The Evaluation on the Credit Risk of Enterprises with the CNN-LSTM-ATT Model

Area of Science:

• Artificial Intelligence
• Machine Learning
• Financial Technology

Background:

• Credit evaluation for SMEs is challenging due to high-dimensional, nonlinear enterprise data.
• Traditional models like logistic regression (LR) and random forest (RF) struggle with accuracy and robustness in large feature spaces.
• Recurrent neural networks (RNNs) face gradient disappearance issues during long sequence training.

Purpose of the Study:

• To develop an advanced compound neural network model for improved enterprise credit evaluation.
• To address the limitations of traditional machine learning and RNN models in handling complex financial data.
• To enhance the accuracy and robustness of credit risk assessment for SMEs.

Main Methods:

• Proposed a compound neural network model integrating Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) with a soft attention mechanism (CNN-LSTM-ATT).
• Employed the attention mechanism to facilitate gradient propagation, mitigating issues in deep learning models.
• Conducted multi-model comparison experiments and validation on three distinct enterprise datasets.

Main Results:

• The CNN-LSTM-ATT model outperformed traditional LR, RF, and standalone CNN, LSTM, and CNN-LSTM models in most experimental scenarios.
• Experimental results indicated higher accuracy of the proposed model in multi-model comparisons.
• Group testing demonstrated the enhanced robustness of the CNN-LSTM-ATT model.

Conclusions:

• The CNN-LSTM-ATT model offers a significant advancement in enterprise credit evaluation, particularly for SMEs.
• The integration of CNN, LSTM, and attention mechanisms effectively addresses data complexity and improves model performance.
• The model's superior accuracy and robustness make it a promising tool for financial institutions.