An algorithmic approach to investment

Abstract

The objective of this research is to develop a framework for an online low-frequency investment algorithm and compare its performance against the equally-weighted and market-cap weighted benchmark portfolios. The research presents an algorithmic framework for a control system that automates quantitative active portfolio management. The control system sequentially constructs a strategic benchmark portfolio to earn systematic risk premiums, and an active tactical portfolio to exploit temporary opportunities to earn excess returns. Portfolio controls are determined through mean-variance optimisation. The online property of the control system is achieved through recursively estimating the mean-variance inputs with an adaptive algorithm. Inputs for the systematic benchmark portfolio are determined using a rational asset pricing model. Active returns are determined with an ad-hoc predictive model. Risk factors are selected by the investor and excess returns are defined against these factors. Tactical bet size is determined by the relative realised predictive accuracy between systematic and active expected return predictions. Portfolios are robust to estimation errors through offline regularisation of the mean and covariance estimates. A specific configuration of the framework is implemented and tested out-of-sample using approximately ten years of weekly data from the JSE. The out-of-sample results did not demonstrate statistically significant out performance over the benchmarks on a gross and net basis. Sharpe-Ratio point estimates over the out-of-sample period indicated benchmark out performance on a gross basis, but benchmark under performance after incorporating transaction costs.