English 
Español 
Português 
Deutsch 
Français No crystal ball, just counting
Strip the marketing from any football prediction model — ours included — and you find structured counting: how often teams score, how often they concede, weighted toward recent matches, adjusted for opposition and venue. From those rates, probability distributions over scorelines follow mathematically.
The word 'AI' adds no magic to this. A model's value lies in disciplined estimation and honest calibration, not in the label. Anyone claiming their algorithm 'knows' results is selling something.
Inside our model, concretely
Our v2 model estimates each team's attack and defence strength with a time-decayed Poisson framework: recent matches weigh more (half-life around 240 days), small samples get shrunk toward league average to avoid overreacting, and teams need a minimum match count before we trust their rates at all.
Those rates produce a scoreline distribution per fixture, hence probabilities for every market we cover. We then blend model probabilities with market-implied ones (roughly 40/60) — an admission, built into the arithmetic, that the market knows things our data does not: injuries, motivation, weather, lineups.
A pick is only published as official when the blended probability clears the price by a meaningful margin with sane odds caps. Most matches produce no official pick; the full methodology, version history and known limitations live on our model page.
What models cannot do
A model cannot know a manager rested his striker, a squad flew back from a cup tie at 4 a.m., or rain turned the pitch to soup. It compresses history into rates; anything outside history is invisible to it.
More fundamentally: even a perfect probability is not a prediction of one match. A 60% call loses four times in ten by construction. Models earn or lose their keep across hundreds of bets, in calibration — do 60% calls land about 60% of the time? — not in any highlight reel.
How to judge anyone's predictions, including ours
Ask for the full settled record (every pick, not a curated month), results in flat-staked units, sample sizes in the hundreds, and — the sharpest test — closing line value: whether picks consistently beat the price the market closed at. CLV separates skill from luck far faster than profit curves do.
We publish all of the above, including losing runs, on our results pages — our early record includes an instructive stretch where longshot picks went 0-for-8 and forced a model revision. Any tipster showing only wins is showing you marketing.
The model rates a home side's attack at 1.6 expected goals against this defence, the away side at 0.9. The Poisson distribution gives home win 52%, draw 24%, away 24% — fair odds 1.92 / 4.17 / 4.17. The market prices the home team at 2.10, implying 47.6%. The blend lands at ~50%, still above the price's implication: positive edge, official pick. Sixty seconds later a leaked teamsheet could invalidate it — which is why edges are estimates, not facts.
- Prediction models are weighted scoring rates → scoreline distributions → market probabilities.
- Ours blends model with market (40/60) and publishes only clear-edge picks — passing on most matches.
- Models are blind to news; probabilities are calibrated over hundreds of games, not single matches.
- Judge any tipster on full flat-staked records and closing line value, never on highlight months.
Frequently asked questions
What accuracy do football prediction models achieve?
On three-way match odds, top models land the correct outcome roughly 50–55% of the time — barely above always-backing-the-favourite. Accuracy is the wrong metric anyway; calibration and value against prices are what pay, which is why we report units and CLV instead.
Why does the model sometimes disagree with obvious form?
Because 'obvious form' is usually five matches of noise, and the model weighs two seasons of rates. When they clash, the longer sample is right more often — not always, just more often, which is all a model claims.
Do you use machine learning?
The core is classical statistics: time-decayed Poisson rates with shrinkage, blended with market prices. It is transparent, debuggable and published. Fancier architectures add opacity faster than they add football accuracy at our data scale.