Football produces too few examples of many important situations for a model to learn them reliably. Generating additional examples is now a common response, and it carries a specific set of risks.
Rare situations are the ones that matter
Penalties, red card scenarios and specific set-piece configurations occur infrequently enough that a season provides very few examples of each. Models trained on them are working with thin evidence.
Waiting for more data is not an option, because the underlying game changes faster than the examples accumulate.
Synthetic generation produces additional plausible instances so that a model has enough to learn a pattern from.
Generation methods differ in what they assume
Simulation-based approaches build sequences from a physical model of movement, encoding assumptions about acceleration, reaction time and decision-making directly.
Learned generators instead train on real sequences and produce new ones resembling them, which requires enough real data to learn from in the first place.
Both approaches can only extend patterns present in their inputs, so neither invents behaviour the game has not shown.
The failure mode is confident circularity
A model trained partly on generated data can learn the generator's assumptions and present them back as findings. The conclusion then appears supported by a large dataset that was mostly manufactured.
Detecting this requires holding back genuine data for evaluation and never allowing synthetic examples into the test set.
That discipline is easy to state and easy to lose in a pipeline where the two sources have been combined upstream.
Privacy is the other motivation
Physiological and positional data about identifiable players carries restrictions that make sharing difficult between clubs and with researchers.
Synthetic datasets preserving statistical structure without corresponding to real individuals allow methods to be developed and compared openly.
Whether such data is genuinely non-identifiable requires care, since rare movement patterns can be distinctive enough to point back at a person.
The honest use is augmentation
Generated data works best as a supplement that helps a model handle situations it would otherwise see too rarely, with real data still anchoring the conclusions.
Used that way it improves robustness without adding information, which is a modest claim and an accurate one.
Treating it as a substitute for observation is where the approach goes wrong, and the resulting models tend to look strongest exactly where they have been tested least.