Kronos is an open-source foundation model specifically designed for financial market analysis, trained on candlestick data (K-lines) from over 45 global exchanges. Written in Python, the repository implements a decoder-only transformer architecture that treats financial time series as a language to be modeled autoregressively. The core innovation lies in its two-stage framework: a specialized tokenizer that converts continuous, multi-dimensional OHLCV (open, high, low, close, volume) data into hierarchical discrete tokens, followed by a large transformer pre-trained on these tokens to handle the high-noise characteristics inherent in financial data.
The project provides a family of pre-trained models with varying computational requirements. The model zoo includes Kronos-mini with 4.1 million parameters and a 2048 token context length, Kronos-small with 24.7 million parameters and 512 token context, Kronos-base with 102.3 million parameters and 512 token context, and Kronos-large with 499.2 million parameters. All models except Kronos-large are open-source and available through the Hugging Face Hub. The repository includes a live demo showcasing 24-hour forecasts for BTC/USDT trading pairs and has been accepted for publication at AAAI 2026.
The codebase provides straightforward interfaces for financial forecasting through a KronosPredictor class that handles data preprocessing, normalization, prediction, and denormalization automatically. Users can generate forecasts from raw historical K-line data in just a few lines of code. The implementation supports both single time series prediction and batch prediction for multiple assets simultaneously, with the batch prediction method leveraging GPU parallelism for efficient processing. Input data requires a pandas DataFrame with open, high, low, and close columns, with volume and amount being optional fields.
The repository includes comprehensive examples and documentation for getting started with forecasting. A prediction example script demonstrates loading data, generating forecasts, and visualizing results by comparing ground truth against model predictions. An alternative example shows how to make predictions using only OHLC data without volume and amount information. The project also provides a complete finetuning pipeline for adapting Kronos to custom datasets, demonstrated through a Chinese A-share market example using Qlib for data preparation.
The finetuning workflow consists of four main steps: configuration setup, data preparation using Qlib, model finetuning for both the tokenizer and predictor components, and backtesting evaluation. The configuration system centralizes all settings including data paths, training hyperparameters, and model checkpoint locations. The pipeline supports integration with Comet.ml for experiment tracking, though this is optional.
According to GitGenius activity tracking, the repository has shown engagement from multiple contributors with stewart050-crypto-beathany-c, loyedison92, and amgadatiata-ui being the most active triagers and contributors. Issue and pull request response latency shows a median of 15.8 hours with a mean of 93.8 hours across 23 tracked items. The repository shares overlapping contributors with several other projects including koala73/worldmonitor, dograh-hq/dograh, and base/node. The project has been classified under job scheduling, distributed systems, fault tolerance, and workflow automation categories, reflecting its role in orchestrating complex financial data processing and model inference tasks.
