Macroeconomic forecasting in times of crises

Summary We propose a parsimonious semiparametric method for macroeconomic forecasting. Based on ideas of clustering and similarity, we partition the series into blocks, search for the closest blocks to the latest block of observations, and forecast with the matched blocks. In a real‐time forecasting exercise, we show that our approach does especially well for labor market and other key macro variables. Our method outperforms parametric linear, nonlinear, time‐varying, and combination forecasts for the period 1999–2015 and particularly in the Great Recession. When adding financial spreads, our method delivers further improvements for labor market variables and capacity utilization.