Abstract

The analysis of trends and cycles, measured through unobserved components (UC) models, is crucial for an understanding of business cycle and labor market fluctuations and their structural characteristics. Additionally, latent variables play a prominent role for constructing economic forecasts that make efficient use of many sources of available information. There, high-dimensional factor models are the first choice among practitioners and researchers due to their excellent predictive accuracy. At the same time, fractionally integrated models have had considerable success since they allow a flexible modeling of the long-run characteristics of economic time series. The latter methods are well explored from a theoretical side and have been found relevant for various modeling tasks in macroeconomics and finance.

We combine both approaches and explore nonstationary macroeconometric models with latent variables in a fractionally integrated setup. We show that the increased flexibility of the long-run behavior is crucial for the results of these models. In the case of UC models, the identification of trends and cycles strongly depends on the dynamic specification of the latter while in the high-dimensional setup the flexibility allows a parsimonious and automatic treatment of many tightly linked but heterogeneous macroeconomic variables. In the first two parts of the proposed project, we tackle the problem of appropriately formulating univariate and multivariate fractional UC models, assess methods for their estimation and present results for univariate variables like real GDP as well as new evidence on the cyclicality of multivariate labor market flows. In the third part, we describe and assess the different possible formulations of high-dimensional prediction models when a fractional integration framework is adapted. We assess their relative performance in a pseudo out-of-sample study. Due to the strong use of factor models for applied forecasts, even moderate improvements in their forecasting ability could draw substantial interest and lead to beneficial results in practice.

In the renewal project we build on These results and extend the classs of fractional UC models to a multivariate setting with stochastic processes of different Integration orders. The multivariate Domain allows for crucial advancements of e.g. analyzing Long-run co-movements, further increasing the reliability of business cycle estimates, and improving the specification of factor models.