A full-stack prediction engine — from raw data to a polished interactive UI — that forecasts the winner, the scoreline and the goalscorers for every match of the 2026 FIFA World Cup.
┌─────────────────────────────────────────────────┐
│ Data Layer │
│ results.csv · FIFA rankings · FC26 ratings │
│ 2026 squads · 2026 fixtures & venues │
└────────────────────┬────────────────────────────┘
│
┌────────────────────▼────────────────────────────┐
│ Feature Pipeline (140+ features) │
│ Elo · FIFA rank · form · H2H · style · squad │
│ situational · intel · club form │
│ ↓ all point-in-time — zero leakage ↓ │
└────────────────────┬────────────────────────────┘
│
┌────────────────────────▼────────────────────────────┐
│ Five Sub-Models │
│ Form · Style · H2H · Squad · Neural (Keras) │
└────────────────────────┬────────────────────────────┘
│
┌────────────────────────▼────────────────────────────┐
│ Elo-Anchored XGBoost Stacking Ensemble │
│ + Dixon-Coles Scoreline Model │
│ + Real-Squad Goalscorer Model │
└────────────────────────┬────────────────────────────┘
│
┌────────────────────────▼────────────────────────────┐
│ Frontend (React + Vite) │
│ Bracket · Awards · About · Run-a-Prediction │
│ → deployed to GitHub Pages │
└─────────────────────────────────────────────────────┘
For any fixture — say Brazil vs Morocco — the system produces:
| Output | Description |
|---|---|
| 🏆 Winner | Home win / Draw / Away win probabilities with honest confidence scores |
| ⚽ Scoreline | The most-likely scoreline (+ alternatives) from a Dixon-Coles Poisson model |
| 👟 Goalscorers | Top likely goalscorers per side, drawn from each nation's real 2026 squad |
| 🎲 Chaos coefficient | A transparency metric — low-confidence matches are labelled honestly |
Every prediction ships with a calibrated confidence score and a chaos coefficient — the system deliberately widens uncertainty on volatile matchups rather than faking certainty.
| Source | What it provides |
|---|---|
results.csv |
~49 000 international results, 1872 → 2024 (martj42) |
goalscorers.csv |
Historical match-level goalscorer records |
fifa_ranking-2026-04-01.csv |
Monthly FIFA rankings, 1993 → 2026 |
FC26_20250921.csv |
EA FC 26 player ratings (finishing, overall, pace, etc.) |
wc2026_squads.csv |
Official 2026 squad lists for all 48 nations |
fixture_venue/ |
Official 2026 schedule: 105 matches, 48 teams, 16 venues |
The pipeline (features/build_features.py) walks every historical match once in chronological order,
building per-team and per-pair features incrementally — so every feature is strictly point-in-time.
| Family | Key signals |
|---|---|
rating_* |
Incremental Elo (K=30, +65 home advantage), point-in-time FIFA rank & points |
form_* |
Recent results, win streaks, goals for/against over sliding windows |
h2h_* |
Head-to-head win rate, goal difference, historical dominance |
style_* |
Tactical proxies derived from scoring patterns (attack vs defence balance) |
squad_* |
Squad quality from FC26 attributes, weighted by position |
sit_* |
Situational context: tournament stage, home/neutral, rest days, travel |
intel_* |
Chaos coefficient, competitive-balance signals |
Each sub-model predicts a W/D/L probability vector:
- Form model — Gradient boosting over recent-form features
- Style model — Tactical-proxy features
- H2H model — Head-to-head history
- Squad model — Random Forest over squad-quality features
- Neural model — Keras feed-forward net over all feature columns
These feed into an Elo-anchored XGBoost meta-learner trained on
time-aware out-of-fold predictions (TimeSeriesSplit), with raw Elo/rank
priors injected directly into the stacker to preserve the strongest signal.
A Maher/Dixon-Coles independent-Poisson model with low-score (ρ) correction.
Attack/defence/home-advantage strengths estimated via recency-weighted
(540-day half-life) fixed-point iteration. Returns expected goals, the full
scoreline matrix, and the most-likely scorelines.
Ranks each nation's players by: position weight × finishing, where finishing
blends FC26 attributes with real goals-per-appearance (last 4 years) and
goals-per-cap. Pulls from official squad CSV → live API fallback → FC26 last resort.
Validation trains on matches before 2022-01-01 and tests on 147 tournament matches (World Cup 2022 + Euro 2024 + Copa América 2024).
| Model | 3-way accuracy | Log-loss |
|---|---|---|
| Raw Elo favourite (baseline) | ~53.7% | — |
| Elo-anchored stacking ensemble | ~51.7% | ~0.995 |
| Dixon-Coles score model | ~48.3% | — |
Why these numbers are good: 3-way international football has a practical accuracy ceiling of ~50–54%. Roughly a quarter of matches are draws, outcomes are low-scoring, and even bookmakers land in a similar band. The ensemble trades a small amount of raw accuracy for better-calibrated, honest probabilities — it keeps draw mass and widens uncertainty on chaotic matches rather than faking confidence.
The frontend is a zero-backend, client-rendered React 18 + TypeScript app built with Vite.
It loads a single pre-computed data.json (exported from the Python pipeline) and renders it
as an interactive, colour-coded experience.
| Panel | What you see |
|---|---|
| 🟣 Bracket | Full knockout tree, R32 → Final, with the predicted champion |
| 🔴 About | How the model works — features, sub-models, blind validation |
| 🟡 Awards | Simulated tournament awards (Golden Boot, Golden Ball, etc.) |
| 🟢 Run | Pick any two teams and get live W/D/L + scoreline + goalscorers |
Built from the four official World Cup 2026 brand stripes:
| Colour | Hex | Role | |
|---|---|---|---|
| 🟣 | Violet | #5A1AE6 |
Bracket, primary accents |
| 🔴 | Red | #E51A14 |
About panel, lock / warning states |
| 🟡 | Lime | #BFE600 |
Awards, score highlights |
| 🟢 | WC Green | #00A85A |
Run panel, success states |
Typography: Unbounded (display) · Inter (body) · JetBrains Mono (numbers & labels)
| Module | What it does |
|---|---|
live/api_football.py |
Cache-first API-Football (v3 free tier, 100 req/day) — fixtures, lineups, injuries, squads. Never crashes: falls back to stale cache or empty result |
live/squads.py |
Resolves official squads for teams missing from the CSV |
live/odds_api.py |
Pulls bookmaker odds for benchmarking |
live/lineup_adjust.py |
Adjusts predictions when confirmed lineups arrive |
update.py |
Match-day rules: lineup confirmation at T-24h, injuries/weather at T-3h, key-player-out penalty, prediction lock at T-1h |
This repository contains the frontend and the pre-computed prediction results only. The full ML pipeline, training code, feature engineering, and model source code are not included in this public release.
What you'll find here:
src/— React + TypeScript frontend source codepublic/data.json— Pre-computed predictions exported from the ML pipelinedocs/— Screenshots and branding assets
# Install (Node 20+)
npm install
# Dev server with HMR
npm run dev
# Production build → dist/
npm run buildPushing to main triggers a GitHub Actions workflow that builds the frontend
and publishes dist/ to GitHub Pages — no manual steps required.
Interested in the full model, the training pipeline, or a collaboration? Feel free to reach out:
@daudibrahimhasan on GitHub
© 2026 TRIONDA · Built by @daudibrahimhasan