More on influence settings per support. I sent the following question a while ago and tonight received this reply. I welcome comments or opinions since this response does raise a question or two,
My question is regarding influence settings.
> I've done some testing with different settings for the target and relateds
> and have observed that the prediction varies greatly depending on changes
> to these settings. I've been struggling to come up with a common sense
> approach for the determination of what settings to use to no avail. Would
> you be kind enough to elaborate in greater depth than exists in the help files
> on how to select effective settings. Can they be selected simply by
> visualizing the relationships between the targets and relateds on comparative
> graphs? I'm really lost on this issue, and the group has not been able
> to come up with any viable solutions or theories.
And Support's Answer
--------------------
One of the great powers of neural nets is its ability to make sense out of what seems like a bunch of uncorrelated data (our logo is "making sense out of chaos" for that same reason.) It is difficult to decide how much data to give the net because most of the time you can't see the relationships that NeuroStock could exploit; give it too few data and the accuracy suffers, give it too much and the training will be too slow.
Influence period is simply the number of days that NS will look back
when making a prediction. Say a particular stock peaks two days after
some other stock does the same, if NS is looking at the history of the other stock for the last (for example a influence period of 10) 10 days then it would be able to see it peaking and issue a sell.
NS has influence factors in all its incomming data ports (raw price, raw volume, pre-processed short and long term filters), so NS can pickup most signals if the influence period is long enough.
I start by finding data that is clearly correlated (a leading
correlation is good, lagging is useless) such as market indicators, industry averages, and other stocks of the same industry. Sometimes smaller companies react (or I should say get reacted on) much faster than bigger companies in the same industry providing a few days of lead. The influence period should be about twice as long as the lead time. Tech analysis indicators make very good related stocks.
|
