On-line portfolio selection

Abstract

On-line portfolio selection, aiming to sequentially determine optimal allocations across a set of assets, is a fundamental research problem in computational finance. This thesis investigates this problem by conducting a comprehensive survey and presenting a family of new strategies using machine learning techniques. The major contributions of this thesis are summarized as follows.

First of all, a new strategy named "CORrelation-driven Nonparametric learning" (CORN) is presented to overcome the limitation of existing pattern matching based strategies which adopt Euclidean distance to measure similarity of two patterns.

Second, unlike most existing approaches based on the trend-following principle, we develop new strategies by applying online learning techniques to exploit ”mean reversion", an important phenomenon in financial markets. In particular, two strategies are presented, that is, “Passive Aggressive Mean Reversion" (PAMR), which is based on the first order passive aggressive online learning method, and ”Confidence Weighted Mean Reversion" (CWMR), which is based on the second order confidence-weighted online learning method.

While the above mean reversion strategies (PAMR and CWMR) are shown to achieve good empirical performance on many real data sets, they implicitly make a single-period mean reversion assumption, which does not always hold, leading to poor performance on some real data sets. To overcome the limitation, we assume multiple-period mean reversion, or so-called “Moving Average Reversion" (MAR), and present a new on-line portfolio selection strategy named ”On-Line Moving Average Reversion'' (OLMAR), which exploits MAR by applying on-line learning techniques. Empirically, OLMAR is able to overcome the drawbacks of the existing mean reversion algorithms by achieving significantly better results, especially on the datasets where the existing mean reversion algorithms fail.

Finally, we conduct an extensive set of empirical studies for evaluating the performance of the proposed algorithms in comparison to the state-of-the-art algorithms. Our empirical results showed that (i) the proposed algorithms generally outperform the state of the art in terms of the cumulative return and risk-adjusted return; and (ii) the proposed algorithms are highly efficient and scalable for large-scale on-line portfolio selection in real-world applications.