The performance staff we work with are skeptical of black boxes — for good reason. A model that can't explain itself doesn't get used, no matter how accurate it is in a backtest.
"Predict injury risk" isn't a decision. "Flag when a player's acute:chronic workload ratio suggests reduced practice load this week" is. The more specific the decision, the easier it is to validate the model against real outcomes.
We've shipped models that were 3-4% less accurate than a more complex alternative, specifically because the simpler model could show its work — which feature moved the risk score, and by how much. Adoption rates were dramatically higher.
Every prediction should be checkable against what actually happened, in a format coaching staff can review without pulling in a data scientist. If they can't independently sanity-check the model, they won't bet a lineup decision on it.
A model in front of performance staff in week 4 of the season, with visible iteration, builds more trust than a "perfect" model that ships in the offseason and asks for blind faith.
Applied ML in sports performance isn't a data science problem first — it's a change-management problem that happens to involve data science.
