1What we model
At the core sits a Dixon Coles goals model, the bivariate Poisson approach published in 1997 and still the academic baseline. We extend it with time decay, an expected goals blend and Bayesian shrinkage, then blend the result with the de overrounded bookmaker consensus, because the market carries information the model cannot see on its own, such as sharp money and late team news. The whole engine reduces to two lines. The first builds one calibrated probability matrix over every scoreline:
where DC is the Dixon Coles probability of the exact scoreline x to y, M is the bookmaker market with its margin removed, C is the fixture context and λ is the blend weight. Dixon Coles sets the shape, the market keeps it honest, context tilts it. The exact value of λ and the context terms are how we are calibrated and stay private. The second line reads every market off that single matrix:
There is no separate model per market, which is why a SharpCast Builder uses the true joint probabilities rather than the bookmaker style assumption that the legs are independent. Term definitions live in the glossary.
2How picks are chosen
For every fixture the model produces a probability and an edge for every market in scope. We rank by probability first, with value as the tiebreak, because the bankroll is built by picking winners, not by chasing edge on coin flips. To be published a pick has to be the favourite of three or stronger, carry a genuine value lean over the fair price, and pass a data quality bar. An edge so large it looks like a data error is quarantined and never published. The Main Outcome, the single headline pick per fixture, carries one extra gate: a minimum price, so a near certain but trivial outcome can never become the headline.
A 0 to 100 display number, capped at 95 as a deliberate humility buffer, that ranks picks the same way the selector does. It reads off as a tier: Strong 80 to 95, Healthy 60 to 79, Modest 40 to 59, Slight 20 to 39, Watch below 20. It is a display summary only. It does not pick the bet; the selector does.
3How the record is kept honest
Every SharpCast Pick is committed before kickoff and stamped with the time it was filed, shown on each fixture. The full record is kept on this site. Nothing is quietly removed: settled outcomes stand, every miss stays visible, and the strike rate is computed from every settled pick, wins and losses alike. Live settled metrics are on the record, and Closing Line Value against the closing price is our gold standard measure of model quality.
The numbers come first. Every written justification on a fixture page is generated from that pick's filed numbers and then AI reviewed for clarity and accuracy before it is shown. The AI never chooses a pick and never moves a probability; the model output is the record, the words only explain it.