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 : NeuroStock -- Ignore unavailable to you. Want to Upgrade?

To: Bill Scoggin who wrote (341)	11/4/1998 1:50:00 AM
From: Jay Hartzok	Respond to of 805

When you do this, do you normally start with an entirely new network, No. I save the net with the verify period under a different name {for example: IOMv.neu becomes IOM.neu}, then I remove the verify period and train the remaining data completely, back prop and annealing. This usually takes another 3 or 4 hours. Then I have two fully trained nets. Just removing the verify period with no additional training would do nothing, because the extra data would not be trained. Also, remember that Neuro never trains the last ten days of data, so you have another verify period in there too. Two more things about current nets: First, I've noticed that in order to get an accurate verify period on most nets, the net must see data showing the market starting to decline. Therefore, I have been setting my training end date after the start of the market decline at 8/3/98. I will move this date ahead as we move further into November allowing the net to see more of the decline. Second, I've noticed that, because Neuro bases all of its daily predictions on data from previous days, most scattered dots in the stats window are usually from the beginning of the training period, simply because Neuro doesn't have the previous data to look at to project accurately for that period. In order to clean the net of these imperfect predictions, I have taken to removing the first few months of data from the net usually ending up with a training start date of 1/2/96 or 2/1/96. I do this after the net has been trained for some time. The net has to see that early data for awhile to establish the predictions for January and February. Once they are established it doesn't seem to need to see that data anymore. Since Neurostock doesn't allow you to see the training set error, theres' not much way of knowing other than observing the verification period colors. Thats really all you need to see. As long as the predictions in verify period match those on the fully trained net, you should be in good shape. These matches don't have to be exact, but should be very close. Jay