Benchmarking zero-shot foundation time series forecasting models for industrial applications

Time-series forecasting foundation models recently emerged with zero-shot capabilities, leveraging generalized training on diverse datasets. This study compares zero-shot foundation models to traditional statistical, machine learning, and deep learning methods using industrial and academic multivariate datasets. Results show foundation models, particularly Moirai large, often outperform traditional methods while reducing dataset-specific tuning needs. These findings highlight their industrial potential by allowing for simpler, yet more accurate forecasting.