lstar vs trendseries

trendseries extracts trends from economic time series through two pipe-friendly functions, backed by an unusually broad method set — Hodrick-Prescott in one- and two-sided variants, Baxter-King, Christiano-Fitzgerald, Hamilton regression, Beveridge-Nelson, unobserved components, plus the moving average and smoothing family. The 1.4 release adds decomposition proper: an exported decompose_series() that splits a series into trend, seasonal, and remainder across five methods and guarantees the components add back to the original values.

The package is moving from breadth of methods to rigour about what those methods produce. Recent work has been about defaults and guarantees rather than new filters: the unobserved components model now derives its signal-to-noise ratios from Hodrick-Prescott lambdas so the default output is economically interpretable, decomposition carries an exact additive identity, and a log transform gives a uniform multiplicative variant across every method. Naming is being tidied in the same spirit, with group_vars deprecated in favour of group_cols. Side-by-side method comparison — passing several methods and getting each one's components as separate columns — suggests an audience that treats method choice as a research question rather than a setting.

Expect the comparison and diagnostic side to keep developing, since the package now produces multiple decompositions of the same series and offers no ranking between them; the entries give no indication of new filters being queued.