Step 1
Only what was known at the time
The model reads past results, but only those that had actually happened before the forecast was made. Anything later is invisible to it, so it can never peek at the future.
Football forecastingTwo modelsWorld Cup 2026 · Premier League 2026–27
This project gives every one of them a probability, from 0–0 to 10–10, before kickoff. Then it publishes how that went.
Two models, one engine
A World Cup is 104 matches between national teams who rarely meet. A Premier League season is 380 matches between clubs who play every week. The same engine runs under both, tuned and tested separately.
Model 01Finished · published
Forecast all 104 matches of the 2026 World Cup before kickoff, then simulated the whole 48-team bracket to price every team's path.
Open the World Cup model →
Model 02Live · 2026–27 season
Prices every remaining fixture of the 2026–27 season each time new results arrive, and runs controlled experiments to find out what actually improves it.
Open the Premier League model →
The engine
Six steps turn past results into a probability for every possible score. Both models follow them; each tunes the details for its own competition.
Step 1
The model reads past results, but only those that had actually happened before the forecast was made. Anything later is invisible to it, so it can never peek at the future.
Step 2
From those results it learns an attack and a defence strength for every team, with an honest measure of uncertainty. Teams with little data are pulled gently towards average.
Step 3
Your attack, their defence, a league-wide base rate and home advantage combine into an expected number of goals for each side.
log λ = μ + home + attack − defence
Step 4
Goals follow Poisson curves, with the Dixon–Coles correction for how often low scores like 0–0 and 1–1 really happen. Every score from 0–0 to 10–10 gets a probability.
Step 5
Home win, draw and away win are sums of regions of the same grid, and so are totals and goal margins. They can never contradict each other.
Step 6
To answer bigger questions, like who wins the tournament or the league, the model simulates every remaining match many thousands of times and counts the outcomes.
Try it
Set each team's attack and defence and switch home advantage on or off. The engine turns them into expected goals, then into the full score landscape and the three outcomes.
Most likely single score:
Illustrative values chosen for teaching: a base rate μ of 0.20 and a home advantage of 0.18 on the log scale. The real models learn these from data.
Same engine, different jobs
The core never changes. What each model learns from, how it treats home advantage, and how it is judged are tuned to its competition.
| World Cup model | Premier League model | |
|---|---|---|
| Learns from | 49,000+ international results | 4,560 Premier League matches, 2014/15 on |
| Elo anchor | K = 40, scaled by match importance | K = 20, league matches only |
| Home advantage | Hosts get 1.4× the learned edge; elsewhere each side gets half | Learned from the data, applied to every home side |
| New or thin teams | Uncertainty widened for teams with little data | Promoted clubs start 75 Elo below average; unseen clubs start from the prior |
| Simulation | 20,000 whole tournaments, extra time and penalties included | 20,000 whole seasons under league tiebreak rules |
| How often it re-fits | About 40 daily refits through the tournament | A new forecast after new results arrive |
| Judged against | Tied plain Elo; the market was better by 0.010 | Slightly ahead of Elo, not proven; the market is ahead by 0.0065 |
| Status | Finished, archived with a DOI | Live, scored match by match |