chronos bolt base

Introduction

Chronos-Bolt is a family of pretrained time series forecasting models designed for zero-shot forecasting. These models are based on the T5 encoder-decoder architecture and have been trained on nearly 100 billion time series observations. They offer significant improvements in speed and memory efficiency compared to the original Chronos models.

Architecture

Chronos-Bolt utilizes the T5 encoder-decoder framework. It processes historical time series data by chunking them into patches, which are then encoded. The decoder uses these encoded representations to generate quantile forecasts for multiple future steps through direct multi-step forecasting. These models are optimized to be up to 250 times faster and 20 times more memory-efficient than their predecessors.

Training

Chronos-Bolt models have been pre-trained on extensive datasets, allowing them to perform zero-shot forecasting effectively. They have shown superior performance in probabilistic and point forecasting metrics like the Weighted Quantile Loss (WQL) and Mean Absolute Scaled Error (MASE) across 27 datasets, outperforming both statistical models and deep learning models that have been trained on those datasets.

Guide: Running Locally

To run Chronos-Bolt locally, follow these steps:

1. Install AutoGluon:

   pip install autogluon
   

2. Load Data and Initialize Predictor:

   from autogluon.timeseries import TimeSeriesPredictor, TimeSeriesDataFrame
   
   df = TimeSeriesDataFrame("https://autogluon.s3.amazonaws.com/datasets/timeseries/m4_hourly/train.csv")
   
   predictor = TimeSeriesPredictor(prediction_length=48).fit(
       df,
       hyperparameters={
           "Chronos": {"model_path": "autogluon/chronos-bolt-base"},
       },
   )
   

3. Make Predictions:

   predictions = predictor.predict(df)
   

For enhanced performance, consider using cloud-based GPU resources such as AWS EC2 or Google Cloud Platform, which can provide the necessary computational power for handling large datasets and models.

License

This project is licensed under the Apache-2.0 License.