|  APR 1991

Using Neural Nets In Market Analysis by Mark B. Fishman, Dean S. Barr and Walter J. Loick

Can neural networks aid traders? By Mark B. Fishman, Dean S. Barr and Walter J. Loick Neural networks are the newest form of artificial intelligence technology to make its way into market analysis. Neural networks, also referred to as neural nets, mimic the human brain's most powerful ability, that of pattern recognition. The neural net can make a decision or a prediction by drawing on an inventory of patterns previously learned, seeking the one most relevant to the current situation. What distinguishes neural nets, moreover, from ""classical"" AI paradigms such as that of rule-based systems is that, in a real sense, t they ""learn"" by example and modify themselves, rather than having to be programmed with specific preconceived rules. Neural nets use layers of real or ""simulated"" electronic neurons with various numeric weights (dynamically modifiable in the ""training"" phase) representing the connection strengths between them to ""learn"" a predictive ability. A neural net, once trained, can be described as a matrix of neurons and connection strengths that represent the ""solution"" to a problem. Commercially available neural net shells, as well as those still being researched, use a variety of algorithms to adjust the connection strengths between the neurons until convergence is achieved and the system is able to make correct predictions with some, usually high, percentage of accuracy. Once the network successfully sets the weights of the connections to produce the correct output, then the trained net is ready to use as a predictive or diagnostic tool. THE FUTURE IS FANCIFUL While this technology may seem fancifully futuristic, some very good and effective neural network software programs are in fact already available to the public at a reasonable cost. We use Brainmaker, created by California Scientific Software. A neural net software package such as Brainmaker should be thought of as a ""shell"" program, similar to an expert system shell, the effectiveness of which is largely dependent on the choice of algorithms by which it ""learns."" What makes this software package a powerful tool for learning relationships is the proprietary back-propagation algorithm used by its designers.

by Mark B. Fishman, Dean S. Barr and Walter J. Loick