Original research · Published
Luck vs skill: how long a track record has to be
We split 49,615 league matches (six leagues, 2005-06 to 2025-26) into what persists and what does not. The part of a result explained by underlying performance carries over to the next season. The part that is not explained almost entirely disappears.
Do teams that win more points than their goal difference suggests keep doing it?
No. The luckiest 10% of teams (199 team-seasons) won 8.1 points more than their goal difference predicted. The next season they averaged +1.0, so about 88% of the surplus disappeared. The unluckiest 10% went from −7.2 to −1.0.
Does last season's points total predict this season's?
Yes, substantially. Across 1,997 clubs that stayed in the same league, points per game correlated 0.70 from one season to the next (95% interval 0.67 to 0.72). Goal difference, the underlying measure, persisted slightly more (0.74). 62% of top-quarter clubs stayed in the top quarter.
How much of the table is luck, by number of games
| Games played | Share of the spread that is luck |
|---|---|
| 5 | 72% |
| 10 | 56% |
| 19 | 41% |
| 38 | 26% |
| 76 (two seasons) | 15% |
Early-season form is a weak guide to the rest of the season: points per game after 5 games explained 25% of the variation in points per game over the remaining matches; after 19 games, 39%.
By league
| League | Team-seasons | Luck share (full season) | Season-to-season correlation |
|---|---|---|---|
| Premier League | 420 | 22% | 0.74 |
| La Liga | 420 | 24% | 0.80 |
| Serie A | 420 | 22% | 0.77 |
| Bundesliga | 378 | 28% | 0.63 |
| Ligue 1 | 414 | 29% | 0.64 |
| Championship | 504 | 36% | 0.33 |
The Championship has the most luck and the least persistence: its clubs are closer in strength, so random outcomes move the table more.
Method
Luck share. For each league-season we compare the observed variance of points per game with the variance that would arise if every match result were a random draw from that league-season's points distribution (Tango's method). The remainder is attributed to differences in team strength. By-games figures rescale the luck variance by the number of matches.
Points beyond goal difference. We regress points per game on goal difference per game across all team-seasons (about 0.62 points per goal) and treat the residual, scaled to 38 games, as points not explained by underlying performance. Its season-to-season correlation is 0.13 (95% interval 0.09 to 0.18).
Limitations. Persistence uses only clubs that stayed in the same division, which excludes relegated and promoted clubs. The luck model treats matches as independent and counts fixture-order and opponent effects as luck. Goal difference is itself partly luck, so these figures understate the total luck in results. Bootstrap intervals resample team-season pairs, which are not fully independent.
Data: full-time results from football-data.co.uk. Analysis: Football Edge Lab, reproducible with python -m epl.luck_skill in the project repository. Cite this page when using the figures.