A fitness score built from your own history cannot predict a new block

These models are curve fits over what you already did, which is why they break the moment you do something different.

The endurance apps all sell some version of the same object. A single number that rises when you train well, plus a projection of what you could race. It is a satisfying thing to look at and I understand the appeal. I also think most athletes misunderstand what the number is doing, and the misunderstanding costs them races.

The number is a fit. It takes your recent training, weights it by recency and intensity, and produces a curve describing your past. Within the range of things you have already done, it interpolates well. That is not a small achievement and it is genuinely useful for spotting a bad month.

The trouble starts at the edges. Ask it about a stimulus you have never given it and it does not say that it does not know. It extrapolates, confidently, from a shape it learned from a different kind of training. An athlete who has spent six months doing steady volume gets a five kilometre projection that assumes a speed reserve they have not built. The projection is not a lie. It is an answer to a question the data cannot support.

There is a second failure, worse because it is invisible. These models generally treat all load as one currency. An hour is an hour, adjusted for heart rate or power. But the body does not keep one adaptation account. Tendon stiffness, aerobic enzyme density, red cell mass, running economy and neuromuscular coordination all have different time constants, and some of them are barely reflected in the inputs at all. Economy in particular can change substantially with no change to the model's inputs whatsoever, because the athlete simply got better at running while doing the same work.

So the score can sit flat through a period where the athlete improved a lot, and rise through a period where they were quietly digging a hole.

My rule is boring. Treat the number as a compliance check, not a prediction. It tells you whether you did the training. Then predict races from races, from time trials, from the sessions you know correlate for this athlete. That is a small dataset and an honest one.

The moment an athlete starts training to raise the score, the score has stopped measuring anything.