How our predictions are built
No "AI" buzzwords. Three transparent layers, blended into one probability per market, then calibrated against the verified outcomes you can download from our track record.
▶ Prefer to watch? Take the 60-second tour1 · Team strength (Dixon-Coles)
Dixon-Coles is the academic baseline for football modelling: a Poisson goals model with attack and defence ratings learned per team, plus a low-score correlation correction. Ratings are refit nightly on every result from the leagues we cover, weighted by recency (45-day half-life, last month's form counts for more than last season's).
The fit is maximum likelihood, run per league. A team's rating drives an expected goals number for every fixture they play.
2 · Match goal distribution (Poisson)
Once we have expected goals for both sides, a bivariate Poisson distribution gives us probabilities for every scoreline (1-0, 2-2, 3-1, etc.) up to 5-5. From the scoreline grid we derive every market we tip: 1X2, BTTS, Over/Under at every line, double chance, draw no bet.
3 · Bookmaker blend
Bookmakers spend millions on odds setting. Ignoring their prices is ego. We blend our model probability with the lowest UK retail bookmaker odds (Sky Bet / Bet365 / Ladbrokes / Coral / Paddy Power / William Hill), implied probabilities stripped of the overround. Weighting depends on sample size for that market: more verified picks = more model weight; thin markets lean on the books.
The lowest UK retail price is deliberate. Sharp prices from Pinnacle or exchanges like Betfair are tighter but most UK punters can't take them. We tip prices users can actually use.
4 · Calibration (the part nobody else does)
A model that says "65% likely" should win 65% of the time across thousands of bets. Most don't. Bookmaker overround means raw implied probabilities are systematically wrong. We track every prediction in prediction_logs/verified_outcomes.csv and bucket them by predicted probability and market type. The buckets feed back into the model as adjustments: if our 70-80% Home Win bucket actually hits 65% in real life, we apply a haircut next time.
Every public hit rate on the site (homepage, /accuracy, /track-record, Telegram footer) is computed live from this CSV. No spin, no carve-outs.
5 · The minimum-odds floor
Below @1.20, a leg adds barely any payout to an accumulator but still carries full break-the-acca risk. We skip them. A 5-leg acca priced at @1.44 isn't an acca. It's marketing. If our model can't find legs that combine to at least @2.00, we don't post one.
What we don't do
We don't use neural nets or transformers. The dataset is too small (low thousands of verified bets), the signal-to-noise ratio is too low, and the marginal gain over Dixon-Coles + Poisson is <1% in serious published research. Anyone selling you on "AI football predictions" is selling theatre.
We don't have a paid tier. We don't sell tipster subscriptions. The site is informational; the affiliate link to the bookmaker is how we pay for the API costs.
Want to dig in?
Every prediction we've ever made is downloadable as CSV from the track record page. Code is closed-source for now, but the methodology won't change without an entry on the changelog.