# Chronos: Learning the Language of Time Series

**Authors:** Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Shchur, Syama Sundar Rangapuram, Sebastian Pineda Arango, Shubham Kapoor, Jasper Zschiegner, Danielle C. Maddix, Michael W. Mahoney, Kari Torkkola, Andrew Gordon Wilson, Michael Bohlke-Schneider, Yuyang Wang  
**arXiv:** [2403.07815](https://arxiv.org/abs/2403.07815)  
**Date:** 2024-03-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2403.07815v1/

## Key points

- Chronos is a framework for pretrained probabilistic time series models
- Tokenizing time series values through scaling and quantization into a fixed vocabulary is part of the approach
- Transformer-based language model architectures are trained on tokenized sequences using cross-entropy loss
- Chronos models are pretrained based on the T5 family with varying parameters on diverse datasets, including synthetic data generated via Gaussian processes
- Benchmark analysis across 42 datasets shows superior performance of Chronos models compared to other methods in training corpus and comparable or superior zero-shot performance on new datasets
- Ability of Chronos models to leverage time series data from various domains effectively enhances zero-shot accuracy for forecasting tasks
- Pretrained models like Chronos can simplify forecasting pipelines by leveraging diverse time series data sources for improved accuracy in unseen prediction scenarios

## Summary

, , , , 

In their paper titled "Chronos: Learning the Language of Time Series," Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, and their team introduce Chronos, a framework for pretrained probabilistic time series models. The approach involves tokenizing time series values through scaling and quantization into a fixed vocabulary, followed by training transformer-based language model architectures on these tokenized sequences using cross-entropy loss. The Chronos models are pretrained based on the T5 family with varying parameters (ranging from 20M to 710M) on a diverse set of publicly available datasets, including a synthetic dataset generated via Gaussian processes to enhance generalization. The authors conducted a comprehensive benchmark analysis across 42 datasets encompassing classical local models and deep learning methods. Their results demonstrate that Chronos models exhibit superior performance compared to other methods on datasets included in the training corpus. Additionally, they showcase comparable or even superior zero-shot performance on new datasets relative to approaches specifically trained on those data points. This highlights the ability of Chronos models to leverage time series data from various domains effectively, enhancing zero-shot accuracy for forecasting tasks. Overall, the findings suggest that pretrained models like Chronos can serve as valuable tools in simplifying forecasting pipelines by leveraging diverse time series data sources for improved accuracy in unseen prediction scenarios. The research team has made their inference code and model checkpoints available at https://github.com/amazon-science/chronos-forecasting for further exploration and application in related fields.