Empirical Significance of Learning in a New Keynesian Model with Firm-Specific Capital

This paper examines the empirical significance of learning, a type of adaptive, boundedly rational expectations, in the U.S. economy within the framework of the New Keynesian model. Two popular specifications of the model are estimated: the standard three equation model that does not include capital, and an extended model that allows for endogenous capital accumulation. Estimation results for learning models can be sensitive to the choice for the initial conditions for agents expectations, so four different methods for choosing initial conditions are examined, including jointly estimating the initial conditions with the other parameters of the model. Maximum likelihood results show that learning under all methods for initial conditions lead to very similar predictions as rational expectations, and do not significantly improve the fit the model. The evolution of forecast errors show that the learning models do not out perform the rational expectations model during the run-up of inflation in the 1970s and the subsequent decline in the 1980s, a period of U.S. history which others have suggested learning may play a role. Despite the failure of learning models to better explain the data, analysis of the paths of expectations and structural shocks during the sample show that allowing for learning in the models can lead to different explanations for the data