PM Atlas PM Atlas home

Concept · information-theory

Brier Score

A proper scoring rule that measures probabilistic forecast accuracy as the mean squared difference between predicted probabilities and binary outcomes (0 or 1). Lower scores indicate better calibration; the universal metric for benchmarking forecasters, models, and markets. Range 0 (perfect) to 1 (maximally wrong on every prediction).

Key insights

In their words

Polymarket's headline Brier score of 0.047 masks category-specific failures like sports markets scoring 0.325 (worse than a coin flip).· Mandloi
GPT-4.5 achieves a Brier score of 0.101 versus superforecasters' 0.081, with LLMs improving roughly 0.016 Brier points per year.· Forecasting Research Institute
Some models game the benchmark by copying prediction market prices rather than reasoning independently.· Forecasting Research Institute

Where it matters

Brier score is the metric every PM "we're accurate" headline reduces to · but it's an aggregate that can hide huge category-level failures. The only honest reporting is decomposed Brier: by category, by horizon, by trade size, by liquidity decile. The LLMs-game-the-benchmark finding is doubly relevant: (1) PMs themselves can be gamed by traders copying other PMs (creating false consensus), and (2) Brier benchmarks used to compare PMs to LLMs/humans are contaminated when one input copies another. For builders, the practical move is to publish per-category Brier curves, and to flag any category where Brier exceeds the random-baseline of 0.25 · that's a market that's actively misinforming.

Connections

Platforms linked to this concept

Related concepts

Sources

Open in the interactive atlas →