Model Overview
Chronos is a foundation model for univariate time series forecasting that leverages transformer architecture trained on a large corpus of time series data. It introduces language-liked tokenization that turns real-valued time series into discrete tokens. Chronos is pretrained via probabilistic forecasting objective. It also utilizes data augmentation techniques such as TSMixup, and Gaussian smoothing kernel to generate synthetic time series data for pretraining.
Key Features
- Foundation Model: Pre-trained on large-scale univariate time series data
- Time Series Tokenization: token-level training and inference
- Multi-domain: Pretrained on various time series domains
