Three Neural Net Evaluations
by John Kean
As interest in using neural nets for trading the markets has grown, so has interest in the products being
offered. We will discuss three of these: @ Brain, a new entry from Talon Development, and Brainmaker
Standard and Brainmaker Professional, two veteran products from California Scientific Software.
For those unfamiliar with this type of software, it's a form of artificial intelligence that, rather than being
taught what to do by human beings, teaches itself. Neural networks are composed of layers of processors
(neurons) connected to each other by variably weighted links. The neural networks with the greatest
application in finance and market prediction typically have three types of layers: First, an input layer
through which data has been selected and processed by the user; second, one or more hidden layers in
between that are connected to both the input layer and the third layer, which is called the output layer. It
is at the output layer that the user receives the estimates generated by the program.
In training a net for market predictions, usually the user builds a file of historical data that he or she
believes can be used to predict a change in stocks, bonds, gold or whatever. Each line of the file contains
historical information along with a subsequent change that has occurred. The number of lines depends on
how many time periods (days, weeks, months) it is designed to train over. Neural nets best learn the
general characteristics of a problem by looking at many examples. As the neural net program trains itself,
it seeks to minimize the number of errors between its estimates (outputs) and what actually occurs
(patterns). After the user has trained and tested a net he's satisfied with, he then feeds in current inputs
and the net produces current outputs (predictions).