Kronos: The Open-Source Foundation Model That Learns the Language of Financial Markets

Financial markets generate vast quantities of structured data. Every trade, every price movement, every volume spike is recorded as a sequence of candlestick patterns 鈥?the iconic OHLCV (Open, High, Low, Close, Volume) data that traders have used to read market behavior for over a century. But here’s the striking thing: while large language models have transformed how we process text, code, and natural language, the language of financial markets has never had its own foundation model 鈥?until now.

Kronos, developed by researchers from NeoQuasar and accepted to AAAI 2026, is the world’s first open-source foundation model specifically designed for financial candlestick (K-line) data. It’s a family of decoder-only Transformer models pre-trained on data from over 45 global exchanges, capable of serving as a unified foundation for diverse quantitative finance tasks 鈥?from price forecasting to risk modeling.

Why Financial Markets Needed Their Own Foundation Model

General-purpose time series foundation models (TSFMs) exist, and some are quite capable. But financial candlestick data has a unique character that makes it fundamentally different from the time series these models were designed for:

• Extreme noise: Market data is notoriously noisy 鈥?price movements are influenced by countless factors, many of which are unobservable or random
• Multi-dimensional structure: A single candlestick isn’t just a single number 鈥?it’s five interconnected values (OHLCV) that must be understood together
• Regime shifts: Market behavior can change dramatically between bull and bear markets
• Global heterogeneity: Different exchanges have different trading hours, tick sizes, and participant compositions, but the underlying “language” of price action has commonalities across all of them

How Kronos Works: The Two-Stage Architecture

The Kronos-Tokenizer is the key innovation. Rather than treating each OHLCV value independently, it creates hierarchical discrete tokens that capture the structural relationships within candlestick data. This tokenizer maps the continuous, multi-dimensional price-volume space into a sequence of discrete tokens that the Transformer can process using the same attention mechanisms that make LLMs so effective at modeling sequential data.

Once tokenized, the Kronos Transformer 鈥?a decoder-only autoregressive model 鈥?is pre-trained on massive quantities of K-line sequences from 45+ global exchanges. This pre-training teaches the model the “grammar” of financial markets.

Model Family: From 4.1M to 499M Parameters

Kronos comes in four sizes:

The open-source versions are all available on Hugging Face.

Forecasting with Kronos

Kronos is designed to be straightforward to use:

from model import Kronos, KronosTokenizer, KronosPredictor

tokenizer = KronosTokenizer.from_pretrained("NeoQuasar/Kronos-small")
model = Kronos.from_pretrained("NeoQuasar/Kronos-small")
predictor = KronosPredictor(model, tokenizer, max_context=512)

pred_df = predictor.predict(
    df=x_df,
    x_timestamp=x_timestamp,
    y_timestamp=y_timestamp,
    pred_len=pred_len,
    T=1.0,
    top_p=0.9,
    sample_count=1
)

Probabilistic Forecasting

Unlike traditional point-estimate forecasting models, Kronos supports probabilistic forecasting 鈥?generating multiple possible future paths and averaging them. By adjusting T (temperature) and top_p (nucleus sampling probability) and increasing sample_count, you can generate a distribution of possible outcomes rather than a single prediction.

Fine-Tuning for Your Own Data

Kronos can be fine-tuned on your own proprietary data using Microsoft’s Qlib. The fine-tuning pipeline includes tokenizer adaptation, model fine-tuning, and backtesting evaluation.

Live Demo: Bitcoin Forecasting in Your Browser

Perhaps the most engaging way to understand what Kronos can do is the live demo hosted on GitHub Pages. The demo shows a 24-hour BTC/USDT forecast 鈥?you can watch the model generate probabilistic price paths and see how the predictions compare to actual market movement.

AAAI 2026 Recognition

The acceptance of the Kronos paper by AAAI 2026 signals that this work is taken seriously by the academic community. The paper describes both the architectural innovations and extensive experimental results across multiple financial tasks and markets.

Important Caveats

Kronos is a research project, and the authors are transparent about its limitations. Financial markets are complex adaptive systems 鈥?past performance does not guarantee future results, and any trading strategy developed using Kronos should be thoroughly backtested and validated before deployment with real capital.

For researchers studying financial time series, quants looking for a powerful base model to build on, and developers interested in the intersection of AI and finance, Kronos is one of the most interesting open-source releases of 2026.

pip install -r requirements.txt

All open-source model weights are available on Hugging Face under the MIT license.