Key takeaways
- Forecast price candles with an AI model built for markets, not for general time series.
- Start from ready-made models on Hugging Face, with several sizes to match your hardware.
- Adapt it: a finetuning pipeline lets you train on your own market data.
- Try the idea first: a live demo shows a BTC/USDT forecast for the next 24 hours.
- The catch: this is research software, not financial advice, and its own backtest is only a demonstration.
What it is
Kronos is a family of decoder-only foundation models, pre-trained on candlestick sequences, the price bars used in market charts. It is open source under the MIT licence and lives on GitHub at shiyu-coder/Kronos.
The problem it solves
Market data is noisy, and general-purpose time-series models were not built for it. Quantitative researchers and students who want to experiment with forecasting have had few open models designed for candlesticks.
Kronos treats price candles like a language: a tokenizer turns continuous candle data into discrete tokens, and a Transformer learns from those tokens. The README says the training data comes from over 45 global exchanges.
What you get
- Ready-made models: pre-trained weights are available from the Hugging Face Hub, in sizes the README lists from 4.1M to 499.2M parameters.
- Forecasts with uncertainty: sampling settings support probabilistic forecasting, so you can generate more than one possible path.
- Many series at once: a batch method predicts on multiple datasets in parallel.
- Your own data: a complete pipeline is provided for finetuning on your own datasets.
- A quick look: a live demo shows a BTC/USDT forecast for the next 24 hours.
The catches
- This is research software, not financial advice. A forecast is not a promise, and acting on one can lose money.
- The finetuning and backtest pipeline is, in the project's words, a simplified example and not a production-ready quantitative trading system.
- The demo signals are raw predictions. A realistic backtest, the README says, must model transaction costs, slippage and market impact.
- Some code comments in the finetuning folder were written by an AI assistant and, the README warns, may contain inaccuracies.
- Maturity: the work is described as accepted by the AAAI 2026 conference, but that is the authors' statement and we did not check it. Setup needs Python 3.10 or newer.
What we did not test
We did not run Kronos, download any model or test its forecasts against real prices. Every claim here comes from a literal quote in the project's README or the GitHub API, and the descriptions of the models are the authors' own.
Verdict
Editors' view: this is for researchers, students and developers curious about forecasting models built for markets. If you want a ready-made trading tool, this is not it. Treat any output as an experiment, and validate it independently before it goes near money.
Sources
Need help adopting this, or have a software project in mind?
Tell us what you are building. The initial consultation is free. Email [email protected] or follow @zolnaltd.
Start a conversation →