How soccer luck is measured, and how well it holds up

The data

2018 and 2022 are built from every shot of every match with its expected goals (xG): the chance that a shot from that spot, in that way, goes in. The expected goals are StatsBomb’s, from their open data, which has every shot of whole tournaments and of one full league season: the World Cups of 2018 and 2022, the Euros of 2020 and 2024, and the 2015-16 Premier League and La Liga. They rate the chance, not the finisher. The open data has no current seasons, which is why the current ones on this site come from another source.

The model’s few fitted numbers (below) were fitted and tested on the same data: 1,517 league matches (the 2015-16 Premier League, La Liga, Serie A and Ligue 1) and 230 tournament matches. So the shot-based pages are measured on matches the model was tuned on; with so few numbers fitted that matters little, but the tests below are not out-of-sample for these seasons, and the page for a new season would be.

2026 is built from openfootball: fixtures, results and, for finished matches, who scored and when. It is in the public domain (CC0) and has no shots, so there is no expected goals and no finishing luck. Volunteers keep its files up to date, so the newest matchdays can take days or weeks to arrive; each page says how far the results run.

The replay with the same day’s form

Each team’s shots of the 90 minutes are resampled: as many as a Poisson draw around the number it took, each drawn from its own shots, each a goal at its expected-goals chance. Own goals come at the league’s rate (0.047 a team a match). 4,000 replays a match.

Score effects. A team that leads sits back and gives up chances it wouldn’t at level; a team that trails pushes. So a shot taken while protecting a lead stands for more of a team’s form and one taken while chasing for less: each counts 0.67 at −2, 0.82 at −1, 1.22 at +1, 1.49 at +2. Measured directly, the effect is confounded (better teams lead more often), so its size is the one that best predicts the rest of a season: after ten matches, r² 0.471 without it, 0.484 with it.

Calibration. The replay’s win, draw and loss chances match how matches ended: ranked probability score 0.150 against 0.228 for the league’s base rates. Holding out each league in turn, the best temperature on the other three is between 0.94 and 1.02, so the replay is used as it comes. Draws are a little underpredicted when one side dominates the chances (22% predicted, 27% drawn), as in every model with independent goals.

Luck shares

The narrowest view keeps every shot and redraws only its outcome. Should all of a team’s finishing above its expected goals count as luck? Some players finish better than the chances they get, but at team level it repeats only weakly. Coin-flip points with 75% of finishing counted as luck predict the rest of a season best after five and ten matches (half edges it after 19): 0.301 after 5, 0.477 after 10, 0.478 after 19, against 0.291, 0.453, 0.469 at 100%. Own goals are luck in full.

The adjust-the-assumptions panel

On each shot-based league match page the panel changes how much of finishing counts as luck, and flips a chance, then reruns the coin-flip replay of the 90 minutes in your browser (the same arithmetic as the engine’s, checked against it). The replay with the same day’s form doesn’t move: it resamples the shots as they were taken, so what you count as luck doesn’t enter it.

Goals alone

The replay. Without shots a match can only be replayed from what its two sides’ goals say. Each club is rated, by ridge regression, on the goal difference of every match of the tournament (a match counts at most 3 goals) and on the goals in its matches; the home edge and the scoring level are read off the tournament itself. The two ratings give each side a Poisson number of goals, with draws made a little likelier than two independent counts say, and the match is replayed 4,000 times from that. A knockout tie level after 90 minutes plays 30 more at a third of the scoring rate, then penalties, a coin flip. Because the ratings are built from the same tournament’s goals, the match itself is part of what rated its teams: the replay is what the goals say, which is a different thing from what we expected beforehand (that is the prediction).

What luck means here. A side’s luck is the gap between the points it has and the points its goals earn (its goal points): it shows up as winning one-goal games, losing heavy defeats, a late winner. Own goals and penalty shootouts are luck as everywhere. It cannot show whether a side made its goals, only whether its goals turned into points. A match’s luck index, lucky or expected, is how far the result was from what the goals say: a win the goals gave 30% scores +70.

Does it work? Not as well as with shots, and we say so. Each measure over a club’s first matchdays against its points a match over the rest of that season (r², 580 club-seasons: ten seasons of the Premier League, La Liga and the Bundesliga):

The points a side’s goals earn predict the rest of a season about as well as the points it has, and a little worse than plain goal difference. So from goals alone, adjusting the table shows how the points were won; it is not a better forecast than the real table, and in the first ten matchdays the gaps are mostly noise. Shots are what make luck-adjusting pay: the seasons built from them (2018 and 2022) predict far better than the real table.

Knockout ties

A tournament’s group matches are replayed like league matches. A knockout tie replayed level after 90 minutes goes to extra time (a third of the 90 minutes’ scoring) and then to penalties, a coin flip; the number on the page is how often each side goes through. Its luck index is measured against who went through.

Win probability

The line on each match page is the chance of winning from the score and the minutes left, with goals for the rest of the match at the league’s rates (1.51 home and 1.13 away a match; the average at a tournament’s neutral grounds), a draw counted half. For the seasons built from goals alone the rates are that season’s own, and a match whose scorers and minutes our source does not list has no line.

Validation

For the seasons built from shots. What predicts the rest of a season: each measure over a team’s first matches against its points per match over the rest (r²), the four leagues pooled.

Calibration of the replay by bin (home win chance against home wins, league matches):

Known limits

• No betting line, so no “before kickoff” or “same teams, fresh game” number from a line.
• From goals alone we cannot see chances: a side that won 1–0 on one shot looks the same as one that won it by dominating. The shot-based seasons can tell them apart.
• Our goals source is kept by volunteers. When it lags, the pages show the matches it has, and say how far they run; the matches in between appear when it catches up.
• Red cards change a match, and the shot replay keeps the shots they produced but doesn’t replay the card itself.
• The shot model is fitted on 2015-16 seasons and four tournaments, the same matches shown on the shot-based pages; the game has changed a little since (more shots from closer in), which the expected goals carry, the shares less so.