> if you make it complex enough, you can engineer quite a few models that can predict natural phenomenon pretty well.

Well, yes. Next thing is to pick the easiest of these models: Occam's razor. But pondering whether such model is "correct" won't bring any further improvements.