Paper: arXiv 2601.00478

Authors: Zongxiao Wu, Ran Liu, Jiang Dai, Dan Luo

Abstract

Credit risk assessment increasingly relies on diverse sources of information beyond traditional structured financial data, particularly for micro and small enterprises (mSEs) with limited financial histories. This study proposes a multimodal framework that integrates structured credit variables, climate panel data, and unstructured textual narratives within a unified learning architecture. Specifically, we use long short-term memory (LSTM), the gated recurrent unit (GRU), and transformer models to analyse the interplay between these data modalities. The empirical results demonstrate that unimodal models based on climate or text data outperform those relying solely on structured data, while the integration of multiple data modalities yields significant improvements in credit default prediction. Using SHAP-based explainability methods, we find that physical climate risks play an important role in default prediction, with water-logging by rain emerging as the most influential factor. Overall, this study demonstrates the potential of multimodal approaches in AI-enabled decision-making, which provides robust tools for credit risk assessment while contributing to the broader integration of environmental and textual insights into predictive analytics.

Complexity vs Empirical Score

  • Math Complexity: 4.5/10
  • Empirical Rigor: 7.5/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper employs established deep learning architectures (LSTM, GRU, Transformer) with SHAP for explainability, representing moderate mathematical complexity, but the focus is heavily on a specific empirical dataset and backtesting methodology.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Integrate climate & text data<br>for mSE default prediction"] --> B
  
  subgraph B ["Data Inputs"]
      B1["Structured Credit Data"]
      B2["Climate Panel Data<br>Physical Risks e.g., Water-logging"]
      B3["Unstructured Textual Narratives"]
  end

  B --> C["Computational Framework<br>Multimodal Learning Architecture"]
  
  C --> D{"Deep Learning Models"}
  D --> D1["LSTM"]
  D --> D2["GRU"]
  D --> D3["Transformer"]
  
  D --> E["Analysis & Explainability"]
  E --> F["Key Findings & Outcomes"]
  
  subgraph F ["Outcomes"]
      F1["Text & Climate models<br>outperform Structured alone"]
      F2["Multimodal integration<br>yields significant improvement"]
      F3["SHAP Analysis:<br>Physical climate risks crucial<br>Water-logging = Top factor"]
  end

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 a deep learning‑based multimodal framework that uses modality‑specific encoders and intermediate representation‑level fusion to combine structured credit variables, climate panel data, and textual loan assessments for credit default classification.

Source sentence
we propose a multimodal learning framework that employs modality-specific encoders and adopts a representation-level intermediate fusion mechanism to combine heterogeneous representations for credit default classification (Tavakoli et al., 2025).

Assumes 6

  • Default is defined as 30 days past payment due. data
    Source sentence
    Default is defined as 30 days past payment due.
  • The dataset is severely imbalanced with a non‑defaulter to defaulter ratio of 64:1. data
    Source sentence
    We note that the dataset is characterised by a severely imbalanced ratio of 64:1 (non-defaulters: defaulters).
  • Missing continuous features are filled with the mean and missing categorical features are replaced with new categories. data
    Source sentence
    In cases where continuous features are missing, they are filled with the means of the features, while missing values for categorical features are replaced with new categories.
  • Weight of evidence (WoE) encoding assumes a log‑odds relationship between feature bins and default risk. model
    Source sentence
    The WoE value in each category/bin for a given feature is calculated as the logarithm of the proportion of non-defaulters to the proportion of defaulters in that particular
Show 2 more
  • Climate risk factors are relevant to borrowers’ repayment capacity in the agricultural setting. other
    Source sentence
    Given that our loan data is based on the agricultural setting, we can likely attribute this superior performance to the high relevance of climate risk factors such as droughts, floods, and temperature extremes to borrowers’ repayment capabilities.
  • Textual loan officer assessments contain valuable qualitative information about borrower behavior. other
    Source sentence
    These results suggest that loan officers’ written assessments may contain valuable insights that reflect borrowers’ repayment intentions and operational capabilities.

Data

  • Agricultural loan dataset · Chinese bank (national basis) alternative proprietary
    agricultural micro and small enterprises (mSEs)
    Source sentence
    The agricultural loan dataset used in this study was provided by a Chinese bank operating on a national basis

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • Models using only climate data outperform structured‑only models, with the GRU achieving an average AUC of 0.658, KS of 0.332 and H‑measure of 0.189. performance
    Source sentence
    Interestingly, while transformer models generally exhibit superior performance in unimodal analyses, they are slightly outperformed by the GRU model, which achieves average AUC, KS, and H-measure values of 0.658, 0.332, and 0.189, respectively.
  • Text‑only transformer models also beat structured‑only models, attaining an average AUC of 0.661, KS of 0.371 and H‑measure of 0.224. performance
    Source sentence
    Among the three models, the transformer model achieves the best performance, with average AUC, KS, and H-measure values of 0.661, 0.371, and 0.224, respectively.
  • Structured‑only models obtain lower predictive performance (average AUC 0.609, KS 0.301, H‑measure 0.157). performance
    Source sentence
    The average AUC, KS, and H-measure for the model relying solely on structured data (S) are 0.609, 0.301, and 0.157, respectively.
  • Integrating climate and text modalities yields significant improvements over any single modality, with transformer‑based multimodal models achieving the highest prediction accuracy. predictive-accuracy
    Source sentence
    Integrating this information into prediction models may help to capture subtle borrower characteristics that structured data may ignore.
  • SHAP analysis identifies water‑logging by rain as the most influential climate factor for default prediction. statistical
    Source sentence
    Using SHAP-based explainability methods, we find that physical climate risks play an important role in default prediction, with water-logging by rain emerging as the most influential factor.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 42 related papers). Coverage is partial, so this is neither replication nor refutation.

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.

Cited References (90)

  1. A multi-view contrastive learning framework for spatial embeddings in risk modelling — → 2511.17954
  2. State of China’s climate in 2024 (2025) — link
  3. Do firms benefit from carbon risk management? Evidence from the credit default swaps market (2025) — link
  4. A ‘divide and conquer’ reject inference approach leveraging graph-based semi-supervised learning (2025) — link
  5. Climate risk and green innovation‐ESG disconnect: Firm‐level evidence from China (2025) — link
  6. Unleashing the power of text for credit default prediction: Comparing human-written and generative AI-refined texts — → 2503.18029
  7. Climate conditions, credit risk cycles, and technological progress: Evidence from China (2025) — link
  8. A Good Sketch is Better than A Long Speech: Evaluate Delinquency Risk through Real-Time Video Analysis (2024) — link
and 82 more
  1. The power of satellite imagery in credit scoring: a spatial analysis of rural loans (2024) — link
  2. Predicting financial distress using multimodal data: An attentive and regularized deep learning method (2024) — link
  3. Enhancing Credit Risk Assessment in the Era of Climate Change: A Machine Learning Approach for Loan and Lease Portfolios (2024) — link
  4. Climate mitigation and adaptation strategies in the transport sector: An empirical investigation in Germany (2024) — link
  5. Attention-based Dynamic Multilayer Graph Neural Networks for Loan Default Prediction — → 2402.00299
  6. Flood, farms and credit: The role of branch banking in the era of climate change (2024) — link
  7. Being Stranded with Fossil Fuel Reserves? Climate Policy Risk and the Pricing of Bank loans (2024) — link
  8. Credit Risk Meets Large Language Models: Building a Risk Indicator from Loan Descriptions in P2P Lending — → 2401.16458
  9. Predicting financial distress using current reports: A novel deep learning method based on user-response-guided attention (2024) — link
  10. Thermal stress and financial distress: Extreme temperatures and firms’ loan defaults in Mexico (2023) — link
  11. Depicting Risk Profile over Time: A Novel Multiperiod Loan Default Prediction Approach (2023) — link
  12. Extracting narrative data via large language models for loan default prediction: when talk isn’t cheap (2023) — link
  13. Impacts of extreme weather events on mortgage risks and their evolution under climate change: A case study on Florida (2023) — link
  14. Understanding and Attribution of Extreme Heat and Drought Events in 2022: Current Situation and Future Challenges (2023) — link
  15. Causes of a Typical Southern Flood and Northern Drought Event in 2015 over Eastern China (2023) — link
  16. Machine learning in bank merger prediction: A text-based approach (2023) — link
  17. Fairness of ChatGPT and the Role Of Explainable-Guided Prompts (2023) — link
  18. Severe Weather and Peer-to-Peer Farmers’ Loan Default Predictions: Evidence from Machine Learning Analysis (2023) — link
  19. Interpretable machine learning for imbalanced credit scoring datasets (2023) — link
  20. Multi-Modal Deep Learning for Credit Rating Prediction Using Text and Numerical Data Streams — → 2304.10740
  21. Unlocking the Power of Voice for Financial Risk Prediction: A Theory-Driven Deep Learning Design Approach (2023) — link
  22. Loan default prediction using a credit rating-specific and multi-objective ensemble learning scheme (2023) — link
  23. Impacts of Drought on Loan Repayment (2023) — link
  24. Designing and Managing Human-AI Interactions (2022) — link
  25. Designing Transparency for Effective Human-AI Collaboration (2022) — link
  26. Who said what: Mining semantic features for success prediction in reward-based crowdfunding (2022) — link
  27. Climate Change Risk and the Cost of Mortgage Credit (2022) — link
  28. Credit default prediction from user-generated text in peer-to-peer lending using deep learning (2021) — link
  29. On the combination of graph data for assessing thin-file borrowers’ creditworthiness (2021) — link
  30. A transformer-based model for default prediction in mid-cap corporate markets (2021) — link
  31. Do Credit Supply Shocks Affect Employment in Middle-Income Countries? (2021) — link
  32. The Influence of Environmental Pollution and Drought on the Satisfaction With Life of Entrepreneurs in Zambia’s Mining Sector (2020) — link
  33. Crowd, Lending, Machine, and Bias (2020) — link
  34. The value of text for small business default prediction: A deep learning approach (2020) — link
  35. Mining Semantic Soft Factors for Credit Risk Evaluation in Peer-to-Peer Lending (2020) — link
  36. Temperature Shocks and Industry Earnings News (2019) — link
  37. A Deep Bidirectional GRU Network Model for Biometric Electrocardiogram Classification Based on Recurrent Neural Networks (2019) — link
  38. Mortgage Finance and Climate Change: Securitization Dynamics in the Aftermath of Natural Disasters (2019) — link
  39. A prediction-driven mixture cure model and its application in credit scoring (2019) — link
  40. Deep learning models for bankruptcy prediction using textual disclosures (2019) — link
  41. Central bank mandates, sustainability objectives and the promotion of green finance (2019) — link
  42. The impact of natural disasters on the banking sector: Evidence from hurricane strikes in the Caribbean (2019) — link
  43. Construction and Application of a Climate Risk Index for China (2018) — link
  44. Predicting Distresses using Deep Learning of Text Segments in Annual Reports (2018) — link
  45. Towards Bankruptcy Prediction: Deep Sentiment Mining to Detect Financial Distress from Business Management Reports (2018) — link
  46. Climate change challenges for central banks and financial regulators (2018) — link
  47. Can agricultural credit scoring for microfinance institutions be implemented and improved by weather data (2018) — link
  48. Loan default prediction by combining soft information extracted from descriptive text in online peer-to-peer lending (2017) — link
  49. Finance and inequality: The distributional impacts of bank credit rationing (2017) — link
  50. Attention is All you Need (2017) — link
  51. A Unified Approach to Interpreting Model Predictions (2017) — link
  52. Description-text related soft information in peer-to-peer lending – Evidence from two leading European platforms (2016) — link
  53. Benchmarking state-of-the-art classification algorithms for credit scoring: An update of research (2015) — link
  54. Default risk in agricultural lending, the effects of commodity price volatility and climate (2014) — link
  55. On the Properties of Neural Machine Translation: Encoder–Decoder Approaches (2014) — link
  56. Financial fragility and natural disasters: An empirical analysis (2014) — link
  57. Access to Credit, Natural Disasters, and Relationship Lending (2012) — link
  58. An Empirical Investigation of Statistical Significance in NLP (2012)
  59. Loan Portfolio Performance and El Niño, an Intervention Analysis (2011) — link
  60. How do International Financial Flows to Developing Countries Respond to Natural Disasters? (2010) — link
  61. Measuring classifier performance: a coherent alternative to the area under the ROC curve (2009) — link
  62. Detecting and understanding the multi-decadal variability of the East Asian Summer Monsoon: Recent progress and state of affairs (2009) — link
  63. The Credit Scoring Toolkit: Theory and Practice for Retail Credit Risk Management and Decision Automation (2007)
  64. Historical comparison of the 2001/2002 drought in the Canadian Prairies (2007) — link
  65. Global observed changes in daily climate extremes of temperature and precipitation (2006) — link
  66. Credit Risk Models and Agricultural Lending (2005) — link
  67. Long Short-Term Memory (1997) — link
  68. Intelligent decision support systems (1992) — link
  69. A Value for n-person Games (1988) — link
  70. The Effect of AI-Enabled Credit Scoring on Financial Inclusion: Evidence from an Underserved Population of over One Million (2024) — link
  71. Adapting to Climate Risk with Guaranteed Credit: Evidence from Bangladesh (2024) — link
  72. Exogenous Shocks, Credit Reports and Access to Credit: Evidence from Colombian Coffee Producers (2021) — link
  73. A Case Study of the Impact of Climate Change on Agricultural Loan Credit Risk (2021)
  74. Customer churn prediction for commercial banks using customer-value-weighted machine learning models (2021) — link
  75. Housing and Mortgage Markets with Climate-Change Risk: Evidence from Wildfires in California∗ (2021)
  76. OUP accepted manuscript (2021) — link
  77. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (2019) — link
  78. Preparing for the Cognitive Generation of Decision Support (2017)
  79. How Collateral Laws Shape Lending and Sectoral Activity (2016) — link
  80. Evaluating link prediction accuracy on dynamic networks with added and removed Edges (arXiv:1607.07330) (2016)
  81. Does weather matter? How rainfall shocks affect credit risk in agricultural micro-finance (2015)
  82. ENTREPRENEURS AND ENTREPRENEURSHIP IN AFRICA: WHAT IS KNOWN AND WHAT NEEDS TO BE DONE (2002)

5 of 90 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.