Real-time financial surveillance via quickest change-point detection methods

We consider the problem of efficient financial surveillance aimed at "on-the-go" detection of structural breaks (anomalies) in "live"-monitored financial time series. With the problem approached statistically, viz. as that of multi-cyclic sequential (quickest) change-point detection, we propose a semi-parametric multi-cyclic change-point detection procedure to promptly spot anomalies as they occur in the time series under surveillance. The proposed procedure is a derivative of the likelihood ratio-based Shiryaev-Roberts (SR) procedure; the latter is a quasi-Bayesian surveillance method known to deliver the fastest (in the multi-cyclic sense) speed of detection, whatever be the false alarm frequency. We offer a case study where we first carry out, step by step, statistical analysis of a set of real-world financial data, and then set up and devise (a) the proposed SR-based anomaly-detection procedure and (b) the celebrated Cumulative Sum (CUSUM) chart to detect structural breaks in the data. While both procedures performed well, the proposed SR-derivative, conforming to the intuition, seemed slightly better.