Are Forecasting Models Usable for Policy Analysis?

Abstract

In one of the early papers describing the applications of vector autoregression (VAR) models to economics, Thomas Sargent (1979) emphasized that while such models were useful for forecasting, they could not be used for policy analysis. Recently this point has been vigorously reasserted, by Sargent himself (1984) and by Edward Learner (1985) among others. The point has required vigorous reassertion because VAR models are used widely, and few who use them stay pure—in the sense of never thinking about their implica- tions for policy—for very long. There are several reasonable ways to generate conditional forecasts from VAR models. Once one has seen how easy it is to obtain a forecast conditional on a certain configuration of policy variables, it is tempting to make such forecasts and difficult to be sure that they do not influence one's opinions on policy choices. Robert Litterman, a leader in demonstrating the value of VAR models in forecasting, has also developed methods to use them to make forecasts conditional on policy choices and even to use them in computing optimal policy rules. (See, for example, Litterman 1982, 1984.)