Demand Forecasting
The Covid-19 pandemic posed an unprecedented challenge to demand forecasting systems, especially those with a simpler design, which suffered catastrophic failures due to a phenomenon known as “drift” or deviation. This drastic change was resulting from erratic and volatile consumer behavior, as well as interruptions in the supply chain from key Asian suppliers, affected by rigorous measures against Covid-19. In this turbulent scenario, Obramat faced the challenge of accurately forecast the demand for highly seasonal products (such as heating and cooling systems) to manage orders from its suppliers efficiently.
At WhiteBox, we innovated with the implementation of a prediction model based on decision trees (Gradient Boosting), replacing the previous and obsolete Holt-Winters model. This new model, developed with open source tools such as Spark for data processing and LightGBM, together with scikit-learn for modeling, is able to forecast stock needs 60 days in advance for each product in all stores. Thanks to the robustness of this system, it was perfectly integrated into Obramat's Google Cloud Platform infrastructure, ensuring accurate and scalable predictions with daily and weekly aggregation.
The renewal of its demand forecasting model has placed Obramat in a privileged position to anticipate seasonal market fluctuations, achieving significant savings and avoiding stockouts. Not only does this model allow Obramat to execute purchase orders in advance, but its modular structure and automated lifecycle management keeps it always up to date, protecting it against future deviations or “drifts”.
- More than 900 references managed by the model.
- 30 establishments distributed throughout Spain.
- Prediction of 27,000 simultaneous time series.
