Prediction-based Mean–variance Portfolios with Risk Budgeting Based on Neural Networks
Yilin Ma,Yudong Wang,Weizhong Wang,Chong Zhang
DOI: https://doi.org/10.1016/j.eswa.2023.120638
IF: 8.5
2023-01-01
Expert Systems with Applications
Abstract:The integration of return prediction and portfolio models has been widely investigated and proven effective. The derived prediction-based portfolios can be classified into five main categories. This paper comprehensively compares the out-of-sample performance of these categories based on the traditional mean–variance (MV) model for the first time, where different neural networks are used for return prediction. The MV model and equal-weighted portfolio are used as comparisons. Due to the high volatilities of these prediction-based portfolios, the risk budgeting method is applied to advance these portfolios, where stocks with higher predicted volatilities own higher proportions of portfolio risk contribution. The equal risk contribution is used as the benchmark. Further, the out-of-performance of these advanced portfolios is investigated after deducting transaction fees since their high turnovers can seriously affect the obtained profits in practical investments. Empirical tests are implemented with component stocks of the Shanghai Stock Exchange (SSE) 50 index, where daily data range from 2011 to 2021. Experimental results indicate that the prediction-based MV portfolio built with forecasted errors (MV-FE) is significantly superior to the others under different risk aversion levels, followed closely by the prediction-based MV portfolio built with forecasted returns (MV-FR). In particular, medium risk aversion is more suitable for these portfolios. Moreover, the risk budgeting method considerably reduces the volatilities of these portfolios and improves their risk-adjusted returns. The risk budgeting method is more efficient than the equal risk contribution method. After subtracting their transaction fees of turnovers, the MV-FR portfolio with risk budgeting, i.e., the MV-FR-RB portfolio, outperforms the others and still obtains satisfying performance. Also, there is little difference between the MV-FR and MV-FE portfolios with risk budgeting.