Estimation de la tendance-cycle avec des m\'ethodes robustes aux points atypiques

TL;DR

This paper compares traditional and robust non-linear methods for estimating trend-cycle components in economic series, especially around shocks, and proposes extensions to improve robustness and confidence interval estimation.

Contribution

It introduces a methodology to extend Henderson and Musgrave moving averages for robustness to shocks and external information, enhancing real-time trend-cycle estimation.

Findings

Robust moving averages reduce revisions around shocks.

Traditional methods sometimes produce significant revisions.

Robust non-linear methods do not always yield satisfactory trend-cycle estimates.

Abstract

Seasonally adjusted series are usually used to analyse the business cycle and turning points. When the irregular is too high, it is preferable to smooth the series in order to analyse the trend-cycle component directly. This study focuses on the real-time estimation of the trend-cycle component around shocks and turning points. The linear moving averages classically used for estimating the trend-cycle, which are sensitive to the presence of atypical points, are compared with robust non-linear methods. We also propose a methodology for extending the Henderson and Musgrave moving averages to take account of external information and thus construct moving averages that are robust to the presence of certain shocks. We describe how to estimate confidence intervals for estimates derived from moving averages, thereby validating the use of these new moving averages. By comparing the methods on simulated and real series, we show that: building robust moving averages makes it possible to reduce revisions and better model turning points around shocks, without degrading the estimates when no shock is observed; robust non-linear methods do not make it possible to extract a trend-cycle component that is satisfactory for economic analysis, with sometimes significant revisions. This study is fully reproducible and all the codes used are available under https://github.com/AQLT/robustMA.