If you go talk to modellers (even in fields other than climate science) they can explain how they generally manage to avoid the trap of over-fitting to historical observations (including tests that include hindcasting over only a somewhat arbitrary subsequence of the available data).