Forecasting for Everyone - Foundation Models and the Future of Time Series Prediction
summary
A foundation-model pipeline trains transformer-based networks on 250 million time series points from many industries, then performs pure inference on unseen problems by inputting a short context (eg last 12 points) and recursively decomposing signals into trend/seasonal components. Evaluated against GluonTS, SageMaker Autopilot and Azure AutoML on ~35,000 series across 40 datasets, the zero-shot approach is best on ~40% of cases while currently producing point forecasts only. A demo lives at forecasting.networks.com.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:25Why forecasting is critical for every company
- 3:52Traditional forecasting challenges
- 5:58Pre-trained to zero-shot
- 8:24Training on 250M data points
- 10:48Zero-shot vs standard solutions
- 13:17Zero-shot limitations and potential
- 14:42Q&A: Financial markets
- 16:48Q&A: Reinforcement learning
related talks
31mFoundational Models for Time Series Forecasting: are we there yet?
30mFoundational Models for Time Series Forecasting: are we there yet
31mFoundational Time Series Models in Practice: The Future of Forecasting, or Just Hype?
33mFrontiers of Foundation Models for Time Series
21mFoundational Models for Time Series Forecasting: are we there yet?
40mMulti-Series Forecasting at Scale with StatsForecast
31mAdvanced machine learning time series....
29mProbabilistic demand forecasting at scale