摘要：Conservatism is a notorious shortcoming of the worst-case robust portfolio selection model. Numerous studies have been done to tackle this issue from the perspective of theoretical and practical. Based on the existing literature, this paper aims to develop less conservative portfolio models. When the assumption of normality for returns is not valid, higher-order moments have been demonstrated effective in improving portfolio performance. Hence, the worst-case mean-VaR optimization portfolio involving the higher-order moments is developed in this work. Additionally, the machine learning-based preselection is also designed and implemented for selecting risky assets to further overcome the potential conservatism. In the numerical experiments, the US 12 industry portfolio data set from Kenneth R. French is used to test and compare the proposed portfolio models and baseline strategies. The out-of-sample results show that the proposed portfolios have better comprehensive performance than benchmarks.
