Kronos

Kronos is an open-source foundation model family for financial candlestick (K-line) data, published in the shiyu-coder GitHub repository. It tokenizes OHLCV sequences and uses an autoregressive Transformer to forecast future market data. Pre-trained models are distributed through the Hugging Face Hub, and the repository includes forecasting examples and a fine-tuning pipeline.

What It Is

Kronos is a set of decoder-only models pre-trained on the "language" of financial markets. The README states it is the first open-source foundation model for K-lines, trained on data from over 45 global exchanges. Unlike general-purpose time-series foundation models, it is designed for the high-noise characteristics of financial data.

It uses a two-stage framework: a specialized tokenizer quantizes continuous multi-dimensional OHLCV data into hierarchical discrete tokens, and a large autoregressive Transformer is pre-trained on those tokens to serve as a unified model for quantitative tasks.

Forecasting Workflow

Users load a tokenizer and a model from Hugging Face, then wrap them in the KronosPredictor class, which handles preprocessing, normalization, prediction, and inverse normalization. Input is a pandas DataFrame with open, high, low and close columns (volume and amount are optional), plus historical and future timestamps. Sampling parameters such as temperature, top-p, and sample count control probabilistic forecasts. The output is a DataFrame of forecasted open, high, low, close, volume, and amount values. A predict_batch method forecasts multiple series in parallel, provided they share the same lookback and prediction length. Example scripts produce plots comparing forecasts with ground truth.

Model Zoo

The released models are Kronos-mini (4.1M parameters, 2048 context), Kronos-small (24.7M, 512), and Kronos-base (102.3M, 512), all open on Hugging Face. Kronos-large (499.2M) is listed as not open-source. A live demo shows a 24-hour BTC/USDT forecast.

Fine-tuning and Backtesting

The repository provides a pipeline that uses Qlib to prepare Chinese A-share data, fine-tunes the tokenizer and then the predictor with multi-GPU torchrun, and runs a simple top-K strategy backtest. The README describes it as a demonstration, not a production-ready trading system. Kronos has been accepted by AAAI 2026, and a paper is available on arXiv. The project is released under the MIT License.