A SYSTEMATIC REVIEW OF DEMAND FORECASTING MODELS FOR RETAIL E-COMMERCE ENHANCING ACCURACY IN INVENTORY AND DELIVERY PLANNING

Abstract

In the rapidly evolving landscape of retail e-commerce, characterized by volatile consumer behavior, diverse product assortments, and fluctuating market conditions, demand forecasting has emerged as a critical strategic capability. Accurate demand forecasting not only underpins effective inventory management and delivery planning but also serves as a cornerstone for optimizing supply chain responsiveness, minimizing stockouts and overstock situations, and enhancing customer satisfaction. As e-commerce platforms increasingly embrace data-driven operations, the role of predictive analytics in shaping demand planning strategies has gained unprecedented prominence. This systematic review presents a comprehensive synthesis of 72 peer-reviewed academic articles, industry reports, and empirical case studies published between 2010 and 2024. The objective is to evaluate the breadth, depth, and evolution of demand forecasting methodologies specifically applied within the context of retail e-commerce. The selected studies were rigorously screened and categorized based on methodological foundations, technological sophistication, and practical applications. Four major categories of forecasting models are examined: (1) traditional statistical approaches, such as AutoRegressive Integrated Moving Average (ARIMA), Seasonal ARIMA (SARIMA), and exponential smoothing models; (2) machine learning techniques, including decision trees, random forests, support vector regression, and ensemble methods; (3) hybrid frameworks that integrate statistical modeling with machine learning or deep learning components; and (4) advanced deep learning architectures, particularly Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRUs), Convolutional Neural Networks (CNNs), and transformer-based models. The findings reveal distinct performance advantages across model categories. Traditional statistical models demonstrate continued relevance in scenarios marked by stable demand patterns, short forecasting horizons, and limited data complexity. However, their limitations become evident when applied to highly nonlinear, sparse, or volatile datasets. In contrast, machine learning models offer enhanced adaptability and accuracy, especially in handling high-dimensional data environments with diverse product lines and unpredictable promotional impacts. Deep learning models further advance this capability by capturing complex temporal dynamics, long-range dependencies, and multivariate input streams. These models are particularly effective for SKU-level forecasting, where product-specific demand patterns fluctuate frequently and require real-time recalibration. Many reviewed studies adopted modular architectures that allow domain-specific tuning and facilitate deployment within enterprise resource planning (ERP) and inventory management systems.