The Science

GOLAZO prices every match with the same toolkit quants use to beat real sportsbooks. Every chart below is live — drag the sliders.

📈 1 · Elo ratings — measuring strength

Every nation carries an Elo rating. The probability that team A beats team B is the logistic curve

E_A = 1 / (1 + 10^(−(R_A − R_B + H) / 400))
67%−600+600Elo gap

Win-expectancy E = 1 / (1 + 10−Δ/400). A +120 edge ≈ 67% expectancy.

H is home-field advantage in Elo points (~70, ≈ 0.4 goals), applied only to the host nations. After each result ratings exchange points, R′ = R + K·G·(W − E): K weights match importance, G dampens blowouts. Drag the slider to see how an Elo gap maps to win-expectancy.

2 · Poisson — how many goals?

A team that is expected to score λ goals actually scores k goals with Poisson probability

P(k) = e^(−λ) · λ^k / k!
20%
0
32%
1
26%
2
14%
3
6%
4
2%
5
0%
6

P(k goals) = e−λ·λk/k!. The most likely scoreline shifts right as λ grows.

Football goals are close to Poisson. The Elo gap sets each side’s λ. Slide λ and watch the goal distribution shift.

🔢 3 · Dixon–Coles — the scoreline matrix

Two independent Poissons under-count 0-0 and 1-1 draws, so Dixon & Coles (1997) multiply the low-score cells by a correction τ(ρ):

P(x, y) = τ_ρ(x, y) · Poisson(x; λ_h) · Poisson(y; λ_a)
0
1
2
3
4
5
0
1
2
3
4
5
away goals → · home goals ↓
Home 53%Draw 26%Away 22%

Most likely score 11 (12.2%). The 6×6 grid is the joint pmf with the Dixon–Coles low-score correction.

The normalised grid is the joint distribution of every scoreline — and from it we read 1X2, over/under, both-teams-to-score and the most likely score. Set the two λ’s and explore.

🎲 4 · Monte-Carlo — simulating the cup

We play all 104 matches thousands of times and count how often each team wins. The estimate of any probability converges as the number of simulations N grows:

P(event) ≈ (# sims with event) / N,   error ∝ 1/√N
true p = 17%1 sim1500
After 0 sims
0.0% ± 100.0%

The estimate wobbles, then homes in on the truth. Monte-Carlo error shrinks like 1/√N.

Watch a single probability being estimated live — noisy at first, then converging on the truth. That’s why the tournament page runs thousands of sims.

🧮 5 · Odds, EV & Kelly — sizing the bet

Fair odds are o = 1/p. If you believe the true probability is p, your edge is EV = p·(o−1) − (1−p), and the growth-optimal stake is the Kelly fraction

f* = (p·(o − 1) − (1 − p)) / (o − 1)
f* = 10%stake 0%100%
EV +0.100/coin · Kelly 10%

Growth peaks at the Kelly fraction f* = (p·b − q)/b. Bet more and long-run growth falls; past 2f* you go broke.

The curve is your long-run log-growth versus how much of your bankroll you stake. It peaks exactly at f*. No edge (EV ≤ 0) → f* = 0, don’t bet. We suggest half-Kelly to tame variance.

🔮 6 · Brier score — grading your judgement

When you commit a probability vector over {home, draw, away}, we grade it with a strictly proper scoring rule you can only beat by being honest:

Brier = Σ_k (p_k − o_k)²        Skill = 1 − Brier_you / Brier_ref

o_k is the actual outcome (one-hot). Averaged over your settled bets and benchmarked against a coin-flip prior, this Brier Skill Score ranks forecasting ability independently of how many coins you wagered. The skill leaderboard rewards calibration, not luck.

References: Elo (1978); Dixon & Coles, Modelling Association Football Scores (JRSS-C, 1997); Kelly, A New Interpretation of Information Rate (1956); Brier (1950); Hvattum & Arntzen, Using ELO ratings for match result prediction (2010).