Skip to content
Get packs

Kronos

Open-source pre-trained models that forecast financial candlestick (K-line) price data.

GitHub Download source
Stars
45.7k
Forks
7.5k
Licence
MIT
Language
Python
Last update
7 months ago
Type
AI model / agent

What it is

Kronos is a family of decoder-only foundation models pre-trained on financial K-line sequences (open, high, low, close, volume data). It first converts price candles into discrete tokens with a specialized tokenizer, then uses a Transformer trained on those tokens to forecast future price candles. The README says the models are on Hugging Face and can be run from Python with a KronosPredictor class.

  • Load historical 5-minute stock K-line data from a CSV and generate a forecast for the next 120 periods.
  • Generate multiple probabilistic forecast paths by adjusting sample_count, T, and top_p.
  • Run batch predictions across several time series that share the same lookback length using predict_batch.
  • Fine-tune a pre-trained Kronos model on your own asset data using the released scripts.

Great for

  • Quantitative researchers who want a pre-trained time-series model for K-line data
  • Developers experimenting with forecasting on stock, crypto, or other exchange data
  • Teams that want to fine-tune a base model on their own market data (fine-tuning scripts are mentioned in the README news)
  • Students learning how foundation models can be applied to financial time series

Not ideal for

  • People looking for a ready-made trading bot or guaranteed returns; the README only describes forecasting, not trading strategies or performance.
  • Users who need forecasts with no Python setup, since the README's usage path is Python code with pip-installed dependencies.
  • Anyone who cannot evaluate forecast quality themselves; the README does not present a benchmark summary in the truncated text, so accuracy should be verified independently.
Step-by-step guide: how to use KronosInstall, first run, examples, tips and FAQs

How to get it

  1. 1

    Clone and build from source

    You'll need: Python 3.10+ · pip

    Install dependencies

    pip install -r requirements.txt
  2. 2

    Run it

    • Clone the repository from GitHub.
    • Install Python 3.10+ and run pip install -r requirements.txt.
    • Load a tokenizer and model with KronosTokenizer.from_pretrained and Kronos.from_pretrained from Hugging Face.
    • Create a KronosPredictor and call predict on a pandas DataFrame with open, high, low, close columns.
  3. 3

    Get the most out of it

    • Keep lookback length at or below 512 for Kronos-small and Kronos-base, as the README recommends.
    • Use sample_count greater than 1 to generate multiple forecast paths and inspect their spread.
    • Use predict_batch to forecast many series at once, keeping all series at the same lookback length.
    • Fine-tune the model on your own data using the fine-tuning scripts added in August 2025.

Downloads

Source code (.zip)master branch · straight from GitHub

Strengths

  • Trained specifically on K-line sequences rather than general-purpose time series, per the README's description.
  • Multiple model sizes are published on Hugging Face, from 4.1M to 102.3M parameters for open models.
  • Clear step-by-step code examples for loading a model, preparing inputs, and generating forecasts.

Trade-offs

  • The README is truncated and does not show accuracy or backtest results in the text provided, so real-world performance is unclear.
  • The largest model (Kronos-large, 499.2M parameters) is listed as not open-source in the model table.
  • Forecasts are probabilistic and depend on input data quality and settings; the README includes a disclaimer-free description and does not cover financial risk.

Stats synced from GitHub. The assessment above was drafted with AI from the project's README on 2026-10-29 and may miss recent changes — check the README before relying on a command. MIT licence: free to use, modify and distribute, including commercially, as long as you keep the copyright notice.