Time Series Forecasting in E-commerce: A Systematic Literature Review of Methods, Data, Evaluation Metrics, and Future Directions (2021–2026)

Abstract

The rapid growth of e-commerce has increased the need for accurate forecasting of product demand and sales to support data-driven business decisions. This study aims to systematically review recent research on time series forecasting in e-commerce and synthesize the forecasting methods, dataset characteristics, evaluation metrics, and future research directions. A Systematic Literature Review (SLR) was conducted using Google Scholar, Scopus, ScienceDirect, IEEE Xplore, SpringerLink, and ACM. The search was limited to English-language journal articles, conference papers, and preprints published between 2021 and 2026. After title, abstract, and full-text screening, 15 studies were included in the final analysis. The studies were analyzed descriptively based on four research questions. The results show that the reviewed forecasting approaches can be grouped into hybrid models, deep learning models, and comparison-based methods, with hybrid approaches being the most frequently identified (9 of the 15 studies, 60%), followed by deep learning models (4 studies) and comparison-based methods (2 studies). Dataset sources varied across studies, while daily data were used in 12 studies (80%). MAE and RMSE were the most frequently reported evaluation metrics, although differences in datasets, forecasting tasks, and evaluation procedures limited direct comparison across studies. Five main future research directions were identified: broader data coverage, richer contextual information, improved model architectures, wider forecasting scope and efficiency, and more consistent evaluation procedures. Overall, the review provides a structured overview of recent e-commerce forecasting research and identifies areas for further investigation.