A method I've used with some success is to take my d/b (14 yrs), split it in 2 parts, develop the idea on one lump and then run test the produced system on the other. That avoids testing a system on the data that was used to develop it.