How to build an artificial trader.

It's hot, this area of hybrid systems. It means training a computer to do the gritty work, as well as determine trades. But first, we must understand its parts.

Editors note: This is the first of a four-part series on using hybrid systems to build a mechanical artificial trader. This piece provides an overview of each artificial intelligence-based technology and shows how it relates to trading. For more information on these technologies, refer to past articles listed on page 58 ("In the library").

A great human trader has a collection of experiences on which to base his trades. A hybrid system uses multiple technologies to solve problems, which allows us to simulate the human thought process.

To design a hybrid system, several technologies can be used. Here we'll discuss eight. Six are from the field of artificial intelligence: neural networks, genetic algorithms, statistical pattern recognition, expert systems, fuzzy logic and machine learning. The other two are trading related: intermarket analysis (see Futures, August 1994) and general technical analysis and charting. We begin by defining each of these technologies.

Neural networks

Neural networks are good for many trading-related applications, such as 1) predicting the future direction of a commodity, 2) selecting between trading systems, 3) predicting the trend of a commodity, 4) making subjective analysis mechanical and 5) predicting tops and bottoms.

Last month's article developed a neural network to predict the S&P 500 five weeks out. But traded as a stand-alone system, it had large adverse movements. By combining neural networks with other technology, or using multiple neural networks, we are able to reduce adverse movement and still garner unleveraged returns of 15% to 30%.

The most popular neural network algorithm is back propagation, which trains neural networks by repeating presentations of input and output pairs representing the relationships being learned. The learning algorithm operates by adjusting weights between processing elements in the network with the general goal of reducing the average error for the data set. Often, neural networks will create a modeling equation unknown to the user.

Neural networks are not black magic and most people do not know how to use them properly. Consequently, most people fail to use them successfully. To make them work, you need a well-defined problem. You need to choose your output(s) before deciding on your inputs. One of the biggest problems with neural networks is poor generalization: The network does well on training data, but not on unseen data.