In retail analytics, slow-moving-inventory (SMI) refers to goods which rarely sell, resulting in very sparse count processes. Forecasting the sales of such goods is challenging, because traditional predictive models rely on large enough sales volumes to be accurate. In this work, we develop modelling, inferential and predictive methods able to learn the dynamics of sparse count processes for SMI products with few to no sales. We flexibly introduce covariates into the self-exciting model for sparse processes of Porter et al., (2012). We extend the model to include a cross-excitation contribution that allows differing series to excite one another, capturing the process of intertwined contemporaneous excitation dynamics. We integrate individual products into a Bayesian hierarchical model that accommodates shrinkage and information passing across differing sparse count process, without requiring the data for each product to exist over the same time period. We illustrate our methods on a retail analytics dataset from a major supermarket chain in the UK.
Joint work with James Pitkin and Gordon Ross