Bankruptcy prediction using the text-based communicative value of earnings call transcripts
研究了将财报电话会议文本的沟通价值纳入机器学习模型能否提升破产预测效果，发现加入正面和不确定性语调变量能显著改善长期预测，且与信用风险理论一致。
Abstract We examine whether incorporating the text-based communicative value ( TCV ) of earnings call transcripts improves the effectiveness of bankruptcy prediction models within machine learning frameworks, utilizing U.S. firm data from 2005 to 2020. We find that the inclusion of earnings call transcripts TCV variables significantly improves the overall bankruptcy prediction effectiveness in addition to Barboza’s et al. (Expert Syst Appl 83:405-417, 2017) financial variables. Notably, the incremental contribution of earnings call transcripts TCV variables is more pronounced in the future longer-term bankruptcy predictions, aligning with the forward-looking nature of earnings call transcripts and complementing the findings of Chen et al. (Expert Syst Appl 233:120714, 2023). Furthermore, the feature engineering results indicate that the improvement is mainly driven by positive tone and uncertainty tone variables, which more directly capture key determinants in structural-form credit risk models (e.g., asset value, volatility, and incomplete information). These tone variables signal a firm’s prospective financial condition and future asset value distribution, thereby reflecting bankruptcy risk. Consistent with theoretical expectations, positive tone is negatively related to bankruptcy risk while uncertainty tone has the opposite effect. Finally, these results remain robust when annual report TCV variables are included as additional benchmark model input variables.