Nowadays, churn prediction models in non-contractual settings are gaining increasing interest. In a non-contractual setting the exact moment of customers dropout is unknown. The popular approach to identify active customers is to fit parametric probability model and then infer the probability of being alive from the model and customer’s datum.
This approach is employed in extension to NBD model (Donald G. Morrison 1988), Pareto/NDB model (David C. Schmittlein 1987) or BG/NBD model (Fader et al. 2005). But despite the respect these models were earned, they can’t utilize time- dependent covariates apart from recency and frequency.
However, in real-life settings, many other time-dependent covariates are available, for example seasonality or scheduled promotional events. We developed the extension of BG/NBD model, which is able to utilize any kind of covariates, including time-dependent variables and monetary values from transactions.
Proposed model demonstrated improvements of churn prediction in comparison with BG/NBD model on a real dataset.
Please fill in the short form below to receive a copy of this white paper