運用機器學習模型分析影響公司風險的ESG因子：以台灣市場為例

Abstract

本研究採用機器學習模型，以模型找出何種ESG因子對於台灣企業的公司風險與股價崩跌有較高的解釋力，拆分企業社會責任對於風險的影響。本次研究用梯度提升決策樹(Gradient Boost Trees，以下稱GDBT)、XGBoost、隨機森林（Random Forest），並使用Refinitiv資料庫中的綜合分數、環境分數、社會分數與公司治理分數下的14個指標作為ESG變數。我採用公司特有風險與兩項股價崩跌風險指標（NCSKEW、DUVOL）作為風險變數，以2010-2019年間的992筆台灣公司股價資料計算而成，期望能探究機器學習模型在ESG評分與公司個別風險的效果。\n而實證結果顯示，使用XGBoost模型與GDBT在對公司風險與股價崩跌風險模型解釋力上比起隨機森林有較佳的表現。進一步透過分析因子重要性後，數據結果顯示ESG分數綜合指標如ESG綜合分數在公司風險的重要度表現較不佳，顯示比起採用社會責任細項指標，投資人若想依照綜合分數作為投資組合風險管理考量，較無效率。社會類別如企業社會責任策略分數、社區分數與人權分數因子中，在公司特有風險與股價崩跌風險當中，皆具有一定程度的影響性，可以作為企業內部風險管理考量上的指標依據。

This study approaches three machine learning models to find out which ESG perfor-mance factors have significant impact on firm specific risk and stock crash risks ( NCSKEW,DUVOL ). Three models such as Gradient Boost Trees (Gradient Boost Trees, hereinafter referred to as GDBT), XGBoost, and Random Forest are applied into analyzing the effect of 14 ESG Performance from Reuters database on firm specific risks and stock crash risks. The sample of the study is mainly based Taiwanese firms in Refinitiv ESG database, ranging from the period of 2010 to 2019.The empirical results show that XGBoost and GDBT have the better performance than Random Forest in ex-plaining the company risk and stock crash risk. Through factor importance analysis, I found that combined ESG score are less important in the part of firm risks. This shows that rather than taking ESG composite score into account, investors should consider in-dividual dimensions of Environmental, Social, and Governance indicators for further portfolio risk management.\nSocial categories, such as corporate social responsibility strategy scores, communi-ty scores, and human rights score, have a certain degree of influence in the firm specific risks and stock crash risks. These scores could be indicators for internal risk manage-ment on the scope of portfolio management and firm risk management.