ESG 事件嚴重度與股價反應：台灣股市之衡量、事件研究與機器學習預測

Abstract

本研究探討台灣股市如何對 ESG 負面事件定價。以 TEJ（台灣經濟新報）之事件雷達分數（Event Radar Score, ERS）作為事件來源，2018 至 2025 年間共辨識出 9,680 筆事件，涵蓋環境（E）、社會（S）、公司治理（G）三個支柱。本研究以一條核心問題貫穿三章：事件如何辨識？市場整體如何反應？哪些公司特徵可事前預測反應幅度？ 第三章以 ERS 每日分數變動量（daily delta）低於 −5 定義事件，並與 TEJ ESG 事件資料庫（ERS 分數之底層事件來源）交叉比對，99.8% 之 ERS 事件可對應至同日之事件紀錄，支柱分類一致性達 84.8%，嚴重度相關係數為 0.461。本研究並建構台灣市場日頻 Fama–French 三因子（MKT−Rf、SMB、HML），樣本期間涵蓋 1,925 個交易日，以取代過去文獻常用之市場調整模型。 第四章以三因子殘差為異常報酬進行事件研究。整體 CAR(0, 20) 顯著為負；分支柱觀察，G 事件之 CAR(0, 20) 為 −0.91%（t = −4.44，p < 0.01），至第 20 日仍未收斂；S 事件呈短暫負向反應，E 事件則於所有窗口均不顯著。嚴重度梯度完整，CAR(0, 20) 絕對值自輕微事件之 0.12% 遞增至嚴重事件之 2.19%，差距約 18 倍。橫斷面迴歸進一步顯示嚴重度與估值（log P/B）為最穩健之解釋變數，週轉率於三因子殘差下亦達顯著。 第五章以機器學習預測 CAR(0, 20)。比較 OLS、Ridge、ElasticNet 與 XGBoost 四模型，僅使用 ERS 特徵（L1）時資訊係數（IC）至多 0.05（線性模型 0.02、XGBoost 0.05）；加入公司層級基本面（L2）後 IC 跳升至 0.14–0.17，再加入產業虛擬變數（L3）幾無增益。此一跳躍證實 ESG 事件嚴重度須搭配公司特徵才具有預測力。SHAP 解釋顯示過去 12 月報酬、P/B、波動率、週轉率為最重要之四項驅動因子；P/B 與週轉率亦同時為第四章迴歸中最顯著之兩項公司特徵變數，呈現跨方法之一致性。閾值、特徵滯後與驗證／測試集時間間隔三項穩健性檢驗均支持主要結論。 本研究之貢獻在於：(1) 提出可重複、且與 TEJ 事件標記層一致之事件辨識方法；(2) 建構台灣市場 FF3 因子並升級事件研究基準；(3) 跨方法確認 G 支柱事件之差異化定價效率；(4) 以機器學習與 SHAP 橋接傳統與資料驅動分析。

This thesis asks how the Taiwan stock market prices negative ESG events. I use the Event Radar Score (ERS) from the Taiwan Economic Journal (TEJ) to identify 9,680 events between 2018 and 2025, covering the Environmental (E), Social (S), and Governance (G) pillars. The thesis is built around one question answered in three steps: how do we identify events, how does the market react on average, and which firm features can predict the size of the reaction? Chapter 3 defines an event as a day when the ERS daily change falls below −5. Cross-checking against the underlying TEJ ESG event database gives a 99.8% same-day match rate, 84.8% pillar-label agreement, and a 0.461 Pearson correlation between my severity ranking and the database's per-event intensity tag. I also build a daily Fama–French three-factor series (MKT−Rf, SMB, HML) for Taiwan over 1,925 trading days, replacing the market-adjusted model used in earlier Taiwan ESG event studies. Chapter 4 runs an event study using FF3 residuals as abnormal returns. The pooled CAR(0, 20) is significantly negative. By pillar: G events have CAR(0, 20) = −0.91% (t = −4.44, p < 0.01) and do not recover within 20 days; S events show a short-lived drop; E events are not significant at any horizon. Severity produces a clean, monotonic gradient, with |CAR(0, 20)| rising from 0.12% (mild) to 2.19% (severe), an 18-fold gap. Cross-sectional regressions confirm severity and valuation (log P/B) as the most robust drivers; turnover becomes significant once systematic factors are removed. Chapter 5 asks whether CAR(0, 20) is predictable. Comparing OLS, Ridge, ElasticNet, and XGBoost across three feature sets: ERS features alone (L1) yield information coefficient (IC) of at most ~0.05 (linear models at 0.02, XGBoost at 0.05); adding firm fundamentals (L2) lifts IC to 0.14–0.17 across all four models; adding industry dummies (L3) adds almost nothing. Event severity predicts the reaction only when paired with firm-level features. SHAP identifies past 12-month returns, P/B, volatility, and turnover as the top four drivers; P/B and turnover are also the two strongest firm-level variables in the Chapter 4 regression. Three robustness checks (event threshold, feature lag, validation–test gap) support the main result. The contributions are fourfold: (1) a reproducible event identification method that is cross-consistent with TEJ's own qualitative event tagging; (2) a Taiwan-market FF3 factor series and an upgraded event-study baseline; (3) cross-method evidence that governance events drive most of the cross-sectional signal; and (4) a bridge between traditional regression and machine-learning analysis using SHAP.