Applications of the inverse infection problem on bank transaction networks

The Domingos-Richardson model, along with several other infection models, has a wide range of applications in prediction. In most of these, a fundamental problem arises: the edge infection probabilities are not known. To provide a systematic method for the estimation of these probabilities, the authors have published the Generalized Cascade Model as a general infection framework, and a learning-based method for the solution of the inverse infection problem. In this paper, we will present a case-study of the inverse infection problem. Bankruptcy forecasting, more precisely the prediction of company defaults is an important aspect of banking. We will use our model to predict these bankruptcies that can occur within a three months time frame. The network itself is built from the bank’s existing clientele for credit monitoring issues. We have found that using network models for short term prediction, we get much more accurate results than traditional scorecards can provide. We have also improved existing network models by using inverse infection methods for finding the best edge attribute parameters. This improved model was already implemented in August 2013 to OTP Banks credit monitoring process, and since then it has proven its usefulness. Copyright Springer-Verlag Berlin Heidelberg 2015