A new method for detecting abrupt shifts in time series

Abstract

Abrupt shifts in time series are a topic of growing interest in a number of research areas. They can be caused by a range of different underlying dynamics, for example, via a mathematical bifurcation, or potentially as the result of an auto-correlated stochastic process (i.e. ‘red’ noise). Here we present a method that detects abrupt shifts by searching for gradient changes that occur over a short space of time. It can be automated, allowing many time series to be analysed by the user at once, such as from high spatial resolution data. Our method detects abrupt shifts regardless of their origin (which it cannot deduce). We present a comparison with the method of abrupt shift detection from the changepoint R package, which is based on changes in mean over the time series. Our method performs better on data with an underlying trend where comparisons of means may fail.