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: alan von weiler who wrote (42)	1/7/1998 7:18:00 PM
From: current trend	Read Replies (1) \| Respond to of 871

Can neural nets reliably predict a couple of days into the future? Wow, that sounds too good to be true! CT

To: alan von weiler who wrote (42)	1/8/1998 2:52:00 PM
From: Optim	Respond to of 871

Alan, Neural nets tend to work better on larger, more liquid stocks. These stocks have the volumes that create the very patterns that a net feeds off of. A smaller penny stock probably won't work as well because there is less 'patterns' to pick up. This is probably why you are achieving such high confidence levels. Because the net hasn't seen a lot of patterns, everything tends to look the same, and it thinks it is working on data that is very similar. This isn't necessarily bad, as those very same patterns may still apply. For example lets say a stock is normally very thinly traded and there is a hugh spike in volume today without the price increasing. The net may compare it to a pattern it found previously where the price moved up 2 or 3 days after the large increase in volume. It would then forecast that the price would move up in the next few days. This is an oversimplified view of the nets, but you may get what I mean. The more discernable patterns that you can feed a net, the better it will be at applying itself to future patterns. The confidence level may decrease, however it may actually do better because the patterns (or non-linear relationships) that it has learned apply better to last few days of history. This is why it isn't necessarily good to have too much history. Patterns that worked in 1993 or 1994 won't always work today, so this might confuse the net and lower your accuracy. But too little history doesn't supply the net with enough patterns to apply itself well to new data. This is why I like to start with a year of data, and never use more than 3 years. Depending on your software you can go as low as 3 or 6 months (ie: Neuroshell Trader) sucessfully. With Neurostock, I find about 1 to 2 years worth of history works best. And you need to slide the windows forward as you add new data. So if you train a net on a period from 01/01/97 to 12/31/97 and then apply it for the month of January, come February you would adjust your training periods to 02/01/97 to 01/31/98 and then apply it to February. A little tedious, but it allows the nets to learn the most recent non-linear relationships amongst your inputs, as these will probably have the most relevance to new data. This is why adding other sources of relevant data (ie. bonds, tbills, etc) can be beneficial. The nets sees patterns that worked recently and have relevance, plus it sees a lot of them to build a robust net. This is where the tough part comes in. Trying to find good secondary inputs (influences) can be difficult, and is probably where you spend most of your time. Optim