Online portfolio selection with parameterized characteristics

Online portfolio selection sequentially allocates wealth among a set of assets and aims to maximize the investor’s cumulative return in the long run. Various existing algorithms in the finance and accounting area adopt an indirect approach to exploit one asset characteristic through the channel of assets’ expected return and thus cannot fully leverage the power of various asset characteristics found in the literature. This study aims to propose new algorithms to overcome this issue to enhance investment performance.

We propose a parameterized portfolio selection (PPS) framework, which directly incorporates multiple asset characteristics into portfolio weights. This framework can update parameters timely based on final performance without intermediate steps and produce efficient portfolios. We further append L1 regularization to constrain the number of active asset characteristics. Solving the PPS formulation numerically, we design two online portfolio selection (OLPS) algorithms via gradient descent and alternating direction method of multipliers.

Empirical results on five real market datasets show that the proposed algorithms outperform the state of the arts in cumulative returns, Sharpe ratios, winning ratios, etc. Besides, short-term characteristics are more important than long-term characteristics, and the highest return category is the most important characteristic to improve portfolio performance.

The proposed PPS algorithms are new end-to-end online learning approaches, which directly optimize portfolios by asset characteristics. Such approaches thus differ from existing studies, which first predict returns and then optimize portfolios. This paper provides a new algorithmic framework for investors’ OLPS.