FootballSeer
Football prediction dashboard in the spirit of BetMines: probabilities for every
goals-based market across 38 leagues, a screener, value bets, a matchup
predictor, and a public track record that compares the model with the bookmakers.
pipeline/ Python: data ingest, Dixon-Coles model, backtest, JSON export
web/ Next.js app that renders the exported JSON as static pages
data/ Prediction ledger + backtest (committed, so the track record is auditable)
How it works
- Data: football-data.co.uk results and
bookmaker odds for 22 European divisions plus 16 extra leagues worldwide. If the
site is unreachable, the pipeline falls back to the newest Wayback Machine
snapshot. For fixtures it falls back to openfootball
schedules.
- Model: one Dixon-Coles model per league. It estimates each team's attack and
defence strength, plus home advantage and a low-score correction. Recent matches
count more (half-life ≈ 1 year), and a ridge penalty pulls teams with little data
towards the league average.
- Markets: the model produces a probability for every scoreline, and every
market is derived from that one table: 1X2, double chance, draw-no-bet,
over/under 0.5–4.5, BTTS, team goals, clean sheets and correct score.
- Value: if model probability × average bookmaker odds − 1 is above 5%, the bet
is flagged.
- Track record: live predictions go into
data/ledger.csv before kick-off and
are settled when results arrive. backtest replays the season week by week,
using only data that existed at the time. /lookback shows every prediction from the
past seven days next to its result, with ESPN final scores filling in until
football-data publishes.
Running it
# pipeline
cd pipeline
python -m venv .venv && .venv/Scripts/activate # or source .venv/bin/activate
pip install -r requirements.txt
python -m pytest -q
python -m footballseer backtest --start 2025-08-01 # one-off, ~10-20 min on first download
python -m footballseer run # daily
# web
cd web
npm install
npm run dev
Basketball
The same product for basketball lives under /basketball (sport switch at the top of the sidebar),
with every football page mirrored: games, live scores, lookback, value bets, daily bets, previews,
favourites, filters, generator, scanner, trends, team and league stats, predictor, leagues and track record.
- Leagues: NBA, WNBA, NCAA men's (Division I), G League and Australia's NBL from ESPN's public
scoreboards; EuroLeague and EuroCup from the EuroLeague's public API. No keys.
- Model (
pipeline/footballseer/basketball/model.py): weighted ridge regression on each side's
points (offence, defence, home advantage, neutral sites, time decay). Margin and total are normal
around the expected values, with spreads measured from residuals; overtime resolves level scores.
Decay and shrinkage are tuned per league (NCAA's sparse schedule needs much less shrinkage).
- Markets: moneyline, spread, total, team totals and winning margin for the game; for each half
and quarter, a three-way result (periods can end level), spread, total and team totals; and
half-time/full-time. Each period gets the league's fitted share of the points and its own spreads.
The 2nd half includes overtime, quarters are regulation only, and NCAA men's (two halves) has no
quarter markets. Spread and totals are priced at the bookmaker's line when ESPN publishes one,
otherwise at the model's fair spread and the league's average total; period lines scale the game's.
Every game carries its lines and period scores, so settlement is exact.
- Odds caveat: ESPN keeps no odds for past games, so the backtest measures accuracy and calibration
but not ROI. Value-bet ROI and model-vs-market comparisons build up from live predictions, which
store the bookmaker's lines and prices in
data/basketball/ledger.csv.
python -m footballseer basketball backtest # one-off, ~3 min on first download (NCAA is ~450 scoreboard days)
python -m footballseer basketball run # daily, writes web/public/data/basketball/
Live basketball comes from /api/basketball/live (ESPN and EuroLeague feeds, same caching as football).
Deploying (≈ $0/month)
- Web: import the repo into Vercel (or Cloudflare Pages) with root directory
web.
Every page is static, so it serves from the CDN. The one exception is live scores: /api/live and
/api/live/[league]/[event] are serverless functions that proxy ESPN with 15–20 s caching, so
however many browsers poll, upstream sees a few requests a minute.
- Pipeline:
.github/workflows/predict.yml runs twice a day on GitHub Actions,
commits the refreshed JSON, and the push triggers a redeploy.
- A database (e.g. Neon Postgres) is only needed once you add accounts, saved
strategies or alerts.
Known limitations / next steps
- The model only sees goals. Adding xG, shots, lineups and injuries is the biggest
accuracy upgrade. A LightGBM layer on top of these ratings is the natural next step.
- Promoted teams start at the league average and take a few weeks to settle.
- Odds-based markets are limited to 1X2 and O/U 2.5 (what football-data publishes).
Use The Odds API or API-Football for BTTS, Asian handicap, corners and so on.
- Cups, European competitions and leagues outside these 38 need another data
source (API-Football covers 1,000+ competitions).
- Gambling regulation differs by country. Check the local rules before adding
affiliate links or charging for tips.
