Multi-Level Recurrent Neural Network and Attention Mechanism for Credit Risk Assessment in Supply Chain Finance

With more globalization and competition in the marketplace, SCF has become an essential part of the modern financial system. CR management in the supply chain has become more complex and important. Most current CR assessment methods use LR and BP neural networks, which make it difficult to fully explore the long-term and short-term dependencies in dynamic time series data. They also have insufficient processing capabilities for high-dimensional data, resulting in low assessment accuracy. This paper innovatively adopts the nonlinear modeling ability and the capture advantage of long-term and short-term dependencies of multi-level time recursive LSTM in multi-level recurrent neural networks (RNNs), combined with the multi-head attention (MHA) mechanism to evaluate the CR of SCF. The study first uses PCA to select and reduce the characteristic indicators of CR in SCF, and constructs a three-layer LSTM model. Each layer performs sequence modeling on the input CR data of SCF and outputs the hidden state of each time step. The hidden state is then used as the input of the MHA mechanism for multi-head parallel computing. Each attention head is calculated according to its specific query, key and value weights, and a different weight is assigned to each time step. Finally, the outputs of all attention heads are concatenated and mapped to the target dimension using a linear transformation, and sent to the fully connected layer and the Softmax layer for CR classification. The experiment was based on the financial data of new energy vehicle manufacturers such as BYD, Great Wall and Tesla in the Wind public dataset. The results showed that the accuracy of the multi-level LSTM-MHA mechanism model reached 0.96, which was 0.16 higher than the LR model. The cumulative explanation of the 15 feature indicators reached 98.30%. Experiments show that the integration of multi-level RNN and MHA mechanism can effectively improve the accuracy of SCF’s CR assessment and greatly promote the SCF’s development.