CatLady,
When I begin training a stock, I have usually gone through the ritual of checking the fundamentals and selecting relateds through visual and computative correlation. At this point I usually run Backpropagation for 30 minutes. If a reasonable diagonal does not begin to form in the scatter plot ( often within 10 minutes ), I will look for better relateds.
After that, I usually run Backpropagation for several hours ( 8-10? ). I let it run until the net doesn't seem to be changing the profitability total very often. At this point, I switch to Annealing and let it run to the completion of all 4 passes. After that I switch back to Backpropagation and run it for a couple of hours.
In what little I know, I believe you can interrupt Backpropagation training and continue it without any real problem. However, I believe that Annealing should be allowed to finished uninterrupted to the end of all four passes ( otherwise you may be left on a shorter mountain in terms of the previous analogy ).
If you check both Backprop and Annealing, NeuroStock seems to ALWAYS begin with Backpropagation when you click Train. It will progress through all four Annealing passes, then go back to Backprop. I often let it run, stopping it after it has spent a couple hours in the second Backprop cycle. If I find that it has started another Annealing cycle, I always let it finish the 4 passes.
I am curious about the relateds that I have seen everyone using. In general, I use as many industry and related stocks as well as indices such as the volatility index as I can. Maybe it's just the way I used them, but I have found that my models are more stable if I choose more stocks over fewer stocks with multiple time influences.
Vic |
