Foundational Time Series Models in Practice: The Future of Forecasting, or Just Hype?
summary
Clear claims and concrete results: patch TST uses nonoverlapping patches to get long lookbacks and reports state‑of‑the‑art leaderboard wins versus vanilla Transformer baselines, while zero-shot accuracy from pretrained foundational models matches strong global models on some benchmarks. Practical tools mentioned include Morai and times FM, and a DataDog case shows adding private domain data and fine‑tuning improves accuracy; retraining can cause stability issues week‑over‑week and inference memory is modest.
AI-written summary · about
chapters (14)
- 0:00Welcome
- 0:24Introduction
- 1:28What is time series forecasting
- 3:38Local versus global models
- 6:07History of forecasting
- 9:41Rise of foundational models
- 12:51How patching works
- 15:01Current landscape
- 17:28Practical advice
- 20:14Fine-tuning and retraining
- 23:05Model stability and uncertainty
- 24:29Future of foundational models
- 26:14Future or hype
- 26:35Q&A
related talks
31mFoundational Models for Time Series Forecasting: are we there yet?
33mFrontiers of Foundation Models for Time Series
30mFoundational Models for Time Series Forecasting: are we there yet
19mForecasting for Everyone - Foundation Models and the Future of Time Series Prediction
31mAdvanced machine learning time series....
54mTransformes for Time Series: Is the New State of the Art (SOA) Approaching?
10mWhy Your Forecasting Transformer Isn’t Working (And How To Fix It...
40mMulti-Series Forecasting at Scale with StatsForecast