Multi-Horizon Demand Forecasting for E-Commerce Fulfillment Using Ensemble Deep Learning

Abstract

Efficient inventory management and assignment of fulfillment capacity, along with labor planning is built on accurate multi-horizon demand forecasting in large-scale e-commerce businesses. This paper introduces an ensemble deep learning approach fusing Long Short-Term Memory (LSTM) networks, Temporal Convolutional Networks (TCN) and Gradient Boosted Trees (GBT) to predict the demand in various categories with different seasonality patterns, demand volatility, promotional effects, and customer purchasing behaviors across multiple horizons. Traditional statistical forecasting methods like Moving Average, ARIMA, and Exponential Smoothing may not effectively model the nonlinear dependencies, rapid changes in demand, and intricate temporal dynamics often seen in today's e-commerce landscape. Advanced AI and deep learning techniques have thus become a promising alternative to enhance the accuracy of forecasts. Proposed framework combines the strength of LSTM network to model the sequence, TCN for extracting long-range temporal dependencies and GCs for feature engineering. The framework achieves this by combining prediction results from these different learners using the technique of weighted ensemble aggregation, which results in increased robustness and generalization for short-term, medium-term, and long-term forecasting. It uses transactional sales data, promo signals, the holidays, customer engagement data, and stock data to provide one-day to ninety-day outlooks. A detailed experimental analysis was performed with representative e-commerce demand datasets with multiple product categories. Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE) and Weighted Absolute Percentage Error (WAPE) were used to evaluate forecasting performance. The experimental results show that ensemble model is always superior than the single forecasting model, and it can decrease the error of the forecast and ensure the stability of the forecasting results in the case of fluctuation demand. The ensemble framework proved to be successful in enhancing the reliability of the forecasts, especially during holidays and promotional periods. The results show that hybrid deep learning-based systems can offer significant advantages for inventory management, warehouse resource management, transportation scheduling, and supply chain decision-making. In addition, multi-horizon forecasting can help companies effectively plan stock replenishment strategies, prevent stockouts, lower over-stocking expenses, and enhance customer satisfaction. The proposed methodology adds to the existing research on intelligent supply chain analytics by offering a scalable, adaptive, and data-driven approach for prediction in the dynamic e-commerce environments. Transformer architectures, reinforcement learning, and real-time demand sensing mechanisms could be combined in future upgrades to further boost forecasting accuracy and operational responsiveness.