應用遷移式自然語言結合 ESG 報章媒體情緒建構企業違約預警模型

Abstract

企業信用風險相關研究一直都是學術界關注的議題，過去已經有不少研究指出信用風險與傳統財務數據相關，例如：帳市比、公司槓桿、股價波動度等等。而近年來各界永續議題的關注度不斷提升，愈來愈多投資者認為環境、社會和公司治理（Environmental, Social, and Governance，簡稱 ESG）議題的表現會影響到企業整體營運狀況，應將企業的 ESG 表現納入投資決策中。然而在信用風險的研究方面，過去的研究主要專注在傳統財務數據等結構化資料上，且較少關注 ESG 因素對信用風險的影響，本研究嘗試結合結構化資料以及非結構化資料，建立機器學習的模型，對信用風險進行預測。結構化資料方面，本研究除了傳統財務數據外，額外加入碳排放量等與 ESG 相關的指標數據；非結構化資料方面，將利用BERT（Bidirectional Encoder Representations from Transformers）模型以及 FinBERT (BERT for Financial Text Mining) 模型，對新聞媒體進行語意分析，從媒體文本中萃取出財務情緒以及 ESG 情緒，最終建立隨機森林模型。本次研究發現，ESG 因子對於信用風險能夠提供有用的資訊，ESG 整體表現愈好的企業，有助於降低信用風險。

Corporate credit risk has been a prominent topic in academia, with previous studies emphasizing the correlation between credit risk and traditional financial data. However, the growing focus on sustainability, particularly Environmental, Social, and Governance (ESG) factors, has led to the need for their inclusion in credit risk research. This study aims to combine structured and unstructured data, incorporating ESG indicators alongside traditional financial metrics. By leveraging machine learning techniques and sentiment analysis on news media using BERT and FinBERT models, a random forest model is developed. The findings reveal that ESG factors provide valuable information, as companies with better ESG performance tend to exhibit reduced credit risk.