Paper: arXiv 2503.18029
Authors: Zongxiao Wu, Yizhe Dong, Yaoyiran Li, Baofeng Shi
Abstract
This study explores the integration of a representative large language model, ChatGPT, into lending decision-making with a focus on credit default prediction. Specifically, we use ChatGPT to analyse and interpret loan assessments written by loan officers and generate refined versions of these texts. Our comparative analysis reveals significant differences between generative artificial intelligence (AI)-refined and human-written texts in terms of text length, semantic similarity, and linguistic representations. Using deep learning techniques, we show that incorporating unstructured text data, particularly ChatGPT-refined texts, alongside conventional structured data significantly enhances credit default predictions. Furthermore, we demonstrate how the contents of both human-written and ChatGPT-refined assessments contribute to the models’ prediction and show that the effect of essential words is highly context-dependent. Moreover, we find that ChatGPT’s analysis of borrower delinquency contributes the most to improving predictive accuracy. We also evaluate the business impact of the models based on human-written and ChatGPT-refined texts, and find that, in most cases, the latter yields higher profitability than the former. This study provides valuable insights into the transformative potential of generative AI in financial services.
Complexity vs Empirical Score
- Math Complexity: 4.0/10
- Empirical Rigor: 7.5/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper uses established deep learning and NLP techniques (like LIME and transformer models) but focuses heavily on empirical validation with a specific loan dataset, backtesting profit impact, and comparing multiple generative AI tools, making it highly practical for implementation.
Research Flowchart
flowchart TD A["Research Goal<br>'Unleashing the power of text for credit default prediction:<br>Comparing human-written and generative AI-refined texts'"] --> B["Data & Inputs"] B --> C["Methodology<br>1. Human-written loan assessments<br>2. ChatGPT-refined versions<br>3. Structured loan data"] C --> D["Computational Process<br>Deep Learning Models<br>(e.g., BERT, LSTM, Fusion Models)"] D --> E["Comparative Analysis<br>Length, Semantic Similarity,<br>Linguistic Features"] E --> F["Key Findings & Outcomes"] F --> F1["AI-refined text significantly<br>enhances prediction accuracy"] F --> F2["ChatGPT analysis of borrower<br>delinquency contributes most"] F --> F3["Business Impact:<br>AI-refined models yield<br>higher profitability"] F --> F4["Context-dependent effect<br>of essential words"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes using ChatGPT‑refined loan assessment texts together with structured data to improve credit default prediction models.
Source sentence
Using deep learning techniques, we show that incorporating unstructured text data, particularly ChatGPT-refined texts, alongside conventional structured data significantly enhances credit default predictions.
Assumes 2
- Missing continuous features are imputed with their mean values. data
Source sentence
In cases where continuous features are missing, they are filled with the means of the features
- Default is defined as 90 days past payment due. other
Source sentence
Default is defined as 90 days past payment due.
Data
- mSE loan dataset alternative proprietary
Source sentence
final sample consists of 2,460 loans with durations ranging from 1 to 8 months, originating between December 2008 and July 2009.
Costs
Not applicable
Finds 5
- ChatGPT‑refined texts improve predictive accuracy compared to human‑written texts when added to structured data. predictive-accuracy
Source sentence
While both texts benefit credit scoring models (compared to models using only structured data), ChatGPT-refined texts can lead to more accurate prediction results in most cases.
- Models using both negative and positive factors from ChatGPT‑refined texts generally outperform those using other text choices. performance
Source sentence
models which utilise the negative+positive text generally outperform those using other text choices.
- AI‑refined text is less effective for PRAUC than human‑written text. negative
Source sentence
the PRAUC results indicate that AI-refined text is less effective at predicting credit defaults compared to human-written texts.
- In most cases, models based on ChatGPT‑refined texts yield higher profitability than those based on human‑written texts. performance
Source sentence
find that, in most cases, the latter yields higher profitability than the former.
- ChatGPT's analysis of borrower delinquency contributes the most to improving predictive accuracy. predictive-accuracy
Source sentence
Moreover, we find that ChatGPT's analysis of borrower delinquency contributes the most to improving predictive accuracy.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 36 related papers). Coverage is partial, so this is neither replication nor refutation.
1 paper in the archive
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.