9730489 Stock This project pursues a research program to develop new methods for forecasting economic time series and to apply these methods to U.S. macroeconomic data. The research program is based on four premises. First, data availability has increased dramatically over the past decade; now literally thousands of series are available for real time forecasting applications. Yet, the small and , arguably, highly structured large models that dominate economic forecasting fail to exploit this vast array of data. Second, recent empirical studies using formal statistical tests confirm conventional wisdom that many economic time series relations are unstable over time. This is an obstacle to many standard forecasting methods, but it also presents an opportunity: properly accounting for this time variation would improve economic forecasts. Third, recent algorithmic and computational advances permit the development of robust forecasting methods based on nonGaussian filtering that can accommodate a wider range of time variation in parameters than conventional linear or state space models. Fourth, economic forecasting is an important practical contribution to the economics profession to society, and as such improved methods for economic forecasting constitute a valid end in themselves. This said, experience over the past several decades has shown that the tools and results of economic forecasting have had significant positive spillovers into applied and theoretical macroeconomics. This project consists oaf two main projects. The first is to use the tools of modern time series analysis to reexamine the construction of diffusion indexes and their use for forecasting. In traditional business cycle analysis, a diffusion index is a measure of the extent to which an expansion or recession has spread across sectors or regions of the economy. Because they are based on many series, these indexes hold out the possibility of providing useful information for forecasting that is not contained in th e main macroeconomic aggregates that dominate modern academic investigations of economic forecasting. They approach developed in this project is to use a dynamic factor model to define diffusion indexes. Preliminary results indicate that, with sufficiently many series, these indexes can be estimated precisely even in the presence of time varying parameters. Promising results are reported here for an initial forecasting experiment, in which factors extracted from 183 macroeconomic time series are used to forecast four main economic aggregates. The second main project on robust forecasting seeks to develop and to implement forecasting procedures that are robust to out-of-sample changes in the process followed by the series and/or to in-sample mispecification. The effort draws on large literatures on robust signal extraction and on nonGaussian filtering. Evidence is presented that these and new, related tools have the potential to result in significant improvements in macroeconomic forecasts relative to linear models with conventional Gaussian time varying parameter uncertainty. ??
