Improving Momentum Strategies using Adaptive Elastic Dynamic Mode Decomposition
Yusuke Uchiyama, Kei Nakagawa · 2021
Dynamic Mode Decomposition (DMD) is a new method proposed in the field of fluid analysis that expresses the dynamics of multivariate time series data by superposition of modes corresponding to stable, neutral and unstable manifolds. DMD does not explicitly require the governing equation of the multivariate time series, and extracts the structure of spatiotemporal dynamics only from the data. DMD is also a dimension reduction method which makes it possible to extract essential low-dimensional spatiotemporal features embedded in high-dimensional dynamics. This feature can be used to extract modes corresponding to trend components from complex time evolutions of multivariate time series. In this study, we propose Adaptive Elastic DMD (AEDMD), which is an extension of DMD, to extract sparse spatiotemporal structure, and use it to improve traditional momentum strategy in finance. Specifically, AEDMD is applied to the price series to estimate the price trend based on the spatiotemporal structure behind them. Buying and selling based on the estimated trend by AEDMD, we demonstrate that it is possible to surpass the momentum strategy based on the simple past trend.