Do the Hamilton and Beveridge–Nelson filters provide the same information about output gaps? An empirical comparison for practitioners

Abstract

I compare the empirical performances of the recently-developed Hamilton and Beveridge–Nelson filters of nonstationary time series, using quarterly data on real gross state product in U.S. states. There is meaningful overlap between the two filters, with average correlation coefficients ranging between 0.60 and 0.97. The Hamilton filter and its more recent modification produce cycles of greater volatility and amplitude than the Beveridge–Nelson filter and appear to outperform in pseudo-out-of-sample forecasting exercises of future GSP growth and inflation (though the outperformance is not generally statistically significant). The Beveridge–Nelson filter is, however, less sensitive to realizations of new data.