Prediction interval tests are applied to the reduced forms of two quarterly models of the U.S. (the "old" FRB-MIT model and the Michigan model). The results illustrate the range of tests one can perform on an estimated simultaneous equation model. In particular, the tests determine whether ex post forecast errors can be attributed to structural deficiencies of the models. The paper examines confidence regions and other aspects of forecast distributions-comparisons between mean forecasts and nonstochastic forecasts, comparisons between, forecast variances from multiperiod endogenous simulations and those from one period simulations, and comparisons between forecast variances and residual variances.
This paper puts forward a method for improving the forecasting accuracy of an existing macroeconometric model without changing its policy response characteristics. The procedure is an extension and formalization of the practice of additive adjustments currently used by most forecasters. The method should be of special interest to forecasters who use models built by other investigators because it does not involve reestimation of the original model and uses only information routinely included in the documentation available to model users. The paper ends with a demonstration of the prediction improvement realized by application of this method to a version of the MIT-Penn-SSRC (MPS) model.
This paper puts forward a method for simulating an existing macroeconometric model while maintaining the additional assumption that individuals form their expectations rationally. This simulation technique is a first response to Lucas' criticism that standard econometric policy evaluation allows policy rules to change but doesn't allow expectations rules to change as economic theory predicts they will. The technique is applied to a version of the St. Louis Federal Reserve Model with interesting results. The rational expectations version of the St. Louis Model exhibits the same neutrality with respect to certain policy rules as small, analytic rational expectations models considered by Lucas, Sargent, and Wallace.
This paper describes recent modifications to the mixed-frequency model vector autoregression (MF-VAR) constructed by Schorfheide and Song (2012). The changes to the model are restricted solely to the set of variables included in the model; all other aspects of the model remain unchanged. Forecast evaluations are conducted to gauge the accuracy of the revised model to standard benchmarks and the original model.
This paper integrates and extends some recent computational advances in Bayesian inference with the objective of more fully realizing the Bayesian promise of coherent inference and model comparison in economics. It combines Markov chain Monte Carlo and independence Monte Carlo with importance sampling to provide an efficient and generic method for updating posterior distributions. It exploits the multiplicative decomposition of marginalized likelihood into predictive factors, to compute posterior odds ratios efficiently and with minimal further investment in software. It argues for the use of predictive odds ratios in model comparison in economics. Finally, it suggests procedures for public reporting that will enable remote clients to conveniently modify priors, form posterior expectations of their own functions of interest, and update the posterior distribution with new observations. A series of examples explores the practicality and efficiency of these methods.
This paper was prepared for the inaugural Colin Clark Lecture, Australasian Meetings of the Econometric Society, July 1994.
This paper develops a vector autoregression (VAR) for macroeconomic time series which are observed at mixed frequencies – quarterly and monthly. The mixed-frequency VAR is cast in state-space form and estimated with Bayesian methods under a Minnesota-style prior. Using a real-time data set, we generate and evaluate forecasts from the mixed-frequency VAR and compare them to forecasts from a VAR that is estimated based on data time-aggregated to quarterly frequency. We document how information that becomes available within the quarter improves the forecasts in real time.
One of the more significant developments in econometric modeling over the past decade has been the invention of the forecasting technique known as Bayesian vector autoregression (BVAR). This paper provides a detailed description of the process of specifying a BVAR model of quarterly time series on the U.S. macroeconomy. The postsample forecasting performance of the model is evaluated at an informal level by comparing the model's performance to certain naive forecasting methods, and is evaluated at a formal level by means of efficiency tests. Although the null hypothesis of efficiency is rejected for the model's forecasts, the accuracy of the model exceeds that of naive forecasting methods, and seems comparable to that of commercial forecasting firms for early quarter forecasts.