Markus Meierer and Patrick Bachmann (University Zurich and ETH Zurich) shared with us their recent research on generalizing a probabilistic model commonly used by marketing analysts for predicting customers' spending behavior (i.e., the Gamma/Gamma mixture model). Their now modeling approach (a) incorporates time-invariant and time-varying contextual factors, (b) accounts for shifted spending patterns and (c) allows for dependence of consecutive transaction values for individual customers. They demonstrate by analyzing 175,000 customers across three retail businesses scenarios that their novel approach can improve the accuracy of out-of-sample predictions significantly. Error rates can be reduced by up to 50%.

Thanks to Markus and Patrick for sharing their insights with out!