A novel risk-control model for the online portfolio selection of high-frequency transactions

Abstract

Online portfolio selection (OPS) is an active area of study that has recently attracted the attention of researchers from a diverse range of fields. However, the existing models applied to financial markets typically do not consider incremental learning, asset volatility, and the removal of system noise at the eigenvector level. We first propose a bidirectional incremental learning method, the bidirectional-broad learning system (Bi-BLS), which is more flexible and can conduct OPS in two incremental directions. Then, we propose a new method for removing the noise based on random matrix theory and principal component analysis, which can remove the noise at the eigenvalue and eigenvector levels. Furthermore, two novel risk-control metrics, the Adjusted Sharpe Ratio and the Adjusted Information Ratio, are also proposed to treat asset volatility as risk. These ratios first use asset volatility as the denominator other than the standard deviation of portfolio weights. Finally, a series of experiments are conducted to assess model performance. The results show that the proposed model produces a steady increase in cumulative return and achieves the best performance in terms of final cumulative return. This model also outperforms other models (Best Stork, UBAH, BCRP, UCRP, ONS, PAMR, CWMR, OLMAR, CORN, LOAD, NEW, ALAIT and SNN) as measured by risk-control metrics, such as the Sharpe Ratio, Adjusted Sharpe Ratio, Information Ratio, and Calmar Ratio. Therefore, the proposed model can produce excess returns while effectively controlling risk.