Kronos

Kronos

A family of decoder-only foundation models for financial market K-line sequences, designed to process high-noise market data for quantitative forecasting tasks.

Made by shiyu-coder on GitHub. These notes are Minda's, written from its README; the project's own files and license are in its repository.

About

What it is: It quantizes continuous market data into hierarchical discrete tokens and uses an autoregressive transformer to generate forecasts for quantitative tasks.

Who it suits: This project suits quantitative researchers and data scientists who need to model market trends using historical price data.

Where to use it: This project runs in local environments where you can execute scripts to process market data and generate predictions.

• Predict future price and volume trends from historical K-line data
• Execute batch forecasting for multiple assets at the same time
• Fine-tune the model on custom financial datasets

How to set it up:

1. Install Python 3.10 or newer and the required dependencies
2. Install Qlib to prepare data for custom tasks
3. Configure your data paths and hyperparameters in the settings file
4. Run the data preparation script to format your market history

pip install -r requirements.txt

pip install pyqlib

How to use it: You use the KronosPredictor class to load a model and tokenizer, ingest historical CSV data, and generate future projections.

| Tool | What it does | |---|---| | KronosPredictor | Manages data preprocessing and inference | | predict | Generates a forecast for a single asset series | | predict_batch | Performs parallel inference on multiple series |

predictor = KronosPredictor(model, tokenizer, max_context=512)

Example: You prepare a CSV of historical prices and run the predict method to return a data frame with forecasted open, high, low, close, volume, and amount values.

For newcomers and experts: Newcomers can think of this as a model that learns patterns in market charts. Experts will recognize the decoder-only transformer architecture that treats financial data as a sequence of hierarchical discrete tokens. It uses a two-stage training approach to capture domain-specific signals.

Good to know:

• The context window is 512 steps for most models.
• It requires column data for open, high, low, and close prices.
• It is licensed under the MIT License.

On GitHub, 4 November 2026

• Stars: 46,545 (503 that day)
• Forks: 7,665
• Watching: 407
• Open issues and pull requests: 295
• Language: Python
• License: MIT
• Trending: #12 repository of the day on GitHub trending (4 November 2026)

Links

• Code: https://github.com/shiyu-coder/Kronos
• Releases: https://github.com/shiyu-coder/Kronos/releases
• Issues: https://github.com/shiyu-coder/Kronos/issues
• Star history: https://www.star-history.com/#shiyu-coder/Kronos&Date

Get the code

git clone https://github.com/shiyu-coder/Kronos.git

Then follow the setup in its README. Nothing in this kit runs by itself.