I think K% and BB% are probably the most predictive, but you can get those anywhere. Seeing the percentile for them is helpful. I have not seen any research indicating that the statcast stats are more predictive than FIP. I have seen studies which have show that the x stats are less predictive than FIP. I have not seen anything about the predictiveness of stats like exit velocity and spin rate, but those were not designed to be predictive. They are certainly interesting as descriptives.