Graph convolutional networks on customer/supplier graph data to improve default prediction

Company default prediction is a top priority for governments, credit institutions, individual investors, and society in general. However, standard financial health analysis is normally based on individual behavior, lacking the holistic view of a company’s financial environment. Usually, such analysis is based on yearly/quarterly companies balance sheets that only reflect a definite financial snapshot, as well as alerts related to the transactional company’s abnormal behavior. Even further, these kinds of models improve when considering the company’s sector and sales, which are also modelled from an individual point of view. However, companies interact on a daily basis with their customers and suppliers, forming networks and holding exposures to other companies financial problems and instabilities. Here, we attempt to overcome these issues by studying how graph analysis techniques may be applied on the top of a customer/supplier graph to improve default prediction. Our experiments performed with real-world data demonstrate that when graph metrics are calculated over standard directed and transformed undirected cocitation customer/supplier graphs and graph convolutional networks are applied, future default patterns can be better predicted.