Accounting for structural breaks in the Hamilton filter

This paper reveals that the regression-based trend-cycle decomposition of Hamilton (2018) introduces spurious cyclical dynamics when time series undergo structural breaks. Because the filter relies on an h-step ahead projection, trend adjustment mechanically lags behind regime shifts, inducing large and persistent biases in the cyclical component. We propose a parsimonious modification using a dummy-variable structure that allows the regression intercept and autoregressive components to adjust to identified breaks. Extensive Monte Carlo simulations show that this approach substantially reduces Root Mean Squared Error (RMSE) across a wide range of data-generating processes. These gains are robust to variations in break magnitude, timing and number; sample size; and signal-to-noise ratio. Notably, our approach remains superior in pseudo-real-time settings, where the benefits from break-adjustment outweigh the costs of recursive break date estimation and revision. An empirical application to the Swiss franc/euro exchange rate demonstrates that our modification avoids the spurious cyclical spikes generated by standard filters during major policy shifts. Both ex-post and recursive real-time estimates confirm that our proposed framework provides a more reliable decomposition for macroeconomic assessment and inference.