No black box. Here's exactly how every probability on this site is built — and where we're honest about its limits.
We fit a Dixon-Coles model on the full history of men's international matches, weighting recent games more heavily (a ~3-year half-life) and down-weighting friendlies hard — B-squads and dead rubbers don't tell us much. Each team gets an attack and defense rating, plus a home-advantage term that we switch off for neutral World Cup venues. A separate Elo rating is fit as a cross-check. We then blend in two signals results-only models miss: recent on-pitch expected goals (each team's xG for & against from Euro 2024, Copa América 2024 and World Cup 2022 — actual chance quality, not just scorelines) and squad club strength (the ClubElo quality of each 26-man squad). Both matter most going into a tournament of fresh rosters, where the bare international record lags reality.
From those ratings we build a full scoreline matrix for any matchup, with the Dixon-Coles correction that fixes how plain models under-count low scores and draws. Every market falls out of that one object: 1X2, Over/Under, Both Teams To Score, Asian handicap, correct score. They're internally consistent by construction.
We simulate the entire 2026 World Cup 10,000 times — all 12 groups, the 8 best third-placed teams, and the knockout bracket — to produce advance, win-group, reach-final and title probabilities. We add a touch of rating uncertainty to each run because the new 48-team format is genuinely high-variance.
A model probability is only useful against the market. We take the sharpest book's odds, strip out the bookmaker's margin (the "vig") to get a fair price, and flag a value bet only where our probability beats that fair price. Stakes are sized with quarter-Kelly to survive variance and model error. We grade ourselves on closing-line value — beating the final price is the truest sign of an edge.
Validated walk-forward on 2,409 international matches the model never saw during fitting — we always train only on earlier games, then predict forward, so there's no hindsight.
Baseline = always predicting the historical home/draw/away frequencies. Beating it by 28.1% on RPS and 19.3% on log-loss is the model earning its keep. We publish our live track record — every pick and its closing-line value — once matches kick off.