REFER: randomized online factor selection framework for portfolio management

Portfolio management is a critical problem in both machine learning and finance communities. To predict the returns of assets, existing studies have been leveraging side information to mine price-sensitive indicators, i.e., factors. However, with the brisk expansion of factor collection, existing factor selection methods face two main issues, i.e., high-cost and low-precision. In this paper, we first formalize the task of online factor selection as an online learning problem where a learner selects a set of factors in each round and aims to minimize the long-term regret. Then, we propose a Randomized onlinE Factor sElection fRamework, named REFER, which not only is particularly devised to address the above two issues, but also can widely serve for any existing multifactor models. Specifically, by studying the regret of existing factor-selecting policies, we propose two randomized policies along with their bandit variants that achieve sublinear regrets. The bandit variants further improve computational efficiency, and achieve a balanced trade-off between cost and precision. Finally, both theoretical analysis and extensive experiments on the real-world dataset demonstrate the effectiveness of our proposed framework and policies in terms of comprehensive evaluation criteria.