Industrial demand forecasting with scalable and adaptive approaches
Industrial forecasting faces increasingly complex data, diverse demand patterns, and fast operational decisions that strain traditional methods. From spare-parts logistics with intermittent demand to real-time utility consumption streams, models must be accurate, scalable, adaptive, and economically relevant. This dissertation develops and evaluates forecasting methods that connect statistical rigor, machine-learning flexibility, and industrial decision-making.
First, it proposes an economically grounded evaluation framework that links forecast quality to operational outcomes such as service levels and inventory costs, showing why model assessment should go beyond error metrics. Second, it studies the trade-off between out-of-the-box machine-learning models and custom forecasting solutions. While both can be useful, the results show that domain-informed custom models often deliver better short-horizon accuracy (e.g., one- and two-day-ahead) and can yield tangible operational benefits.
Third, the thesis examines learning across related time series to improve robustness at scale. Using semi-global schemes, it demonstrates that grouping series by structural similarity can outperform purely local or fully global modeling. Finally, it advances toward real-time adaptability by developing online ensemble methods that update continuously under concept drift.
Overall, the dissertation offers a pathway toward forecasting systems that are statistically sound, operationally interpretable, and responsive to changing industrial environments.