Expected goals (xG)
The sum of model-estimated scoring probabilities for shots. Shot-based xG derived from actual chances is different from a pre-match forecast of goal rates.
Both teams to score (BTTS)
An event that occurs when each side scores at least once. BTTS Yes and No probabilities can be calculated from a joint distribution of final scores.
Poisson distribution
A probability distribution for counts, often used as a baseline for modeling football goals under assumptions about event rates.
Dixon–Coles adjustment
A low-score dependence correction commonly used alongside Poisson scoring models; its parameters must be fitted and evaluated.
Over 2.5 goals
A match total of three or more goals. The half-goal threshold prevents a push on exactly 2.5 goals.
Calibration
Agreement, across many forecasts, between stated probabilities and observed event frequencies. A group predicted at 70% should occur about 70% of the time in a well-calibrated model, allowing for sample uncertainty.
Brier score
The mean squared difference between an event probability and its binary result. Lower is better, evaluated against a suitable baseline and on unseen outcomes.
Log loss
A scoring rule that penalizes assigning low probability to an outcome that occurs. It must be computed using valid forecast probabilities, with special attention to numerical limits.
Backtest
A retrospective evaluation on historical observations. Its credibility depends on avoiding leakage, documenting the model version and separating training from testing.
Prospective track record
A sequence of predictions published or irreversibly recorded before the events they predict, later scored against independently verified outcomes.
Data leakage
Using information in a forecast that would not have been available when the forecast was made. Leakage can make backtests look far better than genuine live performance.
Closing line value (CLV)
The difference between a recorded earlier market price and a market's later closing price, measured with a defined convention. CLV alone does not prove realized profitability.
Expected value (EV)
A weighted average of possible outcomes under assumed probabilities and payoffs. It is only as reliable as the probability, price and cost assumptions used.
Model disagreement
A difference between forecasts for the same event. It can expose uncertainty or competing assumptions; disagreement alone is not a validated signal.
Probability calibration and Brier score
Calibration compares predicted frequencies with observed outcomes. Learn the arithmetic and limitations in our Brier score and calibration guide.