Title: Graph convolutional networks for bankruptcy prediction in P2P Bondora market Authors:  Tomas Plihal - Masaryk University (Czech Republic) [presenting]
Oleg Deev - Masaryk University (Czech Republic)
Abstract: The accurate prediction of bankruptcy in peer-to-peer (P2P) lending markets is a critical endeavour, yet current machine learning models often consider only individual attributes without adequately accounting for relational information among borrowers. The aim is to utilize a novel approach that leverages graph convolutional network (GCN) models for predicting bankruptcies in the P2P Bondora market. The proposed model incorporates borrower-specific features, such as credit history and loan information, and enriches this data by exploiting the high-order relational structures among borrowers through graph networks. The choice of GCN is motivated by its efficacy in capturing localized node features and edge connections, thus providing a comprehensive understanding of both node attributes and the graph topology. To validate the approach, its performance is compared against traditional classification models. The model offers insights into the relative importance of borrower attributes and network features, thereby contributing to the understanding of risk factors in P2P lending markets. By synthesizing node-specific attributes with graph-structured data, the model provides a more nuanced and effective tool for risk assessment. The findings have broad implications for enhancing decision-making processes in P2P lending platforms and offer avenues for future research in integrating graph theories and financial risk modelling.