Yes, this definitely looks like a valid approach once you know which variables you want to control for! I guess it is a similar task to what was done here with multi-calibration dtai.cs.kuleuven.be/sports/blog/...

3️⃣ Interestingly, using features based on the 3 prior actions improves performance without introducing bias. This is likely because the features are team-agnostic.

2️⃣ Other features in VAEP do not seem to introduce this bias as they do not record any global information about the current possession but only keep track of the last 3 actions.

1️⃣ VAEP uses the current score difference in the game as a feature: this tends to slightly favor actions from stronger teams, but also improves predictive accuracy.

All possession value frameworks describe the game state using slightly different sets of features. Some features can subtly act as a proxy for team strength (e.g., possession length), thus biasing action ratings for players on stronger teams. Here’s what we found 👇