100% it happened, and every time you said 0% it did not. You were certain and always right.
• 1 is the worst possible: every time you said 100% it did not happen, and every time you said 0% it did. Confidently wrong on every call.
• 0.25 is the “no information” baseline: it is what you get by always saying 50%. Beating 0.25 means you are adding real signal; scoring worse than it means your confidence is actively hurting you.

So the number to beat is not 0, which is unreachable for genuinely uncertain events. The number to beat is 0.25. A skilled forecaster on hard questions might land somewhere in the low teens as a decimal, say 0.15, and that is already very good.

Why it punishes overconfidence

The squaring is doing the important work. A gap of 0.2 costs 0.04, but a gap of 0.4 costs 0.16, four times as much for only twice the error. Big misses are penalized far more than proportionally, so a single overconfident blunder can wreck an otherwise good score.

That is deliberate. It pushes you toward humility on questions you cannot really call. Saying 95% when you should have said 70% barely helps when you are right and hurts a lot when you are wrong. Over many forecasts, the person who states honest, moderate probabilities beats the person who swaggers to the extremes and occasionally gets burned.

A proper scoring rule

The Brier score is a “proper” scoring rule, and that word carries a strong guarantee. A scoring rule is proper when your expected score is best exactly when you report your true probability. There is no clever number you can state instead of your real belief that improves your score in the long run.

This matters because it makes honesty the optimal strategy, not just the virtuous one. If you privately think an event is 60% likely, then reporting 60% minimizes your expected Brier score. Shading it to 80% to look bold, or to 50% to look safe, both make your expected score worse. The rule is built so that the only way to win is to tell the truth about your uncertainty.

Calibration and resolution

The Brier score can be split into two components that measure different skills. Calibration asks whether your probabilities mean what they say: of all the times you said 70%, did about 70% actually happen? Resolution asks whether you separate events at all: do you push toward 90% and 10% when you can, rather than hedging near 50% on everything?

A good forecaster needs both. Perfect calibration with no resolution is just always guessing the base rate, which is safe but uninformative. Strong resolution with poor calibration is bold but unreliable. The Brier score rewards the combination: confident, well-separated forecasts that also turn out to be true at the rate you claimed.

Calibration is worth understanding on its own, because it is the most intuitive half of the score. See forecast calibration explained for how to read a calibration curve and spot your own biases.

How platforms use it

Forecasting platforms and prediction markets lean on Brier-style scoring to rank people fairly. Because it is proper, they can trust that a leaderboard built on it rewards genuine skill rather than reckless swings. Because it accumulates over many forecasts, one lucky call cannot fake a good record, and one unlucky one cannot sink a real forecaster.

This is also how research identifies the best forecasters in the world. The people who consistently post low Brier scores across hundreds of questions are the ones studies call superforecasters, and their edge shows up as a scoring gap, not a single dramatic prediction.

The practical lesson is simple: if you want to get better, keep score. Track your probabilities, resolve them against reality, and watch the number move. Nothing sharpens judgment faster than being graded honestly, again and again.

Clutch does the bookkeeping for you. Predict real news and sports with in-app credits, and the app keeps a running score of how well calibrated you are, so you can watch it improve. Get the app and start building a track record.