In the November 1996 article "On developing trading systems," I demonstrated that neural network technology no longer works well when used in a manner employing brute force. To demonstrate, I trained a neural network to directly predict price change over the next several bars and achieved poor results, as expected. I speculated that the reason neural network technology no longer works well when applied in a crude fashion is because too many traders took the neural net approach and made the markets efficient relative to such techniques. However, if applied in a manner consistent with its capabilities and targeted at unique problems, neural network technology can be highly effective. Let us, then, explore a different approach to training a neural network. This time, I will train neural networks to emulate my subjective recognition of chart patterns.
When looking at charts, I often feel that I can intuitively identify good entry points based on visually recognizable patterns. One pattern I look for is a pullback in a trend, in which the market is clearly moving in a given direction before it undergoes a small correction -- that is, a pullback -- before continuing. Could a neural net be trained to automatically perform the same process of recognition that I believe I can accomplish by eye? Such an application of the technology might be a more appropriate use of neural networks than trying to train a neural net to directly predict the market.
So I decided to conduct an experiment to determine whether a specific pattern in a given market, one that can be visually detected and marked on charts, can be learned and identified by a neural network. I developed two neural networks, each of which was responsible for identifying a specific pattern in market behavior: One was trained to identify a visually recognizable pattern for entry into long positions and the other to recognize a similar pattern for entry into short positions. I intended this approach to take the subjectivity out of visual recognition of chart patterns while still maintaining the human element in a fully mechanized system.
TRAINING THE NETWORK
How do you train a network to emulate subjective judgment? First of all, you need to provide the neural net with examples of that judgment in operation -- training facts. Training facts would consist of all those instances on a chart that I could identify as pullbacks in trends and, for the purpose of contrast, all those I could not so identify.
PROCEDURE
Figure 3 illustrates several of my marks for pullback-in-trend long entry patterns. Although the chart only shows a period of several months, I examined all data from January 3, 1983, through June 27, 1997, marking every instance of a pullback in a long trend that I could visually identify. After performing this marking process, I had a file (click1.dat) containing the dates and times of every bar that I marked. I then initialized RunMode (see Figure 1) to "False" in the Power Editor.
FIGURE 3: PULLBACKS IN TRENDS.
Here's a TradeStation chart illustrating marked pullbacks in trends for the long side of the market. These were generated by eye, not by a neural network.
In contrast to popular sentiment, neural networks do not have to be a waste of time: When used in an appropriate manner, they may indeed be useful components of sophisticated and highly profitable trading systems.
Jeffrey Owen Katz, Ph.D., is a professional trader and consultant in Selden, NY. His firm, Scientific Consultant Services Inc. (516 696-3333), specializes in custom programming of trading systems and tools, provides expert consultation on systems development and the use of Omega Research's tools, and develops publicly available cutting-edge software for traders. Donna L. McCormick is a writer and vice president of Scientific Consultant Services Inc.