Mitigating the Bullwhip Effect in Demand Forecasting

Mitigating the bullwhip effect is a key strategy to increase supply chain efficiency, and forecasting models play a crucial role in amplifying or mitigating this effect. Existing methods for bullwhip mitigation provide only limited control over the trade-off between forecast accuracy and bullwhip effect mitigation through model evaluation and selection, or rely on reactive corrections such as order smoothing. We propose a novel loss function approach to proactively train machine learning forecasting models to mitigate the bullwhip effect. By combining forecast errors and bullwhip effect into a single loss function and jointly minimizing their weighted sum during model training, our method provides fine-grained control over the trade-off between these two competing metrics. Extensive validation on the M5 dataset shows that our method effectively mitigates the bullwhip effect when the weight given to the bullwhip term is sufficiently large, and that substantial mitigation can be achieved with only a limited loss in forecast accuracy.