HN
Today

Google's 200M-parameter time-series foundation model with 16k context

Google Research has unveiled TimesFM, a 200M-parameter, decoder-only foundation model specifically designed for time-series forecasting, boasting an impressive 16k context length. This pre-trained model aims to revolutionize forecasting by offering a robust, scalable solution, and its availability on GitHub and integration into BigQuery signal a significant step forward for practical AI applications in this domain. Its release addresses critical challenges in handling long temporal sequences and provides sophisticated quantile predictions.

3
Score
0
Comments
#3
Highest Rank
14h
on Front Page
First Seen
Mar 31, 5:00 AM
Last Seen
Mar 31, 6:00 PM
Rank Over Time
934558776811202530

The Lowdown

TimesFM (Time Series Foundation Model) is a groundbreaking development from Google Research, introducing a pre-trained, decoder-only foundation model tailored for time-series forecasting. This initiative democratizes advanced time-series analysis, offering a powerful tool for researchers and developers to predict future values based on historical data with unprecedented accuracy and context.

  • TimesFM is a pre-trained, decoder-only foundation model developed by Google Research. Its primary purpose is highly accurate time-series forecasting.
  • The latest version, TimesFM 2.5, features 200 million parameters, significantly reduced from the previous 500 million, while improving performance.
  • It supports an expanded context length of up to 16,000, a substantial increase from 2,048, allowing for more comprehensive historical data analysis.
  • The model also offers continuous quantile forecasting, with an optional 30M quantile head for predictions up to a 1,000-step horizon.
  • Resources include a research paper (ICML 2024), Hugging Face checkpoints, a Google Research blog post, and direct integration into Google BigQuery as an official product.
  • The open-source GitHub repository provides installation instructions and Python code examples for easy implementation using PyTorch or Flax.

With its advanced capabilities and broad context support, TimesFM represents a significant advancement in time-series prediction, promising to enhance various applications from finance to logistics by providing more precise and robust forecasts.