international Journal of Forecasting 3 (1987) 449-451 North-Holland 449 zyxwvutsrqpo THE SOPHISTICATION OF ‘NAIVE’ MODELING David A. AAKER Unruersuy of Calrfornra, Berkeley, CA 94720, USA Robert JACOBSON University of Washington, Seattle, WA 98195, USA Abstract: The relatively high predictive power of a naive market share model is due to (1) its similarity to the reduced form representation of the underlying model and (2) the comparison with forecasts generated from structural models that are likely to be m&specified. Brodie and de Kluyver should be complimented for raising a number of important issues both in the general area of forecasting and in the specific area of modeling market share. We would like to take this opportunity to offer some further discussion and implications of their very challenging findings. It may seem surprising that linear extrapolations of past values are able to forecast with about the same level of forecast accuracy as structural models. How and why can such a ‘naive’ model yield this degree of predictive accuracy? The answer lies in the fact that these models are not naive or simple. It can be shown that these models are reduced from representations of the underlying structural model [Zellner and Palm (1974)J. Reduced form models may capture very complicated/sophisticated causal mechanisms. A ‘naive’ random walk is the best predictor of stock prices because this time series behavior is the outcome of the complex causal mechanism of efficient markets. The rationale for Box and Jenkins (1976) ARIMA models is that the past history of a series is useful in forecasting its future behavior. Causal influences that gave rise to regular patterns in the past are apt to give rise to these same regular patterns in the future. If it is possible to model this pattern, a degree of forecasting accuracy can be achieved. The time series pattern of a series is a synopsis of the influence of a host of causal factors influencing the series. This pattern reflects the combined effects of many different types of causal/structural phenomena. This is in fact what a reduced form is, i.e., a solution to a structural model. A reduced form will never provide a better representation, in terms of explanatory power, than the underlying structural model. The reduced form and the ‘true’ structural model are observationally equivalent. However. a reduced form model, such as that obtained through time series methods, may provide a better representation than the theorized (as opposed to the underlying) structural model. The ability to model the underlying causal mechanism determining the behavior of a series is limited by (1) knowledge of the phenomena and (2) measures of the underlying causal factors. Because of these limitations, the structural models formulated may be quite different from the underlying model. 0169-2070/87/$3.50 0 1987. Elsevier Science Publishers B.V. (North-Holland)