Review of Context-Aware DL-Based Models to Improve Grocery Retail Forecasting

There a few factors that make grocery retail forecasting particularly challenging: intricate demand patterns are recorded based on many contextual effects such as promotions, seasonality, local events or weather and fickle shopping habits. Nonlinear relationships between the independent and dependent variables are not well captured in conventional statistical models, which can lead to economies of scale not being optimised, and can result in over or under stocking; unnecessary stockouts, waste, and lost sales. Recent progress has been made towards context-aware deep learning (DL) for grocery retail forecasting and we provide an overview in this review paper. We systematically review the ways different DL architectures such as RNNs, LSTM networks, CNNs, Transformers and hybrid models are combined with different types of context information streams in an effort to improve forecast accuracy. The paper classifies contextual information into internal (e.g. pricing, promotions, inventory) and external (e.g. weather, calendar events, social trends). We discuss the methodological advances in the fields of feature engineering and multimodal data fusion, as well as attention mechanisms that compute to what extent each context is relevant dynamically. We also explore the issues of insights’ applicability in large retail chains, related to data quality, computational complexity modelling interpretability, and scalability. Through comparing the performances of models between previous publications and real-world case studies, it is found that context-aware DL approaches have greater potential to overcome traditional methods in dealing with high-dimensional, signal noise and non-stationarity of retail data. The paper also discusses emerging trends, including graph neural networks for product relationship modeling and federated learning for privacy-preserving forecasting, and suggests avenues for future research to close the gap between academic innovation and industry deployment.