Presented by Ross Link at the Institute for International Research Marketing Mix Optimization Conference, October '99, Ft. Lauderdale, Florida

It is common practice today to take the results of marketing mix regression models and use them to forecast sales and profits in a user friendly "simulator". Often, the regression models were estimated with detailed data (e.g. store level) while the simulators use high level data (market or even national level). Unfortunately, store level models, when applied to market level data, give incorrect forecasts. Based on extensive holdout testing, we show the best type of data to use in marketing mix models and simulators in order to maximize forecast accuracy at various aggregation levels. Model specification issues are addressed. Our findings also have implications for the increasingly common practice of "fusing" marketing mix model estimates from different aggregation levels into one overall model.