A Residual-Corrected Hybrid ARIMA–CNN–LSTM Framework for High-Accuracy Tobacco Sales Forecasting in Regulated Markets

• Published in: International journal of computational intelligence systems Vol. 18; no. 1; pp. 1 - 25
• Main Authors: Huang, Shiyu, Zhou, Lili
• Format: Journal Article
• Language: English
• Published: Dordrecht Springer Netherlands 27.07.2025; Springer Nature B.V; Springer
• Subjects: Accuracy; ARIMA; Artificial Intelligence; Artificial neural networks; Autoregressive moving-average models; Computational Intelligence; Control; Deep learning; Engineering; Forecasting; Hybrid deep learning; Machine learning; Mathematical Logic and Foundations; Mechatronics; Optimization; Public health; Residual correction; Robotics; Sales; Time series analysis; Tobacco; Tobacco sales forecasting; Trend analysis; Hybrid deep learning; Time series analysis; ARIMA; Tobacco sales forecasting; Residual correction
• ISSN: 1875-6883; 1875-6891; 1875-6883
• DOI: 10.1007/s44196-025-00930-4

As a common consumer product threatening public health, tobacco not only hinders the development of national public health, but also plays a significant impact on the national economy. The ARIMA model is reliable in learning linear or regular relationships, while the deep learn, such as convolutional neural network (CNN) and long short-term memory network (LSTM), is superior when capturing and learning nonlinear relationships. Combining time-series forecasting models with deep learning technologies, the hybrid architecture could integrate advantages and optimize forecasting effect. In this paper, leveraging 2023 daily sales data from a Southern Chinese tobacco company, this study proposes a new hybrid deep learning framework that integrates ARIMA, CNN, and LSTM models to address these inherent limitations and enhance prediction accuracy. This architecture decomposes forecasting tasks into linear trend analysis and nonlinear residual learning. The ARIMA component learns the linear relationship, and the CNN–LSTM component plays the role in the residual-driven correction. They enable synergistic capture of temporal dependencies and localized anomalies and enhancing the fitting effect. This hybrid model's optimization primarily relies on the residual-driven correction mechanism in the CNN–LSTM component, which significantly enhanced the model interpretability ( R 2 : 0.95, enhance 10.5% compare with ARIMA model, enhance 13.1% compare with CNN-LSTM model). This research not only advances hybrid deep learning methods, but also provides a scalable solution for precise predictions in dynamic markets. This excellent forecasting results could also be practiced in inventory optimization and policy impact studies.