Paper: arXiv 2308.03704

Authors: Yancheng Liang, Jiajie Zhang, Hui Li, Xiaochen Liu, Yi Hu, Yong Wu, Jinyao Zhang, Yongyan Liu, Yi Wu

Abstract

Despite the tremendous advances achieved over the past years by deep learning techniques, the latest risk prediction models for industrial applications still rely on highly handtuned stage-wised statistical learning tools, such as gradient boosting and random forest methods. Different from images or languages, real-world financial data are high-dimensional, sparse, noisy and extremely imbalanced, which makes deep neural network models particularly challenging to train and fragile in practice. In this work, we propose DeRisk, an effective deep learning risk prediction framework for credit risk prediction on real-world financial data. DeRisk is the first deep risk prediction model that outperforms statistical learning approaches deployed in our company’s production system. We also perform extensive ablation studies on our method to present the most critical factors for the empirical success of DeRisk.

Complexity vs Empirical Score

  • Math Complexity: 4.0/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper is heavily implementation-focused, describing a multi-stage pipeline with data preprocessing, separate training, and fine-tuning on real-world financial datasets, with extensive ablation studies and production comparison, showing high empirical rigor. While it references deep learning architectures like Transformers, the mathematical exposition is relatively light, focusing more on the framework design and practical techniques rather than dense theoretical derivations.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Develop Deep Learning Framework<br>to Outperform Statistical Models<br>for Credit Risk Prediction"] --> B["Input Data<br>Real-World Financial Data<br>(High-dimensional, Sparse, Noisy, Imbalanced)"]
  B --> C["Key Methodology<br>DeRisk Framework<br>Deep Neural Network Model"]
  C --> D["Computational Process<br>Training on Financial Data<br>Handling Imbalance & Sparsity"]
  D --> E["Key Outcomes<br>DeRisk Outperforms<br>GB/RF Models in Production<br>Critical Factors Identified via Ablation Study"]

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 DeRisk, a deep learning framework for credit risk prediction that outperforms existing statistical learning models in production.

Source sentence
we propose DeRisk, an effective deep learning risk prediction framework for credit risk prediction on real-world financial data.

Assumes 2

  • A multi‑stage process with feature selection and deep‑learning specific engineering is critical to success on low‑quality real‑world financial data. model
    Source sentence
    we observe that a multi-stage process with feature selection and DL-specific engineering processing can be critical to the overall success of our framework.
  • Real‑world financial data are high‑dimensional, sparse, noisy and extremely imbalanced, making deep neural networks challenging to train. data
    Source sentence
    real-world financial data are high-dimensional, sparse, noisy and extremely imbalanced, which makes deep neural network models particularly challenging to train and fragile in practice.

Data

  • Yanqianguan Users credit report dataset other proprietary
    August 2020 to July 2021 · 582,996 users
    Source sentence
    We sample 582,996 Yanqianguan Users and use their credit report data and repayment behavior from August 2020 to July 2021 as the dataset.

Costs

Not applicable The paper reports no trading or execution result.

Finds 4

  • DeRisk outperforms the statistical learning approaches deployed in the company’s production system. performance
    Source sentence
    DeRisk is the first deep risk prediction model that outperforms statistical learning approaches deployed in our company's production system.
  • Oversampling improves AUC and accelerates optimization for the sequential model. predictive-accuracy
    Source sentence
    oversampling (1) improves AUC. (2) accelerates optimization.
  • MLM pre‑training of the sequential model improves performance and speeds up optimization. predictive-accuracy
    Source sentence
    MLM pre-training of the sequential model (1) improves performance. (2) accelerates optimization.
  • Joint models trained on both data formats outperform the separated non‑sequential or sequential models. performance
    Source sentence
    With more data, the joint models outperform either separated models

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 13 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.