Predictions
Custom Soccer Prediction Model
Your factors, your weights, real match history. Nothing is filled in with a guess.
Sign in to build a model
Predictions are computed from synced match history, which is readable only by an account. Signing in is what gives the calculation something to read. A free account keeps 2 models and runs the same engine on the same data as a paid one — the method below is identical whoever runs it, and it is readable without signing in at all.
How the model works
Expected goals come first
Everything else is derived from two numbers: how many goals each side is expected to score. Seven factors each produce their own answer to that question, in goals per match — what the two sides have scored and conceded lately, what they have done at this venue this season, what a typical match in the competition produces, and how each side compares with that average at either end of the pitch. Home and away baselines are kept apart throughout, because the gap between them is home advantage.
Your weights decide what counts
The prediction is the weighted average of those seven answers, and you set the weights. A factor with no data behind it, or with too few matches to read a rate from, is dropped and its weight shared out among the rest — reported on the prediction, never quietly replaced with a league average. If less than half the model’s weight has real data behind it, the prediction is refused rather than produced.
Every scoreline, then the markets
Two expected goal rates give the probability of each scoreline. A home win is the sum of the cells where the home side scores more; over 2.5 is the sum of the cells totalling three or more; both teams to score is the sum of the cells where neither number is zero. Each probability converts back into a price, which is what the bet is worth before a sportsbook adds its margin.
Scored against what actually happened
Every prediction is stored with the data it was computed from, so changing a weight tomorrow cannot alter what the model said yesterday. Once the match is played, the prediction is scored with Brier score and log loss — both measure how well-calibrated the probabilities were, and both are only meaningful over a run of predictions rather than one.
Only soccer, for now. The scoreline model assumes goals arrive like a Poisson process, which is a reasonable description of soccer and a poor one of basketball or American football scoring. Those leagues are in the match history but not in this model. Compare odds for them instead.