Marrying Theory and Practice: Product Network-Inspired Deep Learning for Sales Forecasting

Abstract

Problem definition: Accurate sales forecasting is increasingly critical in operations management (OM), particularly in e-commerce operations. While deep learning (DL) techniques improve predictive accuracy, they are often criticized for their "black-box" nature and lack of connection to OM theoretical foundations. This study explores how integrating OM theories and domain knowledge into DL can improve forecast accuracy.

Methodology/results: We propose a novel DL approach, Product Network-enhanced Sales Forecasting (PNet-SF), inspired by substitutable and complementary product relationships modeled as product networks. PNet-SF integrates a long short-term memory layer, a graph convolution layer, and an attention layer, allowing it to capture both historical time series patterns and interdependencies among products. Empirical analysis using data from JD.com demonstrates its effectiveness and performance superiority over conventional DL models that do not incorporate product networks. We further show that PNet-SF is robust and generalizable.

Managerial implications: This study contributes to the OM literature by demonstrating the value of integrating OM theories and domain knowledge into DL for sales forecasting. The proposed PNet-SF approach is easy to implement with enhanced interpretability and provides implications for e-commerce platforms regarding what information to share with merchants to improve operational effectiveness.

Keywords: sales forecasting, product network, deep learning, attention mechanism