Colloquium Coputer Science - Lukas Hans, TH University Köhln

Title: Mastering Complexity in Industrial Demand: The Development of Scalable and Adaptive Forecasting Frameworks

Abstract:

How can resource-constrained organizations build forecasting systems that are both accurate and operationally useful? This talk synthesizes four papers that address core design choices for scalable, adaptive forecasting in small and medium-sized enterprises and public institutions.

(1) Intermittent demand. Zero-inflated, bursty, and sometimes obsolescent series challenge conventional methods. We develop a multi-criteria decision analysis that makes trade-offs explicit—balancing statistical accuracy, economic consequences (e.g., stockouts, waste, working capital), and computational cost/latency—so practitioners can select fit-for-purpose models rather than chase a single “best” metric.

(2) Build vs. buy. When do bespoke, domain-tailored models justify the engineering effort? We compare them with tuned defaults and AutoML pipelines and show when simple pipelines suffice and when custom modeling pays off.

(3) Local vs. global structure. We test whether pooling information across series improves accuracy, including semi-global clustering strategies, and illustrate results on a large water- demand forecasting benchmark.

(4) Non-stationarity and drift. Using water-consumption data with rapid distributional change, we compare rolling-window, expanding-window, and incremental (online) learning in terms of accuracy, stability, and compute/latency budgets.

Together, these studies yield a practical playbook: treat intermittency explicitly; start simple and escalate to bespoke models only when the marginal gains cover their complexity; prefer global/semi-global models when cross-series structure is strong; and adopt incremental updates when drift is fast. The goal is decision-ready guidance that teams can operationalize with realistic time, tooling, and talent constraints.

Short Bio:

Hi, I’m Lukas Hans (33). I’m a data scientist at REWE Group. My background is in statistics and data science: a B.Sc. in Business Administration (focus on statistics), an M.Sc. in Applied Statistics, and an M.Sc. in Data Science. At REWE, I build AI-driven demand forecasts for logistics to cut waste and improve product availability. My research and day job both focus on making forecasting work under real-world constraints. In my free time I enjoy data competitions and recently won the International Cherry Blossom Prediction Contest.