Predictions
World Cup 2026 Predictions: How Our Model Picks Winners
Most 'expert picks' you'll see this month are guesses with a tie. Here's the model that drives every prediction on Football of Nations, what it's good at, and where it's still wrong.
Why publish the methodology
The fastest way for a sports site to lose your trust is to predict a team will win, refuse to say why, and then change its mind quietly when the team loses. Every prediction on Football of Nations comes from a single model whose inputs and weights are public. If you disagree with a pick, you can also disagree with the model and tell us where.
What the model uses
Three layers, run in this order:
- Team strength — a rolling Elo seeded from the last two World Cup cycles plus 18 months of competitive matches (qualifiers, Nations League, continental finals). Friendlies are excluded; they’re noise.
- xG-based form — non-penalty xG for and against, last 10 matches, weighted recent-most. Caps adjustments at ±0.25 standard deviations to stop a single rout from dominating.
- Tournament context — rest days, travel distance, opponent’s confederation diversity, home-country advantage where it applies. Each adjustment is documented in the model’s GitHub repo.
The output is a per-match win/draw/loss distribution and a tournament- level simulation (50,000 paths) that produces the headline numbers you see on the predictions hub: title odds, group standings, knockout bracket probabilities.
What it doesn’t use
- Pre-tournament friendlies. Too few, too noisy, often missing key players.
- Player-level data beyond starting XI minutes. Without club-and-country load tracking the signal degrades fast.
- Betting-market odds. We treat them as a benchmark to check the model, not as an input. The whole point is to be different from the market.
What the 2022 backtest said
Re-running the same pipeline on Qatar 2022 (with that cycle’s data only) the model’s top-3 expected winners included two of the semi-finalists. It missed Morocco’s run entirely — that’s the kind of result it will keep missing, because the model has no input for “emotional team chemistry” or “tournament momentum”. We’re OK with that: it stays calibrated, and the misses are honest.
Where you’ll see the model output
- Title race tracker — updated weekly on the predictions hub.
- Per-match preview — embedded on each match page in the schedule once the tournament starts.
- Group standings prediction — alongside the live standings, so you can see where the model agrees and disagrees with reality.
What’s next
Pre-tournament we’ll publish:
- Group-by-group predictions (one piece per group).
- The golden boot model: stats-based top 10 contenders.
- A piece on dark horses — teams the model likes that the betting market doesn’t.
If you spot a bug or a missing factor, mail it to me.