Hybrid Modeling for Sales Prediction Using SARIMA, CNN, LSTM, and Stacking Ensemble

Abstract

Accurately forecasting sales in dynamic supply chain environments is essential for optimizing inventory management, resource allocation, and operational efficiency. This study addresses the challenge of achieving precise demand predictions by developing a hybrid modeling framework that integrates SARIMA, CNN, LSTM, and stacking ensemble methodologies. Ineffective sales forecasting often leads to overstocking, understocking, increased operational costs, and diminished customer satisfaction, adversely affecting global supply chain stakeholders. The research evaluates the effectiveness of combining traditional statistical models with advanced machine learning techniques for demand forecasting. SARIMA models effectively captured seasonal and linear trends, while CNN and LSTM architectures identified non-linear and temporal dependencies. However, integrating SARIMA with aggregated weekly data and CNN and LSTM models using daily granular data posed significant challenges. This mismatch excluded SARIMA from the initial stacking ensemble (XGBoost) integration. To address this limitation, a hybrid SARIMA-XGBoost model was subsequently developed and evaluated for performance. Limited time for fine-tuning CNN and LSTM models presented another challenge, leading to SARIMA outperforming both CNN and LSTM in predictive accuracy. The SARIMA-XGBoost hybrid model demonstrated superior performance compared to standalone CNN and LSTM models but was slightly less effective than SARIMA alone. The hybrid model excelled at capturing seasonal patterns, external variables, and irregular trends within the dataset. Historical sales data from 45 Walmart stores, augmented with external variables such as the Consumer Price Index (CPI), unemployment rates, temperature, and holiday indicators, formed the basis of the study. The findings revealed SARIMA’s robustness in handling linear and seasonal trends under constrained conditions, while the SARIMA-XGBoost hybrid model provided enhanced predictive accuracy.

This study concludes that hybrid frameworks hold substantial potential for improving demand forecasting, particularly in addressing diverse temporal granularities and resource constraints. Future research should focus on integrating additional external variables and optimizing deep learning models to refine the hybrid framework’s applicability across industries. Such advancements can empower supply chain managers with actionable insights to reduce costs, enhance operational efficiency, and improve customer satisfaction.