Kronos is a groundbreaking family of decoder-only foundation models specifically designed for the "language" of financial markets, represented by K-line sequences. Unlike traditional time series forecasting models, Kronos is engineered to effectively manage the inherent high-noise characteristics of financial data.
Key features and capabilities include:
- Novel Two-Stage Framework: Kronos employs a unique two-stage process. First, a specialized tokenizer quantizes continuous, multi-dimensional K-line data (OHLCV) into hierarchical discrete tokens. Second, a large, autoregressive Transformer is pre-trained on these tokens, enabling it to serve as a unified model for a wide array of quantitative tasks.
- Financial Forecasting: The
KronosPredictorclass simplifies the forecasting process, handling data preprocessing, normalization, prediction, and inverse normalization, allowing users to generate forecasts with minimal code. - Batch Prediction: For efficiency, Kronos offers a
predict_batchmethod that supports parallel prediction across multiple datasets simultaneously, ideal for forecasting numerous assets or time periods. - Model Zoo: A range of pre-trained models with varying capacities (e.g., Kronos-mini, Kronos-small, Kronos-base) are available on the Hugging Face Hub, catering to diverse computational and application needs.
- Finetuning Pipeline: Kronos provides a comprehensive pipeline for finetuning the model on custom datasets, demonstrated with an example using Qlib for the Chinese A-share market.
- Live Demo: A live demo is available to visualize Kronos's forecasting results, showcasing predictions for trading pairs like BTC/USDT over specified periods.
Kronos is ideal for researchers and developers in quantitative finance seeking a robust, open-source foundation model for financial market analysis and prediction.
