World Cup 2026 Prediction Model

A statistical soccer prediction engine built around the 2026 FIFA World Cup. Uses the Dixon-Coles Poisson model to forecast match outcomes, calibrates probabilities against historical data, and sizes bets using the Kelly criterion.

Built as a learning project to explore sports modelling and quantitative betting theory.

How It Works

1. Dixon-Coles Model (`dixon_coles.py`)

Fits per-team attack (α) and defense (β) ratings from 15 years of international match data using maximum likelihood estimation. Key features:

Time decay — recent matches weighted more heavily (2-year half-life)
Match importance tiers — World Cup finals weighted 4× more than friendlies
Low-score correction (τ) — fixes Poisson's tendency to underestimate 0-0 and 1-1 results
Home advantage (γ) — applied only to the host nation (Canada)
L2 regularisation — prevents overfitting on teams with few matches
Priors — FIFA-ranked fallbacks for teams with sparse data

2. Probability Calibration (`dixon_coles.py → fit_calibration`)

Raw model probabilities are temperature-scaled (T ≈ 0.65) to correct a systematic bias discovered during testing: the model underestimates strong favourites by ~6–8 percentage points.

Probability range	Raw model	After calibration	Actual hit rate
60–70%	64.6%	~70%	70.4%
70–80%	74.4%	~79%	80.8%
80–90%	84.4%	~89%	89.6%

3. Kelly Criterion (`kelly.py`)

Sizes bets to maximise long-run bankroll growth:

Half-Kelly (50% of full Kelly) to reduce variance
2% slippage buffer on market odds
2.5% edge floor — ignores bets with thin or negative expected value
15% daily cap — total exposure capped at 15% of bankroll per matchday

4. Tournament Simulator (`simulator.py`)

Runs 100,000 Monte Carlo simulations of the full WC bracket (group stage → R16 → QF → SF → Final), outputting win probabilities for all 32 teams.

Setup

pip install -r requirements.txt   # numpy, pandas, scipy
python data_fetcher.py            # download 15k real international results

Daily Usage

Fetch live odds and run bet sheet

python -c "
import json, urllib.request, pandas as pd
req = urllib.request.Request(
    'https://site.api.espn.com/apis/site/v2/sports/soccer/fifa.world/scoreboard',
    headers={'User-Agent': 'Mozilla/5.0'})
with urllib.request.urlopen(req, timeout=10) as r:
    data = json.loads(r.read())
rows = []
for event in data.get('events', []):
    for comp in event.get('competitions', []):
        comps = comp.get('competitors', [])
        if len(comps) < 2: continue
        home = comps[0].get('team',{}).get('displayName','')
        away = comps[1].get('team',{}).get('displayName','')
        od = (comp.get('odds') or [{}])[0] or {}
        ml = od.get('moneyline') or {}
        def dec(node, key='close'):
            v = ((node.get(key) or {}).get('odds') or (node.get('open') or {}).get('odds'))
            try:
                v = int(str(v).replace('+',''))
                return round(v/100+1,3) if v>0 else round(100/abs(v)+1,3)
            except: return None
        dn = od.get('drawOdds') or {}
        h = dec(ml.get('home') or {}); d = dec(dn,'close') or (lambda v: round(int(str(v).replace('+',''))/100+1,3) if v else None)(dn.get('moneyLine')); a = dec(ml.get('away') or {})
        if h and d and a:
            rows.append({'home_team':home,'away_team':away,'home_odds':h,'draw_odds':d,'away_odds':a})
pd.DataFrame(rows).to_csv('espn_wc2026_odds.csv', index=False)
print(f'{len(rows)} fixtures saved')
"

python main.py --fixtures real_fixtures.csv --ref-date $(date +%Y-%m-%d) --host Canada --no-sim --odds espn_wc2026_odds.csv

Update bankroll after results

python -c "from kelly import update_bankroll; update_bankroll(-300)"  # net P&L in ETB

Run full tournament simulation

python main.py --fixtures real_fixtures.csv --host Canada

Backtest Results

Walk-forward evaluation across 11,849 matches from 2014–2026. Elo is used as the market proxy.

Metric	Value
Matches tested	11,849
Total bets placed	14,713
DC Brier score	0.549
Elo (market proxy) Brier	0.555
Edge over market	+1.1%
Starting bankroll	2,000 ETB
Final bankroll	108,841 ETB
Win rate	32.9%

Win rate is low because the model primarily bets draws and underdogs at high odds — it doesn't need to win often, it needs the wins to pay enough. Backtest is against Elo prices, not a real sharp bookmaker, so real-world performance will be lower.

Live Results — WC 2026

🟢 LIVE — Real money bets placed on melbet-et.com using automated execution via morning.py + bet_executor.py.

Starting bankroll: 2,000 ETB | Current bankroll: 2,134 ETB | Overall: +134 ETB (+6.7%)

June 12–19 results are real tracked bets. From June 20 onwards, bets are placed live on melbet-et.com via the automated workflow. Bankroll reflects actual account balance.

Note on sample size: ~31 settled bets is still noise. The model's 1.1% edge only becomes statistically detectable after 200–500 bets. The EPL 2026–27 season (starting August) is where real validation happens — 380 matches per season.

Matchday 1 — June 12, 2026

Match	Bet	Odds	Stake	Result	P&L
Canada vs Bosnia and Herzegovina	Canada	1.87	196 ETB	LOST (1–1 draw)	−196 ETB
USA vs Paraguay	Paraguay	3.85	104 ETB	LOST (USA 4–1)	−104 ETB

Day P&L: −300 ETB | Bankroll: 1,700 ETB

Matchday 2 — June 13, 2026

Match	Bet	Odds	Stake	Result	P&L
Qatar vs Switzerland	Qatar	14.0	25 ETB	LOST (1–1 draw)	−25 ETB
Brazil vs Morocco	Draw	3.7	85 ETB	WON (1–1)	+230 ETB
Haiti vs Scotland	Haiti	6.0	48 ETB	LOST (0–1)	−48 ETB

Day P&L: +157 ETB | Bankroll: 1,857 ETB

Matchday 3 — June 14, 2026

Match	Bet	Odds	Stake	Result	P&L
Australia vs Türkiye	Australia	5.0	123 ETB	WON (2–0)	+492 ETB
Ivory Coast vs Ecuador	Draw	2.85	23 ETB	LOST (1–0)	−23 ETB
Ivory Coast vs Ecuador	Ecuador	2.45	8 ETB	LOST (1–0)	−8 ETB
Netherlands vs Japan	Japan	3.65	49 ETB	LOST (2–2 draw)	−49 ETB
Sweden vs Tunisia	Tunisia	4.30	50 ETB	LOST (5–1)	−50 ETB
Germany vs Curaçao	Draw	16.0	18 ETB	LOST (7–1)	−18 ETB
Germany vs Curaçao	Curaçao	31.0	8 ETB	LOST (7–1)	−8 ETB

Day P&L: +336 ETB | Bankroll: 2,193 ETB

Matchday 4 — June 15, 2026

Match	Bet	Odds	Stake	Result	P&L
Spain vs Cape Verde	—	—	—	0–0 draw (no bet)	—
Belgium vs Egypt	Draw	3.95	74 ETB	WON (1–1)	+218 ETB
Belgium vs Egypt	Egypt	5.75	104 ETB	LOST (1–1)	−104 ETB
Saudi Arabia vs Uruguay	Draw	4.50	75 ETB	WON (1–1)	+263 ETB
Saudi Arabia vs Uruguay	Saudi Arabia	8.00	30 ETB	LOST (1–1)	−30 ETB
Iran vs New Zealand	Draw	3.50	16 ETB	WON (2–2)	+40 ETB
Iran vs New Zealand	New Zealand	4.70	37 ETB	LOST (2–2)	−37 ETB

Day P&L: +350 ETB | Bankroll: 2,543 ETB

Matchday 5 — June 16, 2026 (model not run — no bets)

Match	Result
France vs Senegal	France 3–1
Norway vs Iraq	Norway 4–1
Argentina vs Algeria	Argentina 3–0

Matchday 6 — June 17, 2026 (model not run — no bets)

Match	Result
Austria vs Jordan	Austria 3–1
Portugal vs Congo DR	1–1 draw
England vs Croatia	England 4–2
Ghana vs Panama	Ghana 1–0
Colombia vs Uzbekistan	Colombia 3–1

Matchday 7 — June 18, 2026

Match	Bet	Odds	Stake	Result	P&L
Czechia vs South Africa	South Africa	4.80	130 ETB	LOST (1–1 draw)	−130 ETB
Switzerland vs Bosnia-Herzegovina	Switzerland	1.556	142 ETB	WON (4–1)	+79 ETB
Canada vs Qatar	Canada	1.294	74 ETB	WON (6–0)	+22 ETB

Day P&L: −29 ETB | Bankroll: 2,513 ETB

Matchday 8 — June 19, 2026 (model not run — no bets)

Match	Result
USA vs Australia	USA 2–0
Morocco vs Scotland	Morocco 1–0
Brazil vs Haiti	Brazil 3–0
Türkiye vs Paraguay	Paraguay 1–0

Matchday 9 — June 20, 2026

Match	Bet	Odds	Stake	Result	P&L
Netherlands vs Sweden	Netherlands	1.714	—	MISSED (5–1)	—
Germany vs Ivory Coast	Draw	4.60	119 ETB	LOST (2–1)	−119 ETB
Germany vs Ivory Coast	Ivory Coast	6.50	59 ETB	LOST (2–1)	−59 ETB
Ecuador vs Curaçao	Draw	8.50	24 ETB	WON (0–0)	+180 ETB

Day P&L: +2 ETB | Bankroll: 2,515 ETB

Matchday 10 — June 21, 2026

Match	Bet	Odds	Stake	Result	P&L
Spain vs Saudi Arabia	Draw	10.5	22 ETB	LOST (4–0)	−22 ETB
Spain vs Saudi Arabia	Saudi Arabia	34.0	8 ETB	LOST (4–0)	−8 ETB
Belgium vs Iran	Draw	4.52	13 ETB	WON (0–0)	+46 ETB
Belgium vs Iran	Iran	7.6	51 ETB	LOST (0–0)	−51 ETB
Uruguay vs Cape Verde	Uruguay	1.455	203 ETB	LOST (2–2)	−203 ETB
New Zealand vs Egypt	New Zealand	5.75	46 ETB	LOST (1–3)	−46 ETB
New Zealand vs Egypt	Draw	4.04	33 ETB	LOST (1–3)	−33 ETB

Day P&L: −317 ETB | Bankroll: 2,197 ETB

Uruguay 2–2 Cape Verde was the decisive blow. The model gave Uruguay 95.5% — one of its highest-confidence calls — and it drew. That's a 1-in-20 outcome by the model's own estimate. Pure variance, but it cost 203 ETB in a single bet.

Matchday 11 — June 22, 2026

Match	Bet	Odds	Stake	Result	P&L
Argentina vs Austria	Argentina	1.46	182 ETB	WON (2–0)	+83.72 ETB
France vs Iraq	Draw	9.7	42 ETB	LOST (France 3–0)	−42 ETB
France vs Iraq	Iraq	28.0	26 ETB	LOST (France 3–0)	−26 ETB
Norway vs Senegal	Senegal	3.35	17 ETB	LOST (Norway 3–2)	−17 ETB
Jordan vs Algeria	Jordan	6.6	62 ETB	LOST (Algeria 2–1)	−62 ETB

Day P&L: −63.28 ETB | Bankroll: 2,133.99 ETB

Argentina won comfortably but the 4 losing long shots (draw/Iraq/Senegal/Jordan) outweighed it. France 3–0 Iraq and Norway 3–2 Senegal were the model's misses — both went to the stronger team.

Running Total

Day	P&L	Bankroll	Note
Start	—	2,000 ETB	—
June 12	−300 ETB	1,700 ETB	Canada draw & Paraguay wrong
June 13	+157 ETB	1,857 ETB	Brazil/Morocco draw hit
June 14	+336 ETB	2,193 ETB	Australia 5.0 win; 5 losers
June 15	+350 ETB	2,543 ETB	Three draws hit (Belgium, Saudi/Uruguay, Iran/NZ)
June 16	—	2,543 ETB	Model not run
June 17	—	2,543 ETB	Model not run
June 18	−29 ETB	2,513 ETB	South Africa drew; Switzerland & Canada won
June 19	—	2,513 ETB	Model not run
June 20	+2 ETB	2,515 ETB	Ecuador 0–0 draw saved the day; Netherlands missed
June 21	−317 ETB	2,197 ETB	Uruguay 2–2 Cape Verde at 1.455 odds — 95.5% model call lost
June 22	−63 ETB	2,134 ETB	Argentina won; France/Iraq, Norway/Senegal, Jordan/Algeria all went to the stronger team

Overall (settled): +134 ETB (+6.7%) across ~45 settled bets

⚠️ Uruguay drawing at 1.455 odds is exactly the kind of low-probability high-cost event Kelly sizing cannot protect against — the stake was large because the model was confident. This is not model failure, it's the tail risk baked into betting 95%+ favourites at short odds with Kelly. 100 ETB single-bet cap now in place to limit future exposure.

Why the Model Bets Underdogs and Draws

The model only bets when Expected Value > 0:

EV = (model_probability × payout) − stake

A heavy favourite at short odds (e.g. Switzerland at 1.211) requires winning 83%+ just to break even. If the model says 77%, the EV is negative — you lose money long-term even though Switzerland probably wins.

Draws and underdogs are where bookmakers consistently misprice:

Draws: bettors dislike them, so bookmakers price them slightly worse than true probability
Underdogs: public bias toward recognisable teams inflates their implied probability, leaving value on the other side

The core rule: it's not about who wins, it's about whether the odds pay you enough for the risk.

Limitations

50 games is too few: The model's 1.1% edge over the market only becomes statistically meaningful after 500+ bets. A single tournament is dominated by variance.
Soft market proxy: The backtest beats Elo, not a real bookmaker like Pinnacle. Real-world edge is likely smaller.
Training data skew: Model is trained on all international matches — WC results may differ from qualifiers and friendlies.
Not financial advice: This is a demo project. Bet responsibly.

Club Football Model

The same Dixon-Coles engine extended to Premier League, La Liga, Bundesliga, Serie A, and Ligue 1 — where 500+ bets per season make edge statistically measurable.

Key difference from the WC model: home advantage (γ) applies to all home teams, not just the tournament host.

# Download 11 seasons of top-5 league data (~19,000 matches, run once):
python club_main.py --fetch

# Generate today's bet sheet for any league:
python club_main.py --league epl
python club_main.py --league laliga
python club_main.py --league bundesliga
python club_main.py --league seriea
python club_main.py --league ligue1

# Provide your own odds CSV:
python club_main.py --league epl --odds my_odds.csv

Project Structure

├── main.py              # WC entry point — fit model, simulate, generate bet sheet
├── club_main.py         # Club football entry point (EPL, La Liga, Bundesliga, etc.)
├── dixon_coles.py       # Dixon-Coles Poisson model + temperature calibration
├── simulator.py         # Monte Carlo tournament simulator (100k runs)
├── elo.py               # Elo model (market proxy for backtesting)
├── kelly.py             # Kelly criterion bet sizing
├── backtest.py          # Walk-forward backtest (DC vs Elo)
├── calibrate.py         # Sweep time-decay parameter PHI
├── data_loader.py       # CSV validation and loading
├── data_fetcher.py      # Download real international match data from GitHub
├── club_data_fetcher.py # Download club football data from football-data.co.uk
├── odds_scraper.py      # Scrape historical odds from ESPN
├── generate_fixtures.py # Generate synthetic test fixtures
├── tests.py             # Unit tests
└── requirements.txt

CLI Reference

# Download real match data
python main.py --fetch

# Fit model + simulate tournament
python main.py --fixtures real_fixtures.csv --host Canada --n-sims 100000

# Bet sheet from odds file (no simulation)
python main.py --fixtures real_fixtures.csv --no-sim --odds my_odds.csv

# Calibrate time-decay half-life
python calibrate.py --fixtures real_fixtures.csv

# Walk-forward backtest
python backtest.py --fixtures real_fixtures.csv --bankroll 2000

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

World Cup 2026 Prediction Model

How It Works

1. Dixon-Coles Model (`dixon_coles.py`)

2. Probability Calibration (`dixon_coles.py → fit_calibration`)

3. Kelly Criterion (`kelly.py`)

4. Tournament Simulator (`simulator.py`)

Setup

Daily Usage

Backtest Results

Live Results — WC 2026

Matchday 1 — June 12, 2026

Matchday 2 — June 13, 2026

Matchday 3 — June 14, 2026

Matchday 4 — June 15, 2026

Matchday 5 — June 16, 2026 (model not run — no bets)

Matchday 6 — June 17, 2026 (model not run — no bets)

Matchday 7 — June 18, 2026

Matchday 8 — June 19, 2026 (model not run — no bets)

Matchday 9 — June 20, 2026

Matchday 10 — June 21, 2026

Matchday 11 — June 22, 2026

Running Total

Why the Model Bets Underdogs and Draws

Limitations

Club Football Model

Project Structure

CLI Reference

About

Uh oh!

Releases

Packages

Uh oh!

Contributors

Uh oh!

Languages

Name		Name	Last commit message	Last commit date
Latest commit History 18 Commits
__pycache__		__pycache__
.gitignore		.gitignore
README.md		README.md
auth_state.json		auth_state.json
backtest.py		backtest.py
calibrate.py		calibrate.py
club_data_fetcher.py		club_data_fetcher.py
club_main.py		club_main.py
data_fetcher.py		data_fetcher.py
data_loader.py		data_loader.py
dixon_coles.py		dixon_coles.py
elo.py		elo.py
espn_wc2026_odds.csv		espn_wc2026_odds.csv
fixtures.csv		fixtures.csv
generate_fixtures.py		generate_fixtures.py
historical_odds.csv		historical_odds.csv
june14_odds.csv		june14_odds.csv
june15_odds.csv		june15_odds.csv
june18_odds.csv		june18_odds.csv
june20_odds.csv		june20_odds.csv
kelly.py		kelly.py
main.py		main.py
morning.py		morning.py
odds_scraper.py		odds_scraper.py
real_fixtures.csv		real_fixtures.csv
requirements.txt		requirements.txt
simulator.py		simulator.py
tests.py		tests.py
to-do-or-implement.md		to-do-or-implement.md
to-run.txt		to-run.txt
weekly-check.md		weekly-check.md
yesterday_odds.csv		yesterday_odds.csv

Folders and files

Latest commit

History

Repository files navigation

World Cup 2026 Prediction Model

How It Works

1. Dixon-Coles Model (dixon_coles.py)

2. Probability Calibration (dixon_coles.py → fit_calibration)

3. Kelly Criterion (kelly.py)

4. Tournament Simulator (simulator.py)

Setup

Daily Usage

Backtest Results

Live Results — WC 2026

Matchday 1 — June 12, 2026

Matchday 2 — June 13, 2026

Matchday 3 — June 14, 2026

Matchday 4 — June 15, 2026

Matchday 5 — June 16, 2026 (model not run — no bets)

Matchday 6 — June 17, 2026 (model not run — no bets)

Matchday 7 — June 18, 2026

Matchday 8 — June 19, 2026 (model not run — no bets)

Matchday 9 — June 20, 2026

Matchday 10 — June 21, 2026

Matchday 11 — June 22, 2026

Running Total

Why the Model Bets Underdogs and Draws

Limitations

Club Football Model

Project Structure

CLI Reference

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

1. Dixon-Coles Model (`dixon_coles.py`)

2. Probability Calibration (`dixon_coles.py → fit_calibration`)

3. Kelly Criterion (`kelly.py`)

4. Tournament Simulator (`simulator.py`)

Packages