Injury prediction fails because injuries are not one thing

Models get built to forecast a category that lumps together a torn hamstring, a rolled ankle and a broken hand, and then they underperform.

The task is usually written down like this: predict whether a player will be injured in the next seven days. It sounds like a well-formed question. It is not, and the way it is framed guarantees the model will disappoint whoever paid for it.

The outcome variable is the first problem. A hamstring strain from a sprint, an ankle rolled in a tackle, a stress reaction in a foot and a fractured hand from falling all get the same label. Their causes have almost nothing in common. One is a tissue overload with a long buildup, one is a random collision, one is a slow accumulation of bone stress, one is bad luck on a wet surface. Ask a model to predict all of them at once from training load and you have asked it to find a single pattern behind four unrelated processes.

The second problem is that the events are rare. A squad might record a few dozen incidents in a season across thirty players and hundreds of sessions. Machine learning is not good at rare outcomes with small samples and dozens of candidate variables, and it fails in a specific, seductive way: it finds structure in that particular season that does not exist in the next one. The model looks excellent on the data it was built with and mediocre the moment it meets a new year.

Then there is the intervention paradox, which I find the most interesting part. If the model works and staff act on its warnings, the predicted injuries do not happen, and the model now looks wrong. Its success destroys its own evidence. Nobody has a clean way of scoring a system whose whole purpose is to falsify itself.

None of this means quit. It means narrow the question. A model built only for non-contact hamstring injuries in sprinters, using sprint exposure, prior hamstring history and time since last maximal effort, is a tractable problem with a coherent mechanism behind it.

Contact injuries should be dropped from the outcome entirely. You cannot forecast somebody else's studs.

Small questions with mechanisms behind them beat large questions with more data. That is true well beyond sport, and sport keeps proving it the hard way.