Silicon Investor (SI) -- The First Internet Community

STOCKTALK

We've detected that you're using an ad content blocking browser plug-in or feature. Ads provide a critical source of revenue to the continued operation of Silicon Investor. We ask that you disable ad blocking while on Silicon Investor in the best interests of our community. If you are not using an ad blocker but are still receiving this message, make sure your browser's tracking protection is set to the 'standard' level.

Strategies & Market Trends : Neural Nets - A tool for the 90's -- Ignore unavailable to you. Want to Upgrade?

To: current trend who wrote (41)	1/7/1998 7:17:00 PM
From: alan von weiler	Respond to of 871

My understanding is that if you have more neurons, the program will be able to pick up more patterns within the stocks. I don't think speed is the result. To increase speed, you would need a faster computer. Regards, Alan

To: current trend who wrote (41)	1/8/1998 3:10:00 PM
From: Optim	Respond to of 871

>By running more "neurons" are you increasing computing power -- >speed of output, or would your end results be different with >variable numbers of neurons? What limits the number of neurons >besides software design? Generally, the more neurons your neural net software has the more patterns or non-linear relationships it can detect. This is assuming you can feed it enough patterns to utilize all those neurons. It is sort of a diminishing returns theroy. Having a large number of neurons allows your software to adjust itself to detect highly complex relationships in you data. But it also has the effect of 'memorizing' your data, which cause poor results in out of sample (or new) data. Software that is currently on the market usually have some sort of mechanism to miminize the memory effect. You want a neural net to learn complex relationships in your data, while still generalizing enough so as to be able to apply itself well to out of sample tests. For example if you feed a general neural net program (not Neurostock!) raw data such as price, high, low, volume it will usually give a prediction that is very close to what you fed it. This is why you 'pre-process' your data so as to enhance the patterns available in the data. For example if you feed a net the momentum indicator, or percent change in price, you will usually get better predictions. This is because the net 'sees' a 5% change in price instead of a jump from 50 to 52.50 in price. By enhancing the actual change, you let the net see a more discernable pattern that it can apply better in the future. Generaly you will also find that more neurons slow down the training process. Again this is because the patterns the nets are learning are much more complex. Optim