An injury model that warns you every week is not a warning

Predictive telemetry keeps promising to see the tear coming, but the base rates turn almost every alert into a false alarm.

The pitch is always the same. Sensors on the limbs, a model on the laptop, a red flag next to a name on Thursday, a hamstring saved on Saturday. I want it to be true. The arithmetic makes it hard.

Soft tissue injuries are rare events at the level of the individual session. A given player in a given week is overwhelmingly likely to be fine. When you build a classifier for a rare event and it is even moderately imprecise, most of the names it flags will be players who were never going to get hurt. Push the sensitivity up so you catch the real ones and the alert list grows until it includes half the squad, at which point the manager stops reading it. Push it down so the list is short and credible, and you miss the injuries, which is the only thing you were built to prevent.

That is not a flaw in any particular vendor. It is what happens when you predict something uncommon. Medicine has fought this argument for decades over screening programmes and largely lost it in public, because the emotional pull of catching one case is stronger than the statistical cost of a hundred unnecessary interventions.

There is a second problem, subtler. The intervention contaminates the evidence. If the model flags a player and the staff rest him and he does not get injured, you cannot tell whether the model was right or whether he was always going to be fine. Every successful prevention destroys its own proof. Clubs end up believing in the system through a run of quiet months and abandoning it after two bad ones, which is faith, not evaluation.

So what is the telemetry actually good for? Description. It is excellent at telling you what a player did, which is not nothing. If a midfielder's high-speed output in the final twenty minutes has fallen off across six weeks, that is a real observation about a real body, and a coach can act on it without pretending it is a prophecy. Monitoring change in an individual against his own baseline is defensible. Ranking twenty-five players against a population model and calling the top three at risk is mostly theatre.

I would keep the sensors. I would sack the traffic lights.

The honest version of the technology sounds much less impressive in a boardroom, which is exactly why the dishonest version keeps getting bought.