Multiple change point detection based on Hodrick-Prescott and $l_1$ filtering method for random walk time series data

TL;DR

This paper introduces new methods combining Hodrick-Prescott and l1 filtering to detect multiple change points in random walk time series, addressing parameter selection challenges and validating with simulated and real stock data.

Contribution

It presents novel change point detection techniques tailored for random walk processes, including parameter selection strategies and maximum change point estimation.

Findings

Effective detection of change points in simulated data

Successful application to SNP stock data

Comparable or improved performance over existing methods

Abstract

We propose new methods for detecting multiple change points in time series, specifically designed for random walk processes, where stationarity and variance changes present challenges. Our approach combines two trend estimation methods: the Hodrick Prescott (HP) filter and the l1 filter. A major challenge in these methods is selecting the tuning parameter lambda, which we address by introducing two selection techniques. For the HP based change point detection, we propose a probability-based threshold to select lambda under the assumption of an exponential distribution. For the l1 based method, we suggest a selection strategy assuming normality. Additionally, we introduce a technique to estimate the maximum number of change points in time segments using the l1 based method. We validate our methods by comparing them to similar techniques, such as PELT, using simulated data. We also demonstrate the practical application of our approach to real-world SNP stock data, showcasing its effectiveness in detecting change points.