A football model that performed well two seasons ago will usually perform worse today without anything having broken. The sport changes underneath the model faster than most domains do.
Tactical trends move the relationships
Football goes through periods where a particular approach spreads rapidly through a competition, altering how often certain situations arise and what follows from them. A pressing trend changes where turnovers happen and therefore what a turnover is worth.
A model trained before such a shift has learned relationships from a game that is no longer being played, even though the data format is unchanged.
The degradation is gradual and invisible in the input data, which is what makes it dangerous.
Rule and format changes are abrupt
Adjustments to substitution allowances, handball interpretation or timekeeping change behaviour immediately across an entire competition. There is no transition period during which a model can adapt gradually.
Historical data collected under previous rules describes a different set of incentives, and blending it with current data blurs both.
Practitioners handle this by weighting recent seasons more heavily or by cutting the training window at the change, both of which reduce the data available.
Squad turnover breaks player-level continuity
Player models depend on individuals whose abilities change with age, injury and role, and squads are substantially rebuilt over a few seasons.
Estimates learned about a player two years ago may describe someone in a different physical state playing a different position.
This is a form of drift that no amount of retraining on team-level data addresses, because the underlying subject has changed.
Competition between models accelerates the decay
Where models inform recruitment, the clubs using them compete for the same undervalued players, and that competition removes the mispricing the model identified. The model is accurate about a condition it helped to end.
An approach that works becomes widely adopted and stops working, which is a dynamic familiar from any market with informed participants.
Football's analytical edge is therefore temporary by construction rather than by poor implementation.
Monitoring matters more than initial accuracy
The practical defence is measuring performance continuously against fresh outcomes rather than validating once at deployment. A model that is checked only when it is built has no way of reporting its own decline.
Tracking the distribution of inputs also gives early warning, since a shift in how often situations occur usually precedes a fall in accuracy.
Organisations that treat a model as a finished object rather than as something requiring maintenance discover the decay through a bad decision instead of through a chart.