媒體情緒於企業違約預警：基於公開資訊語意分析

Abstract

本文建立基於公開新聞資訊媒體情緒量化指標的企業違約預警模型。採用美 國多數著名報社之新聞文本，我們以 VADER 的文字探勘技術萃取攸關企業信用 風險之資訊內涵，進而構建基於情感傾向、強度以及新聞報導量的媒體情緒量化 指標 (SENTI)。羅吉斯迴歸模型之下的實證結果顯示，納入媒體情緒量化指標能 有效提升模型違約預警的準確度。特別地，本文發現公開新聞資訊之負向報導有 助於降低財務危機企業被誤判為不具財務危機企業的可能（其模型之型一誤差從 而降低），進一步的數值結果更表明，誤差極小化之下所求取的最適違約判別臨 界值，能有效降低型一誤差從而產生更優化的實質損失分類預測效果，此研究發 現呼應了 Begley et al. (1996)的實證結果。

This paper proposes a corporate default prediction model where media sentiment is derived from public news. Using prevalent news media of several major newspaper publishers in the U.S., we apply the VADER (Valence Aware Dictionary for sEntiment Reasoning) text mining technique to extract information that associate with the firms’ default risk, and the SENTI indicator—characterized by the news contents’ emotion tendency, intensity, and coverage—is then derived to quantify media sentiment. Our logistic regression results show that, incorporating SENTI can enhance the accuracy performance of corporate default prediction. In particular, with negative media sentiment, a lower probability of the model in predicting default firms as non-default ones can be observed - resulting in the model’s Type-I forecasting error being decreased accordingly. Further numerical evidence confirms that, when adopting an optimal threshold subject to minimized errors, a significant decrease in Type-I error can be arrived at, giving rise to the best classification forecasts of default loss scenarios. This finding is consistent with that of Begley et al. (1996).