How the model works
Every forecast is built from past results only. For any given match, the model can see games played before it and nothing else, so it never gets to peek at the answer.
Team strength
A rating for every club, updated after each match, aware of home advantage.
Recent form
Points and goal difference over each side's last five and ten games.
Head to head
How these two clubs have gone in their recent meetings.
Rest days
How long each side has had since its last match.
What confidence gap means
For every match the model gives three numbers that add up to 100: the chance of a home win, a draw and an away win. The confidence gap is simply the distance between the biggest of those three and the second biggest.
Say a match comes out 45 home, 30 draw, 25 away. Home is the favourite, but only by 15 points, so it is close to a coin flip. Another match comes out 78 home, 14 draw, 8 away: a gap of 64 points. The second match is the one worth paying attention to. Picks are ranked by that gap, which is why the daily list is short.
Where the model is weak
- Draws are genuinely hard. A draw is rarely the most likely single outcome, even when it is the most likely single scoreline. Any model built this way under-calls draws, and roughly a quarter of matches end level. Expect those to be the losses.
- New signings and injuries are invisible. The model reads results, not team sheets. A side missing three starters still carries last month's rating.
- Promoted teams and early season. With few matches on record, a rating is a guess. August and September are the shakiest months.
- Cup ties and dead rubbers. Rotation and nothing riding on the game both break the assumption that teams are trying to win.
Bookmaker odds are hard to beat
Bookmakers price matches with more data than this model has, and they take a margin on top. A price that looks like 50/50 usually pays less than even money, which is how the house stays ahead. Being right more often than not is not the same as making money.
For that reason we store the bookmaker price for every pick with the margin stripped out, so you can see where the model disagreed with the market and whether that disagreement paid. Treat the picks as one opinion among several, not as instructions.
What we publish
Picks are written to a permanent record before kickoff and cannot be edited or removed afterwards. Every settled pick, won or lost, shows on the full record, along with how well the stated probabilities matched reality. Anyone promising certainty is selling something else.