Predict future price movements from candlestick data
Input historical candlestick data to obtain predictions on how future price movements will unfold. This is the first open-source foundation model designed with consideration for the high noise characteristics of financial markets. It was trained on data from over 45 global exchanges and was accepted to AAAI 2026.
Structure for converting candle data into tokens
The model uses a dedicated tokenizer that converts continuous OHLCV data into hierarchical discrete tokens. An autoregressive transformer then learns these tokens to perform various quant tasks. Unlike general time-series models, it is designed to handle the unique noise of financial data.
Model scale and release scope
Pre-trained models of various capacities are available on the Hugging Face Hub. The smallest, Kronos-mini, has 4.1M parameters and a context length of 2048. Kronos-small (24.7M) and Kronos-base (102.3M) have a context length of 512. The largest, Kronos-large (499.2M), is not released as open source.
Running predictions and batch processing
Install dependencies with pip install -r requirements.txt in a Python 3.10 or higher environment. Using the KronosPredictor class allows you to complete everything from data preprocessing to prediction in a few lines of code. Input data must be in the form of a pandas DataFrame containing open, high, low, and close columns, while volume and amount are optional. To predict multiple assets or periods simultaneously, use the predict_batch method to leverage GPU parallel processing. In this case, all series must have the same historical length and prediction length.
Considerations before practical application
This repository is released under the MIT license. The provided fine-tuning pipeline is a simplified example for demonstration purposes and is not a production-ready quant trading system. Signals generated by the model are raw predictions, so actual operation requires processes such as portfolio optimization or risk factor neutralization. Additionally, backtesting must accurately model transaction costs, slippage, and market impact to evaluate realistic performance. Some code comments are AI-generated, so the code itself should be the final reference for logic.
