A1067
Title: A prompt-based deep learning method for leveraging textual information in enhancing default prediction
Authors: Zongxiao Wu - University Of Edinburgh (United Kingdom) [presenting]
Yizhe Dong - Univerisity of Edinburgh (United Kingdom)
Yaoyiran Li - University of Cambridge (United Kingdom)
Baofeng Shi - Northwest Agriculture & Forestry University (China)
Abstract: As digitalisation technologies flourish, financial institutions tend to incorporate vast amounts of unstructured data, particularly textual assessments, to mitigate information asymmetries in lending decision-making processes. A novel, prompt-based deep learning method is proposed to extract information from multiple textual assessments provided by loan borrowers and loan officers. Using a micro-small enterprise dataset, the effectiveness of two modes of the proposed method (off-the-shelf prompting and fine-tuned prompting) is explored in enhancing default prediction. Importance measures are then employed to examine the feature importance of textual variables against standard variables. The results show that both modes can effectively extract information from textual assessments, with the fine-tuned prompting mode displaying superior performance. Although texts alone are surprisingly powerful at predicting default, combining standard data and texts yields even stronger results. The feature importance of textual variables is also found to considerably surpass that of standard variables. Overall, the study underscores the substantial potential of texts in improving default prediction and provides a series of recommendations for collecting loan assessments. The proposed prompt-based learning approach also contributes methodologically to multidisciplinary research utilizing text mining.