Live since World Cup 2026 · self-updating daily
Calibrated football probabilities,
with the receipts published.
An ensemble of Elo ratings and a Dixon-Coles goals model, scored on every prediction it makes. Its prediction error sits within a fraction of a percent of the bookmakers' closing line — and the same backtests prove it doesn't beat it. That honesty is the product: well-calibrated probabilities you can verify, not picks that promise profit.
World Cup 2026, live
—
matches predicted pre-kickoff
Winner picked
—
Prediction error, live
—
Brier score: 0 = perfect, 0.2222 = pure guessing
Club season log
starts Aug
5 leagues, same live scoring
Does a stated 60% happen 60% of the time?
Reliability curve, 2025-26 walk-forward backtest (… matches) — each match predicted blind from ratings that exclude it.
Against the hardest benchmark there is
Prediction error (Brier score — how far stated probabilities land from what actually happened; lower is better) vs the bookmakers' closing line on the same matches, 2025-26. The market stays slightly sharper, and this chart shows by exactly how much.
The live log — written before kickoff, scored after
live recordWorld Cup 2026, complete: every probability below was logged before the match and scored when the result arrived. A backtest can be rerun until it flatters; this cannot.
| date | match | result | p(H/D/A) stated | picked | error (brier) |
|---|
What this is not
Not a tipster and not a money-printer. The backtests on this page show the closing line is sharper than the model — that is exactly why they are published. Informational only; nothing here is betting advice.
Ask the model
Type a question in plain English — get the model's numbers and its read
My bet log
Track your own bets and real ROI — saved in your browser, never uploaded
Fixtures & probabilities
Upcoming matches with model probabilities and derived markets
Match analyzer
Full ensemble + AI write-up for any fixture