amazon-science/chronos-forecasting

Chronos: Pretrained Models for Time Series Forecasting

What it solves

Chronos provides pretrained foundation models for time series forecasting, eliminating the need to train custom models from scratch for every new dataset. It enables zero-shot forecasting, meaning the models can make accurate predictions on unseen data without further training.

How it works

Chronos treats time series forecasting as a language modeling problem. It transforms time series data into a sequence of tokens through scaling and quantization, allowing it to leverage language model architectures. The project offers three main model variants:

  • Chronos-2: The latest version supporting univariate, multivariate, and covariate-informed forecasting.
  • Chronos-Bolt: A high-performance variant that uses a patch-based approach to chunk historical data, enabling direct multi-step forecasting for significantly faster inference and lower memory usage.
  • Chronos (Original): The base family that uses cross-entropy loss on quantized tokens to generate probabilistic forecasts by sampling future trajectories.

Who it’s for

Data scientists and ML engineers who need reliable, fast, and scalable time series forecasting across various domains without the overhead of manual model tuning or training.

Highlights

  • Zero-shot capabilities: High performance on unseen datasets across univariate and multivariate tasks.
  • High efficiency: Chronos-Bolt is up to 250x faster and 20x more memory-efficient than original models of the same size.
  • Probabilistic forecasting: Generates quantile forecasts to provide prediction intervals rather than just single-point estimates.
  • AWS Integration: Direct deployment paths via AutoGluon-Cloud and Amazon SageMaker JumpStart.

Related

  • Project
    Nixtla/nixtla

    A foundation model for time series forecasting and anomaly detection that provides zero-shot inference and fine-tuning capabilities across multiple domains.

  • Project
    DataDog/toto

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

  • Project
    PriorLabs/tabpfn-time-series

    A zero-shot time series forecasting library that frames forecasting as a tabular regression problem using the TabPFN foundation model.

  • Project
    ant-intl/Falcon-TST

    A family of large-scale time series foundation models, including univariate and multivariate versions, for efficient and accurate forecasting.

  • Project
    SalesforceAIResearch/uni2ts

    A PyTorch library for the pre-training, fine-tuning, and evaluation of Universal Time Series Transformers, enabling zero-shot forecasting across diverse datasets.