What is Kronos?
Kronos is a family of decoder-only foundation models trained specifically on financial candlestick sequences. Its tokenizer converts continuous market data into hierarchical discrete tokens, which an autoregressive Transformer uses to generate forecasts. Inputs include historical open, high, low and close values, with optional volume and traded amount data.
The models run through Python code, with pretrained models and tokenizers available from Hugging Face Hub. The KronosPredictor class handles preprocessing, normalization, forecasting and inverse normalization, returning predictions as pandas DataFrames. The repository also provides visualization examples and a fine-tuning workflow using Qlib for data preparation and backtesting.
Kronos is a modeling component rather than a production-ready trading system. Its example strategy requires additional portfolio construction, risk management and realistic transaction-cost modeling before operational use. The model catalogue identifies mini, small and base variants as open source, while the large variant is marked as not open source.
The project describes Kronos as the first open-source foundation model for financial candlesticks, trained on data from over 45 global exchanges, and reports that the work was accepted by AAAI 2026.
Key capabilities
- Candlestick forecasting Generates future market-data sequences from historical price observations and timestamps.
- Probabilistic sampling Exposes temperature, nucleus sampling and forecast-path sampling controls.
- Batch prediction Processes multiple series with GPU parallelism when historical windows and forecast lengths match.
- Data processing Handles normalization and inverse normalization, with optional volume and amount inputs.
- Model adaptation Provides separate fine-tuning scripts for the tokenizer and forecasting model, including multi-GPU training.
- Research evaluation Includes Qlib-based data preparation, a demonstration backtest and forecast comparison plots.
Where Kronos fits
Kronos fits quantitative research workflows that turn historical candlestick data into forecasts or experimental trading signals. Developers can run inference on individual assets, process multiple series in batches or adapt pretrained models to their own datasets. The Qlib example demonstrates data preparation, fine-tuning and strategy evaluation on Chinese A-share market data.
Before deployment, teams should verify input columns, timestamp alignment, model-tokenizer pairing and available compute. Kronos-small and Kronos-base have a context limit of 512 observations, with longer inputs truncated automatically. Batch inputs require matching historical and prediction lengths. Strategy evaluation should account for transaction costs, slippage, market impact and risk-factor exposure rather than treating raw forecasts as a complete investment strategy.
Install
pip install -r requirements.txtReviews of Kronos
Be the first to share your experience with Kronos.
Write a review
Loading review form…
Alternatives to Kronos
HummingbotHummingbot is an Apache 2.0 framework for automated crypto trading strategies. It runs locally or in Docker, connects to centralized and decentralized exchanges and offers scripts, controllers, executors, paper trading and the hbot CLI.
OctoBotOctoBot is a GPL-licensed, self-hosted crypto trading bot by Drakkar-Software. It automates DCA, grid, basket, LLM and TradingView strategies on CCXT exchanges, with a simulator, backtesting and a new node mode.