FIFA Money Printer

Predict international football match outcomes (aimed at the 2026 World Cup) and turn calibrated probabilities into betting signals across 1X2, totals, and Asian-handicap markets. The model is a logistic-regression + Dixon-Coles goals-model blend with Elo, form, head-to-head, and Glicko features.

See SUMMARY.pdf for the full story (methodology, results, and every idea that was tested — both what worked and what didn't).

How the model works (best configuration)

Validated SOTA: 0.839 out-of-sample log loss (3-way), beats the Elo baseline in 16 of 17 walk-forward folds, and well-calibrated (predicted draw rate 0.221 vs 0.220 actual).

Data. International matches since 2000. Elo / form / H2H / Glicko are computed over the full history; only friendlies are dropped from the model's training rows (kept in the ratings).

Features (12, all causal pre-match snapshots):

• elo_diff — competition-weighted Elo (the dominant signal)
• is_neutral
• rolling form (last 10 games): points, plus an offense/defense split (goals scored / conceded), home & away
• head-to-head vs the specific opponent: shrunk win rate, goal difference, log(#meetings)
• glicko_exp — an uncertainty-shrunk Glicko win expectancy (discounts the rating gap when a team is rarely rated)

Two models, blended 50/50:

1. Logistic regression on the 12 features, trained on competitive matches with exponential time-decay weights. Linear, calibrated, and — proven by experiment — hard to beat here.
2. Dixon-Coles goals model: per-team attack/defense + home edge by time-weighted maximum likelihood from goals (fit on all history with a ~5-year half-life, since international teams play rarely), with the low-score correction. Produces the full scoreline matrix.

The blend wins because the two are decorrelated (a feature classifier vs a goals process): averaging captures both, and the goals model repairs the draw structure the classifier misses. 1X2 log loss went 0.848 (LR alone) → 0.842 (blend) → 0.839 (Dixon-Coles on expanding history).

Secondary markets (over/under, total-goals buckets, 3-way handicap) are derived from the scoreline matrix reconciled to the blend's better 1X2 — so they inherit the blend's calibration plus Dixon-Coles' "win by 1 vs win by 2+" margin detail. wc=True applies a ~+8% World-Cup goal-scale (a real, measured effect) that sharpens totals/handicap at WC finals.

Live prediction uses a synthetic-row trick: the fixture is appended to the results and the exact training feature pipeline is re-run, so live and training features can never drift.

Setup

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Get the data (not stored in git)

1. data/results.csv — martj42 "International football results" (Kaggle). download.py fetches it (needs a Kaggle API token in ~/.kaggle/kaggle.json).
2. Optional, for the market-value feature: the salimt football-datasets dump (kaggle datasets download -d xfkzujqjvx97n/football-datasets) into data/.

The whole data/ folder is git-ignored — regenerate it from the downloads above.

Run (from src/)

cd src
python walk_forward.py        # walk-forward backtest (LR, no market value)
python run_with_mv.py         # backtest WITH market value (--injuries, --model both)
python dixon_coles.py         # Dixon-Coles vs LR vs blend
python predict.py             # demo predictions (LR / DC / blend + secondary markets)
python bet.py fixtures_example.csv   # betting portfolio from a fixtures+odds CSV

Always run from src/ (modules import each other by name and use ../data/...). Use the venv's Python (../.venv/bin/python) — a different numpy build prints harmless matmul warnings.

Predict a fixture (Python)

from predict import MatchPredictor
pr = MatchPredictor("../data/results.csv")          # or (..., "country_market_value.csv") for MV
pr.predict("Brazil", "Argentina", neutral=True)     # {'home_win', 'draw', 'away_win'}
pr.total_buckets("Brazil", "Argentina", wc=True)    # 0-1 / 2-3 / 4+  (wc=True: WC goal-scale)
pr.handicap("Brazil", "Argentina", line=-1)         # Asian handicap +/-1 (with push)
pr.predict("France", "Spain", unavailable={"France": ["Kylian Mbappé"]})   # drop a player

Bet from a CSV — bet.py

python bet.py fixtures.csv --bankroll 1000 --kelly 0.25 --edge 0.03
python bet.py fixtures.csv --mv country_market_value.csv   # use the market-value feature

CSV columns (fixtures_example.csv is a template):

column	meaning
home, away	team names (martj42 spelling)
neutral	1/0 (default 1)
wc	1 => apply World Cup goal-scale (default 0)
odds columns	decimal odds; blank where you have none

Odds columns: home_win, draw, away_win, over_2.5, under_2.5, goals_0_1, goals_2_3, goals_4plus, hcap_home_-1, hcap_away_+1, hcap_home_+1, hcap_away_-1.

Output shows, per bet: your model probability, the devigged fair market probability, vs_fair (model − fair; >0 means you disagree with the sharp market, not just the vig), the edge, and a fractional-Kelly stake.

Modules (src/)

module	role
wc_pipeline.py	features (Elo, form, H2H, Glicko) + build_features / build_dataset
dixon_coles.py	Dixon-Coles goals model → scoreline matrix → 1X2 + secondary markets
walk_forward.py	walk-forward backtest, model registry, IS/OOS gap
wc_squads.py / wc_market_value.py / wc_squad_dataset.py	competitive filter + market-value feature
predict.py	MatchPredictor (blend, secondary markets, absence override)
portfolio.py	devig + push-aware fractional-Kelly staking
bet.py	CSV → portfolio CLI
run_with_mv.py	one-command runner
chemistry.py	causal chemistry feature (tested null; reference tool)

Honest notes

• The model is at its information ceiling (~0.84 log loss, beats Elo 16/17). Many "beyond strength" feature ideas were tested and ruled out — see SUMMARY.pdf.
• Betting reality: you supply the odds; the system finds where your model disagrees with the market. Without historical odds we can't yet measure realized profit (CLV / betting backtest) — that's the one open dependency.
• Stakes default to conservative (¼-Kelly, 5% max/bet, 25% max exposure). Don't crank to full Kelly — calibrated ≠ infallible.