Joint Linear Trend Recovery Using L1 Regularization

TL;DR

This paper analyzes the effectiveness of L1 trend filtering for recovering joint linear trends and detecting change points in time series, providing theoretical guarantees and empirical validation.

Contribution

It offers new sufficient conditions for L1 trend filter performance and shows that change point detection does not require the weak irrepresentable condition.

Findings

Almost optimal rate for mean estimation.

High probability recovery of slope change points.

L1 trend filter performs well in finite sample simulations.

Abstract

This paper studies the recovery of a joint piece-wise linear trend from a time series using L1 regularization approach, called L1 trend filtering (Kim, Koh and Boyd, 2009). We provide some sufficient conditions under which a L1 trend filter can be well-behaved in terms of mean estimation and change point detection. The result is two-fold: for the mean estimation, an almost optimal consistent rate is obtained; for the change point detection, the slope change in direction can be recovered in a high probability. In addition, we show that the weak irrepresentable condition, a necessary condition for LASSO model to be sign consistent (Zhao and Yu, 2006), is not necessary for the consistent change point detection. The performance of the L1 trend filter is evaluated by some finite sample simulations studies.