What is the Brier score?
For each binary forecast, score = (p − y)², where p is a probability from 0 to 1 and y is 0 or 1. Average those scores across the sample. This follows the binary convention in the official scoringrules documentation. Some multiclass conventions sum both outcomes and double the binary score; compare like with like.
Read the example correctly
The invented rows 70,1; 20,0; 90,0; 40,1 have scores 0.09, 0.04, 0.81 and 0.36. Their mean is 0.325. A constant 50% forecast scores 0.25 for every binary outcome. The example’s skill relative to that baseline is 1 − 0.325 ÷ 0.25 = −30%. This does not measure any platform’s historical results.
Choose a defensible sample
Record one probability per outcome at a predefined forecast horizon. Preserve market identifiers, outcome side, observation time and final resolution separately. Repeated observations from one event are not independent events. Report exclusions, category coverage and sample size before making performance claims.
Score quality is different from trading profit
A forecast can score well without producing profitable trades after executable prices, fees and liquidity. Read how to evaluate Polymarket accuracy and use the expected-value calculator to separate forecast probability from trade cost.