A0311
Title: Simultaneous component decomposition and anomaly detection in financial time series
Authors: Subin Jeong - Chungnam National University (Korea, South) [presenting]
Minsu Park - Chungnam National University (Korea, South)
Abstract: Anomaly detection algorithms in financial time series have been developed through various studies. In time series data, not only seasonality and trend but also unknown fluctuations such as noise are included. An algorithm that can detect even skewed points is proposed with high frequandwell as removing components such as seasonal-trend in time series data. The proposed algorithm goes through two processes. First, a noise-robust signal decomposition method is applied using the statistical, empirical mode decomposition technique to decompose signals into intrinsic mode functions with unique frequencies, filtering out low-frequency signals such as seasonality and trend. Second, a generalized outlier detection approach that can be applied to skewed distributions was used through the first intrinsic mode function among the decomposed signals. Through various real data and simulated data, the proposed algorithm properly detects the influence values generated in the sparsely dense part of the asymmetric distribution, and the smoothing spline-based empirical mode decomposition method clearly decomposes the signals between high and low frequencies, resulting in a good performance. Through this method, the proposed algorithm is expected to be effectively applied to detect anomalies in nonlinear, non-stationary, and skewed time series data with trend and seasonal variations.
