Kronos: The First Open-Source Foundation Model for Financial K-Line Data

What if we could treat financial candlestick data the way GPT treats natural language? That's exactly what Kronos does.

Built by researchers from Tsinghua and Nanjing University, Kronos is the first open-source foundation model specifically designed for financial market K-line (candlestick) data. It was accepted at AAAI 2026 and has gained 14,800+ GitHub stars.

The Problem with General-Purpose Models

General-purpose time series foundation models (TSFMs) like TimeMoE or Moirai don't perform well on financial data. Financial time series have:

Low signal-to-noise ratio — markets are inherently noisy
Strong non-stationarity — statistical properties change over time
Higher-order dependencies — complex patterns beyond simple autocorrelation

Kronos addresses these challenges with a domain-specific approach.

Two-Stage Framework

Stage 1: Custom Tokenizer

Kronos converts continuous multi-dimensional OHLCV (Open, High, Low, Close, Volume) data into hierarchical discrete tokens using Binary Spherical Quantization (BSQ). This preserves both price dynamics and trading activity patterns.

Stage 2: Autoregressive Transformer

A decoder-only transformer (GPT-style) is pre-trained on 12 billion+ K-line records from 45 global exchanges. It covers:

Asset types: Stocks, futures, forex, options, cryptocurrency
Timeframes: 1min, 5min, 15min, 30min, 60min, daily, weekly, biweekly

Benchmark Results

The zero-shot performance improvements are significant:

Metric	Improvement	Compared To
Price prediction RankIC	93%	TimeMoE (SOTA TSFM)
Volatility prediction MAE	9%	Previous TSFM
Synthetic data fidelity	22%	Previous models
Return prediction (full-shot)	60%	DLinear

Quick Start

from model import Kronos, KronosTokenizer, KronosPredictor

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

pred_df = predictor.predict(
    df=x_df,
    x_timestamp=x_ts,
    y_timestamp=y_ts,
    pred_len=120
)

Three model sizes are available on Hugging Face:

Kronos-mini (4.1M params, context 2048) — lightweight, CPU-friendly
Kronos-small (24.7M params, context 512) — general prediction
Kronos-base (102.3M params, context 512) — highest accuracy

All under MIT license.

Fine-Tuning

Fine-tuning scripts were released in August 2025:

# Tokenizer fine-tuning
python finetune_tokenizer.py --data_path ./data/your_market/ --epochs 50

# Model fine-tuning (multi-GPU)
torchrun --nproc_per_node=4 finetune_model.py \
    --model_name NeoQuasar/Kronos-base \
    --data_path ./data/your_market/ \
    --distributed