Neural Nets - A tool for the 90's Message Board

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

Public Reply Prvt Reply Mark as Last Read File

Previous 10 Next 10 Previous Next

To: tulsainvest001 who wrote (71)	4/9/1998 4:07:00 PM
From: Optim	Read Replies (1) of 871

>I did notice on my first attempt at QCOM that the prediction >train 1 yr return" column was much higher than the "train 1 yr eval" >which could indicate overtraining (???). Maybe this will be clearer >in V1.4. Some of the tutorials seem to indicate that it is >especially good if the eval %return is better than the train return, >don't know if that means the inverse as well(eval %return <<train >%return = bad). What you are looking for in any net is for the eval test to be consistantly good. What the eval period means is that a net is trained on a period of time, and then applied to the following period which the net has never seen before. For example, if I train a net on a period from 01/01/95 to 12/31/96 and then 'backtest' it on the period 01/01/97 to 12/31/97 I will be testing the net on a period of data that it has never seen before. If the net fairs well during that out of sample period, it usually means that my inputs are good, and the net was able to generalize well enough to apply itself on data that it never used during training. This indicates that the net stands very good chance of continuing to work in real world testing and should be profitable. A good trading strategy can make the most of a bad net however. The results you see for the eval period in NS Trader are based on a simple buy above zero/sell bellow zero trading strategy. This doesn't include any sort of stops, commisions, or screening criteria. >Do you know how the training works when you add a new stock to a >chart? It looks like it is retraining but is it training the same >inputs on each charts data and keeping a separate prediction for >each or making one net which is applied to each stock? Trader actually creates a new net for each stock. These are all contained in the one file though, which explains why they can get to be pretty big! Keep me posted on how your nets are doing. It's great to share successes and failures with others to get an idea of what works and what doesn't. Maybe we can echange some nets in the future... Optim

Report TOU Violation

Share This Post

Public Reply Prvt Reply Mark as Last Read File

Previous 10 Next 10 Previous Next