Research · Jun 9, 2026
New machine-learning research on football forecasting
My new football-forecasting study, Dominant Below, Drowning Above, uses five major European leagues from 1993/94 to 2025/26 to study promoted clubs as a domain-shift problem. The main result is that machine-learning models forecast established clubs well, but promoted clubs remain much harder to distinguish before the season begins.
Update
My new football-forecasting research, Dominant Below, Drowning Above, is now available.
Why this matters
Promotion is a clean forecasting problem: newly promoted clubs are the strongest teams from the lower division, but immediately become among the weakest teams in the league above. The study uses this boundary to ask whether a model trained on the whole league can read promoted clubs as well as it reads established clubs.
Main result
The same model performs well for established clubs but much less well for promoted clubs. In the study, survival forecastability is about 0.80 AUROC for established clubs and about 0.57 for promoted clubs, close to the boundary where prediction becomes weak.
Read more
The full interactive research page includes the story, data, model design, forecast page, and playground: fatih.ai/football.
