Colincot/wc-fever-2026

WC Fever 2026

Built this during the actual World Cup 2026 because I wanted something more than a scoreboard — a dashboard that actually models what's happening.


[image]

What it does

The match predictor takes two teams and a stage, runs them through an XGBoost classifier trained on WC 2006–2022 results, and outputs win/draw/loss probabilities. Gets about 52% accuracy on the test split — way above the 33% random baseline. During a code review pass I caught a dead _xg_model reference that was silently falling back to a placeholder, fixed that to wire up the real StatsBomb pipeline.

The xG shot map pulls ~15k shots from StatsBomb open data (World Cup 2018 and earlier) and renders them on an SVG pitch coloured by expected goals value. You can filter by team and see a per-shot log alongside the pitch — distance, angle, pressure, and model xG for each attempt. Rebuilt this module after catching the dead model bug above.

The group stage tracker fetches live standings from football-data.org and runs 10,000 Monte Carlo simulations per group to estimate each team's advancement probability. I initially had a naive round-robin loop that generated schedules incorrectly — switched to a proper Berger tournament schedule algorithm and the simulated standings matched the real historical results much more closely.

Stack

Run it locally

# Backend
cd backend
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env        # add your football-data.org key (optional)
python app.py               # Flask on :5001

# Frontend (separate terminal)
cd frontend
npm install
npm run dev                 # Vite on http://localhost:5173

First call to /api/predict trains and caches the model (~5 sec). After that it loads from models/outcome_predictor.pkl.

How the models work

Match Predictor XGBoost classifier (objective="multi:softprob", 200 trees, depth 4). Features: FIFA ranking diff, goals scored/conceded in last 5, head-to-head win rate, tournament stage, and host-nation flag. Trained on bundled WC 2006–2022 match results. Test accuracy 52% vs 33% random baseline.

xG Model Logistic regression on StatsBomb shot data. Features: shot distance from goal, angle to goal centre, and whether the shooter was under pressure. Falls back to a simple analytic formula (distance * angle / constant) if the StatsBomb library isn't installed — same API surface either way.

Monte Carlo group tracker Generates all round-robin fixtures using a Berger tournament schedule (not a naive nested loop — that was the original bug). Simulates each match using the outcome predictor's probabilities, runs 10,000 iterations, and caches advancement probabilities per group so repeated calls don't re-run the full simulation.

API

Data sources

football-data.org — free tier, no credit card. Sign up at football-data.org/client/register and paste the key in backend/.env. The app falls back to cached fixtures if the key is missing.

StatsBomb open data — completely free, no key needed. Installed via pip install statsbombpy. About 15k shots from WC 2018 and earlier tournaments used to train the xG model.

What I'd improve

• —More WC data — only 5 tournaments of training data. Euro/Copa América results would help a lot.
• —Real-time StatsBomb feed — they have a live data product but it's commercial. For now the xG model is trained on historical shots only.
• —Player-level xG filtering — the shot map aggregates by team. Per-player breakdowns would be a lot more interesting.
• —Mobile layout — the pitch SVG and probability bars don't really work on small screens. Would need a different layout strategy entirely.

Screenshots

[image] [image] [image]

Built by Tanishk Tiwari · VIT Bhopal · WC 2026