|
PureBytes Links
Trading Reference Links
|
> I recently spent thousands on purchasing neural network & intermarket > analysis software from a company in Florida but I can't seem to get any > benefit from them. They appear to be nothing more than glorified lagging > indicators. > > Is there anyone out there who has had experience with such programs ? > Could you perhaps enlighten me on how to obtain some benefit from the > supposedly "predictive" capabilities of neural networks and intermarket > analysis. Any advice will be very much appreciated.
I have some experience with that "company in Florida" and that particular software package. Suffice it to say that after some initial success that faded quickly, my experience with it mirrors your own. The concept however, now that's a different story:
My trading vehicle for the S&P is based solely on a "home-made" neural net, "Trendy." To quote from the Trendy web site:
Trendy's Neural Nets
Trendy's Neural Net Engine actually produces five distinct, custom designed nets each evening, although only one, NeuralTrend, is at present used in the trading algorithms.
The five nets are: 1). Coefficient of NeuralTrend [Ntrend], the overall market rise or fall projected out 2 days; 2). ShortTerm Adaptive Projection [SAP], a 5-day Adaptive Moving Average projected out 2 days; 3). MediumTerm Adaptive Projection [MAP], a 10-day Adaptive Moving Average projected out 4 days; 4). Range Projection High [RangeH], a projection of the next day's high; and 5). Range Projection Low [RangeL], a projection of the next day's low.
Trendy's neural nets were designed and trained using the NeuroShell 2 software package from Ward Systems Group, Inc. The nets use financial data from a dozen domestic and international markets as their daily input. The data is heavily pre-processed by a free-standing file utility program before it is presented to the nets. Some of the pre-processing elements include acceleration components, change components, lagged acceleration components, lagged change components, adaptive moving average deltas, etc. When the data is finally presented to the neural nets, it is in fact devoid of all absolute price values, containing instead hundreds of relative measures and comparisons. Several columns of categorical data are also a part of the pre-processed neural net input, such as day of the week, date relative to the end of the calendar month, and date relative to contract expiration.
The nets were constructed using General Regression Neural Networks [GRNN] architecture, and were trained using Genetic Adaptive Calibration and a breeding pool size of 300. Training durations were quite long. The Training Set consisted of more than four-years worth of data, which ended on the last day of June, 1996. The Trendy S&P Position Model is up over 300 S&P points in the past 12 months of "out-of-sample" trading.
The failed experience with the "Florida company" and their alleged neural net led to the development of our own neural net which has formed the basis of a robust and profitable trading system. So this sounds like an ad, you say? So be it. Damn it, this is the first thing that has worked for me in the S&Ps in 5 years of futures trading and another 5 before that in the OEX. I don't mind telling folks about. And if you think ads are banned in this forum, let me suggest that no matter what the content of a post, if the author provides his web site address or business address in his signature, my friends, you are reading an ad. We are all grown-ups here, we can assess the value of products and services without the protection of big brother, whether in the form of government, do-gooders, moderators, or whatever. There are some quality folks here offering excellent services and the reason we all know that is that they have been "advertising" in this forum from day one.
In any case, neural nets exist that are effective in trading the S&P. I'll leave it each of you to figure out where to look for at least one of them........
Allan P. Harris http://www.TrendySystems.com
|