Kronos is a pre-trained foundation model specialized for financial candlestick (K-line) sequences, addressing the limitations of general-purpose time series models on high-noise market data. It provides a Python library for generating OHLCV forecasts, fine-tuning on custom datasets, and evaluating performance via backtesting scripts, though it is not a production-ready trading system.
Project overview
It claims to be the first open-source foundation model for financial K-lines, pre-trained on data from over 45 global exchanges.
Deployment
Refer to project documentation
License
MIT
Best for
Data teams, researchers, and AI engineers exploring a specialized open-source foundation model for financial K-line forecasting and custom fine-tuning.
Key capabilities
Generates future forecasts for financial candlestick data (OHLCV) using the KronosPredictor class.
Provides a pipeline to fine-tune the tokenizer and predictor models on user-provided datasets.
Includes scripts to evaluate the finetuned model's performance using a simple top-K strategy backtest.
Limitations and risks
The provided finetuning pipeline is a simplified example and not a production-ready quantitative trading system.
Raw signals generated by the model are not pure alpha; real-world workflows require sophisticated techniques like portfolio optimization to achieve stable returns.
Code comments within the finetune directory were generated by an AI assistant and may contain inaccuracies.
Getting started
Installation requires Python 3.10+ and running pip install -r requirements.txt. Users then download a pre-trained Kronos foundation model (such as Kronos-mini, Kronos-small, or Kronos-base) and corresponding tokenizer (such as Kronos-Tokenizer-base or Kronos-Tokenizer-2k) from Hugging Face Hub, instantiate KronosPredictor, and call the predict method.