BFF

Abstract

As data acquisition is rapidly increasing, fast monitoring of data streams is necessary. Properties of these unending data sequences are unknown, where they may exhibit change points, trends and seasonality. Methods applied must be robust to such behaviour. In this work we focus on the challenge of sequential change point detection in the presence of trend. The majority of existing methods are not capable of dealing with such behaviour resulting in a large number of false positives. This paper proposes a Bayesian formulation of an adaptive estimation procedure to learn the distribution of the data stream with sequential updating forms for reduced computational burden. Additionally, a change point procedure is proposed where the estimated posterior distribution is used. Our procedure is capable of detecting change points in the presence of trend which we justify mathematically. We demonstrate the optimal performance of our proposed methodology to competitive sequential and batch procedures on large simulated and real data that exhibit trend.