A knowledge-based methodology for tuning analytical models

Many computer-based analytical models for decision-making and forecasting have been developed in recent years, particularly in the areas of economics and finance. Analytic models have an important limitation which has restricted their use: a model cannot anticipate every factor that may be important in making a decision. Some analysts attempt to compensate for this limitation by making heuristic adjustments to the model in order to "tune" the results. Tuning produces a model forecast that is consistent with intuitive expectations, and maintains the detail and structure of the analytic model. This is a very difficult task unless the user has expert knowledge of the model and the task domain. This paper describes a new methodology, called knowledge-based tuning, that allows a human analyst and a knowledge-based system to collaborate in adjusting an analytic model. Such a methodology makes the model more acceptable to a decision-maker, and offers the potential of improving the decisions that either an analyst or a model can make alone. In knowledge-based tuning, subjective judgments about missing factors are specified by the analyst in terms of linguistic variables. These linguistic variables and knowledge of the model error history are used by the tuning system to infer a specific model adjustment. A logic programming system was developed that illustrates the tuning methodology for a macroeconometric forecasting model that empirically demonstrates how the predictability of the model can be improved. *Department of Computer Science, Polytechnic University **Department of Computer Science, St. John's University. To appear in IEEE Transactions on Systems, Man, and Cybernetics, Volume 21, Number 2, March, 1991. A KNOWLEDGE-BASED METHODOLOGY FOR TUNING ANALYTICAL MODELS 2 Biosketch of Authors Roy S. Freedman received his BS and MS in Mathematics in 1975, his MS in Electrical Engineering in 1978, and his PhD in Mathematics in 1979, all from Polytechnic University. He joined the Hazeltine Corporation Research Laboratories in 1979, and was responsible for the software engineering of several well-known AI systems. In 1985 he returned to Polytechnic University as an Associate Professor. As a consultant, Dr. Freedman has done extensive work for a broad range of clients in the financial services and telecommunication industries, including the New York Stock Exchange, Equitable Life, MCI, Chemical Bank, Intelligent Technology Inc. (Tokyo), Grumman Aerospace, and Hazeltine Corporation. In 1989, he founded Inductive Solutions, Inc., a firm that produces tools for building embedded knowledge-based systems. Dr. Freedman has published over thirty papers in artificial intelligence and software engineering as well as two books, Programming Concepts with the Ada Language and Programming with APSE Software Tools, both now published by McGraw-Hill. He has been an Associate Editor of the journal IEEE Expert for the past four years, and is a member of the IEEE Computer Society, AAAI, ACM, and SIAM. Dr. Freedman can be reached at the Department of Electrical Engineering and Computer Science at Polytechnic University, 333 Jay Street, Brooklyn, NY 11201 or at Inductive Solutions, Inc., 380 Rector Place, Suite 4A, New York, NewYork 10280. Gerald J. Stuzin is an Associate Professor of Computer Science at St. John's University. He obtained his Ph.D. from Polytechnic University in 1989, and an MBA in Economics from New York University in 1972. Since 1969, Dr. Stuzin has worked as an economic consultant at American Can Company, CBS, Inc., Union Carbide, and the International Forecast Group. He is presently a consultant at Business Planning Systems, an economics consulting firm. His current interests are in developing intelligent analytic tools. Dr. Stuzin can be reached at the Department of Computer Science at St. John’s University, Jamaica, New York 11471. A KNOWLEDGE-BASED METHODOLOGY FOR TUNING ANALYTICAL MODELS