The model behind our predictions
Not the exact formula. This page explains what the model actually looks at, how it checks itself, and what we deliberately do not publish.
What goes into a prediction
Ten categories of signal feed every fixture. Each one is scored on its own before anything is combined.
Recent form
The last several matches for each side, weighted more heavily the closer they are to kickoff, and split into home and away form rather than treated as one number.
Head-to-head history
How these two sides have actually matched up in the past. Some pairings run consistently differently than either team's overall form would suggest.
Home advantage
Measured per team and per competition rather than assumed to be the same everywhere. Some sides defend a home ground far better than the league average, others barely at all.
Squad availability
Who is actually expected to play, weighted by how much that player normally contributes to the team, not just a count of names on an injury list.
Referee tendencies
Cards, penalties and stoppage time follow patterns tied to the individual referee appointed to the match, not only to the two teams on the pitch.
Rest and schedule congestion
Days since each side's last match, travel involved, and how packed the recent fixture list has been. Fatigue affects results in ways the league table never shows.
Underlying performance
Shot quality and other match data that predict future results more reliably than the scoreline alone, since a team can win badly or lose well.
Market-specific modelling
Over/under, both teams to score and correct score are modelled in their own right for each fixture, not derived after the fact from the match-winner probability.
Weather conditions
Temperature, wind and precipitation at kickoff, since they change how a match is actually played, not just how comfortable it is to watch.
Positional matchups
How one side's specific strengths line up against the other's specific weaknesses by position on the pitch, rather than comparing the two squads as a whole.
How the signals come together
Each signal above is scored on its own, then combined into one probability for every market on the fixture. The exact weighting between them, and where a pick crosses from a coin flip into genuine confidence, is what the model actually is, so those specifics are not published here. What is published is the result of using them.
What confidence has actually meant
Every prediction carries a confidence score. Here is how each band has performed, measured only on matches that have already finished.
| Confidence score | Graded predictions | Actual accuracy |
|---|---|---|
| High (70%+) | 2,805 | 63.7% |
| Medium (50-69%) | 91,027 | 49.8% |
| Low (<50%) | 66,099 | 43.4% |
Checked and adjusted, continuously
Every graded match feeds back into the system. When a signal's real-world performance drifts from what it predicted, its influence is adjusted going forward on a rolling window, not by hand and not after the fact. A model that is never corrected just gets confidently wrong for longer.
What this page will not tell you
The exact formula, the specific weights between these signals, and the thresholds that separate a high-confidence pick from a coin flip are what the model actually is, so those specifics stay unpublished. What is published, in full and without exception, is the graded record.
See the full graded recordFor entertainment purposes only. Must be 21+ where required. Not available in all states.
Follow the competitions you care about
Create a free account to follow your competitions and teams, and see only their fixtures and predictions on your overview.
Create a free accountFree, no card required, and you can delete your account at any time.