How our World Cup predictions work
Cup26 AI is a transparent statistical model — not a black box and not copied from the bookmakers. Here is exactly how the numbers are made, and how well they have done on real matches.
The model, step by step
1 · Team strength (Elo)
Every nation gets an Elo rating that starts from long-run strength and is then calibrated on hundreds of recent real internationals. Wins against strong opponents in important matches move a rating more than friendlies, and recent form counts more than old form.
2 · Each match (Dixon-Coles Poisson)
Two teams' ratings become expected goals, which feed a Dixon-Coles bivariate Poisson model to produce win / draw / loss probabilities. The Dixon-Coles correction fixes a known flaw of plain Poisson: it under-counts low-scoring draws (0-0, 1-1) that are common in football.
3 · The whole tournament (10,000 simulations)
We play out all 104 matches through the real bracket 10,000 times with Monte Carlo simulation. Counting how often each team reaches each round gives the championship and advancement probabilities you see across the site.
4 · Live updates
Once the tournament starts, real results are locked in: eliminated teams drop to 0%, qualified teams' paths are recomputed, and only the remaining matches are simulated. The model re-runs every few hours so the odds stay current.
How accurate is it?
We tested the model the honest way — "walk-forward". Going through 920 real internationals (2023-10-12 → 2026-05-28) in time order, each match was predicted using only the data available before it was played, then the result was revealed and the ratings updated. These are out-of-sample results on 770 matches — no peeking, no curve-fitting.
For context on the same matches: blindly picking the home side is right 49% of the time, and picking the higher-rated team 60%. A pure coin-flip scores a Brier of 0.67; our 0.54 means the probabilities carry real information.
Football is high-variance and draws are genuinely hard to call — no model gets these right every time, and we make no claim to beat the betting market. Treat the percentages as well-calibrated estimates, not certainties. Upsets are part of the game (and part of the fun).
What we do — and don't — do
- ✓It is a statistical model, not a chatbot. The maths above (Elo + Poisson + Monte Carlo) is the same family used by academics and professional modellers — it is fully deterministic and reproducible, not a language model guessing.
- ✓It is independent of the bookmakers. We do not copy or blend in betting odds; the numbers come only from match data. That is why our picks sometimes differ from the market.
- ✓It is information, not financial advice. Predictions are for analysis and entertainment. If you bet, that is your decision — 18+, and please play responsibly.
See it in action
Browse the live championship probabilities, or build your own bracket in the simulator.