Training Data & the Low Sample Size Problem

Football is an inherently low-sample-size sport for most individual skills. A striker might take 3 shots per game; over a 30-game season that's 90 shots — barely enough to profile finishing quality. A midfielder's long-range shots might total 15 in a season — statistically meaningless. Training data, particularly from 11v11 sessions (~90 additional minutes per week), dramatically increases sample size for skill profiling. But training data has a validity caveat: undefended training shots are not the same as defended game shots.

Two approaches to the sample size problem: (1) aggregate multiple seasons of match data (3 seasons of Coutinho gives 226 long-range shots — usable but slow); (2) use training-level event data to supplement. When using training data, always filter by validity — only include training situations that approximate game conditions (e.g., 11v11 live play, not unopposed finishing drills). Report sample size prominently in any skill profile; flag profiles under 50 instances as unreliable.

1. Using a single season of data for rare-event metrics: One season of long-range shots is almost always too small. Aggregate multiple seasons or acknowledge the limitation.
2. Treating all training data as equally valid: Unopposed drills, small-sided games, and 11v11 live play have very different validity coefficients.
3. Reporting rates without sample sizes: A 35% long-range conversion from 6 shots is meaningless.