Probabilistic Demand Forecasting with Graph Neural Networks

TL;DR

The paper addresses scalable probabilistic demand forecasting in settings with thousands of interrelated articles. It introduces GraphDeepAR, an end-to-end architecture that couples a GNN encoder with a DeepAR decoder and constructs graphs from article attribute similarity, avoiding fixed graph structures; forecasts are produced as a $t$-distribution with parameters $(\mu, s^2)$. Key contributions include (i) integrating relational graphs into both training and inference for probabilistic forecasts, and (ii) a scalable similarity-based graph construction that yields meaningful article embeddings. Empirical results on three real-world datasets show GraphDeepAR consistently outperforms a non-graph DeepAR baseline in RMSE, MAE, and WMAPE, with pronounced gains on top-selling articles and useful embeddings for downstream tasks, albeit with higher training time. The findings demonstrate practical value for retailers seeking improved inventory and logistics planning through graph-aware forecasting and provide a foundation for further exploration of dynamic graphs and alternative relation signals.

Abstract

Demand forecasting is a prominent business use case that allows retailers to optimize inventory planning, logistics, and core business decisions. One of the key challenges in demand forecasting is accounting for relationships and interactions between articles. Most modern forecasting approaches provide independent article-level predictions that do not consider the impact of related articles. Recent research has attempted addressing this challenge using Graph Neural Networks (GNNs) and showed promising results. This paper builds on previous research on GNNs and makes two contributions. First, we integrate a GNN encoder into a state-of-the-art DeepAR model. The combined model produces probabilistic forecasts, which are crucial for decision-making under uncertainty. Second, we propose to build graphs using article attribute similarity, which avoids reliance on a pre-defined graph structure. Experiments on three real-world datasets show that the proposed approach consistently outperforms non-graph benchmarks. We also show that our approach produces article embeddings that encode article similarity and demand dynamics and are useful for other downstream business tasks beyond forecasting.

Figures (4)

• Figure 1: GraphDeepAR Architecture. The figure uses a single article for illustration.
• Figure 2: Graph Illustration. Article graphs for retail (left) and e-commerce (right).
• Figure 3: GNN article embeddings for retail for week 67 (left) and 75 (right).
• Figure 4: Example predictions of DeepAR and GraphDeepAR on the Retail dataset.