Sports analytics glossary

Twenty-four terms that come up in sports forecasting and model evaluation, each in one or two plain sentences.

Probability and prices

Probability

How likely something is, from 0 (never) to 1 (always). A 0.6 probability means it should happen about six times in ten.

Implied probability

The chance a quoted price claims. For decimal odds it is 1 divided by the price, so odds of 2.50 imply 40%.

Fair odds

The price that matches a probability exactly, with no margin added. A 40% chance has fair odds of 2.50.

Margin (overround)

The amount by which the implied probabilities of all outcomes add up to more than 100%. It is the price-maker's built-in cut.

Two-way and three-way markets

A two-way market has two outcomes (win or lose). A three-way market adds a draw. Three-way markets usually carry a larger margin.

Margin-free probability

An implied probability after the margin is taken out, usually by dividing each one by their total so they add to 100%.

Models and ratings

Model

A set of rules or equations that turns past data into a forecast, such as a win probability or a predicted score.

Rating

One number per team that summarises its strength. The gap between two ratings feeds the forecast for their game.

Elo rating

A rating that moves up after a win and down after a loss, by more when the result was a surprise.

Power ranking

Teams listed from strongest to weakest by rating, rather than by wins and losses alone.

Home advantage

The average lift a team gets from playing at home. Models estimate it from past results and add it to the home side.

Predicted score

A model's expected points or goals for each side. It is an average over many possible games, not a promise about one.

Feature

One input to a model, such as rest days or recent scoring rate.

Overfitting

When a model learns the noise in old data and then forecasts new games worse than a simpler model would.

Checking whether forecasts are any good

Calibration

Whether stated chances match what happens. If things given a 70% chance come true about 70% of the time, the forecasts are well calibrated.

Reliability diagram

A chart of forecast probability against observed frequency. Perfect calibration sits on the diagonal line.

Brier score

The average squared gap between the forecast probability and the result (1 or 0). Lower is better.

Log loss

A score that punishes confident forecasts that turn out wrong much more than cautious ones. Lower is better.

Out-of-sample test

Scoring a model only on games it never saw while it was being built. It is the honest way to judge a forecast.

Sample size

How many games a figure is based on. Small samples swing widely, so a result on 50 games proves much less than one on 5,000.

Favourite-longshot bias

The pattern, seen in many markets, where quoted prices overstate an outsider's chance by more than they overstate a favourite's: the margin sits more heavily on the outsider.

Baseline

A simple forecast used for comparison, such as "the home team wins 55% of the time". A model has to beat it to be worth having.

Confidence interval

A range that shows how uncertain an estimate is. A wide interval means the data cannot pin the number down.

Open data

Data anyone may download and reuse, usually with a licence that asks only for a credit.