V.10:4 (160-168): Developing Neural Network Forecasters For Trading by Jeffrey Owen Katz, Ph.D.

Developing Neural Network Forecasters For Trading by Jeffrey Owen Katz, Ph.D.

Neural networks, for many, still exist only in the realm of theory, without having real, practical, everyday uses — yet. But neural network applications need not be confined solely to theory and simulation for trading purposes. Ambitious traders can build neural nets for themselves. There are pitfalls along the way, but Jeffrey Owen Katz explains what they are and how to avoid them.

For many traders, neural networks exist only as fantasy, surrounded by an aura of mystery, unknown and unexplainable. For some, the term "neural networks" is more or less a recognized one, but the words do not mean much. A few traders have purchased neural network toolkits, thinking they can immediately begin to develop their own neural trading systems, but they have not been particularly successful. Let's shed some light on neural networks themselves by looking at a successful neural network forecaster and then seeing how to develop and train neural net forecasters.

The concept of neural network forecasting systems interests traders because neural nets appear to possess the potential to fulfill trading fantasies. The neural network that produced the output in Figure 1 signaled four out of five tradeable bottoms in the Standard & Poor's 500, almost as if it were a "perfect" oversold/not-oversold oscillator; this is especially impressive, considering the data depicted were collected almost a year after the network had last been trained. The network applied what it had previously learned to the new data and then produced forecasts that would have allowed a trader to enter long positions with excellent timing in most instances.

In addition, neural networks compel the interest of traders for a number of other reasons. First, neural nets can cope with "fuzzy" patterns (those easily recognized by eye but difficult to define using precise rules—for example, the head and shoulders formation) and deal with probability estimates in uncertain situations; second, neural nets are able to integrate large amounts of information without becoming stifled by detail; third, neural nets can "learn" from experience (the design of such systems is not dependent on having an already-expert trader on whom to base the rules of an "expert system"); fourth, neural nets may be retrained and thereby adapt to changing market behavior; and finally, under the correct circumstances, neural nets, with proper training, would be able to recognize almost any pattern that might exist in any market.