文章结合机器学习中的交叉验证、在线学习和集成学习方法, 对基于不同高维协方差估计量的投资策略权重进行动态组合, 以获得优于传统投资组合策略的样本外表现. 基于这一目标, 文章对机器学习中比较前沿的在线加权集成(online weighted ensemble, OWE) 算法的样本更新方式、学习模型和目标函数进行了替换和修改, 改进后的mixed-OWE算法能够更好地适用于多组合的动态混合策略投资. 通过数值模拟, 文章将mixed-OWE应用在基于二次效用目标函数的投资问题上, 结果表明其样本外表现优于传统静态方法. 随后, 文章进一步使用A股近10年的数据作为样本对mixed-OWE进行了全局最小方差组合投资, 经过一定的参数调整后, mixed-OWE策略实现的组合方差优于其成分组合以及等权重组合.

To acquire better out-of-sample performance than traditional portfolio strategies, this paper mainly adopts machine learning techniques such as cross-validation, online learning and ensemble learning to dynamically combine different portfolio weights based on several large-dimensional covariance estimators. In order to achieve this goal, we select a state-of-art online weighted ensemble (OWE) algorithm in machine learning domain and adjust its way of updating data stream, learning models and objective function. The new mixed-OWE algorithm can be better used to perform investment based on multi-portfolio dynamic combination. Through numerical simulation, we adopt the mixed-OWE on solving a portfolio investment problem based on quadratic utility target function. The result shows that the out-of-sample performance of mixed-OWE is slightly better than traditional static methods. Then, we use real data of A-share over the last ten-years to test the performance of mixed-OWE and the result shows that after some parameter adjustments, it can outperform its component portfolios and 1/$N$ portfolio.