On the stability of global forecasting models (2506.05776v1)

Abstract: Forecast stability, that is the consistency of predictions over time, is essential in business settings where sudden shifts in forecasts can disrupt planning and erode trust in predictive systems. Despite its importance, stability is often overlooked in favor of accuracy, particularly in global forecasting models. In this study, we evaluate the stability of point and probabilistic forecasts across different retraining frequencies and ensemble strategies using two large retail datasets (M5 and VN1). To do this, we introduce a new metric for probabilistic stability (MQC) and analyze ten different global models and four ensemble configurations. The results show that less frequent retraining not only preserves but often improves forecast stability, while ensembles, especially those combining diverse pool of models, further enhance consistency without sacrificing accuracy. These findings challenge the need for continuous retraining and highlight ensemble diversity as a key factor in reducing forecast stability. The study promotes a shift toward stability-aware forecasting practices, offering practical guidelines for building more robust and sustainable prediction systems.