基于负相关集成学习的投资组合优化

Portfolio Optimization Based on Negative Correlation Ensemble Learning

主要利用集成学习中的负相关学习思路构建了一种集成全局方差最小化(EGMV)的投资组合策略.作为机器学习领域最常用的工具之一,集成学习一般用于提升一些弱学习器的表现,其中负相关学习通过主动制造多个具有差异性的弱学习器并进行加权来提升预测效果.借鉴这一思路,在原始GMV优化问题的基础上引入了制造差异性的惩罚项,从而使EGMV能够同时输出多个差异化GMV权重估计量,随后进行等权加权以对冲估计误差.基于A股2000-2019年全样本股票日收益率数据,将EGMV与多个经典投资组合策略进行了夏普率和换手率的对比,发现EGMV相比于原始GMV能够实现比较明显的样本外提升,表明集成学习框架能够被用于提升经典投资组合策略的表现.

This paper utilizes the idea of negative correlation ensemble learning to construct an ensemble global minimum variance(EGMV)portfolio strategy.As one of the common tools in machine learning domain,ensemble learning techniques are often used to boost the performance of some weak learners.Among them,negative correlation learning can increase weak learner’s forecasting ability by intentionally creatig multiple diversified weak learners and averaging.Using this idea,this paper introduces penaly term to create diversity based on the original GMV optimization problem,so that EGMV can simultaneously output multiple diversified GMV weight estimators and average them to hedge the estimation error.Based on full stock panel daily return data from Chinese A share market,this paper compares EGMV with several classical portfolio strategies in terms of Sharpe’s ration and turnover.The empirical result shows that EGMV clearly has better out-of-sample performance compared to original GMV,indicating that ensemble learning framework can be applied to imporve the performance of classical portfolio strategies.