How We Predict the World Cup
OddsFlow uses advanced statistical models and AI to generate World Cup 2026 predictions. This page explains every step of our pipeline — from raw data to the probabilities you see on each group page.
Prediction Pipeline
1. The Dixon-Coles Model
At the core of our predictions is the Dixon-Coles model (Dixon & Coles, 1997), a Poisson-based statistical framework designed specifically for football match prediction. Unlike basic Poisson models that assume home and away goals are independent, Dixon-Coles introduces a critical correction.
How it works
- 1.Every team gets two parameters: attack strength (how many goals they tend to score) and defense strength (how many they tend to concede)
- 2.A home advantage factor adjusts for the known benefit of playing at home
- 3.The rho (ρ) parameter — Dixon-Coles' key innovation — corrects for the observed correlation in low-scoring matches. In football, 0-0 and 1-0 results happen more often than a basic Poisson model predicts. The ρ correction fixes this
- 4.A time-decay weighting gives recent matches more influence than older results, so the model adapts to current form
For each match, the model outputs a full probability distribution over all possible scorelines (0-0, 1-0, 0-1, 1-1, 2-0, ... up to 10-10). From this matrix we derive win/draw/loss probabilities, expected goals, most likely score, Over/Under 2.5, and BTTS predictions.
2. Shin De-Vigging — Removing Bookmaker Margin
Bookmaker odds contain a built-in margin (overround) that makes the implied probabilities sum to more than 100%. To extract the true underlying probabilities, we use the Shin method.
Why Shin, not simple normalization?
Simple normalization divides each implied probability by the total overround — treating favourites and longshots equally. But bookmakers don't apply margin uniformly. They pad more margin on longshots(where bettors are less price-sensitive) and less on favourites. Shin's method (1991, 1993) models this asymmetry, producing more accurate true probabilities — especially for underdogs and draws, which are critical in World Cup group stages where upsets decide qualification.
3. Isotonic Calibration — Making Probabilities Honest
A well-calibrated model means that when we say "Team X has a 70% chance of winning," they should actually win roughly 70% of the time across many similar predictions. Raw model outputs are rarely perfectly calibrated, so we apply isotonic regression.
How isotonic calibration works
- 1.Group historical predictions into buckets by predicted probability
- 2.Compare predicted probability vs actual outcome frequency in each bucket
- 3.Fit a monotonically non-decreasing function that maps raw predictions to calibrated probabilities
- 4.Apply this mapping to all future predictions
Isotonic regression is non-parametric — it doesn't assume any specific shape for the calibration curve. This makes it more flexible than Platt scaling (logistic calibration) and better suited to football's complex probability landscape.
4. Monte Carlo Simulation — 10,000 Tournaments
Individual match predictions aren't enough to forecast tournament outcomes. Who wins Group A depends on the results of all 6 group matches. Who reaches the Final depends on the entire bracket. We solve this with Monte Carlo simulation.
The process
- 1.For each of the 10,000 simulations, randomly generate a result for every group match using the Dixon-Coles probability distributions
- 2.Calculate group standings (points, goal difference, goals scored) and determine which 2 teams from each group advance
- 3.Simulate the entire knockout bracket (R32 → R16 → QF → SF → Final) using the same Dixon-Coles match model
- 4.Record every team's furthest stage reached in this simulated tournament
After 10,000 iterations, we count how often each team reached each stage. If France advanced past the group stage in 8,270 out of 10,000 simulations, their R32 probability is 82.7%. This approach naturally captures path dependencies and bracket effects that single-match models miss.
Why 10,000?
With 10,000 simulations, the standard error for a 50% probability event is ±0.5% — precise enough for meaningful comparisons. Increasing to 100,000 would improve precision to ±0.16% but at 10x the computational cost, with negligible practical benefit.
5. Data Sources & Parameters
Our model incorporates multiple data streams to estimate team strength:
Historical Results
International match results including World Cup, continental tournaments, qualifiers, and friendlies — weighted by recency and match importance
FIFA Rankings
Official FIFA ranking points as a prior for team strength, especially for teams with limited head-to-head history
Bookmaker Odds
Pre-match odds from multiple bookmakers, de-vigged using Shin method, as a market-consensus signal
Team Ratings
Attack, midfield, defense, and goalkeeper ratings derived from squad composition and recent performance
6. What You See on Group Pages
Each group prediction page displays:
Group Stage
- • Advance to R32 probability
- • Group winner / runner-up / 3rd / eliminated %
- • Expected points (from 3 matches)
Tournament Path
- • R32 → R16 → QF → SF → Final → Champion
- • Full knockout progression probabilities
- • Based on 10,000 simulated brackets
Head-to-Head Matchups
- • Win / Draw / Loss probabilities
- • Expected goals per team
- • Most likely scoreline
- • Over 2.5 goals & BTTS
Team Analysis
- • OVR / ATK / MID / DEF / GK ratings
- • Key players spotlight
- • Championship probability
7. Limitations & Transparency
No prediction model is perfect. We believe in transparency about what our model can and cannot do:
- ✗Injuries & suspensions — The model uses team-level strength, not individual player availability. A key injury (e.g., a star striker) may not be fully reflected
- ✗Tactical changes — New managers, formation shifts, or strategic surprises are not modeled
- ✗Weather & venue — Climate differences across USA/Mexico/Canada host cities are not factored in
- ✓Probabilities, not certainties — An 80% prediction means the outcome should NOT occur ~20% of the time. Upsets are expected
We track prediction accuracy in real-time on our AI Performance page so you can verify our model's calibration against actual World Cup results.
See the Predictions in Action
Browse all 12 World Cup 2026 groups with AI-powered qualification probabilities and match predictions.
View All Groups →References
- Dixon, M.J. & Coles, S.G. (1997). "Modelling Association Football Scores and Inefficiencies in the Football Betting Market." Journal of the Royal Statistical Society: Series C, 46(2), 265-280.
- Shin, H.S. (1991). "Optimal Betting Odds Against Insider Traders." The Economic Journal, 101(408), 1179-1185.
- Shin, H.S. (1993). "Measuring the Incidence of Insider Trading in a Market for State-Contingent Claims." The Economic Journal, 103(420), 1141-1153.
- Barlow, R.E. et al. (1972). Statistical Inference under Order Restrictions: The Theory and Application of Isotonic Regression. Wiley.