Home

How the AI works

OddsIQ’s tournament forecast is built on 50,000 Monte Carlo simulations of the World Cup 2026, run daily. Here’s how we arrive at each team’s probability.

The ingredients

Elo ratings

Each team’s strength is measured by Elo — a standard chess/football rating that updates after every match. Stronger opponents mean more points for beating them. We start with public Elo ratings and apply confederation adjustments (e.g. South American teams historically overperform in knockouts).

Recent form

We look at each team’s last 10 international matches, weighting recent performances more heavily. A team on a 5-match winning streak gets a small Elo boost; a team in a slump gets a small penalty.

Dixon-Coles Poisson model

For each match, we model the goals each team is likely to score using a Poisson distribution. The Dixon-Coles adjustment corrects for the fact that 0-0, 1-0, and 1-1 scores are more common than pure Poisson would predict — especially in international football.

Bracket draw

Who plays who matters. We simulate each team’s actual path: group opponents, then potential knockout matchups based on final standings. A team in an “easy” group advances more often.

The simulation

For each of the 50,000 simulations, we:

  1. Simulate all 72 group-stage matches using the Dixon-Coles model
  2. Determine group winners and runners-up (plus best third-placed teams)
  3. Seed the Round of 32 based on final standings
  4. Simulate each knockout round through to the final, including extra time and penalties when scores are level

We count how often each team reaches each stage. A team that wins 6,500 out of 50,000 simulations has a 13% chance of winning the World Cup.

What the model doesn’t know

No model is perfect. Ours doesn’t account for:

  • Injuries to key players in the days before a match
  • Weather or altitude specifics (Mexico City vs Miami)
  • Manager changes mid-tournament
  • Referee decisions

We update daily, but a single red card or injury mid-tournament will always arrive faster than our next refresh.

Track record

We publish our full prediction history and calibration — see how accurate our model has been →