Combined Mean Reversion Strategy for On-line Portfolio Selection

Abstract

Online portfolio selection has attracted increasing interests in machine learning community and information theory recently. Empirical evidences show that stock price relatives are likely to follow the mean reversion phenomenon in the long term. While Anticor algorithm is shown to achieve good empirical performance on many real datasets, it makes the mean reversion assumption during two time windows, which is not always satisfied, leading to not enough good performance in some real datasets. To overcome the limitation, this article also considers the mean aversion during two time windows and proposes an on-line combined mean reversion strategy (OLCMR), which fully exploits the property of the stock price fluctuation in applying on-line learning techniques. Our empirical results show that OLCMR can overcome the drawbacks of Anticor algorithm and achieve significantly better results.