Sadly, many ML comps involve exactly this. Making a predictor for something, where the prediction has no real world value, and even dubious value at a thought e

Sadly, many ML comps involve exactly this. Making a predictor for something, where the prediction has no real world value, and even dubious value at a thought experiment level. I think one of the biggest skills to have in the ML space, is knowing what is worth training a model to know, and what isn't. Just like in engineering, the most successful products are those that solve a real world problem, no matter how elegantly the others might have been made.