Boosting Moving Average Reversion Strategy for Online Portfolio Selection: A Meta-learning Approach

• Published in: Database Systems for Advanced Applications Vol. 10178; pp. 494 - 510
• Main Authors: Lin, Xiao, Zhang, Min, Zhang, Yongfeng, Gu, Zhaoquan, Liu, Yiqun, Ma, Shaoping
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2017; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Portfolio Optimization; Portfolio Selection; Price Relative; Temporal Heterogeneity; Transaction Cost
• ISBN: 3319556983; 9783319556987
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-319-55699-4_30

In this paper, we study the online portfolio selection problem from the perspective of meta learning for mean reversion. The online portfolio selection problem aims to maximize the final accumulated wealth by rebalancing the portfolio at each time period based on the portfolio prices announced before. Mean Reversion is a typical principle in portfolio theory and strategies that utilize this principle achieve the superior empirical performances so far. However there are some important limits of existing Mean Reversion strategies: First, the mean reversion strategies have to set a fixed window size, where the optimal window size can only be chosen in hindsight. Second, most existing mean reversion techniques ignore the temporal heterogeneity of historical price relatives from different periods. Moreover, most mean reversion methods suffer from noises and outliers in the data, which greatly affects the performances. In order to tackle the limits of previous approaches, we exploit mean reversion principle from a meta learning perspective and propose a boosting method for price relative prediction. More specifically, we generate several experts where each expert follows a specific mean reversion policy and predict the final price relatives with meta learning techniques. The sampling of multiple experts involves mean reversion strategies with various window sizes; while the meta learning technique brings temporal heterogeneity and stronger robustness for prediction. We adopt online passive-aggressive learning for portfolio optimization with the predicted price relatives. Extensive experiments have been conducted on real-world datasets and our approach outperforms the state-of-the-art approaches significantly.