Demand forecasting for fast-moving products in grocery retail

Abstract

Demand forecasting is a critically important task in grocery retail. Accurate forecasts allow the retail companies to reduce their product spoilage, as well as maximize their profits. Fast-moving products, or products with a lot of sales and fast turnover, are particularly important to forecast accurately due to their high sales volumes.

We investigate dynamic harmonic regression, Poisson GLM with elastic net, MLP and two-layer LSTM in fast-moving product demand forecasting against the naive seasonal forecasting baseline. We evaluate two modes of seasonality modelling in neural networks: Fourier series against seasonal decomposition. We specify the full procedure for comparing forecasting models in a collection of product-location sales time series, involving two-stage cross-validation, and careful hyperparameter selection. We use Halton sequences for neural network hyperparameter selection.

We evaluate the model results in demand forecasting using hypothesis testing, bootstrapping, and rank comparison methods. The experimental results suggest that the dynamic harmonic regression produces superior results in comparison to Poisson GLM, MLP and two-layer LSTM models for demand forecasting in fast-moving products with long sales histories. We additionally show that deseasonalization results in better forecasts in comparison to Fourier seasonality modelling in neural networks.