Online Hoeffding Bound Algorithm for Segmenting Time Series Stream Data

In this paper we introduce the ISW (Interval Sliding Window) algorithm, which is applicable to numerical time series data streams and uses as input the combined Hoeffding bound confidence level parameter rather than the maximum error threshold. The proposed algorithm has two advantages: first, it allows performance comparisons across different time series data streams without changing the algorithm settings, and second, it does not require preprocessing the original time series data stream in order to determine heuristically the reasonable error value. The proposed algorithm was implemented in two modes: off line and online. Finally, an empirical evaluation was performed on two types of time series data: stationary (normally distributed data) and non stationary (financial data).