Against Time-Series Foundation Models

Or: My Experience In Modern Forecasting

Time-series foundational models (TSFMs) are currently engaged in a knife fight to prove their worth against statistical models that have been around for half a century. TSFMs borrow similar architectural patterns from LLMs, and take a similar approach to curating large foundational datasets and pretraining. Only instead of words, the primitive is a time-series. Why are they struggling to prove their worth? This is obviously not what happened in other domains where foundational models have emerged. Is it some property of reality that forecasting time-series is harder? Or are these models…

saved by

• Oscar Marret

related reading

• The Unreasonable Difficulty of Time Series Forecastingsuzyahyah.github.io
• The AI Superforecasters Are Here - by Scott Alexanderastralcodexten.com
• Time Series - From Analyzing the Past to Predicting the Future | Towards Data Sciencetowardsdatascience.com
• On the Opportunities and Risks of Foundation Modelsarxiv.org
• Training Imperativesdan.io
• TimeGPT-1 | Hacker Newsnews.ycombinator.com
• GitHub - google-research/timesfm: TimesFM (Time Series Foundation Model) is a pretrained time-series foundation model developed by Google Research for time-series forecasting.github.com
• 2402.03885arxiv.org
• Training LLMs to Predict World Events (Guest Post with Mantic) - Thinking Machines Labthinkingmachines.ai
• Forecasting is Way Overrated, and We Should Stop Funding It — LessWronglesswrong.com
• ML Blog - Everything you need to know about Time Series Forecastingmrtunguyen.github.io
• Is Capability a Liability? More Capable Language Models Make Worse Forecasts When It Matters Mostarxiv.org