Paper: arXiv 2508.02738

Authors: Yumeng Shi, Zhongliang Yang, DiYang Lu, Yisi Wang, Yiting Zhou, Linna Zhou

Abstract

Corporate credit rating serves as a crucial intermediary service in the market economy, playing a key role in maintaining economic order. Existing credit rating models rely on financial metrics and deep learning. However, they often overlook insights from non-financial data, such as corporate annual reports. To address this, this paper introduces a corporate credit rating framework that integrates financial data with features extracted from annual reports using FinBERT, aiming to fully leverage the potential value of unstructured text data. In addition, we have developed a large-scale dataset, the Comprehensive Corporate Rating Dataset (CCRD), which combines both traditional financial data and textual data from annual reports. The experimental results show that the proposed method improves the accuracy of the rating predictions by 8-12%, significantly improving the effectiveness and reliability of corporate credit ratings.

Complexity vs Empirical Score

  • Math Complexity: 4.0/10
  • Empirical Rigor: 8.5/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper uses standard deep learning architectures (GRU, FinBERT) without complex mathematical derivations, but demonstrates strong empirical rigor with a purpose-built dataset (CCRD), specific accuracy improvements (8-12%), and clear implementation details for NLP feature extraction.

Research Flowchart

  flowchart TD
  A["Research Goal: <br> Improve Credit Rating Accuracy <br> by Integrating Annual Report Insights"] --> B

  subgraph B ["Methodology & Data"]
      direction LR
      B1["Dataset: CCRD<br>Financial + Annual Reports"] --> B2["Feature Engineering<br>FinBERT for Text Features"]
  end

  B --> C["Computational Process<br>Deep Learning Model<br>Integrated Financial & Text Features"]

  C --> D["Key Findings & Outcomes"]
  D --> D1["8-12% Accuracy Improvement"]
  D --> D2["Enhanced Reliability of Ratings"]
  D --> D3["Unlocks Value of Unstructured Data"]

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 novel framework that integrates financial data with features extracted from corporate annual reports using large language models.

Source sentence
In response, we propose a novel framework that integrates financial data with features from corporate annual reports.

Assumes 5

  • Annual reports are divided into N batches for processing by the Transformer encoder. model
    Source sentence
    For an annual report, it is assumed to be divided into N batches.
  • Corporate credit rating is treated as a multi‑class classification problem over seven rating categories. model
    Source sentence
    Corporate credit rating is formulated as a multi-class classification task.
  • The original 23 rating grades are consolidated into 7 major categories for the experiments. data
    Source sentence
    the 23 rating grades were consolidated into 7 major categories: AAA, AA, A, BBB, BB, B, and CCC.
  • Financial numerical features are extracted using existing CNN, GNN, or RNN models and then combined with annual report features. model
    Source sentence
    we try to supplement the aforementioned financial numerical feature extraction models MFNF with the features of the annual report extracted by large models
Show 1 more
  • The dataset contains 2,307 samples with 20 financial attributes and annual report documents. data
    Source sentence
    This process resulted in a new dataset containing 2,307 samples, which integrates 20 financial attributes and annual report documents.

Data

  • Comprehensive Corporate Rating Dataset (CCRD) · constructed by authors using Kaggle financial data and annua fundamentals public
    2010-2016 · U.S. publicly listed companies
    Source sentence
    we have developed the Comprehensive Corporate Rating Dataset (CCRD), a large-scale dataset that combines traditional financial data with annual report data.

Recognised sources: Kaggle

Costs

Not applicable The paper reports no trading or execution result.

Finds 3

  • Adding annual report features (ARF) improves overall rating accuracy by approximately 8% to 12% across all models. predictive-accuracy
    Source sentence
    After adding annual report features (ARF), the accuracy improved by approximately 8% to 12% across all models.
  • The BB rating category shows the largest precision improvement, exceeding 20% when ARF are incorporated. performance
    Source sentence
    Among the seven rating categories, the BB category showed the largest precision improvement, exceeding 20%.
  • GNN‑based baseline models benefit the most from ARF integration, with performance gains close to 12%. performance
    Source sentence
    Among the three baseline models, the GNN-based models benefited the most from the integration of ARF features, with an improvement close to 12%.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 17 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 (28)

  1. Mining Emotions of the Public from Social Media for Enhancing Corporate Credit Rating (2018) — link
  2. Social Media Analytic-Based Corporate Credit Rating Forecasting (2024)
  3. Investment Value Evaluation of Listed Companies Based on Machine Learning (2023)
  4. Predicting distresses using deep learning of text segments in annual reports (2019)
  5. Every corporation owns its image: Corporate credit ratings via convolutional neural networks (2020)
  6. A novel end-to-end corporate credit rating model based on self-attention mechanism (2020)
  7. Every corporation owns its structure: corporate credit rating via graph neural networks (2022)
  8. Adversarial semi-supervised learning for corporate credit ratings (2021) — link
and 20 more
  1. Contrastive pre-training for imbalanced corporate credit ratings (2022)
  2. Multi-task Envisioning Transformer-based Autoencoder for Corporate Credit Rating Migration Early Prediction (2022)
  3. SparseGraphSage: A Graph Neural Network Approach for Corporate Credit Rating (2024)
  4. Credit rating analysis with support vector machines and neural networks: a market comparative study (2004)
  5. Application of deep neural networks to assess corporate credit rating (2020) — link
  6. Predicting corporate credit rating based on qualitative information of MD&A transformed using document vectorization techniques (2020)
  7. Multi-Modal Deep Learning for Credit Rating Prediction Using Text and Numerical Data Streams — → 2304.10740
  8. A hybrid KMV model, random forests and rough set theory approach for credit rating (2012)
  9. A comparative study on base classifiers in ensemble methods for credit scoring (2017)
  10. The impact of news articles and corporate disclosure on credit risk valuation (2016)
  11. The information value of credit rating action reports: A textual analysis (2016)
  12. An empirical analysis of changes in credit rating properties: Timeliness, accuracy and volatility (2009)
  13. Feature selection in corporate credit rating prediction (2013)
  14. Credit ratings and credit risk: Is one measure enough? (2017)
  15. ZETATM analysis A new model to identify bankruptcy risk of corporations (1977)
  16. Multivariate adaptive regression splines (1991)
  17. On the predictive ability of narrative disclosures in annual reports (2010)
  18. FinLlama: Financial Sentiment Classification for Algorithmic Trading Applications — → 2403.12285
  19. Finbert: A pre-trained financial language representation model for financial text mining (2021)
  20. Bert: Pre-training of deep bidirectional transformers for language understanding (2018) — link

2 of 28 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.