Neural Networks in Finance and Investing

Using Artificial Intelligence to Improve Real-World Performance

Book Review
By Lou Mendelsohn

By Robert R. Trippi and Efraim Turban
Probus Publishing Co.
Hard-cover; 513 pages, $65.00.

Neural networks are perhaps the most significant forecasting tool to be applied to the financial markets in recent years. Robert Trippi and Efraim Turban have compiled a collection of articles from industry and academic experts alike which discuss the application of this predictive technology to real-world finance and investing. In addition to an introductory section on artificial neural networks, this book examines their application to the analysis of financial conditions, business failure prediction, debt risk assessment, security market analysis, and financial forecasting. This work is a good introductory survey for technicians interested in familiarizing themselves with neural network applications in the finance industry. Throughout the book, various issues are discussed which have a common thread across neural network applications, including choice of inputs, data transformations, selection of network architecture, and evaluation of results. Technicians will be particularly interested in Parts Five and Six of the book which deal specifically with security market applications and financial forecasting.

The book’s first part, titled “Neural Network Overview” contains three chapters which address the basics of artificial neural networks. The first chapter, written by Larry Medsker, Efraim Turban and Robert Trippi, offers an overview of the structure, inner workings and application of neural networks. The second chapter by Delvin Hawley, John Johnson and Dijjotam Raina briefly discusses neural network applications for corporate finance, financial institutions and professional investors. The third chapter in this section written by Casimir Klimasauskas, is concerned specifically with the application of neural networks to the domain of credit approval. Of course, as is common with survey books of this type, these introductory selections contain some overlap of material.

Part Two, “Analysis of Financial Conditions,” contains four chapters involving the application of neural networks to several financial domains. Two of these chapters describe how hybrid neural network/ expert systems can solve problems that would be much more difficult or impossible to solve using one or the other technology exclusively. The first chapter, by Robert Marose, describes a system that forecasts the credit worthiness of corporate loan applicants. The second chapter, by Don Barker, discusses a tool that integrates these two technologies for determining the financial health of a small business. In their chapter, R. Berry and Duarte Trigueiros describe a system that uses neural networks to extract knowledge from accounting reports. After describing several traditional methods used for this type of knowledge extraction, the authors examine a neural approach to the problem that outperforms these traditional approaches. George Klemic’s chapter examines the use of neural networks by the IRS to identify debtors who are most likely to become delinquent.

Part Three, titled “Business Failure Prediction,” includes six chapters, four of which involve bankruptcy prediction, while the other two involve prediction of bank and thrift failure. Wullianallur Raghupathi, Lawrence Schkade, and Bapi Raju’s chapter discusses the use of neural networks in the domain of bankruptcy prediction, with the long term goal of incorporating a neural network as the underlying model in a decision support system. Another chapter, by Marcus Odom and Ramesh Sharda, compares neural networks to a traditional method of predicting bankruptcy. In this study, neural networks outperformed the traditional method of prediction in all tests. The chapter by Eric Rahimian, Seema Singh, Thongchai Thammachote, and Rajiv Virmani also compares traditional and neural network models for bankruptcy prediction. Kevin Coleman, Timothy Graettinger, and William Lawrence examine the use of neural networks as part of a hybrid system that predicts bankruptcies, utilizing an expert system to recommend a course of action. The last two chapters in Part Three, concerning prediction of bank and thrift failures, support the use of neural networks through comparative studies. Kar Tam and Melody Kiang’s chapter concerning bank failures compares a variety of models including traditional techniques, machine learning, and neural networks. The remaining chapter by Linda Salchenberger, Mine Cinar, and Nicholas Lash compares neural networks to traditional models in making thrift failure predictions. In each of these studies, neural networks were found to outperform other modeling methods.

Part Four, “Debt Risk Assessment,” contains four chapters that explore the use of neural networks to assess risk. The first two chapters concern the task of bond rating. In their chapter, Soumitra Dutta and Shashi Shekhar explore the use of neural networks to predict such ratings and present a comparison of their results with more traditional linear regression models. The following chapter, by Alvin Surkan and Clay Singleton, also examines bond rating, emphasizing the use of an atypical network architecture in an effort to achieve superior performance. The last two chapters in Part Four concentrate on risk assessment in mortgage underwriting. The first of the two, written by Edward Collins, Sushmito Ghosh, and Christopher Scofield, examines the development and use of a neural network system that utilizes a hierarchical organization of multiple neural nets, where the nets at each level act as a panel of judges to generate risk classifications. The next chapter, written by the same authors with the addition of Douglas Reilly, is a brief update on the performance of the system introduced in the previous chapter.

Part Five, titled ” Security Market Applications,” is the longest section in the book and will be of most interest to technical analysts just getting familiar with this technology. Of the eight chapters presented in this section, four address the task of making stock market predictions, three concern predictions in the futures markets and one deals with testing arbitrage pricing theory. The first chapter by Halbert White chronicles a case study using neural networks to predict daily stock returns for IBM. He discusses the use of statistical methods to help determine optimal network architectures. The next chapter, by Youngohc Yoon and George Swales, compares a neural network approach with Altman’s multiple discriminant analysis (MDA) model to predict stock price performance. Here, Yoon and Swales show that the neural network approach yielded better performance than MDA. They also offer some insight into the effects of various network architectures on performance. Another chapter, by Takashi Kimoto, Kazuo Asakawa, Morio Yoda, and Masakazu Takeoka, examines the use of modular neural networks in a stock market trading system. The authors discuss work by Fujitsu and Nikko Securities to develop a TOPIX buying and selling prediction system that utilizes a number of neural network modules to generate signals. The chapter discusses many of the aspects of training and testing networks, including the types of prediction simulations used as well as ways to extract information from the trained networks. Ken-ichi Kamijo and Tetsuji Tanigawa contributed the next chapter which focuses on stock price pattern recognition. The researchers use a less familiar type of neural network, especially designed for problems that are dependent on a temporal context. They successfully used these networks to recognize triangle patterns on candlestick charts.

The next three chapters in Part Five address the futures markets. The first chapter, by W. E. Bosarge, Jr., explores various theoretical issues behind prediction in the markets. He explains that adaptive processes, such as neural networks, maybe used to exploit nonlinear structures, like those of financial markets. After a short discussion of chaos and the efficient market model, Mr. Bosarge describes the application of neural networks to these nonlinear, yet not necessarily random, structures. The following chapter, by Karl Bergerson and Donald Wunsch II, presents a trading model based on a hybrid neural network/expert system. In this chapter, the authors discuss a method of selecting “quality” training data for their neural networks coupled with the use of an expert system to handle risk management. The next chapter, by J. E. Collard, discusses the applications of neural networks to recognize buy/sell patterns for the live cattle market. The article includes training results and brief trading performance. The final chapter in Part Five, by Hamid Ahmadi, examines the use of neural networks in testing arbitrage pricing theory.

Part Six, titled “Neural Network Approaches to Financial Forecasting, ” contains three chapters. The first chapter, by Leory Marquez, Tim Hill, Reginald Worthley, and William Remus, offers neural networks as an alternative to regression analysis, one of the most popular quantitative methods used in finance. The next chapter, by Ramesh Sharda and Rajendra Patil, presents an empirical study comparing neural networks to the Box-Jenkins forecasting model for predicting time series. The researchers also discuss a testing method used to determine which parameters and architecture to use for neural networks. The remaining chapter in Part Six, by A. N. Refenes, examines constructive learning and an application of that type of learning to currency exchange rate forecasting. The author then compares the results of this type of learning with fixed architecture networks and statistical forecasting.

Although only an introductory book on the subject, Neural Networks in Finance and Investing covers a broad cross-section of real-world neural network applications in the financial industry in sufficient depth to stimulate the reader to delve further into this emerging field. Readers will be well-advised to explore related technologies including genetic algorithms, fuzzy logic, and expert systems in fashioning quantitative forecasting systems that can synthesize technical, intermarket, and fundamental data inputs within a synergistic framework. Such hybrid approaches, which go beyond traditional methods of single-market analysis proposed in the 1980’s yet still widely used today, will afford a competitive advantage to those traders who use these technologies in the globalized interrelated markets of the 1990’s.

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