Hierarchical neural additive models for interpretable demand forecasts

Demand forecasts are the basis for numerous business decisions, ranging from inventory management to strategic facility planning. While machine learning approaches offer accuracy gains, they notoriously lack interpretability and acceptance. To address this dilemma, we introduce hierarchical neural additive models (HNAMs) for time series. HNAMs expand upon neural additive models by introducing a time-series-specific additive model consisting of level and covariate effects. Covariates may interact only according to a user-specified hierarchy. For example, given the hierarchy weekday, holiday, promotion, weekday effects are estimated independently, whereas a holiday effect depends on the weekday, and a promotional effect is conditioned on both the weekday and holiday. Thereby, HNAMs clearly attribute additive effects to their respective covariates, enabling intuitive forecasting interfaces with which analysts can interact. We provide benchmarks against established machine learning and statistical models on real-world data to reveal HNAMs’ competitive accuracy.