Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Explore fundamental limitations of foundational time series forecasting models in this 43-minute conference talk that challenges conventional approaches to training and evaluation. Examine why training on increasingly large datasets is not always beneficial for time series forecasting performance, and discover how common but inadequate evaluation practices can mask these critical limitations. Learn about the speaker's argument that multimodality represents the key solution to addressing these challenges in foundational model development. Gain insights into the theoretical and practical constraints that affect current foundational models when applied to time series data, and understand alternative approaches that could improve forecasting accuracy and reliability. The presentation draws from research in statistical inference, Bayesian methods, and machine learning to provide a comprehensive analysis of where current foundational time series models fall short and how the field can move forward through innovative multimodal approaches.
