TABLE OF CONTENTS Capital and expense budgeting, capacity forecasting and planning, and marketing campaign designs by telecommunications companies all typically require useful sales forecasts for at least the next 12 to 24 months. Even experienced carriers are frequently surprised by actual demand growth over this horizon. Enthusiastic customer responses to a new marketing initiative (e.g., introduction of the Digital One Rate price plan by AT&T Wireless Services in 1998) can easily exceed anticipated growth and swamp existing and planned capacity. Conversely, demand forecasts that overestimate actual demand can lead to excess and possibly to stranded capacity and can create a financial strain on the company that invested in it. The Iridium satellite network is a recent example. Marketing pressures to forecast what individual customers are likely to do and to identify which customers are most (and least) likely to respond to specific product offers also drive a need for improved prediction methods (Schober, 1999; Strouse, 1999). Increasingly, predictions of likely customer behaviors must be based on only a small amount of historical data, both because new products do not have long histories and because data warehouse administrators retain only a limited number of months worth of data. Forecasting probable future product-purchasing behaviors down to the level of individual households or accounts, based largely or entirely on short-term data, with enough accuracy to be useful in target marketing requires a fundamentally different approach to predictive modeling. The aggregate time series trending and allocation models long used by telecommunications...
