Real time anomaly detection in multivariate time series.

Real world datasets rarely obey the idealised assumptions we make when modelling their behaviour. Heavy tails, complex dependencies between time series, and non-trivial time varying properties all contribute to breakdowns in model performance. This breakdown is extremely prevalent in detecting anomalous behaviour, as any method which is not robust to these will either raise these as anomalies, or lose power against any real anomalous behaviour.

This project, jointly funded by EPSRC and Morgan Stanley aims to move beyond these idealised assumptions, and to detect anomalous structure in the presence of all of these complicated behaviours. Advancements in this direction would vastly widen the applicability of anomaly detection techniques to practitioners in fields where the data does not obey our simplifying assumptions, including finance, atmospheric physics and other complex domains such as audio or sensor data.