Bob,
The commercial versions have a "Confidence Factor" enabled, and I think a few extra trading strategy's. I've noticed that I start getting better verification period recommendations when the number of neurons gets above 30 or so. Lately, I've been starting networks out with 20 or 30 (sometimes 15) neurons and letting them train for several days with the automatic neuron addition feature checked. Usually after several cycles of Simulated Annealing, the number of neurons will increase. This usually enhances the verify period's accuracy and the confidence factor usually increases. I say "usually" because it doesn't always happen. Sometimes its best just to start over with a new network. If I don't like what I see after a few hours of training a new one, I'll often change an influence factor, etc. and start over. Usually after a few tries, a network that looks good in the verify period will be found and I continue training it for a few days.
I don't know how the Confidence Factor is calculated, but when its high (80% or so), the networks' predictions are often amazingly accurate - it is probably worth the commercial version for that feature alone.
Jay and Len (see other posts) have been using this software longer than I have, but I think they agree (and its what they recommended to me a few months ago) that most of thier accurate networks had 40 or more neurons. You should get their opinions - look back at some of their earlier posts to me and others - most of these same questions have been asked.
The way to see verify period performace is to move your training period (BEFORE TRAINING STARTS) date back a couple of months, and observe the red, green, or blue lines that appear in the verify period. Ideally, you would want green (Buy) lines near the start of up-trends, red (Sell) lines near the start of down-trends, and blue lines (Holds) anytime there is sideways movements or continuations of up or down trends. Remember, the verify period data is not presented to the network during training - thats why its important to set up your training period first.
There has been a lot of discussion lately on the "influence period" settings and exactly what they do for us - lots of questions and no real good answers yet, at least in my opinion. I do know that they have a large affect on how sensitive the networks are to recomendation changes. The best way I've found to pick them is to look at the historical charts of the target, and try to see if there is a time period that seems to encompass major movements, or periodic movements, and I try first setting the influence period just long enough to cover this time-frame. There is no good reason at this time for that, but it seems to work pretty well. Neurostock recommends 20 days for the target, and often that works as well as anything else I've tried.
One thing that I really dislike about this package is that the training set total error cannot be seen - as it can in most neural net software that I've played with. Using the scattergraph (which can also displys the verify period results by checking and unchecking the appropriate boxes) and looking at the verify period are the only means we have of judging the networks accuracy and fitness.
I've only been using this software about 4 months or so, including the demo version you have. Since I bought the commercial version, I've found about three companies that I've gotten really good results from. Actually, I've found quite a few that good networks have been created, but their price hasn't moved very much and I don't know for sure whethor or not the networks will predict it when it does - I'm not going to put any money on them until they prove themselves. As far as the others, I've been basically "Paper Trading" them for about a month and a half now (due to the time of the year and being low on cash!) and I think, as best as my shoddy record keeping is, that I would be ahead by about 30% over the last 6 weeks or so if I'd followed the Neurostock Advice, just on those 3 companies alone.
All in all, I'm quite pleased with the software. Its much easier to use that most of the other demos I've played with due to the automatic download ability. It does NOT give the user much lee-way in deciding what inputs are applied to the network - but if the results are accurate - Who cares?
Anyway, sorry about getting long winded. Maybe Jay and Len or some of the others will share their opinions to you.
Bill |
