Online Portfolio Selection Using Signature Feature Set

On-line portfolio selection is an important and challenging question in computational finance and attracts increasing attention in the machine learning and data mining fields. Previously an on-line portfolio selection strategy called On-Line Moving Average Reversion (OLMAR) (Li et al., 2012) exploits the multi-period Moving Average Revision (MAR) pattern of the markets, which is provable to be superior than other benchmark methods on several empirical datasets. However, OLMAR would fail when the market does not exhibit the multi-period MAR pattern. In this paper, we propose a new method to give a more robust and accurate prediction for the next price relatives using the signature feature set, and design the OLSFS strategy by combining the signature based estimator with the OLMAR to improve the performance of the existing OLMAR so that it can be applied to a much more general market condition. We examined both synthetic datasets and empirical datasets to show the robustness and superior performance of our proposed OLSFS