Forecasting intermittent demand with dynamic Poisson mixtures: a score-driven approach

Forecasting intermittent demand has become increasingly important in retail, leading to the development and adoption of new forecasting methods like the score-driven (SD) approach. In this study, we propose a dynamic SD Poisson mixture model for intermittent time series forecasting, formulated within a general dynamic mixture model framework. The proposed model accommodates overdispersed data while preserving parameter recursions analogous to those of the standard SD Poisson model. We evaluate the model using real-world Walmart sales data, benchmarking its performance against existing SD-based forecasting models from the literature. Our results indicate that while the mixture model outperforms the basic SD Poisson model, it does not surpass the predictive accuracy of the SD negative binomial model.