Background Accurate sales forecasting is a key factor in effective business management, especially under conditions of increasing competition and the rapid development of e‐commerce. Sales time series are often characterized by trends, seasonality, and random fluctuations, which complicates the selection of an appropriate forecasting method. Therefore, a comparative analysis of classical exponential smoothing models and modern hybrid approaches is highly relevant for identifying the most accurate and practically applicable sales forecasting methods. Method This study presents a comparative evaluation of five forecasting models—Simple Exponential Smoothing (SES), Holt’s model, Holt–Winters’ model, Theil–Wage model, and SutteARIMA—applied to real retail sales data. The objective is to determine which model provides the most accurate forecasts and under which conditions, thereby supporting more informed planning and inventory control. The analysis is based on the Superstore dataset (2015–2018), sourced from the Kaggle platform, which contains detailed e‐commerce sales data across three product categories: Furniture, Office Supplies, and Technology. This dataset was selected for its high quality, representative seasonal structure, and relevance to practical business forecasting scenarios. Monthly sales quantities were extracted and used to construct time series for each category. Each model was optimized for its respective parameters (e.g., α , β , γ for exponential smoothing and λ , ϕ , θ for SutteARIMA) using grid search to minimize error metrics. Model performance was then evaluated using mean absolute error (MAE), mean squared error (MSE), and mean absolute percentage error (MAPE). Results The results reveal that while SES, Holt’s, and Theil–Wage models produced moderate accuracy (MAPE up to 16%), the Holt–Winters and SutteARIMA models demonstrated significantly higher performance. For example, Holt–Winters achieved a MAPE of 2.88% and MAE of 5.21 in the Technology category, confirming its strength in capturing seasonal trends. SutteARIMA outperformed all models, achieving the lowest forecasting error in all categories—with a MAPE of 2.64% for Technology, 3.79% for Office Supplies, and 5.34% for Furniture—demonstrating excellent short‐term adaptability and precision. These findings underline the importance of aligning forecasting models with the specific structural characteristics of sales data, such as trend and seasonality. The study also confirms that improper parameter selection, particularly in SutteARIMA, can lead to substantial error increases, highlighting the need for careful optimization. Conclusion This research provides a practical and data‐driven foundation for selecting appropriate sales forecasting models in retail. By integrating real‐world data, comparative accuracy metrics, and clear model recommendations, it supports evidence‐based decisions in business operations and inventory planning. The results also offer pathways for extending model application to other domains and integrating them into hybrid forecasting systems.
Boyko et al. (Thu,) studied this question.