Temporal point processes for purchase categories forecasting

Abstract

Forecasting events in banking sphere is very popular topic nowadays because there is much amount of data collected by different information systems such as banks and the availability of data brings the opportunity to make many interesting applications for predictive models based on this data. For example, you can personalize offers based on forecasting customer needs, by predicting the expected time and category of next purchase taking into account the transaction history of a customer. One of the approach to make a predictive and generative model for this task is temporal point processes, which is widely used to make predictions of random events. The problem that comes with building such models is that while learning on a set of clients the resulting model generalizes the behavior of all sources and makes the averaged predictions such that we lose any individuality of a client that we make a prediction on. In this study, we propose the models, based on the temporal point processes framework with some ideas to solve the issue described earlier - first is to learn on different activity populations, and the second is based on model individual scaling. Each of the innovations provides an improvement in the error metrics compared to the original model.