|
Competitive pressures and the speed of technological development have
reduced the length of the typical product lifecycle in many markets to a
matter of months. In addition to predicting sales of mature products,
forecasting effectively in such markets entails addressing the transitional
phases of a product's life--the periods after introduction and prior to
phase-out. Unfortunately, the sales patterns that emerge in the course of
such transitions are not usually amenable to forecasting with conventional
methods. This paper describes a forecasting system currently in use at Sun
Microsystems, Inc. (a major manufacturer of network computer products) that
seeks to address the challenges posed by diminishing product lifecycles.
The system combines a diffusion model to describe transitional sales with
very general constructs for time series analysis known as dynamic linear
models (DLMs). The latter helps represent time-series artifacts such as
seasonality and drift which are frequently found in practical forecasting
situations, and which are only scantily addressed by diffusion models
published to date. The model uses Bayesian statistical techniques, so that
it is able to incorporate judgmental information relating to elements of a
particular forecasting context, and to use records of actual sales for
related products as precedents in forecasting.
|
