TimesFM-3: A zero-shot foundation model for multivariate forecasting

A new time series forecasting model has been introduced that can predict multiple related data sequences simultaneously in a single computation step, rather than generating predictions one sequence at a time. Previously, similar models could only forecast based on the history of individual data sequences in isolation, but this model incorporates related sequences and known future information like scheduled events or promotions to improve forecast accuracy. The model was trained on over a trillion time points and uses a transformer architecture with alternating attention mechanisms that capture both temporal patterns within sequences and relationships across different sequences. This matters because real-world forecasting problems typically involve multiple interconnected data sequences and external factors, so a model that accounts for these dependencies can produce more accurate predictions without requiring customization for specific tasks.