Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Leveraging Spatiotemporal Graph Neural Networks for Multi-Store Sales Forecasting

Manish Singh, Arpita Dayama

TL;DR

This work tackles multi-store sales forecasting by explicitly modeling inter-store dependencies with a spatiotemporal graph neural network (STGNN). The STGNN employs a learnable adjacency, a dilated temporal convolutional backbone, and graph convolution with residual connections to predict log-differences $Y_{\text{diff}}$ and reconstruct dollar sales, trained on 100 epochs with AdamW. Across 45 Walmart stores, STGNN achieves the lowest Normalized Total Absolute Error, best 90th percentile MAPE, and minimal variance in MAPE compared to ARIMA, LSTM, and XGBoost, demonstrating the value of relational structure in interconnected retail environments. The results reveal meaningful functional store clusters and influential nodes discovered without geographic metadata, underscoring the practical impact of relational modelling for robust, scalable demand forecasting and offering a pathway to dynamic graphs and attention-enhanced architectures in retail analytics.

Abstract

This work evaluates the effectiveness of spatiotemporal Graph Neural Networks (GNNs) for multi-store retail sales forecasting and compares their performance against ARIMA, LSTM, and XGBoost baselines. Using weekly sales data from 45 Walmart stores, we construct a relational forecasting framework that models inter-store dependencies through a learned adaptive graph. The proposed STGNN predicts log-differenced sales and reconstructs final values through a residual path, enabling stable training and improved generalisation. Experiments show that STGNN achieves the lowest overall forecasting error, outperforming all baselines in Normalised Total Absolute Error, P90 MAPE, and variance of MAPE across stores. Analysis of the learned adjacency matrix reveals meaningful functional store clusters and high-influence nodes that emerge without geographic metadata. These results demonstrate that relational structure significantly improves forecast quality in interconnected retail environments and establishes STGNNs as a robust modelling choice for multi-store demand prediction.

Leveraging Spatiotemporal Graph Neural Networks for Multi-Store Sales Forecasting

TL;DR

This work tackles multi-store sales forecasting by explicitly modeling inter-store dependencies with a spatiotemporal graph neural network (STGNN). The STGNN employs a learnable adjacency, a dilated temporal convolutional backbone, and graph convolution with residual connections to predict log-differences and reconstruct dollar sales, trained on 100 epochs with AdamW. Across 45 Walmart stores, STGNN achieves the lowest Normalized Total Absolute Error, best 90th percentile MAPE, and minimal variance in MAPE compared to ARIMA, LSTM, and XGBoost, demonstrating the value of relational structure in interconnected retail environments. The results reveal meaningful functional store clusters and influential nodes discovered without geographic metadata, underscoring the practical impact of relational modelling for robust, scalable demand forecasting and offering a pathway to dynamic graphs and attention-enhanced architectures in retail analytics.

Abstract

This work evaluates the effectiveness of spatiotemporal Graph Neural Networks (GNNs) for multi-store retail sales forecasting and compares their performance against ARIMA, LSTM, and XGBoost baselines. Using weekly sales data from 45 Walmart stores, we construct a relational forecasting framework that models inter-store dependencies through a learned adaptive graph. The proposed STGNN predicts log-differenced sales and reconstructs final values through a residual path, enabling stable training and improved generalisation. Experiments show that STGNN achieves the lowest overall forecasting error, outperforming all baselines in Normalised Total Absolute Error, P90 MAPE, and variance of MAPE across stores. Analysis of the learned adjacency matrix reveals meaningful functional store clusters and high-influence nodes that emerge without geographic metadata. These results demonstrate that relational structure significantly improves forecast quality in interconnected retail environments and establishes STGNNs as a robust modelling choice for multi-store demand prediction.

Paper Structure

This paper contains 38 sections, 16 equations, 6 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Dataset and Feature Engineering
  4. Dataset Description
  5. Temporal and Calendar Features
  6. Lag Features and Rolling Statistics
  7. Sales Transformations
  8. Data Cleaning and Store-Level Imputation
  9. Final Feature Set
  10. Methodology
  11. Traditional and Machine Learning Baselines
  12. ARIMA (1,0,1) with Exogenous Inputs
  13. LSTM with Log Transformation
  14. XGBoost Regression
  15. Spatiotemporal Graph Neural Network (STGNN)
  16. ...and 23 more sections

Figures (6)

  • Figure 1: Distribution of weekly store-level sales after aggregation.
  • Figure 2: Weekly sales time series for a representative store
  • Figure 3: Reordered learned adjacency matrix showing functional store clusters and high-influence nodes identified by the STGNN
  • Figure 4: MAPE by model for selected stores
  • Figure 5: MAPE distribution per model across all stores
  • ...and 1 more figures