PropBetEdgeLearnSports Intelligence Academy
PropBetEdge
PropBetEdgeLearnSports Intelligence Academy

Sport academies

Field & court

Individual

Racket

Combat

Network

PropBetEdge.ai PBE All Access
Lessons, glossary and tools from Learn · players, teams, tools and stories from PropBetEdge Search↑↓ to move · Enter to open

Model Literacy · Lesson 5 of 9

Brier Score and Log Loss

How do you grade a probability, not a pick?

Intermediate10 min readUFC / MMAWNBA

Brier score

Brier = average of (forecast − outcome)²      // outcome is 1 or 0
lower is better; 0 is perfect
always forecasting 50% on binary outcomes → 0.25
Worked example · three forecasts
  1. 70% and it happened: (0.7 − 1)² = 0.09.
  2. 20% and it did not: (0.2 − 0)² = 0.04.
  3. 90% and it did not: (0.9 − 0)² = 0.81.
Brier = (0.09 + 0.04 + 0.81) / 3 = 0.313 — one confident miss dominates.

Log loss

log loss = −average of [ y·ln(p) + (1 − y)·ln(1 − p) ]
coin flip (50%) → ln 2 = 0.6931
ForecastLog loss if it happenedLog loss if it did not
5%2.9960.051
30%1.2040.357
50%0.6930.693
70%0.3571.204
95%0.0512.996

Log loss punishes confident misses much harder than Brier. A model that says 95% and is wrong pays about 3.0 on that forecast.

Why "proper" scoring rules lead

Both are proper scoring rules: a forecaster gets the best expected score by reporting their true belief. Win rate is not proper — it rewards pushing every probability toward the favourite. That is why PropBetEdge model cards lead with Brier and log loss rather than a hit rate.

Reading a real model card

Example: PBE Fight Model V1 (UFC), as published on its model card

Out of sample over 7,047 bouts (2013–2026): Brier 0.2363 against a coin-flip 0.2500, log loss 0.6651 against 0.6931, accuracy 59.58%. A historical win-rate baseline scores 0.2447 / 0.6826.

Read it as the card does: a modest, honest improvement over the baselines across thousands of fights — not a machine that is usually right about individual fights.

Example: WNBA model v1, holdout

On the 2025–2026 holdout the WNBA model card reports a Brier of about 0.212 and log loss of about 0.611. It also reports that on games with a sportsbook price, the de-vigged market predicted outcomes better than the model. The card says so plainly, and its PBE Edge figure is described as disagreement with the market, not proven value.

Limitations

Scores are only comparable on the same games. A Brier of 0.21 on basketball games and 0.24 on UFC fights says nothing about which model is "better" — the sports have different base rates and different irreducible randomness. Compare a model with its own baselines, and with the market on the same events.

Go deeper: PBE Fight Model card · WNBA model card