Distributional Regression for Demand Forecasting in e-grocery

E-grocery offers customers an alternative to traditional brick-and-mortar grocery retailing. Customers select e-grocery for convenience, making use of the home delivery at a selected time slot. In contrast to brick-and-mortar retailing, in e-grocery on-stock information for stock keeping units (SKUs) becomes transparent to the customer before substantial shopping effort has been invested, thus reducing the personal cost of switching to another supplier. As a consequence, compared to brick-and-mortar retailing, on-stock availability of SKUs has a strong impact on the customer's order decision, resulting in higher strategic service level targets for the e-grocery retailer. To account for these high service level targets, we propose a suitable model for accurately predicting the extreme right tail of the demand distribution, rather than providing point forecasts of its mean. Specifically, we propose the application of distributional regression methods - so-called Generalised Additive Models for Location, Scale and Shape (GAMLSS) - to arrive at the cost-minimising solution according to the newsvendor model. As benchmark models we consider linear regression, quantile regression, and some popular methods from machine learning. The models are evaluated in a case study, where we compare their out-of-sample predictive performance with regards to the service level selected by the e-grocery retailer considered