What the model reads
Years of results, fixtures and prices across every competition we cover, clubs and national teams alike. Only what was known before kick-off goes in, so a prediction is never helped by hindsight, and every change to the model is tested on matches it has never seen before it goes live.
What it cannot know is news that breaks after the prediction is written: a late injury, a rotated line-up, a red card in the warm-up. Predictions freeze before kick-off and are never revised to catch up.
How the tips are chosen
- Best Tip. Every match gets a probability in every market. The Best Tip is the call the model rates most likely among those worth backing at the price on offer — one per match, or none. A match with no call worth backing has no Best Tip that day, and is never graded as one.
- Banker. A Best Tip that cleared a higher bar. Shown with its own badge; the Bankers lead the Bet of the Day.
- Trust. The model's probability of the call in tenths — Trust 8 means about 80% — and a contract: the measured hit rate of every Trust level is printed beside it, so a number that stops meaning what it says is caught in public.
- Value bet. A selection whose price is above what the model thinks it should be. It loses more often than it wins and pays when it lands, so it is labelled by its edge, never by Trust, and graded on its return. Value Bets are for signed-in readers — a free account.
Nothing is chosen by hand. The pipeline writes each call to the ledger before kick-off and grades it after full time, so every hit rate on the track record is computed by the same code as every other number — never typed in.
How good it is
We publish the answer instead of claiming one: hit rate, Brier score and calibration per market, per league and per model, over every prediction ever settled; the return of the Value Bets on their own page. Every match page shows our probability beside the market's, and the record shows who was right. Every number carries its sample size, because a hit rate without one is a story.
Known limits
- New and promoted teams have little history at their level, so their early-season probabilities are the model's weakest.
- Postponed and rearranged matches keep their frozen prediction until they are played; the page says so.
- Small samples. A competition or a Trust level with few graded calls has a record that can swing either way; read the sample size before the percentage.
- Probabilities are estimates. A 70% call loses three times in ten by definition. Nothing here is betting advice.
The audit trail
Every prediction is written to our database before kick-off with the model that made it and a timestamp, and graded automatically after full time. The database role that writes them has no permission to update or delete a prediction — append-only by construction, not by promise — and the full ledger is an open dataset.