Forecasting market shares with disaggregate or pooled data: A comparison of attraction models

The objective of this paper is to compare market share forecasts using parameter estimates from three different arrangements of single source scanner data. The data arrangements are coupled with two different attraction models and a well known naive model. Each attraction model is specified with either autocorrelated errors or with a lagged attraction term on the right hand side. Forecasts are compared using data aggregated across stores and disaggregated by store. In the latter case, parameters for each store are estimated and a composite (average) share forecast is formed or the data are pooled and a single set of estimated parameters provide the forecast. The study finds that the full cross effects model with an autocorrelated error structure fits the data better than alternative models. However, the differential effects model with an autocorrelated error structure provides the best forecasts. With respect to the alternative data arrangements, the data aggregated across stores (i.e. chain level data) provides the best forecasts.