An open-source statistical model that forecasts 2026 FIFA World Cup matches and title odds — Elo ratings → Dixon-Coles bivariate Poisson → Monte Carlo simulation. No machine-learning black box, no scraped bookmaker odds: just transparent, reproducible football maths.

▶ Live predictions (full 48-team, 50,000-simulation model): https://cup26matches.com · How it works / methodology · Live insight feed · Interactive bracket simulator

🔴 The tournament is LIVE (Jun 11 – Jul 19). The production model now conditions on real results: finished matches are locked, eliminated teams collapse to 0%, the actual bracket (incl. the new best-third qualification, solved with bipartite matching) is used, and only the remaining matches are simulated — re-run automatically within minutes of every full-time whistle.

This repo open-sources the core match model + our honest backtest so you can run, inspect and reproduce the numbers.

It's tested the honest way — walk-forward, out-of-sample on 913 real internationals (Oct 2023 – Jun 2026). Every match is predicted using only data available before kickoff, then scored against the actual result — with proper scoring rules (RPS, log-loss, Brier), not just accuracy, because accuracy alone rewards lucky guessing. Reproduce it yourself in one command:

node backtest.mjs

A forecaster is honest when the things it calls "70%" happen about 70% of the time. Pooling every probability the model issued across the out-of-sample matches:

Changelog — Jun 11, 2026: Monte Carlo raised to 50,000 trials (5× lower tail noise); in-tournament conditioning is live; backtest extended with RPS + a reliability curve + ECE; data refreshed through Jun 2026. · Jun 7: goal-model variance denominator 350→400; per-team strength priors applied on the live site on top of this core model.

No model is a crystal ball — football is high-variance and draws are genuinely hard. These are well-calibrated estimates, and we make no claim to beat the betting market.

The model's call on every finished match of the tournament, updated as it happens:

20/32 correct picks (63%) · avg RPS 0.152 (coin-flip ≈ 0.245) · updated 2026-06-20

Every call is listed — hits and misses. Probabilities are the model's frozen pre-match numbers (ratings don't re-fit mid-tournament), so nothing here is retro-fitted. Reproduce with node track-record.mjs.

Run a blog, forum or fan site? The live model is embeddable — free, auto-updating all tournament:

<!-- Live title-race board (top-10 championship odds, 50k sims) -->
<iframe src="https://cup26matches.com/embed/title-race/" width="100%" height="430"
  style="border:0;border-radius:12px" loading="lazy" title="World Cup 2026 title odds"></iframe>

<!-- Real-time next-match strip (live W/D/L, rotates at kickoff) -->
<iframe src="https://cup26matches.com/embed/next-match/" width="100%" height="92"
  style="border:0;border-radius:10px" loading="lazy" title="Next World Cup 2026 match"></iframe>

More widgets + copy-paste snippets: cup26matches.com/en/widgets

Open data (CC BY 4.0 — free to use/quote/chart with a link back): the full per-team tournament probabilities, regenerated after every match — probabilities.json · probabilities.csv

No dependencies. Node 18+.

git clone https://github.com/Hicruben/world-cup-2026-prediction-model.git
cd world-cup-2026-prediction-model

node predict.mjs brazil argentina      # head-to-head probabilities
node predict.mjs usa mexico usa        # 3rd arg = home team (host bonus)
node backtest.mjs                      # reproduce the accuracy numbers
node calibrate.mjs                     # rebuild ratings from data/results.json

Example:

$ node predict.mjs spain germany

  spain (Elo 2074)  vs  germany (Elo 1927)   [neutral]

  spain            win   53.2%  ████████████████
  draw                   26.8%  ████████
  germany          win   20.0%  ██████

1. Team strength (Elo). Each nation starts from a long-run prior, then is calibrated on recent real internationals — wins over strong sides in important games move a rating more than friendlies, and recent form outweighs old form. See calibrate.mjs.
2. Each match (Dixon-Coles Poisson). Ratings → expected goals → a Dixon-Coles bivariate Poisson gives win/draw/loss probabilities. The Dixon-Coles correction fixes plain Poisson's well-known under-count of low-scoring draws (0-0, 1-1). See elo.mjs.
3. The tournament (Monte Carlo). The live site plays all 104 matches 50,000 times through the real bracket to get championship & advancement odds — and, now the tournament is underway, locks every finished result (real standings, real qualifiers, real bracket slots) and simulates only what's left. Full write-up: cup26matches.com/methodology.

MIT — see LICENSE. Built by Cup26 AI. If you use it, a link back is appreciated. ⭐ the repo if you find it useful!