Skip to content

Latest commit

 

History

History
113 lines (83 loc) · 3.73 KB

chronos.md

File metadata and controls

113 lines (83 loc) · 3.73 KB

Overview

Chronos is a family of pretrained time series forecasting models developed by Amazon Science. Based on language model architectures (T5 transformers), Chronos transforms time series into token sequences for training, enabling powerful zero-shot forecasting capabilities.

How It Works

Chronos uses a novel approach to time series forecasting:

  1. Tokenization: Time series are transformed into sequences of tokens via scaling and quantization
  2. Training: A language model is trained on these tokens using cross-entropy loss
  3. Forecasting: Probabilistic forecasts are obtained by sampling multiple future trajectories given historical context

Model Variants

Chronos-2 (Latest)

  • Parameters: 120 million
  • Architecture: Encoder-only time series foundation model
  • Release: October 2025
  • Capabilities: Univariate, multivariate, and covariate-informed forecasting in a single architecture
  • Performance: State-of-the-art zero-shot accuracy among public models

Chronos-Bolt

  • Accuracy: 5% lower error than original Chronos
  • Speed: Up to 250x faster than original Chronos
  • Memory: 20x more memory efficient
  • Use Case: Production deployments requiring high throughput

Original Chronos

  • Architecture: Based on T5 transformer models
  • Variants: Multiple sizes (tiny, mini, small, base, large)
  • Foundation: Established the tokenization approach for time series

Benchmark Performance

State-of-the-Art Results

Chronos-2 achieves state-of-the-art zero-shot accuracy on:

  • fev-bench: Leading performance
  • GIFT-Eval: Top scores among public models
  • Chronos Benchmark II: Highest accuracy

Adoption

  • 600+ million downloads from Hugging Face (Chronos and Chronos-Bolt combined)
  • Widely adopted in industry and research

Key Features

Zero-Shot Forecasting

  • No fine-tuning required on target datasets
  • Generalizes across diverse time series domains
  • Handles various frequencies and patterns

Probabilistic Forecasts

  • Generates multiple possible future trajectories
  • Provides uncertainty quantification
  • Supports risk-aware decision making

Flexible Input Types

  • Univariate time series (single variable)
  • Multivariate time series (multiple related variables)
  • Covariate-informed forecasting (using external factors)

Use Cases

  • Demand forecasting for retail and supply chain
  • Financial market prediction
  • Energy consumption forecasting
  • Weather and climate predictions
  • Server load and capacity planning
  • Healthcare metrics forecasting
  • IoT sensor data prediction

Deployment Options

Cloud Services

  • Amazon SageMaker JumpStart: Deploy with just a few lines of code
  • AWS Integration: Native support for AWS ecosystem

Open Source

Frameworks

  • Compatible with standard PyTorch workflows
  • Integration with time series libraries
  • REST API deployment options

Architecture Details

Chronos-2 uses an encoder-only architecture optimized for:

  • Efficient processing of time series sequences
  • Support for variable-length inputs
  • Handling missing values and irregular sampling
  • Multi-task learning across forecast types

Training Approach

Models are trained on:

  • Diverse time series datasets from multiple domains
  • Synthetic time series for improved generalization
  • Real-world forecasting benchmarks
  • Various temporal patterns and frequencies

Licensing

Apache 2.0 license - open source for research and commercial use.

Pricing

Free and open source. AWS deployment costs apply when using SageMaker or other AWS services.