chronos bolt base

Introduction

Chronos-Bolt is a family of pretrained time series forecasting models designed for zero-shot forecasting, leveraging the T5 encoder-decoder architecture. These models are highly efficient, providing faster and more memory-efficient forecasts compared to the original Chronos models.

Architecture

Chronos-Bolt models utilize the T5 encoder-decoder architecture, which processes time series data by chunking it into patches for input into the encoder. The decoder generates quantile forecasts across multiple steps, achieving significant improvements in speed and accuracy over previous models.

Training

These models were trained on nearly 100 billion time series observations. They excel in zero-shot settings, outperforming both statistical and deep learning models in terms of forecasting accuracy. The models are available in various sizes, from tiny to base, with parameters ranging from 9 million to 205 million.

Guide: Running Locally

• With AutoGluon:

  1. Install AutoGluon: pip install autogluon
  2. Use the TimeSeriesPredictor to perform zero-shot inference on a dataset.

• With Inference Library:

  1. Install the Chronos Forecasting library: pip install chronos-forecasting
  2. Use the BaseChronosPipeline to load a pretrained model and perform inference.

Consider using a cloud GPU service such as AWS SageMaker for accelerated performance.

License

This project is licensed under the Apache-2.0 License.