A GNN-Based Framework for Bank Creditworthiness Prediction Using Interbank Networks

The purpose of this work is to evaluate whether a model that also incorporates information on the structure of the interbank market performs better with respect to models that focus only on financial statement data on predicting banks’ cred itworthiness. This study also introduced the theory behind the Interbank Money market (IMM) and previous literature. Then it presents the concept of graphs and, in particular, how to use them to represent the banking system. Subsequently, the Multi Layer Perceptron (MLP) and Long Short-Term Memory (LSTM) models were introduced that will be used as baseline to compare them with the Transformed Graph Attention Representation (TGAR) model. This model combines a Graph Attention Network (GAT) network with a Transformer architecture, which together enable the extraction of information about the topology of the data. The results showed that during periods of high market stress, evaluated using historical data of the Volatility Index (VIX), the TGAR model outperforms the baseline. This result is very important for rating agencies and regulators as they should focus both on the financial structure of the analyzed banks, but also on the financial institutions with which they are interacting both directly and indirectly.