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: Len Giammetta who wrote (212)	10/9/1998 1:48:00 PM
From: Optim	Read Replies (1) \| Respond to of 805

What is a "usual" relationship? Wouldn't the net learn from this "failure" and be better able to predict when it was able to relate this pattern the next time it appeared? Also, what signal do you think Neural would give if this happened... I think neutral. At the risk of over simplifying, let say bonds and stocks are highly correlated. When one goes up the other goes up as well. This may be correct historically, but some external influence may come in and cause this relationship to reverse. If the net has been trained on data that had the two moving in tandem, when the relationship changes that model may get stuck in a long or short position, thinking that one of the issues will revert to its previous relationship. This is a very simplified version, but as you introduce more elements into a model, you have more points of failure, or degrees of freedom. This applies to statistics as well, such as a simple linear regression. Do you have any theories on appropriate influence settings. Do you think it's possible to look at a historical chart of the target and the related and know from that observation which influence settings are best, or would you suggest simple experimentation until we get an acceptible accuracy? Ahhh! This has been my main problem with NeuroStock, and the reason I tried other neural network packages. Because NeuroStock doesn't tell you what it feeds a model such as the number of inputs, hidden layers, or how well a given influence is working, it is pretty much a trial and error issue. I try to keep my networks as small as possible. This is the inverse to what the help file tells you, but it tends to work well when trying out of sample. For example, I might force NS to use only 20 neurons, and limit my inputs to the issue and a related index, plus one major index. This may not capture all the moves in the market, but it will be consistent in its results. I haven't tried adjusting the influence settings according to the underlying dominant cycle of the issue, but it might be something worth trying. I don't know if NS adjusts its lookback automatically, or if a 20 day lookback is fixed. If it is fixed some experimentation would be need to find an optimal lookback. A great feature to add to NS would be the ability to show the profit and trading results for the verification period. This would help you to determine if the net had memorized or curve-fit the training set. It would also be great if NS showed which inputs contributed to the model most. I also think that I much more in depth help file would be the greatest improvement he could make. Optim