Segmentation algorithm for non-stationary compound Poisson processes

TL;DR

This paper presents a non-parametric segmentation algorithm for non-stationary regime switching processes, specifically targeting compound Poisson processes with variable regimes, and demonstrates its superior performance on financial market data.

Contribution

The paper introduces a generalized segmentation algorithm that outperforms previous methods for identifying regimes in non-stationary compound Poisson processes.

Findings

The new algorithm more than triples the number of detected regimes in stock market data.

It outperforms the original algorithm in accuracy for regime detection.

Application to London Stock Exchange data shows improved segmentation results.

Abstract

We introduce an algorithm for the segmentation of a class of regime switching processes. The segmentation algorithm is a non parametric statistical method able to identify the regimes (patches) of the time series. The process is composed of consecutive patches of variable length, each patch being described by a stationary compound Poisson process, i.e. a Poisson process where each count is associated to a fluctuating signal. The parameters of the process are different in each patch and therefore the time series is non stationary. Our method is a generalization of the algorithm introduced by Bernaola-Galvan, et al., Phys. Rev. Lett., 87, 168105 (2001). We show that the new algorithm outperforms the original one for regime switching compound Poisson processes. As an application we use the algorithm to segment the time series of the inventory of market members of the London Stock Exchange and we observe that our method finds almost three times more patches than the original one.