Extracting narrative data via large language models for loan default prediction: when talk isn’t cheap

We aim to examine how the Large Language Model (LLM) can contribute to loan default prediction by extracting narrative data. Based on a Chinese FinTech lending platform dataset, we employ four LLMs to predict the probability of default (PD-LLM) based on the narrative data and use the PD-LLM as an additional feature to predict default loans. The empirical results show that the narrative data contain some extra credit information and can hardly be regarded as ‘cheap talks’. The extracted information via LLMs processes some predictive capability to predict default loan applications in both in- and out-of-sample analysis. The out-of-sample results indicate that including PD-LLM can significantly improve out-of-sample forecasting performance. At the same time, the rule-based linguistic characteristics and Word-Frequency-based Models hinder out-of-sample forecasting.