Differential Interpretation in the Survey of Professional Forecasters

Abstract

In this paper we estimate a simple Bayesian learning model to expectations data from the Survey of Professional Forecasters. We reformulate the model in terms of forecast revisions, which allows to abstract from differences in priors and to focus the analysis on the relationship between revisions and signal. The model depends on two parameters, the forecaster’s belief about the signal bias and its weight. The empirical analysis shows that there is significant heterogeneity in the parameters among forecasters, in particular at longer forecast horizons. The cross-sectional distribution of the estimated bias parameter has a median close to zero, while its dispersion decreases with the horizon. A similar result is obtained for the weight parameter, with the median across forecasters increasing toward one and the dispersion decreasing when approaching the target date. The exception to this pattern is CPI inflation for which we find that dispersion, in particular for the estimated weight, increases closer to the target date. Furthermore, the results indicate that a possible explanation for the persistence in dispersion, even at short horizons, is the heterogeneity of interpretation of new information, in the sense that agents with optimistic (relative to the other forecasters) priors are also likely to believe that the signal underestimates the future realization of the variable, and the opposite for forecasters with pessimistic views. We also find that the latter type of forecasters are more likely to assign a low weight to the signal, while optimistic forecasters incorporate the new information faster.