How the model works
Every forecast comes from historical results only. Features for a match are built from matches played strictly before it, so nothing the model sees at training time leaks from the future.
Elo rating difference
Updated after every match, home-advantage aware, margin-weighted.
Rolling form
Points and goal difference across each side's last five and ten matches.
Head-to-head
How the home side has fared in recent meetings between these two clubs.
Rest days
Days since each side last played, capped to avoid outliers.
Confidence gap, not confidence
A 45/30/25 match is a coin flip even though home win is the favourite. A 78/14/8 match is a genuine signal. Sure Bets rank on the distance between the top outcome and the second, which is why a short, selective list beats a long one.
What we publish
The accuracy figures on the board come from a time-based holdout: the model trains on the earlier 80% of your data and is scored on the later 20% it has never seen. Expect roughly 55–75% on real leagues depending on how selective the slice is. Anyone promising more than that is selling something else.