Macroeconomic Forecasting in Times of Crises

Abstract

We propose a parsimonious semiparametric method for macroeconomic forecasting during episodes of sudden changes. Based on the notion of clustering and similarity, we partition the time series into blocks, search for the closest blocks to the most recent block of observations, and with the matched blocks we proceed to forecast. One possibility is to compare local means across blocks, which captures the idea of matching directional movements of a series. We show that our approach does particularly well during the Great Recession and for variables such as inflation, unemployment, and real personal income. When supplemented with information from housing prices, our method consistently outperforms parametric linear, nonlinear, univariate, and multivariate alternatives for the period 1990 - 2015.