Dynamic relationship identification for abnormality detection on financial time series

초록

In this paper, we propose a novel strategy that identifies the dynamic relationship pattern for abnormality detection on financial time series. In particular, we select the basis indices that affect financial time series to discover the spurious relationships and construct a dynamic relationship matrix to model these. Then, we propose a graph embedding model by modifying the structural deep network embedding model to map these relationships into an embedding space. The abnormality is detected by using the outlier detection methods. To evaluate the proposed model, we have conducted the experiments by using the real financial time series (e.g., Apple, Amazon, Coke, Starbucks, and McDonald's). The results showed that the proposed model achieved higher accuracy than the baselines by 4%. © 2021 Elsevier B.V.

키워드