How the model works
In short: the model turns match history, team strength and the circumstances of a particular match into probabilities. No pundits and no gut feeling.
Football
The base is each team's attack and defence strength from goals and expected goals over recent seasons, with recent matches weighted more. On top of that come strength ratings, home advantage, line-ups and missing players, fixture congestion and the referee (for cards). The expected goals of both teams give a grid of probabilities for every score, and from it the result, totals, handicaps and both teams to score.
Hockey
The model plays the score minute by minute: teams slow down when level late on, the trailing side pulls its goalie, and overtime and shootouts follow their own rules. Separately it looks at the starting goalie and back-to-back games. That is why hockey shows both the regulation result and the winner including overtime.
Basketball, NBA (beta)
Every player's strength in attack and defence — from his box score and from how the team plays with him on the floor, starting from last season's ratings. Before a game the model takes the players who can play (rosters and the injury report) with their usual minutes. On top: home court, the altitude of Denver and Utah, the second night of a back-to-back, a busy schedule, travel and time zones, both teams' pace. The expected points give a joint score distribution with overtimes: the winner, handicaps, totals and the regulation 1X2 come from it. The model is most accurate on the winner; on handicaps and totals at the bookmakers' lines the market is still stronger.
Tennis (beta)
A player's strength ratings — overall and on the surface, from wins and from the share of games won — plus the ATP/WTA ranking, head-to-head record, rest and recent workload. From them the model gets the chances to win, then the chances in every game and set — hence games totals and the score in sets; how often players end up in tiebreaks is taken into account too.
How we check ourselves
On history the model saw only past matches each time and estimated the next one — 48,148 matches were checked this way. Since launch we store the last estimate before each match and never change it afterwards. All results are on the model accuracy page.
Questions
Is this a recommendation?
No. We publish a model’s statistical estimates and do not encourage any action. Any probability below 100% means the outcome may not happen.
What does "60%" mean?
Take many matches where the model gave an outcome 60%: it happens in about 60 of 100 and does not in 40. We check this openly on the accuracy page.
How often are the numbers updated?
Every 15 minutes. The model re-estimates when line-ups, squad news or other results arrive. The match page shows the time of the latest estimate.
Why is tennis marked "beta"?
The tennis model is younger than the football and hockey ones: newcomers to the main tour have little history, and the model has only started collecting Challenger and ITF results and serve statistics, so its estimates are still rougher.