Kronos is an open-source foundation model for stock-market candlestick data (K-line bars), trained to predict the next price bar the way a language model predicts the next word. It was built by researchers at Tsinghua University's Institute for Interdisciplinary Information Sciences (Yu Shi, Zongliang Fu, Shuo Chen, Bohan Zhao, Wei Xu, Changshui Zhang and Jian Li), and the accompanying paper was accepted at AAAI 2026. It is aimed at quantitative researchers, machine-learning engineers who work with market data, and anyone who wants a pre-trained starting point for price modelling instead of training a time-series network from scratch. It is not a trading bot and not an app: it is model weights plus code that you run on your own machine.
What it does
Kronos takes a sequence of OHLCV bars (open, high, low, close, volume) and continues it. The authors describe three uses for that single capability:
- forecasting future price bars over a chosen horizon,
- predicting volatility,
- generating synthetic price series that are meant to look like real market data.
It ships as a family of models rather than one checkpoint, from Kronos-mini at 4.1M parameters up to Kronos-large at 499.2M parameters. Mini, small and base are released openly; large is not.
The headline numbers are the authors' own, reported in their paper and not verified by anyone independent: 93% higher RankIC (a forecasting-accuracy metric) than the strongest prior time-series foundation model, 87% higher than the best baseline that was not pre-trained, 9% lower MAE on volatility forecasting, and a 22% improvement in the realism of the synthetic price data it produces. Treat them as a claim from the people who built the model, which is how the project itself presents them.
How it works
The idea borrowed from language modelling is tokenization. Instead of words, Kronos converts each price bar into discrete tokens, then trains a Transformer to predict the next token in the sequence. Sampling from the trained model gives you a continuation of the chart, which is why the same model can be read as a forecaster or as a generator of synthetic markets.
What separates it from a generic sequence model is the training set. According to the paper, it was pre-trained on more than 12 billion K-line records from 45 or more global exchanges, covering 30-plus countries and seven time frequencies, from one-minute bars up to weekly ones. The argument is that a model fed only market data learns market behaviour, rather than the generic shapes that a general-purpose time-series model learns from traffic counts, electricity load and sales figures.
Getting started
The code lives at github.com/shiyu-coder/Kronos and the weights are published on Hugging Face. There is no hosted service and no API key: you clone the repository, download a checkpoint and run inference locally, which also means the compute bill is your own hardware rather than a subscription. The code is MIT licensed. Nothing in the project's materials names an operating system, a GPU, a VRAM floor or a minimum amount of memory, so treat hardware as unlisted; the parameter counts are the only guide you get, and the openly released models are small by current standards, with the largest open option well under half a billion parameters. The paper is on arXiv as 2508.02739 if you want the training details and the evaluation setup before you pull the weights.
When to use it / when not
The main catch is the one the project states plainly: Kronos-large, the 499.2M-parameter version, stays closed source, so the published results were not all produced by something you can download. Behind that sits a second caveat: RankIC and MAE are accuracy metrics, not returns, and the repository does not claim a trading strategy.
Use it instead of training your own candlestick Transformer from scratch when you want a pre-trained baseline for research or feature generation and you have the data pipeline to feed it bars. Don't use it if you are looking for a finished signal, a backtester or a product: this is a model, and everything around it is yours to build.
Alternatives
The obvious comparison, and the source of constant confusion, is Amazon's Chronos: a well-known general-purpose time-series foundation model built on a T5 architecture and used in AWS forecasting tooling. The names are nearly identical and the projects are unrelated. Chronos is the broader tool, covering time series of any kind; Kronos is narrow by design, finance only, and trained specifically on market bars. If your series is not a price chart, Chronos is the one you want. The other alternative is the one most quant teams already use: a conventional, non-pretrained model trained on your own data, which Kronos claims to beat by 87% on RankIC in its own evaluation.
Anyone doing research on financial time series should take this seriously, if only to have a pre-trained reference point: the training corpus is unusually large for this niche, the licence is permissive, and three of the four model sizes are downloadable today. Traders hoping to plug it in and get a signal will be disappointed, and so will anyone who wants the version that produced the best numbers.