Enhancing Demand Forecasting in Multi-Channel Inventory Management: A Design Science Approach

In the hyper-competitive e-commerce landscape, inaccurate demand forecasting leads to significant financial losses from inventory distortion, comprising costly overstocks and sales-damaging stockouts. This thesis addresses this critical challenge by detailing the design, development, and validation of an AI-driven demand forecasting system for Stockpilot, a multichannel inventory management platform. Employing the Design Science Research (DSR) methodology, this work systematically investigates the problem context, stakeholder needs, and data characteristics inherent to e-commerce operations. Analysis of historical sales data revealed a prevalence of intermittent demand patterns, rendering traditional forecasting models ineffective. Through a rigorous empirical evaluation, the Teunter-Syntetos-Babai (TSB) model was identified as the optimal choice, demonstrating superior accuracy and computational efficiency for SKU-level predictions improving the Mean Absolute Percentage Error (MAPE) by nearly 19 percentage points over the previous method. The resulting artifact is a scalable and secure forecasting microservice built with a layered architecture. It operates independently with its own database, ensuring tenant data isolation and preventing performance degradation of the core platform. The system is designed to integrate seamlessly into Stockpilot’s ecosystem via a RESTful API and event-driven data synchronization. Validation through performance and load testing confirmed the system's robustness, meeting the non-functional requirement of sub-second response times under projected peak load and demonstrating effective horizontal scalability. This research provides a validated blueprint for integrating specialized AI forecasting into a multi-tenant SaaS platform, offering a practical solution to enhance inventory accuracy, reduce costs, and support strategic business growth in the dynamic e-commerce sector.