Heuristic Segmentation of a Nonstationary Time Series

TL;DR

This paper evaluates a heuristic segmentation algorithm for nonstationary time series, demonstrating its effectiveness in identifying stationary segments with power law distributed sizes, and analyzing factors influencing its accuracy.

Contribution

It systematically tests the segmentation algorithm on surrogate data, revealing conditions under which it accurately detects stationary segments and how various parameters affect its performance.

Findings

Algorithm accurately segments power law distributed stationary periods.

Performance is influenced by minimum segment size and fluctuation ratios.

Uncorrelated noise and weak correlations do not significantly impair segmentation.

Abstract

Many phenomena, both natural and human-influenced, give rise to signals whose statistical properties change under time translation, i.e., are nonstationary. For some practical purposes, a nonstationary time series can be seen as a concatenation of stationary segments. Using a segmentation algorithm, it has been reported that for heart beat data and Internet traffic fluctuations--the distribution of durations of these stationary segments decays with a power law tail. A potential technical difficulty that has not been thoroughly investigated is that a nonstationary time series with a (scale-free) power law distribution of stationary segments is harder to segment than other nonstationary time series because of the wider range of possible segment sizes. Here, we investigate the validity of a heuristic segmentation algorithm recently proposed by Bernaola-Galvan et al. by systematically analyzing surrogate time series with different statistical properties. We find that if a given nonstationary time series has stationary periods whose size is distributed as a power law, the algorithm can split the time series into a set of stationary segments with the correct statistical properties. We also find that the estimated power law exponent of the distribution of stationary-segment sizes is affected by (i) the minimum segment size, and (ii) the ratio of the standard deviation of the mean values of the segments, and the standard deviation of the fluctuations within a segment. Furthermore, we determine that the performance of the algorithm is generally not affected by uncorrelated noise spikes or by weak long-range temporal correlations of the fluctuations within segments.