\( \newcommand\R{\mathbb{R}} \newcommand\Z{\mathbb{Z}} \newcommand\LL{\mathbb{L}} \newcommand{\E}[1]{\mathbb{E}\left[#1\right]} \newcommand{\V}[1]{\mathbb{V}\left[#1\right]} \newcommand{\ps}[2]{\left\langle #1 \,,\, #2 \right\rangle} \newcommand\1{\mathbb{1}} \newcommand\N{\mathbb{N}} \newcommand\Norm{\mathcal{N}} \newcommand{\transp}[1]{{}^t\!#1} \newcommand\ud{\,\mathrm{d}} \DeclareMathOperator*{\argmax}{argmax} \DeclareMathOperator*{\argmin}{argmin} \DeclareMathOperator{\e}{e} \DeclareMathOperator{\Cov}{Cov} \DeclareMathOperator{\Determinant}{det} \newcommand{\determinant}[1]{\Determinant\left(#1\right)} \)

Abstract

This paper focuses on different approaches to build moving averages for real-time trend-cycle estimation and fast turning point detection. We propose a comparison of the main methods, based on a general unifying framework to derive linear filters. We also describe two possible extensions to local polynomial filters: the addition of a timeliness criterion to control the phase shift (delay in the detection of turning points) and a way to locally parameterize these filters. The empirical comparison of the methods shows that: the optimization problems of the filters from the Reproducing Kernel Hilbert Space (RKHS) theory increase the phase shift and the revisions of the trend-cycle estimates; modeling polynomial trends that are too complex introduces more revisions without decreasing the phase shift; for polynomial filters, a local parameterization reduces the phase shift and the revisions.

Keywords: time series, trend-cycle, seasonal adjustment, turning points.

JEL Classification: E32, E37.