Case file 04 · Model · Sports analytics
OddsLab
A Premier League match-outcome model built from 24 seasons and 9,120 matches. It's trained offline and runs in the page, so you can price a fixture yourself and compare the answer with what the bookmakers thought.
- Year
- 2026
- Role
- Solo: data, modelling, build
- Read
- 4 min
Context
A match-outcome model for the Premier League, built from a quarter-century of top-flight data: 24 seasons and 9,120 matches. It prices a fixture as home win, draw or away win, with a scoreline distribution underneath.
The point was never to beat the market. I wanted to find out honestly whether a model built from public data and form features could get anywhere near the people who do this for a living, and to show the working.
The problem
I'm a football fan, and a Manchester United one at that. Football matches are notoriously hard to predict, but I wanted to run my own tests and see how close I could get to the bookmakers using only publicly available data from past matches. There are two ways to get that wrong. A model that can't beat “always pick the home team”, which is right 43.5% of the time, is worthless. And a model that seems to beat the bookmakers by a lot almost certainly has a mistake in it.
The mistake to worry about was leakage: information from the future slipping into what the model learns from, so it looks clever by knowing things it couldn't have known at the time. Form figures are worked out from surrounding matches, so it's very easy to let a season's later games inform a prediction about its earlier ones, and end up with a number that looks brilliant yet means nothing.
What I built
Twenty inputs for each fixture, a test that only uses seasons the model has never seen, and four models compared against each other, both on overall accuracy and on predicting draws, which is the hardest result to call. Two versions run on the site. One uses 17 measures of form (recent results, goals for and against, head-to-head record) plus the bookmakers' own odds, turned into probabilities with their profit margin taken out. The other swaps the odds for Elo ratings, a running strength score for each team, so it can price a fixture that has no odds yet. Having both also shows how much of the accuracy comes from the bookmakers' odds and how much from the football.
The trained model is built into the page and runs in your browser, so it isn't a write-up of a model, it's the model. Pick two teams and it gives a predicted score, the chance of a home win, draw or away win, and what's driving that price. You can type in a bookmaker's odds to see where the two disagree. Around that sits the full report. It replays 367 matches from 2025-26, a season the model never saw, with its price next to the bookmakers' and the real result. It has a section on leakage, scores all four models side by side, and shows which inputs the model leans on. There's a map of every team's season grouped by playing style, and a simulator that plays out next season 1,000 times for title and relegation odds. It ends with a list of what the model can't do.
Decisions & iterations
When I first tried this, last year, the first accuracy figure I got was about 66%. That meant one of two things. Either I'd found a system that beat the bookmakers and was about to be a millionaire or, far more likely, something in my code was wrong. So the decisions that mattered were all about not fooling myself. The model is only ever tested on seasons that come after the ones it learned from, the way it would have to work in real life. Shuffling the matches at random would let it learn from games that hadn't been played yet. The site has a whole section on leakage: what it is, the three places it could have crept in here, and what it does to the numbers. Even on the finished version, leaving one leak in pushes accuracy up to 57.5%. With it fixed, 55.8%.
The comparison is made the way it should be: against the bookmakers on exactly the same games, not a flattering subset, and next to an “always home” baseline so you can see the floor. Four models are shown side by side instead of one declared winner, because the gaps between them are small enough that a single number would overstate the result. The most accurate one never predicts a draw at all, so the site reports how each model handles draws next to its accuracy.
Outcome
On two seasons it never saw, 713 matches, the model gets 56.7% right, against the bookmakers' 56.4% on the same games and 43.5% for “always home”. That isn't beating the bookmakers. Their odds are among the model's strongest inputs, and a gap of 0.3 points is two matches out of 713. The fairer test is the version with no odds in it. From past results alone it gets 53.2%, and 54.1% once Elo ratings are added. So public data on its own lands about ten points above the naive baseline and about two points short of the bookmakers. That's the honest headline, and it's why the site talks about limits, not edge.



