Default Prediction Using Consumption Data

Current practice in the development of credit risk assessment models relies on a large amount of client data, such as account balance information, socio-demographic data, tenure features, or credit history of the client. With the increasing number of different data sources, it is possible to employ nonconventional information to assess clients’ credit risk. Consumption data are an example of such a source, as they provide an overview of clients’ spending habits. However, considering this information at a level of individual companies where clients’ purchases took place inherently leads to sparse data, as customers typically shop at only a small subset of the available retailers. We demonstrate that commonly used models suffer from poor performance caused by the sparsity of the consumption data, and applying dimensionality reduction techniques such as PCA does not offer a meaningful improvement. In this paper, we propose two word2vec based models that leverage the contextual information of clients’ spending habits. Both of the developed models achieved significantly higher Gini scores, outperforming the baseline models by as much as 72%.