A Note on the Use of Markov Chains in Forecasting Store Choice

The Markov model showed only slight predictive advantage over the no-change model for short-term forecasting of supermarket choices for a sample of 45 families. While this does not imply a blanket rejection of the Markov technique for forecasting, it is important to recall that this case held to a minimum many of the problems facing Markovian analysis-aggregation of dissimilar units, relatively low purchase rates, and requirement of such long sample periods to build up an adequate sample of events that the critical Markovian assumption of stable probabilities is almost certainly violated. Under these circumstances, the simpler model which says that nothing changes, performs almost as well as the more refined Markov formulation. It is possible, of course, that the slight advantage of the Markov model will outweigh the increased cost of using such a model, but the no-change model has advantages both with regard to simplicity and to applying control - chart types of procedures to track series for stability over time. The usual qualifications about representativeness of geographic areas, panels, samples of panel members and time periods, of course, apply to this analysis