First, to say that how a team did last year tells you nothing about how they did this year is just wrong. I don't think you're saying that, but if you are I just don't think that's right and there's a lot of evidence to support me. My guess is that a simple model taking last year's record and using it to predict the next season's record would perform pretty well. It would have a lot of holes, but on average it would probably work.

Second, I believe the model takes into account coaching changes, transfers, etc. I'm not privy to all the details but I know from reading about it that it does more than just naively assume this season is the same as last. So it's actually accounting for the things you say it's not accounting for.

Last, the proof is in the pudding. I went and pulled thousands of their predictions over the last several seasons. I then matched their predictions to the actual game outcomes. And what I found is that the model is well calibrated. It performs well on multiple metrics used to evaluate such models. And among similar models it routinely scores near the top.

You too can go look at the numbers. Don't take my word for it. Others have done similar analysis. FPI performs well. This is just an objective fact.