ant-intl/Falcon-TST

Official implementation of Falcon-TST

What it solves

Falcon-TST provides a suite of large-scale foundation models designed to improve the efficiency and accuracy of time series forecasting. It addresses the need for robust models capable of handling both univariate (single-variable) and complex multivariate (multi-variable) time series data.

How it works

The project offers three distinct model architectures:

  • Falcon-2.0: A univariate, encoder-only model that utilizes an incremental training approach called ORBIT.
  • Falcon-X: A multivariate model that manages heterogeneous data by using a shared latent prototype space and dual-dependency modeling.
  • Falcon-1.0: A hierarchical mixture-of-experts model that combines patch-wise expert specialization with sample-wise routing.

Users can interact with these models via a unified Python client, which handles quantile predictions and supports input masking for missing data.

Who it’s for

This library is intended for data scientists, researchers, and engineers working on time series analysis who need pre-trained foundation models to generate accurate forecasts without building models from scratch.

Highlights

  • Provides a unified API for accessing multiple time-series foundation models.
  • Supports both univariate and multivariate forecasting tasks.
  • Includes built-in support for handling missing data via input masks.
  • Built on PyTorch and compatible with standard Python environments.

Related

  • Project
    google-research/timesfm

    TimesFM is a pretrained decoder-only foundation model from Google Research for high-performance, zero-shot univariate and multivariate time-series forecasting.

  • 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
    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
    skforecast/skforecast

    A Python library for time series forecasting that provides a unified API for scikit-learn compatible ML models, statistical methods, and zero-shot foundation models.