Dave,
I'm reading through old posts and something you said back here brings up a question.
"As for a verification period filtering out noise, I think this is just another way of describing the ability to generalize. That is, you don't really want your net to follow every little bump, but rather, to catch the main trend of your timescale of interest as it changes direction. "
I think I've seen several nets that achieved their best verification fit after just a few of hours training, but the training data fit kept improving for many more hours, while the verification fit got worse. To me, it seems that it is possible to overtrain a net despite what the documentation says on that point.
I'm using about 3 years of training data and 4 to 6 months of verification data with my nets, sometimes more, sometimes less. If it looks good at that setting, I'll retrain it with a shorter verification set.
I agree with your earlier statements, in that I don't want to rely on a net's signals unless the verification fit is sound, regardless of how the training set appears.
How long have you trained your successful nets? How do you decide when you've trained them long enough?
CL |
