1 GENETIC ALGORITHM FORECASTING FOR TELECOMMUNICATIONS PRODUCTS STEPHEN D. SLOAN, RAYMOND W. SAW, JAMES J. SLUSS, JR., MONTE P. TULL, AND JOSEPH P. HAVLICEK School of Electrical & Computer Engineering The University of Oklahoma, Norman, Oklahoma ABSTRACT In this paper, we describe genetic algorithms (GA’s) for forecasting long-term quarterly sales of products in the telecommunications technology sector using widely available economic indicators such as Disposable Personal Income and New Housing Starts as independent variables. Individual chromosomes indicated inclusion or disinclusion of specific economic variables, as well as operational rules for combining the variables. Population evolution utilized random crossover mating, mutation, and inversion. Several features beyond those of the canonical GA were also incorporated, including evolution of individuals in distinct ecosystems with a specified level of intermarriage between ecosystems, the capability for a single gene in an individual’s chromosome to indicate a subroutine call to the complete chromosome of an individual from a previous generation, and hill-climbing applied to improve the most fit offspring produced by a generation. At a forecast interval of eight quarters, individuals exhibiting maximal fitness achieved RMS forecast errors below the the average two-week sales figure. 1. INTRODUCTION The ability to accurately forecast long-term future sales of specific products is a highly desirable capability for many companies operating in the increasingly volatile telecommunications technology sector. Such a capability could allow companies to avoid surpluses and shortages in manufacturing resources, including materials, capital equipment, and personnel. Here, by long-term, we mean forecasting of quarterly sales at a forecasting interval ranging from a few quarters up to a few years. Recently, such long- term forecasting has proven to be an extremely difficult problem due to increased market volatility brought on by numerous factors including deregulation, the Telecommunications Act of 1996, and ever expanding global competition. Whereas simple heuristic location estimation techniques, including, e.g., exponential smoothing, were in the past at least marginally adequate for developing long-term predictions in this market, we have found them to be wholly inadequate in recent years. Moreover, due to the highly nonstationary, evolutionary, and indeed sometimes even chaotic nature of telecommunications product sales, the effort required to continuously reformulate more sophisticated parametric methods including linear regressions, classical Box-Jenkins ARMA models, and Kalman filters can rapidly constitute an insurmountable burden. Recently, artificial neural networks (ANN’s) have been applied to a variegated array of forecasting problems (Donaldson and Kamstra, 1989), (Saravanan, 1993), (Kuan and White, 1994), (Wan, 1994), (Masters, 1995). One of the primary advantages of ANN’s is that they potentially have the capability to capture complex and nonlinear relationships