From Data to Decisions: Enterprise Demand Forecasting with Machine Learning

This is my Ph.D. dissertation. I earned my Ph.D. in Business Analytics from the Haslam College of Business, University of Tennessee, USA, in May 2025. My primary advisor was Dr. Chuanren (Charles) Liu.

• Full dissertation (PDF)
• Official version (University of Tennessee TRACE)
• Blog post

Abstract

Accurate demand forecasting is critical for operational efficiency and strategic decision-making in large-scale enterprises. This dissertation presents a machine learning (ML)-driven demand forecasting framework implemented at a Fortune-500 company HP Inc., focusing on three key areas: ML-based predictive modeling, MLOps and deployment scalability, and Human-in-the-loop forecasting integration. Additionally, we explore how predictive optimization enhances decision-making through end-to-end learning.

The first contribution involves the development of a scalable ML-based forecasting system, leveraging tree-based models (LightGBM), feature engineering, and advanced time-series methodologies. The model captures complex demand drivers, including macroeconomic trends, product life cycle effects, and channel inventory dynamics. By transitioning from traditional statistical models to ML-based approaches, the framework improves forecasting accuracy in key metrics while adapting to evolving market conditions.

The second contribution addresses MLOps and enterprise-scale deployment challenges, ensuring model reliability, automation, and reproducibility. The research outlines best practices in model monitoring, version control, model deployment, and continuous learning pipelines, demonstrating how systematic ML deployment reduces technical debt and maintains forecast accuracy over time.

The third contribution integrates ‘Human-in-the-Loop’ forecasting, ensuring that ML predictions are refined through expert-driven consensus mechanisms. The system incorporates business intelligence inputs such as sales insights, promotional strategies, and market conditions, balancing data-driven automation with human expertise to enhance interpretability and trust in forecasts. Through this closed-loop process, we are able to improve the overall forecast accuracy by 34% (wMAPE) and reduce inventory by 28% while maintaining same service levels.

Finally, this dissertation presents a predictive optimization framework that transforms ML-based predictions into actionable strategies. We showcase how perfect predictions still don’t lead to perfect decisions through a simulation study. Subsequently, we propose an end-to-end learning paradigm that simultaneously addresses demand forecasting, inventory allocation, procurement planning, and production scheduling in the supply chain.