Kronos

Overview

Kronos is the first open-source foundation model for financial candlesticks (K-lines), trained on data from over 45 global exchanges. It is a decoder-only foundation model family designed to handle the high-noise characteristics of financial market data through a novel two-stage framework combining a specialized tokenizer with a large autoregressive Transformer.

The project has reached 35,062 GitHub stars and was accepted by AAAI 2026. Kronos provides multiple model sizes from 4.1M to 499.2M parameters, enabling both lightweight prototyping and large-scale quantitative research.

Features

• Specialized Tokenizer: Quantizes continuous, multi-dimensional OHLCV (Open-High-Low-Close-Volume) data into hierarchical discrete tokens
• Multi-Size Model Family: Mini (4.1M), Small (24.7M), Base (102.3M), and Large (499.2M) variants
• Unified Quantitative Tasks: Single model architecture handles forecasting, classification, and pattern recognition
• Finetuning Pipeline: Complete training pipeline with Qlib framework for A-share markets
• Paper Published: Accepted at AAAI 2026 (arXiv: 2508.02739)

Installation

Prerequisites:

• Python 3.10+

git clone https://github.com/shiyu-coder/Kronos.git
cd Kronos
pip install -r requirements.txt

Quick Start

from model import Kronos, KronosTokenizer, KronosPredictor

# Load pretrained tokenizer and model
tokenizer = KronosTokenizer.from_pretrained("NeoQuasar/Kronos-Tokenizer-base")
model = Kronos.from_pretrained("NeoQuasar/Kronos-small")
predictor = KronosPredictor(model, tokenizer, max_context=512)

# Generate price predictions
pred_df = predictor.predict(
    df=x_df,
    x_timestamp=x_timestamp,
    y_timestamp=y_timestamp,
    pred_len=120,
    T=1.0,
    top_p=0.9,
    sample_count=1
)

Model Sizes

Model	Parameters	Context Length	Status
Kronos-mini	4.1M	2048	Available
Kronos-small	24.7M	512	Available
Kronos-base	102.3M	512	Available
Kronos-large	499.2M	512	Training

Core Concepts

• K-Line Tokenization: Financial OHLCV data is tokenized hierarchically, preserving temporal structure while reducing dimensionality
• Two-Stage Framework: First stage (tokenizer) discretizes continuous data; second stage (Transformer) performs autoregressive prediction on discrete tokens
• Noise Robustness: Architecture specifically handles the high-noise signal-to-noise ratio typical of financial markets

Advanced Features

Finetuning with Qlib:

# Data preparation
from qlib.data import D
df = D.query_data(...)

# Tokenizer finetuning
tokenizer.train(data=train_data)

# Predictor finetuning
trainer = KronosTrainer(model, tokenizer)
trainer.fit(train_loader, val_loader, epochs=100)

Examples

Bitcoin Forecast (Live Demo): The project includes a live demo at shiyu-coder.github.io/Kronos-demo/ for BTC/USDT forecasts.

Pattern Classification:

# Classify a trading pattern
tokens = tokenizer.encode(candlestick_data)
output = model(tokens)
classification = predictor.classify(output)

Pros

• ✅ First open-source foundation model specifically for financial K-lines
• ✅ Multiple model sizes for different resource constraints
• ✅ Trained on 45+ global exchanges
• ✅ Complete finetuning pipeline with Qlib
• ✅ AAAI 2026 publication
• ✅ MIT license

Cons

• ❌ Finetuning demo is explicitly not production-ready
• ❌ Kronos-large model not yet released
• ❌ Context window limited to 512 for most variants
• ❌ Specialized domain (finance only)

When to Use

Kronos is ideal for:

• Quantitative research teams exploring ML for financial markets
• Data scientists building price prediction and pattern recognition systems
• Researchers studying foundation models for financial time series
• Educational purposes for learning about tokenization of financial data

Resources

• GitHub Repository
• Live Demo
• Paper (arXiv)
• Hugging Face Models