Specifying the Forecast Generating Process for Exchange Rate Survey Forecasts

This paper contributes to the literature on the modeling of survey forecasts using learning variables. We use individual industry data on yen-dollar exchange rate predictions at the two week, three month, and six month horizons supplied by the Japan Center for International Finance. Compared to earlier studies, our focus is not on testing a single type of learning model, whether univariate or mixed, but on searching over many types of learning models to determine if any are congruent. In addition to including the standard expectational variables (adaptive, extrapolative, and regressive), we also include a set of interactive variables which allow for lagged dependence of one industry’s forecast on the others. Our search produces a remarkably small number of congruent specifications-even when we allow for 1) a flexible lag specification, 2) endogenous break points and 3) an expansion of the initial list of regressors to include lagged dependent variables and use a General-to-Specific modeling strategy. We conclude that, regardless of forecasters’ ability to produce rational forecasts, they are not only “different,” but different in ways that cannot be adequately represented by learning models.