機器學習下建構ESG股息波動因子投資組合

Abstract

因應投資人需求，全球資產管理規模成長迅速，其中ESG (Environmental, Social and Governance, ESG)、股息與波動為長期投資人熱門選擇標的。本文使用美國證券市場於2003到2019年的資料，透過機器學習XGBoost與歷史因子投資法預測未來股息波動因子特性，並以粒子群最佳化 (Particle Swarm Optimization, PSO) 建構限制資產數量與權重的最佳化投資組合，本文探討議題與實證結果歸納為以下四點：(1) 比較歷史因子投資與機器學習兩種方法之預測能力，兩者皆具相當程度的預測能力，且機器學習預測能力較佳，其中機器學習之重要特徵變數為過去殖利率、波動度、本益比；(2) 分別針對歷史因子投資與機器學習預測法建構Markowitz投資組合，機器學習下之因子投資最接近正確股息波動投資組合表現；(3) 利用PSO配置限制資產數量的投資組合，能夠達到Markowitz全樣本投資組合之績效；(4) 比較全體與ESG資料集結合股息波動因子表現，ESG結合股息波動因子對於投資組合的績效表現有正向關係。

In response to the needs of investors, the scale of global asset management has grown rapidly. ESG, high dividends, and low volatility are popular choices for investors in long-term. In the study, data from U.S. securities market from 2003 to 2019 are used to predict the characteristics of future dividend and volatility factors through machine learning XGBoost model and historical factor investing method. Furthermore, PSO is used to construct optimized portfolio with limits of the number of assets, maximum and minimum weight. The empirical results and main topics are summarized into the following three points: (1) Compare the predictability of dividend and volatility between historical factor investing and machine learning methods, both have great predictive ability and ability of machine learning is better. The important characteristic variables of machine learning prediction are historical dividend, volatility, and price-to-earnings ratio. (2) The performance of portfolio with dividend yield and volatility by machine learning is closer to correct data than historical factor investing method. (3) Using PSO to construct portfolio with a limited number of assets can achieve the performance of Markowitz`s full sample portfolio. (4) ESG combined with high dividend and low volatility has a positive relationship with portfolio performance.