ant-intl/Falcon-TST

Official implementation of Falcon-TST

What it solves

It provides a family of large-scale foundation models designed to make time series forecasting more accurate and efficient across different types of data, including both univariate and multivariate time series.

How it works

The project offers three distinct model architectures:

  • Falcon-2.0: A univariate encoder-only model based on Omni-Range Bootstrap Incremental Training (ORBIT).
  • Falcon-X: A multivariate model that uses a shared latent prototype space and dual-dependency modeling to handle heterogeneous variates.
  • Falcon-1.0: A hierarchical mixture-of-experts (MoE) model that combines patch-wise expert specialization with sample-wise hierarchical routing.

Who it’s for

Data scientists and ML engineers working with time series forecasting who need pre-trained foundation models that can be applied to various datasets without extensive custom training.

Highlights

  • Diverse Model Family: Supports univariate (Falcon-2.0) and multivariate (Falcon-X) forecasting.
  • Quantile Prediction: Provides capabilities for quantile-based forecasting via a unified client API.
  • Scalable Architectures: Employs advanced techniques like Mixture-of-Experts and ORBIT training regimes.

Related

  • Project
    DataDog/toto

    A foundation model for multivariate time series forecasting optimized for observability metrics, providing zero-shot and probabilistic predictions.

  • Project
    amazon-science/chronos-forecasting

    A family of pretrained foundation models for zero-shot time series forecasting that treats time series data as tokens for language model architectures.

  • Project
    Nixtla/mlforecast

    A scalable framework for time series forecasting that allows any scikit-learn compatible machine learning model to be used for forecasting millions of series.

  • Project
    AIStream-Peelout/flow-forecast

    An open-source deep learning framework for time series forecasting, classification, and anomaly detection featuring a wide array of state-of-the-art models.

  • Project
    microsoft/aurora

    Aurora is a foundation model for the Earth system that predicts atmospheric variables like temperature, air pollution, and ocean waves by adapting a general-purpose pre-trained model to specialized forecasting tasks.