Home / Wiki / B / Brier Score

Brier Score

Trading and SpeculationUpdated 16 Nov 2026

The Brier score is a measure of how accurate a set of probability forecasts has been. Each forecast states the chance of an event, such as a 70 percent chance of rain, and each outcome is recorded as 1 if the event happened and 0 if it did not. The score is the average of the squared differences between forecast and outcome. Lower is better: a perfect forecaster scores 0, and one who is always certain and always wrong scores 1.

The measure was proposed by the American meteorologist Glenn W. Brier in 1950 for the verification of weather forecasts and is still a standard tool in meteorology. It has since spread to any field in which predictions are stated as probabilities, including medicine, credit risk, machine learning, forecasting tournaments and the assessment of prices in a prediction market.

01

Formula and example

For events with two possible outcomes, the score is:

BS=1N∑i=1N(fi−oi)2BS = \frac{1}{N} \sum_{i=1}^{N} (f_i - o_i)^2
  • NN is the number of forecasts.
  • fif_i is the probability forecast for event ii, between 0 and 1.
  • oio_i is the outcome, 1 if the event occurred and 0 if it did not.

Suppose a forecaster gives probabilities of 0.9, 0.7 and 0.2 to three events, and only the first occurs. The squared errors are (0.9−1)2=0.01(0.9 - 1)^2 = 0.01, (0.7−0)2=0.49(0.7 - 0)^2 = 0.49 and (0.2−0)2=0.04(0.2 - 0)^2 = 0.04. Their sum is 0.54, and the Brier score is 0.54/3=0.180.54 / 3 = 0.18.

A useful reference point is the forecaster who says 50 percent every time. Each squared error is then 0.25, so the score is 0.25 whatever happens. A score above 0.25 on two-outcome questions is worse than offering no opinion. Brier's original version summed the squared errors over both outcomes and so ranged from 0 to 2. The single-outcome form above is now the more common, and the two should not be mixed when comparing published scores.

02

A proper scoring rule

The Brier score is a proper scoring rule, meaning that a forecaster obtains the best expected score by reporting what he or she truly believes. Shading a probability toward 0 or 1 to appear confident, or toward 50 percent to appear cautious, makes the expected score worse. This property is the reason the score is used to reward forecasters. Because errors are squared, a confident miss is punished heavily: a forecast of 0.95 that fails costs about 0.90, against 0.36 for a forecast of 0.6 that fails. The penalty is bounded, unlike that of the logarithmic score, which becomes infinite when an event given zero probability occurs.

03

Calibration and resolution

The score can be broken down into three parts, a result due to Allan Murphy in 1973:

  • Reliability, or calibration. Whether events forecast at 70 percent happen about 70 percent of the time. Smaller is better.
  • Resolution. Whether the forecaster separates likely from unlikely events, instead of always quoting the long-run average. Larger is better.
  • Uncertainty. The inherent unpredictability of the events, which is highest when they occur half the time. It is outside the forecaster's control.

Because of the uncertainty term, raw scores on different sets of questions are not comparable. Forecasting rare events gives a low score with little effort, since predicting that they will not happen is nearly always right. The Brier skill score corrects for this by comparing the score with that of a reference forecast, usually the historical frequency:

BSS=1−BSBSrefBSS = 1 - \frac{BS}{BS_{\text{ref}}}

A skill score of 1 is perfect, 0 means no better than the reference and a negative value means worse. With a score of 0.18 against a reference of 0.25, the skill score is 1−0.18/0.25=0.281 - 0.18 / 0.25 = 0.28.

04

Use in finance and prediction markets

The price of a yes-or-no contract, such as a binary option or an event contract, can be read as the market's probability of the event. Researchers and platform operators calculate Brier scores on settled contracts to test how well prices predicted outcomes, and to compare a market with polls, expert panels or statistical models. The same test is applied to a decentralized prediction market and to individual traders, whose records can be ranked by score. Forecasting tournaments, including those studied by the psychologist Philip Tetlock, use it to identify consistently accurate participants.

In lending, the score is one of the checks applied to a credit scoring model that outputs a probability of default, alongside measures of how well the model ranks borrowers. In machine learning it is a common loss measure for classifiers, used in fraud detection and trading models. Weather and climate forecasts verified by Brier scores feed into the pricing of a weather derivative and of insurance against natural hazards, and analysts sometimes apply it to probabilistic forecasts of recessions or rate decisions, where a consensus estimate serves as the benchmark.

05

Limitations

The score needs many forecasts to be meaningful. A handful of questions says little, since luck dominates. It treats all questions as equally important and takes no account of how hard each was or of the money at stake. It rewards being close to the outcome without regard to whether a forecast was useful for a decision. And for very rare events, where the difference between a probability of 1 in 100 and 1 in 10,000 matters greatly, the squared error barely distinguishes the two. Analysts therefore report it together with calibration charts and other scores.

Last updated 16 Nov 2026