UNDERSTANDING ARTIFICIAL INTELLIGENCE APPLICATIONS

By: Lou Mendelsohn

Tactical and Strategic
Computing Technologies
Systemsware Corp. 973C Russell Avenue
Gaithersburg, MD 20879
301-948-5391
Editor: Roy S. Freedman
Published 1993, 302 pages, $65

Institutions and individuals both are finding a wide range of uses for artificial intelligence (Al) technologies such as expert systems, neural networks, genetic algorithms and fuzzy logic, to name only a few. For the second year, the Conference on Artificial Intelligence Applications on Wall Street has provided a forum for the discussion and evaluation of Al technologies related to market analysis, asset allocation and regulation. The proceedings from this year’s conference are divided into 11 basic sections, with two to four separate articles in each section.

UNDERSTANDING ARTIFICIAL INTELLIGENCE
The first section, entitled “Understanding News,” contains four articles addressing financial and economic news and data. Two discuss methods of extracting and disseminating financial news, while a third discusses reasoning based on financial news. The fourth article analyzes recent US Supreme Court and lower federal court decisions involving copyright protection of data compilations. “Technology Transfer,” the second section, discusses the use of Al for computer-supported cooperative work and the application of expert systems to technology transfer.

“AI and asset allocation and trading signals,” the third section, focuses first on an expert system that adds weights to time-critical information to enhance portfolio performance and examines a mathematical model that can act as a stand alone system or be used with other AI tools to rank buy/sell signals. Another article introduces STAR, a rule-based system for asset allocation. The system’s objective is to maximize the mean return while limiting turnover and sensitivity to short-term market moves. The final article of this section examines the use of a neural network system for generating trading signals. The article touches on many issues that must be addressed when applying neural networks to this type of task and also discusses the use of a logical network architecture and a rapid training algorithm.

The next section, on stock market prediction, should interest stock traders. All four articles examine the application of neural networks. The first discusses a typical neural network approach to prediction of buy and sell signals for the Milan, Italy, stock exchange. The second article begins with a discussion of the pitfalls involved in utilizing neural networks for prediction of financial time series and goes on to discuss a system, based on the assumption that markets are nonlinear dynamic systems, which utilizes different neural networks depending on the market’s condition. The third article concentrates on various uses of neural nets in investment management and also discusses the use of sensitivity analysis and approximate rule extraction in feed forward neural networks. The last article examines the use of dual adaptive structure neural nets for stock market prediction. Unlike the traditional static architectures used in most of the neural net systems presented at the conference, these were capable of generating and annihilating hidden neurons during training.

The fifth section, titled “Trading workstation support,” would be of primary interest to brokerage firms and brokers. All four articles discuss the application of expert systems to improve trading workstations. The first article concentrates on a system to help identify potential customer problems more quickly, so that support staff can recognize and resolve problems even before the customer becomes aware of them. The next article examines some key elements necessary in an intelligent broker station and the implications that AI has on these elements. The third article identifies some major weaknesses inherent in traditional database technology as applied to expert systems for stock trading and presents a real-time object-oriented database model to overcome these weaknesses. The final article discusses various advanced interface design and embedded AI applications and points out that such embedded systems can greatly aid in trading floor support.

“AI and modem portfolio theory” examines applications of genetic algorithms, fuzzy logic and stochastic search heuristics. The first article examines the use of fuzzy logic and mean-variance optimization in tactical asset allocation. A rule-based stock market predictor that utilizes these technologies is introduced. The next article discusses the application of genetic algorithms to such problems as optimal allocation, portfolio insurance and performance prediction for financial portfolios. The third article concentrates on anomalous risk behavior detection in portfolio management strategies. The system presented utilizes a fuzzy nonparametric anomaly detecting algorithm to determine if a portfolio manager is at significant variance from his/her peers, providing a means of identifying possible high risk or inappropriate management. The last article in this section compares different stochastic search heuristics to the task of finding the optimum portfolio. The heuristics examined are genetic algorithms, simulated annealing and dynamic search space reduction.

“Marketing and business strategies” also contains four articles, the first of which examines the use of Al in strategic business planning. The three remaining articles examine the use of neural nets to select sales promotion instruments, model direct mail responses and predict television audiences. Most of these articles will be of interest to those individuals in the financial services fields or marketing and management who wish to apply AI to these fields.

The first article in “Fixed income and bond ratings” examines the ability to forecast returns in the Italian spot and futures bond markets with neural networks. One interesting part discusses the generation of trading rules by utilizing a genetic algorithm to perform unsupervised learning with recurrent neural networks. The next article discusses a “neural” decision support system for predicting currency exchange rates. The third and fourth articles examine the application of neural nets to Gross National Product (GNP) prediction and bond rating.

The ninth section, on qualitative analysis, contains articles that address such issues as combined approximate reasoning in the mergers and acquisition domain, a knowledge-based system for financial statement analysis, an expert system applied to credit evaluation and the use of AI for technology transfer.

“Discovering stock selection rules” covers such topics as a hybrid case-based and rule-based reasoning system that is capable of identifying nearly satisfied rules and automatically supplying examples of similar, previous cases. Other articles examine machine learning and rule-based refinement, optimal screening for stock portfolio creation and case-based reasoning for financial database mining.

Finally, “Fuzzy financial time series” concentrates on applying fuzzy logic to time-series data analysis. One article analyzes financial time-series data by generating a fuzzy logic model and then optimizing the model through the use of genetic algorithms. The next article utilizes a genetic algorithm to adapt a fuzzy associative memory to model changes in Treasury bill interest rates. The final article introduces a hybrid neural network and fuzzy logic system for stock selection.

Of course, as was the case last year, neural networks and knowledge-based systems were the staple of this year’s conference. Some of the more interesting developments to come out of this year’s meeting included a much greater emphasis on genetic algorithms and fuzzy logic systems. Although not yet as prevalent, chaos theory and dynamic non-linear systems have begun to gain more acceptance and should make an important contribution to next year’s conference. Especially interesting was the degree to which various technologies are being combined.

As evinced by the conference, the application of these AI technologies will come to represent perhaps the most significant advancement in market analysis that the financial industry has ever seen. In today’s interrelated global environment, such tools, capable of synthesizing seemingly disparate technical, intermarket and fundamental data, will become indispensable. For traders interested in keeping abreast of current developments in this arena, this annual conference represents an excellent barometer of research and trading applications being developed with these emerging technologies.

Lou Mendelsohn is president of Market Technologies, Wesley Chapel, FL, a research, development and consulting firm involved in the application of artificial intelligence to financial market analysis.

Reprinted from Technical Analysis of
Stocks & Commodities magazine. (C) 1993 Technical Analysis, Inc.,
4757 California Avenue S.W., Seattle, WA 98116-4499, (800) 832-4642.