Operational Demand Forecasting
A forecasting system designed to predict operational demand and service-related metrics, combining historical patterns, calendar effects and machine learning to support planning and resource allocation.
Operational teams need to anticipate demand in order to plan capacity and allocate resources effectively. The challenge was to build forecasts that reflect recurring patterns while remaining useful across different operational units and service types.
The modelling dataset was built from historical operational records and aggregated around unit, service type, date and time interval. Calendar information and critical dates were incorporated alongside historical demand patterns.
The feature set combines calendar variables, holiday-related information and explicit historical lag features. Rather than relying on positional shifts, historical observations are aligned by date, interval and weekday to preserve the real temporal structure of the problem.
A historical baseline was established first and then compared with an XGBoost-based model. The evaluation uses a temporal split, reserving the most recent period for validation to better reproduce how the model would behave on future observations.
The objective was not simply to obtain a lower error, but to determine whether the additional complexity of Machine Learning provided a meaningful improvement over a strong historical baseline.
The project evolved from a forecasting prototype into a more complete ML workflow involving feature engineering, temporal validation, experiment tracking and model evaluation. The results also highlighted an important lesson: a model should not be promoted to production simply because it is more complex than its baseline.