Detecting Breaks in Real Time: A Panel Forecasting Approach

Abstract

We develop a new Bayesian panel regression approach to estimating an unknown number of breaks and forecasting future outcomes in the presence of scarce information from new regimes. Our approach allows the parameters to be heterogeneous across units but assumes that the timing of breaks is common. Exploiting information in the cross-section greatly increases our ability to detect breaks more rapidly in real time compared with univariate time-series approaches applied separately to each cross-sectional unit. Rapid break detection along with shrinkage of the parameter estimates towards a common prior turn out to be key to generating accurate out-of-sample forecasts. In an empirical application to inflation dynamics in 28 EU countries, we find that our heterogeneous panel forecasting method detects breaks with little delay in real time and that this translates into significantly improved out-of-sample forecasts relative to a range of existing methods.