Anomaly Detection on Financial Time Series by Principal Component Analysis and Neural Networks

TL;DR

This paper introduces a novel anomaly detection method for financial time series using principal component analysis for feature extraction and neural networks for scoring, improving risk model accuracy.

Contribution

It combines PCA and neural networks in a new way for anomaly detection, with a data-driven cutoff calibration, outperforming existing algorithms.

Findings

High and stable detection performance on synthetic and real data

Reduced value-at-risk estimation errors with anomaly correction

Effective anomaly detection without hand-set parameters

Abstract

A major concern when dealing with financial time series involving a wide variety ofmarket risk factors is the presence of anomalies. These induce a miscalibration of the models used toquantify and manage risk, resulting in potential erroneous risk measures. We propose an approachthat aims to improve anomaly detection in financial time series, overcoming most of the inherentdifficulties. Valuable features are extracted from the time series by compressing and reconstructingthe data through principal component analysis. We then define an anomaly score using a feedforwardneural network. A time series is considered to be contaminated when its anomaly score exceeds agiven cutoff value. This cutoff value is not a hand-set parameter but rather is calibrated as a neuralnetwork parameter throughout the minimization of a customized loss function. The efficiency of theproposed approach compared to several well-known anomaly detection algorithms is numericallydemonstrated on both synthetic and real data sets, with high and stable performance being achievedwith the PCA NN approach. We show that value-at-risk estimation errors are reduced when theproposed anomaly detection model is used with a basic imputation approach to correct the anomaly.