This Time is Different: An Observability Perspective on Time Series Foundation Models
Ben Cohen, Emaad Khwaja, Youssef Doubli, Salahidine Lemaachi, Chris Lettieri, Charles Masson, Hugo Miccinilli, Elise Ramé, Qiqi Ren, Afshin Rostamizadeh, Jean Ogier du Terrail, Anna-Monica Toon, Kan Wang, Stephan Xie, David Asker, Ameet Talwalkar, Othmane Abou-Amal
2025-05-22
Summary
This paper talks about Toto, a new AI model designed to predict future trends from time-based data, like weather patterns or stock prices, using a special architecture that looks only at the most recent information.
What's the problem?
Predicting what will happen next in complex time series data is really tough, especially when there are lots of different factors to consider at once, and many existing models can't handle this well on a large scale.
What's the solution?
The researchers built Toto using a decoder-only design, which means it focuses on the latest data to make predictions, and they showed that it works better than other models on big, real-world datasets that involve many variables.
Why it matters?
This matters because better time series forecasting can help in many areas, like planning for disasters, managing finances, or improving healthcare, by giving people more accurate information about what might happen next.
Abstract
Toto is a time series forecasting foundation model using a decoder-only architecture, demonstrating state-of-the-art performance on large-scale benchmarks with multivariate observability data.