There are two broad ways to forecast a football match with software. A statistical model learns from past results and turns team strength into probabilities. An AI language model reads about the match, from form to injuries, and gives a judgement. We run both on matches in selected leagues and grade them side by side, so this comparison is based on our own setup and our own public record rather than on anyone else’s claims.
The two approaches at a glance
| Statistical model | AI analyst (language model) | |
|---|---|---|
| What it reads | Past results, dates, home and away | Text: team news, previews, the model’s own numbers |
| How it decides | Fitted formulas (Elo, Dixon-Coles goals model) | Reasoning over the text it is given |
| Output | A full grid of score probabilities | Probabilities plus a written explanation |
| Repeatable | Yes: same data, same numbers | Not exactly: answers can vary between runs |
| Calibrated | Checked against past seasons | Only if tested over many graded matches |
| Knows about injuries | Only through results | Yes, if given dated sources |
| Main failure mode | Slow to react to news it cannot see | Confident mistakes and invented facts |
Where the statistical model is stronger
A model fitted to thousands of past matches knows how often a side of a given strength wins at home, how often games end 1–1 and how much a recent run of form really matters. Its probabilities are consistent: two similar matches get similar numbers, and the three outcomes always add up correctly. Most importantly it can be backtested, run on past seasons using only the data available at the time, so we know roughly how well calibrated it is before a single forecast is published.
Where the AI analyst can add something
Results only tell a model about a missing striker after the team has started losing. A language model reading a dated club statement knows on the day. That is the gap AI is meant to fill: news that the statistical model cannot see yet, such as key absences, heavy squad rotation or a new manager. AI is also far better at explaining a forecast in plain English.
The risk is that the same fluency makes mistakes hard to spot. That is why our AI tools only work from a dossier of sourced facts, may only nudge the expected goals within fixed limits, and have every written claim checked by an editor.
What our own record shows so far
The AI analysts only run on matches in selected leagues, so a fair comparison uses only the matches where every layer has a forecast. On those 11 matches (ranked probability score, lower is better):
- Published forecast (statistical model): 0.1916
- Claude, shadow: 0.1926
- Codex, shadow: 0.1936
- Combined AI layer, shadow: 0.1919
Across all 72 graded matches, the published forecast scores 0.1929 and the base rate (always the league average) scores 0.2414.
At this sample size the gaps between the model and the AI layers are small, so read them alongside their confidence intervals before drawing any conclusion. The AI layer section of the track record shows the same numbers with confidence intervals, and the models page shows which conditions for giving the AI a weight have been met.
Why we combine them the way we do
We publish the statistical model’s numbers and let the AI layer prove itself first. It gets a weight only if, over at least 300 matches and six weeks, it beats the model on both log loss and RPS, with a 95% confidence interval for the RPS gap that excludes zero and an improvement in at least two thirds of leagues. Those rules were written before the first forecast, so a lucky month cannot change them.
The bottom line
- Use a statistical model for the probabilities: it is consistent, testable and calibrated.
- Use AI for what results cannot show yet, from sourced news, and check what it says.
- Trust neither without a public, locked record over hundreds of matches.
For a shorter take on chatbots in particular, see Can ChatGPT predict football?
