基于AdaBoost的投资组合优化  

Portfolio Optimization Based on AdaBoost

作　　者：钱龙 韦江 赵慧敏[3] 倪宣明 QIAN Long;WEI Jiang;ZHAO Huimin;NI Xuanming(School of Economics&Management,Tsinghua University,Beijing 100084;Students'Affairs Office,Tsinghua University,Beijing 100084;School of Business,Sun Yat-Sen University,Guangzhou 510275;School of Software&Microelectronics,Peking University,Beijing 100871)

机构地区：[1]清华大学经济管理学院,北京100084 [2]清华大学学生处,北京100084 [3]中山大学管理学院,广州510275 [4]北京大学软件与微电子学院,北京100871

出　　处：《系统科学与数学》2022年第2期271-286,共16页Journal of Systems Science and Mathematical Sciences

基　　金：国家自然科学基金(71991474)资助课题。

摘　　要：文章利用AdaBoost集成学习技术提升均值-方差(MV)策略的表现.首先,文章对二次期望效用损失函数进行了分歧分解,从理论上表明集成学习技术有助于提升投资组合策略的表现.其次,文章将收益率均值和协方差压缩估计量中的压缩强度设定为样本外绩效驱动,并利用迭代有效集法和梯度下降法最大化效用的值函数,从而构建了参数化的MV策略作为文章策略AdaBoost.PT的弱学习器.在实证方面,文章利用A股近25年和美股近40年的全股票样本数据,考察了集成投资组合策略在夏普率、标准差、换手率和最大回撤方面的样本外表现,并利用假设检验对夏普率差异的显著性进行验证.基于因子组合数据集的实证结果显示,基于收益率均值压缩估计量的集成策略在4个评估指标下和差异性统计检验中能够取得优于基准策略的结果,此外,使用行业组合数据集的稳健性检验同样显示出一致的结果.This paper adopts the AdaBoost ensemble learning technique to boost the performance of mean-variance(MV) strategy.Firstly,this paper conducts an ambiguity decomposition on the quadratic cost function of expected utility,which proves that ensemble learning can boost the performance of portfolio strategies.Secondly,we parameterize the shrinkage intensity of the mean and covariance shrinkage estimator of return to be out-of-sample driven,and use iterative active set and gradient descent algorithms to maximize the value function,constructing parameterized MV strategy as the weak learner of proposed AdaBoost.PT.In terms of empirical study,we utilize the full panel stock data of A shares in near 25 years and American shares in near 40 years,and examine the performance of ensemble portfolio strategies in terms of Sharpe ratio,standard deviation,turnover and maximum drawdown,and then conduct a hypothesis test to check the significance of Sharpe ratio’s difference.The empirical results show that the ensemble strategies based on return shrinkage estimator are superior to the baseline strategies under all four indices and statistical tests,and the robust tests based on industrial portfolios also show the same results.

关 键 词：投资组合优化 集成学习 分歧分解 

分 类 号：F831.51[经济管理—金融学] F224