GNN-GAT-LSTM: A Graph Neural Network-Based Early Warning Model for Enterprise Financial Risk in Dynamic Inter-Firm Networks

Abstract

Traditional enterprise financial early warning schemes mainly focus on the iandividual financial data of the enterprise itself, ignoring the risk transmission effect in complex economic relationship networks such as supply chain and guarantee chain between enterprises, leading to a drop in early warning accuracy. This paper constructed an enterprise financial risk (EFR) early warning scheme (GNN-GAT-LSTM) drawing on graph neural network (GNN), GAT (graph attention network), and LSTM (long short-term memory), aiming to comprehensively consider the complex correlation between enterprises and thus boost the precision of early warning. To capture local structural information in enterprise association graphs, this study first employs a Graph Neural Network (GNN) architecture with a hierarchical message-passing mechanism to aggregate information from the target enterprise's first-and second-order neighbor nodes, encoding the local topology into a 64-dimensional feature vector. Subsequently, a Graph Attentional Transformer (GAT) with an 8-head self-attention mechanism is introduced to dynamically quantify the risk contributions of different associated enterprises to the target enterprise using differentiable attention coefficients, achieving differentiated weighted fusion of neighbor node information. Finally, considering the time-series characteristics of financial data, 128-dimensional node embeddings generated from three consecutive time slices are sequentially constructed as sequence inputs. Using Long Short-Term Memory (LSTM) units for modeling, this approach captures the dynamic evolution of risk characteristics, forming a risk representation that combines structural and temporal awareness.The empirical outcomes display that the precision, recall, SPEC, and RSE (Relative Squared Error) of the GNN-GAT-LSTM model are 0.9842, 0.9752, 0.9658, and 0.032 respectively; by comparing the warning error rates under different risk levels, the average error rates of the scheme in the three categories of financial health, mild financial crisis and severe financial crisis are 0.97%, 2.8% and 1.9% respectively. In the risk warning of financial industry enterprises, the AUC (Area Under the Curve) value was 0.9624, which was 8.86% higher than the traditional GNN model (0.8841); compared with the latest heterogeneous hypergraph neural network HHGNN (0.9206), it still maintained a 4.54% advantage improvement.The GNN-GAT-LSTM scheme recommended in this article can effectively identify the complex relationships and dynamic changes between enterprises and maintain high recognition accuracy under different false alarm rates.