Machine Learning in Football · 02
Every match, predicted — then marked
A Dixon-Coles goals model turns 150 years of international results into a probability for every scoreline of every World Cup match. The catch is accountability: the model is frozen the day before kickoff, so no result can leak back into it, and then it's graded against what actually happens. See how it rates any match — then watch it mark its own homework as the tournament unfolds.
How the model sees a match
Pick a fixture
Win / draw / loss probabilities, expected goals, the most likely scorelines, and the full score grid — the Dixon-Coles model's complete read on the game. Host nations get a venue edge only when they play in their own country.
Even a 60% favourite loses two times in five. A forecast isn't a prediction of what will happen — it's an honest spread of what might.
Marking the homework
How the model is doing against reality
Every played match, scored against the frozen forecast: did it call the right result, how close was the scoreline, and does it beat a coin-flip baseline on Brier and log-loss? The reliability curve asks the deeper question — when the model says 60%, does it happen 60% of the time?
—
The prediction log
Each result so far, with the model's pre-tournament call beside it.
| Date | Match | Model | W/D/L | Actual | ✓ |
|---|
—
Tune the model
What if it weighed things differently?
The same Dixon-Coles engine, with its dials exposed. Turn each factor on or off, slide its weight, and watch a match's odds move — recomputed live from the genuinely fitted strengths, not a mockup. Every number is real model output.
Everything the baseline odds are built from is listed below. Switch an ingredient off, or slide its weight. Turn them all off and the match falls to a dead heat — no side has an edge; turn them all on and you're back to the published model.
Each ingredient is a real term in the Dixon-Coles model — attacking and defensive strength, the fitted home/host edge, whether recent results are weighted more heavily, and the low-score draw correction. Strengths come from genuine refits; everything recomputes live with the exact model math. Manager is deliberately absent — no clean free data, and a fabricated term would defeat the point.
A model is a stack of assumptions. The honest move isn't to hide them — it's to hand you the dials.
The bettor's question: "What if?"
What if you'd bet the same way on every match?
Pick a stake and a strategy, and walk it through all — played group games in order. The odds come from the model's own probabilities (decimal odds = 1 ÷ the model's chance) — we don't have captured bookmaker prices, so this is a clean test of the strategy itself rather than a real-money return. Flip on a bookmaker's margin to see what the house cut does to it.
—
—