基于集成学习的加权投资组合优化

Weighted Portfolio Optimization Based on Ensemble Learning

文章将机器学习中的分歧分解框架引入投资组合优化问题中,利用将多个子策略权重进行组合的思想,在全局最小方差(GMV)策略的基础上提出了全局最小方差集成(EGMV)策略.具体地,文章利用机器学习领域中对二次损失函数所进行的常见分解方式-分歧分解,对GMV策略的优化问题进行了修改,在其基础上引入了两个额外的参数,即子策略个数和多样化系数,从而构成了新的EGMV投资组合策略.当多样化系数大于1时,EGMV策略能够输出具有多样化权重的多个子策略,从而对冲各资产权重的估计误差,提高加权策略的样本外绩效表现.为了验证EGMV策略的有效性,文章在A股和美股市场上对EGMV策略,GMV策略和其他多个常见策略进行了实证比较.结果显示,在A股市场中,EGMV策略能够在夏普率和换手率上取得平均意义上优于GMV策略的绩效表现,且这一结论在160个不同参数组合下同样成立,这表明EGMV策略具有较好的稳健性.

By introducing the framework of ambiguity decomposition in machine learning domain into the optimization problem and combining multiple portfolio weights,this study proposes a new portfolio strategy called ensemble global minimum variance(EGMV) strategy,which outperforms the classical global minimum variance(GMV) strategy.Specifically,based on the common method of decomposing quadratic error function in machine learning domain,the ambiguity decomposition,we change the optimization problem of GMV,and introduce two extra parameters,the number of sub-strategies and diversity parameter to build the new strategy,EGMV.When the diversity parameter is larger than 1,EGMV can generate more diversified component weights to hedge against the estimation error in different assets,improving the out-of-sample performance of the weighted strategy.In order to verify the effectiveness of the EGMV approach,we test the EGMV strategy in the Chinese and US stock markets,and compare it with other common portfolio strategies.The empirical results show that in Chinese A-share market,EGMV can outperform GMV on average,in terms of the out-of-sample Sharpe ratio and turnover,and remains robust under 160 parameter combination sets,indicating good stability of EGMV.