Modeling Similarities Among Multi-Dimensional Financial Time Series

Pairs trading is one of the most successful strategies for stock investment. The performance of pairs trading heavily depends on modeling how similarity of two paired financial signals. Conventional methods measure similarity based on one-way or two-way signal, ignoring multiple information sources. In this paper, we propose a tensor-based framework to capture the intrinsic relations among multiple factors. Equities data is represented by tensors in firm-time-trading modes, on which tensor decomposition method is applied to seek a set of multilinear patterns for each mode. In this process, structural information is preserved which provides supplementary information for pairs trading. Experiments on stocks data of all constituent firms of S&P500 demonstrate the superior performance of the proposed framework when compared with some state-of-the-art methods.