John Cook - Google+ - Applied statistics can be incredibly ad hoc. When I wro…
John Cook - Google+ - Applied statistics can be incredibly ad hoc. When I wro…
If you test your model using data generated by its assumptions, that’s a bare minimal standard: if you don’t do well in that case, hang it up and go home. But you could do robustness tests, modeling your data to have one distribution but testing it with data from another.
If you use real data and randomly divide it into a training set and a validation set, that’s great. But if you simply say “Look how my method did on this real data set (i.e. the whole thing)” then that’s not impressive.
