Finance Research Seminar – Professor Yizhe Dong

Title: Predicting from Heterogeneous Texts: A Source-preserving Prompt-based Approach

Date: 19 November 2026

Time: 13:30 to 15:30

Venue: NUBS.4.06

If you would like to attend, please register using the following link

Predicting from Heterogeneous Texts: A Source-preserving Prompt-based Approach

Speaker: Professor Yizhe Dong, University of Edinburgh

Yizhe Dong – University of Edinburgh Research Explorer

Abstract:

Textual data used in decision-making is often produced by multiple agents with distinct informational roles and incentives. Existing approaches typically either analyze each text source in isolation or combine heterogeneous texts into shared representations that obscure source-specific meaning and reduce interpretability. We propose a source-preserving, prompt-based framework that converts each text source independently into an interpretable propensity score using cloze-style masked language modeling. We apply the framework to a unique dataset of micro and small enterprise loans containing 15 heterogeneous text variables. Across off-the-shelf and fine-tuned settings, our approach consistently outperforms structured variable-only models and established natural language processing benchmarks. We find that loan officer narratives contain greater predictive information than borrower self-reports, with direct implications for how lenders collect and weight soft information. Profit analysis shows that the predictive gains translate into economically meaningful improvements in lending decisions. These findings contribute to the literature on text-based financial decision-making by showing how heterogeneous narratives can be integrated without collapsing their source-specific information, and by providing an interpretable design for high-stakes predictive analytics.