From ARIMA to LSTM: Evaluating Traditional and AI-Based Models for Accurate Retail Sales Forecasting

Abstract

Accurate forecasting of sales in the retail industry is fundamental to managing inventory, optimizing the supply chain, reducing costs, and improving customer satisfaction. Traditional statistical approaches are effective in detecting seasonality and linear patterns, but often unsuitable to employ when modeling non-linear trends, external factors and live data. This paper addresses the development of machine learning (ML) algorithms for the previously-stated analytical problems in the context of retail sales forecasting models, using various approaches involving complementing traditional time series models with machine learning. We will model in time series as ARIMA, SARIMA, and expontial smoothing; while employing different machine learning sounding approaches using techniques of: Random Forest, XGboost, LSTM and Prophet. Within the modelling, we will establish the performance as an integrated hybrid approach. The findings show while you can achieve what might be considered superior non-traditional forecasting using machine learning models - specifically LSTM networks and deep learning - on complex and high dimensional retail datasets, any machine learning approach would outperform similar models of traditional statistics.

Introduction

Retail sales forecasting is critical for managing revenue, logistics, and customer service, but consumer demand is volatile due to seasonality, promotions, holidays, and economic factors. Traditional statistical models like ARIMA and Exponential Smoothing work well for stable, linear patterns but struggle with complex, multi-variable fluctuations. Machine learning (e.g., Random Forest, XGBoost) and deep learning models (e.g., LSTM) can handle non-linear dependencies, multivariate inputs, and long-term temporal patterns, offering more accurate and adaptive forecasts.

This study used a multi-year retail dataset including sales, stores, items, promotions, holidays, and external factors. Data preprocessing included handling missing values, encoding categorical variables, normalization, and feature engineering for seasonality and temporal effects. Forecasting models were evaluated using MAE, RMSE, and MAPE. Results showed traditional methods perform adequately in stable scenarios, but ML and LSTM models outperform them, especially in complex, multivariate contexts. Historical sales and promotional indicators were the most important predictors. While ML models are computationally intensive, they provide more accurate and robust forecasting for modern retail environments.

Conclusion

This study highlights the increasing importance of using machine learning and deep learning methods in forecasting retail sales. While traditional time series models (for example, ARIMA and ETS) will continue to be important because of their simplicity and interpretability, they are increasingly limited in terms of using them for complex, non-linear, multivariate data. Tree-based machine learning methods, such as XGBoost, have been shown to be highly effective at identifying relative short- and medium-term patterns, particularly when applied with reasonably strong feature engineering. For long-term sequential dependency, LSTM networks were clearly superior and exhibited high predictive accuracy in dynamic retail settings. However, advanced applications raise practical issues, such as greater computational cost and requirement for high-quality data. Ultimately, the forecast model used depends on the particular business scenario, length of the forecasting period, and operation risks and constraints. Careful attention to data preparation, feature engineering, and model interpretability is critical to ensuring these models are successfully used in retail forecasting applications.

References

[1] M. M. Rahman, “riskAIchain: AI-Driven IT Infrastructure—Blockchain-Backed Cyber Risk Management,” Journal of Risk and Financial Management, vol. 12, no. 12, p. 206, Dec. 2024. https://doi.org/10.3390/jrfm12120206 (MDPI) [2] P.-O. Goffard and S. Loisel, “Collaborative and Parametric Insurance on the Ethereum Blockchain,” arXiv preprint arXiv:2412.05321, Dec. 2024. [Online]. Available: https://arxiv.org/abs/2412.05321 (arXiv) [3] M. Hao, K. Qian, and S. C.-K. Chau, “Privacy-Preserving Blockchain-Enabled Parametric Insurance via Remote Sensing and IoT,” arXiv preprint arXiv:2305.08384, May 2023. [Online]. Available: https://arxiv.org/abs/2305.08384 (arXiv) [4] S. Feng, W. Wang, Z. Xiong, D. Niyato, P. Wang, and S. S. Wang, “On Cyber Risk Management of Blockchain Networks: A Game Theoretic Approach,” arXiv preprint arXiv:1804.10412, Apr. 2018. [Online]. Available: https://arxiv.org/abs/1804.10412 (arXiv) [5] J. Anglen, “AI Agents in Insurance Risk Evaluation: Benefits, Challenges, Use Cases & Future,” Rapid Innovation, 2024. [Online]. Available: https://www.rapidinnovation.io/post/the-future-of-personalized-risk-evaluation-in-insurance-with-ai-agents (Rapid Innovation) [6] J. Jada, “The Impact of Artificial Intelligence on Organisational Cybersecurity,” ScienceDirect, vol. 4, pp. 1–10, 2024. https://doi.org/10.1016/j.scidir.2024.100037 (ScienceDirect) [7] J. Viji Gripsy, “Biological software for recognition of specific regions in organisms,” Bioscience Biotechnology Research Communications, vol. 13, no. 1, pp. —, Mar. 2020. doi: 10.21786/bbrc/13.1/54. [8] J. Viji Gripsy and A. Jayanthiladevi, “Energy hole minimization in wireless mobile ad hoc networks using enhanced expectation-maximization,” in Proc. 2023 9th Int. Conf. Adv. Comput. Commun. Syst. (ICACCS), Mar. 2023, pp. 1012–1019. doi: 10.1109/ICACCS57279.2023.10112728 [9] J. Viji Gripsy and A. Jayanthiladevi, “Energy optimization and dynamic adaptive secure routing for MANET and sensor network in IoT,” in Proc. 2023 7th Int. Conf. Comput. Methodol. Commun. (ICCMC), Feb. 2023, pp. 1283–1290. doi: 10.1109/iccmc56507.2023.10083519. [10] S. Karpagavalli, J. V. Gripsy, and K. Nandhini, “WITHDRAWN: Speech assistive Tamil learning mobile applications for learning disability children,” Materials Today: Proceedings, Feb. 2021. doi: 10.1016/j.matpr.2021.01.050. [11] J. Viji Gripsy, “Trust-based secure route discovery method for enhancing security in mobile ad-hoc networks,” Int. J. Sci., Eng. Technol., vol. 13, no. 1, Jan. 2025. doi: 10.61463/ijset.vol.13.issue1.147. [12] J. Viji Gripsy, N. A. Selvakumari, L. Sheeba, and B. Senthil Kumaran, “Transforming student engagement through AI, AR, VR, and chatbots in education,” in Chatbots in Educational Leadership and Management, Feb. 2025, pp. 73–100. doi: 10.4018/979-8-3693-8734-4.ch004.

Copyright

Copyright © 2025 D. Suganthi, G. Hema Sree, C. Monika, K. Nithya Shree. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.