Detecting novelties in time series through neural networks forecasting with robust confidence intervals

Abstract

Novelty detection in time series is an important problem with application in a number of different domains such as machine failure detection and fraud detection in financial systems. One of the methods for detecting novelties in time series consists of building a forecasting model that is later used to predict future values. Novelties are assumed to take place if the difference between predicted and observed values is above a certain threshold. The problem with this method concerns the definition of a suitable value for the threshold. This paper proposes a method based on forecasting with robust confidence intervals for defining the thresholds for detecting novelties. Experiments with six real-world time series are reported and the results show that the method is able to correctly define the thresholds for novelty detection.