Redesigning the classical automatic selection of X-11 seasonal filters

The classical X-11 seasonal adjustment method for monthly and quarterly time series is equipped with routines for data-driven selections of both Henderson trendcycle filters and 3 × k seasonal moving averages, currently involving up to three candidate filters in either case. Although these routines have a long-standing tradition that can be traced back at least to 1960, they have not been adopted in a recent JDemetra+ implementation of a modified X-11 method tailored to the specifics of infra-monthly time series, such as the coexistence of multiple seasonal patterns with potentially fractional periodicities. Focusing on seasonal moving averages, we seek to fill this gap by suggesting a generic redesign of the legacy selection concept based upon the so-called moving seasonality ratio. This blueprint utilises a broader set of candidate seasonal filters and, unlike the original setting, a set of common approaches for deriving the requisite asymmetric variants. Considering intersections of multiple approach-specific selection rules stabilises the final filter choice and, what is more, naturally provides the warranted thresholds controlling the potential recalculation of the moving seasonality ratio from suitably shortened detrended observations. Our proposed redesign is illustrated using one specific rule based upon threshold quartiles and real-time data for three German macroeconomic time series sampled at quarterly, monthly, and daily intervals. The last example also highlights the need for additional intermediate steps in the calculation of the moving seasonality ratio when the data contain complex seasonal dynamics.