Glossary

Expected goals measure the quality of chances rather than the goals that happened to go in. Every shot gets a probability of scoring from its position, angle, assist type and body part; adding up those probabilities over a match gives each team's xG. A team with 2.1 xG created chances worth about two goals, whatever the final score.

xG is a better guide to future results than goals because it strips out finishing luck and goalkeeping variance. Over a season a team's goal difference tends to move towards its xG difference.

On this site, the goal expectations on a match page are the model's own attack-and-defence estimates, not shot-based xG. The methodology page explains the difference.

An xG table re-orders a league by expected goals for and against instead of goals scored and conceded. It shows which teams are over- or under-performing their chance quality: a side sitting 3rd with the 8th-best xG difference is riding good finishing or a goalkeeper in form, and is likely to drift back.

Read it together with the real table. Big gaps between actual and expected goal difference usually close over the next ten to fifteen matches.

Expected points convert a match's probabilities into a points figure: win probability × 3 plus draw probability × 1. A team with a 50% win chance and a 25% draw chance has 1.75 expected points from that fixture.

Summed over a season, xPTS says how many points a team's performances merited. A club with 20 points and 15 xPTS has been fortunate; one with 15 points and 20 xPTS has been playing better than its position suggests.

ELO is a rating that moves after every match: the winner takes points from the loser, and the amount depends on how surprising the result was. Beating a much stronger side moves the ratings a lot; beating a weaker one barely moves them. The rating difference between two teams maps directly to a win probability.

ELO is simple and robust, which is why it is the usual baseline for football models. Our model uses a team-strength rating in the same spirit as one of its inputs; the methodology page lists them all.

Dixon-Coles (1997) is a statistical model that gives every team an attack rating and a defence rating, plus a home advantage term. The expected goals of a fixture follow from those ratings, and a Poisson distribution turns them into probabilities for every scoreline.

Its two refinements over a plain Poisson model are a correction for low-scoring draws (0–0 and 1–1 happen more often than independence predicts) and time decay, so recent matches weigh more than old ones.

This is the model behind the 1X2 probabilities on this site. The ratings are refitted as results come in.

The Poisson distribution describes how many times a rare event happens in a fixed window when it has a known average rate. Goals fit it well: if a team is expected to score 1.5 goals, Poisson gives the probability of exactly 0, 1, 2, 3 or more goals.

Combining the two teams' goal distributions gives a probability for every scoreline, and adding scorelines up gives the home win, draw and away win probabilities, the over/under 2.5 line and both teams to score.

A probability says how often an outcome happens across many similar matches, not what will happen in this one. A 60% pick is expected to lose four times in ten; a run of five losses in a row is normal, not a sign that the model is broken.

The useful question is not whether a single pick came in but whether picks published at 60% land about 60% of the time over hundreds of matches. That is what calibration and the Brier score measure.

1X2 is the full-time result market: 1 for a home win, X for a draw, 2 for an away win. The three probabilities on a match page add up to 100%, and the model's pick is the most likely of the three.

Draws are the hardest outcome to call; a model rarely makes a draw its pick even when the draw probability is close to 30%.

The over/under 2.5 line asks whether a match will have three or more goals (over) or two or fewer (under). It is scored on the final score and does not depend on who wins.

The probability comes straight from the scoreline distribution. Leagues differ: Serie A sits below the big-league average, the Bundesliga above it, and the match page shows how the two teams' recent matches have gone against the line.

Both teams to score asks whether each side scores at least once. A 2–1 or 1–1 is a yes; a 2–0 or 0–0 is a no. It is settled from the final score.

The probability is the share of scorelines in which neither team is kept to zero. It rises when both attacks are rated well and both defences poorly.

The Brier score measures how good a set of probabilities was, not just whether the pick won. For each match the published probabilities are compared with the outcome (1 for what happened, 0 for the rest) and the squared differences are added; the Brier score is the average over all matches.

0 is perfect, 0.667 is a uniform one-third guess across three outcomes, and a typical well-fitted football model lands around 0.60. Lower is better. It is shown on the performance page and on every league's track record.

A model is calibrated when its probabilities match reality: picks published at 55% should land about 55% of the time, those at 70% about 70%. The calibration chart on the performance page groups settled picks by published confidence and plots the observed hit rate against it.

Points above the diagonal mean the model was too modest, points below mean it was overconfident. Our confidence figures pass through a calibration curve fitted on the settled record, so the shown confidence is the calibrated one.

A prediction algorithm turns past results into probabilities for future matches in four steps. First it rates every team's attack and defence from its results, weighting recent matches more. Second it combines the two teams' ratings with home advantage into an expected number of goals for each side. Third it turns those expectations into a probability for every scoreline, usually with a Poisson distribution. Fourth it sums the scorelines into the markets: 1X2, over/under 2.5 and both teams to score.

What makes the output trustworthy is not the model itself but the record: every probability stored with a timestamp before kick-off, settled automatically, and measured with hit rate, Brier score and calibration. That record is public on this site.