Foundational Models for Time Series Forecasting: are we there yet?
summary
Transformers get repurposed for forecasting by quantize to discrete tokens (Chronos uses GPT‑2/T5 with mean scaling and binning) or by patching continuous signals (TimesFM outputs 128‑step patches), trained on mixed public and private corpora; benchmarks show synthetic mixing (5–20% synthetic data) improves performance, public corpora are small (Monash <1B observations) while TimesFM reports training on 100 billion datapoints, and models aim to be zero-shot forecasters but require careful dataset construction and fine‑tuning.
AI-written summary · about
chapters (6)
- 0:00Introduction
- 3:46Foundational models for time series
- 8:23Data scarcity and synthetic data
- 13:16Time series foundational models overview
- 16:26Training data and architectures
- 19:13Benchmarks and future directions
related talks
31mFoundational Models for Time Series Forecasting: are we there yet?
30mFoundational Models for Time Series Forecasting: are we there yet
33mFrontiers of Foundation Models for Time Series
19mForecasting for Everyone - Foundation Models and the Future of Time Series Prediction
54mTransformes for Time Series: Is the New State of the Art (SOA) Approaching?
31mAdvanced machine learning time series....
10mWhy Your Forecasting Transformer Isn’t Working (And How To Fix It...
31mML time series forecasting methods