Robust online signal extraction from multivariate time series

We introduce robust regression-based online filters for multivariate time series and discuss their performance in real time signal extraction settings. We focus on methods that can deal with time series exhibiting patterns such as trends, level changes, outliers and a high level of noise as well as periods of a rather steady state. In particular, the data may be measured on a discrete scale which often occurs in practice. Our new filter is based on a robust two-step online procedure. We investigate its relevant properties and its performance by means of simulations and a medical application.