Kronos is a novel pre-training framework specifically designed for financial K-line (candlestick) data, addressing the limitations of existing Time Series Foundation Models (TSFMs) in handling the complexities of financial markets and crucial downstream tasks like volatility prediction and synthetic data generation.

The core methodology of Kronos involves a specialized tokenizer and an autoregressive pre-training paradigm. The tokenizer is engineered to discretize continuous market information, transforming it into token sequences. This process is crucial as it preserves not only the raw price dynamics but also intricate trade activity patterns inherent in K-line data, enabling the model to capture a richer representation of market behavior.

Kronos is pre-trained using an autoregressive objective, which allows it to learn temporal dependencies and relationships within the data. This pre-training is performed on an exceptionally large and diverse corpus, comprising over 12 billion K-line records sourced from 45 global exchanges. This massive dataset enables Kronos to learn nuanced temporal and cross-asset representations, making it highly adaptable across various financial instruments and markets.

The efficacy of Kronos is demonstrated across a diverse set of financial tasks in a zero-shot setting. For price series forecasting, Kronos significantly boosts the RankIC (Rank Information Coefficient), achieving a 93% improvement over the leading TSFM and an 87% improvement over the best non-pre-trained baseline. In volatility forecasting, it exhibits superior performance with a 9% lower MAE (Mean Absolute Error). Furthermore, Kronos achieves a 22% improvement in generative fidelity for the creation of synthetic K-line sequences, highlighting its capability in producing realistic financial data.

These results position Kronos as a robust and versatile foundation model for end-to-end financial time series analysis. The pre-trained model is publicly available at https://github.com/shiyu-coder/Kronos.