[Shazam]

Originally Posted by Gordies Elbow

The key point you make is around "any accurate Bayesian model" - I'm not sure that these are those.

In mathematics, you need to diagnose your model, evaluating the basic assumptions and comparing model approaches, and prepare the model for a proper defense.

The primary arguments when defending are around assumptions. Some of the assumptions added to these models are subjective (e.g. what constitutes a turnover, what is a proper zone entry.) Some make assumptions about specificity (e.g. shot location makes a shot harder or easier to save.) These may seem obvious, but are they statistically significant given the small numbers of minutes of performance?

In short, models can be descriptive (e.g. relative performance of a machine.) or predictive (e.g. economic forecast.) Descriptive models can be tested against current performance, predictive against future events. In order to be validated, they should be tested. Are any of these?

I (unfortunately) somewhat agree with Brian Burke, where he said “Analytics are like a lamppost to a drunk. They’re good for support, but not for illumination.”

These advanced stats aren’t models. There’s no predictive analysis. People could even take past data and model it forward and match it against past real data but I’ve never seen that done.

I’ve seen some terribly naive analysis done (mostly by the Nation) where they attempt to find things out like when players start to decline but it was done all wrong.