Kronos

A decoder only foundation model trained on K line data from over 45 global exchanges, positioned as the first open source model for financial candlesticks.

The opening line makes a claim: Kronos is the first open source foundation model for financial candlesticks, or K lines, trained on data from over 45 global exchanges. It is described as a foundation model for the language of financial markets. The project is Python, MIT licensed, and its news section tracks progress: acceptance at AAAI 2026, the release of fine tuning scripts to adapt the model to new tasks, and the paper becoming available on arXiv.

Kronos is a family of decoder only foundation models pre trained specifically for K line sequences. Financial data is high noise, so the design is a two stage framework. A specialized tokenizer first quantizes continuous multi dimensional K line data into hierarchical discrete tokens, and then a large autoregressive Transformer is pre trained on those tokens. That division of labor is the novel part for the domain.

A live demo visualizes forecasts for the BTC/USDT trading pair over the next 24 hours. The model zoo releases a family of pre trained models with varying capacities, sized for different computational and application needs, all accessible from the Hugging Face Hub. For predictions, the README points to a KronosPredictor class that handles preprocessing, normalization, prediction, and inverse normalization, with the max context for Kronos-small and Kronos-base sitting at 512.

Kronos is positioned as a unified model for diverse quantitative tasks built on financial market data. The README stresses that the raw signals from the demo are not alpha; a real quantitative workflow would feed them into portfolio optimization with constraints to neutralize risk factor exposure. Fine tuning is demonstrated with a pipeline for the Chinese A-share market, using Qlib to prepare data and run a backtest, with a disclaimer that the pipeline is a demonstration rather than a production trading system.

The fine tuning pipeline runs in four stages: configuration of paths and hyperparameters, data preparation and splitting with Qlib, fine tuning the tokenizer and predictor models, and backtesting the result. Both training scripts are designed for multi GPU training with torchrun. The README notes that many code comments in the finetune directory were generated by an AI assistant, Gemini 2.5 Pro, and should be treated as potentially inaccurate, with the code itself as the source of truth.

• Source 1 · Project repository
• Source 2 · README