Hecht-Neilsen, American Express Test Neural Model For Forex Trades

Can a neurocomputer help traders predict short-term movements in foreign exchange prices? HNC Inc., San Diego-based supplier of virtual neural networks, and American Express Corp., well known collector of investment banks, decided to find out. Hecht-Neilsen claims to have successfully demonstrated feature extraction of foreign currency trading data.

"It's proprietary and we can't discuss it," is an all too familiar refrain to those who follow the development of artificially intelligent systems for trading. The closed-door policy is frustrating for journalists, but no less so for vendors and consultants seeking exposure for their work.

"Customers that are having great success won't let us talk about what they're doing," says Anthony Materna, former vice president of marketing at Hecht-Neilsen.

HNC got fed up with just saying "no comment" and set out to find a partner in the trading business willing to share the results of a development effort with the public. The company found Ted Markowitz, vice president and manager of the corporate technology strategy unit at American Express.

Markowitz'group is responsible for guiding and monitoring evaluation of strategic technologies at several American Express subsidiaries, including the American Express Bank and Shearson Lehman Hutton.

According to sources close to the project, the relationship between HNC and American Express doesn't involve initial funding for the pattern recognition prototype. Rather, American Express contributed foreign currency trading data in exchange for a right of first refusal to take HNC's prototype into production.

"I had some spot FX data we contributed for exploratory work looking at foreign exchange trading support areas," says Markowitz. "Neural models have an interesting quality of adaptive learning, but it's only one of many systems we've looked at -- including knowledge- based systems -- for FX trading support."

The model developed by HNC using the American Express-supplied data isn't intended to guide structural position taking. It's more suited to traders with a time horizon of minutes or hours.

A neural network is a cognitive information processing structure based on models of brain function; it consists of a group of processors that work together and communicate with each other. The output of every processor serves as input to every other processor in the network. Each processor contains a simple algorithm that converts all inputs to a single output. The algorithm, or "learning law," decides how important each input is and calculates and assigns weighting.

Neural networks are inherently parallel, yet they do not require decomposition of problems. They are, therefore, capable of high speed processing. Neural networks are also adaptive because their individual processing elements are self-adjusting.

Computerized neural models are trained to identify "events" or "features" -- recognizable and repeating patterns in data -- in the same way that the brain learns to recognize patterns in incoming data. The brain, composed of trillions of connections between billions of N-ary switches known as neurons, develops unique patterns of electro-chemical weighting in response to stimuli.

Virtual neural networks such as those supplied by HNC in the form of microcomputer expansion boards use serial emulation to mimic the pattern recognition locus of the brain -- region 17.

The pattern recognition skills of neural networks are intended to amplify human pattern recognition skills, and reduce the envelope of uncertainty enough to give traders a competitive advantage.

HNC's foreign exchange prototype trained by sifting through the 'noise' of archived foreign exchange price data for recognizable technical features. A price history for the dollar versus the yen was used to train the model.

The prototype developed a feature vocabulary that includes many of the price movement patterns widely recognized by technical analysts -- head and shoulders, double tops, wedges, and pennants. It also identified patterns heretofore unknown to chartists.

A common problem in the application of pattern recognition systems to financial markets is the identification of ghost patterns -- noise masquerading as an event. Figure 1 shows the validated dollar/yen features identified by HNC's neurocomputer.

HNC's neural trading assistant takes the most recent four-hour frame of price movement and goes through a curve-fitting exercise to determine which feature is most likely to occur next. An example of the neural network's predictive skill is shown in Figure 2. The figure compares actual trade data with the neurocomputer's predictions on an event-by-event basis. The actual trade data represented is a novel event -- not used in the training period.

HNC engineers tested the features learned in yen trading against both D-mark and pound sterling trade data. The results, while encouraging, indicated that neurocomputers will deliver a higher degree predictive accuracy when trained in, and confined to, a single pair of currencies.

Industry experts are not convinced of the utility of neural network models for predictive purposes. At the recent Expert Systems In Trading conference sponsored by Waters Information services, Harvey Goodman, manager of consulting for Gold Hill Computers Inc. said, "neural networks are not ready to do anything very sophisticated. There are products available that can be plugged into a PC and used to develop a fairly intelligent pattern recognition system that can be used as input to an expert system that monitors trading. But anything more sophisticated, such as projections and predictions, I don't think are possible now."

"We regard it as an infant technology;" says Markowitz of American Express, "we're nowhere near ready to trade on it."