Brier Score & Log Loss Calculator

Enter what you predicted and what happened. This scores the set two ways and compares both with the "know nothing" baseline.

Why the baseline is the whole point#

A scoring number on its own says nothing. If 80% of your fixtures ended in a home win, predicting "80% home" every single time scores well without containing a single piece of insight. The skill score is the honest reading: it asks whether your probabilities beat that lazy baseline.

Anything at or below zero means they did not — no matter how good the raw Brier looked.

Two metrics that disagree on purpose#

Both Brier and log loss are proper scoring rules: each is minimised by stating what you actually believe, so neither rewards hedging toward 50%. Where they part company is confidence.

Brier is quadratic and forgiving — being certain and wrong costs a bounded amount. Log loss is logarithmic and brutal about it, which is what you want from a model whose probabilities feed a staking decision, because a confidently wrong 95% is exactly the forecast that empties a bankroll.

Try it: set a forecast to 0.99 and mark it as not happening. Brier barely moves. Log loss jumps.

Scoring is not the same as profit#

A well-calibrated model can still lose money, and a badly calibrated one can win for a while. Scoring tells you whether the probabilities are honest; closing line value tells you whether they were better than the market's. You need both, and you need enough of a sample for either to mean anything.

This is the same arithmetic behind Tofiko's own Model Accuracy view — read how our models work for what goes into them, and calibration for what a score like this can and cannot detect.