Measuring business cycles using VARs

We propose to measure business cycles using vector autoregressions (VARs). Our method builds on two insights: VARs automatically decompose the data into stable and unstable components, and variance-based shock identification can extract meaningful cycles from the stable part. This method has appealing properties: (1) it isolates a well-defined component associated with typical fluctuations; (2) it ensures stationarity by construction; (3) it targets movements at business-cycle frequencies; and (4) it is backward-looking, ensuring that cycles at each date only depend on current and past shocks. Since most existing filters lack one or more of these features, our method offers a valuable alternative. In an empirical application, we show that the two shocks with the largest cyclical impact effectively capture postwar U.S. business cycles and we find a tighter link between real activity and inflation than previously recognized. We compare our method with standard alternatives and document the plausibility and robustness of our results.