Kronos — Overview
Kronos is an open-source foundation-model family specifically designed for the "language" of financial markets — namely multi-dimensional K-line sequences (OHLCV). Unlike general-purpose time-series foundation models, Kronos targets the particular challenges of market data (high noise, heavy tails, heteroskedasticity) by combining a domain-aware discrete tokenizer with a large autoregressive Transformer backbone.
Core ideas and architecture
- Hierarchical discrete tokenizer: Kronos first quantizes continuous OHLCV inputs into hierarchical discrete tokens, enabling the Transformer to model market signals in a tokenized sequence space.
- Decoder-only autoregressive Transformer: After tokenization, Kronos trains a standard autoregressive Transformer to predict future tokens, which can be decoded back to price/volume forecasts.
- Two-stage framework: (1) tokenizer training/finetuning; (2) predictor pretraining/finetuning. This separation lets users adapt the tokenizer to new markets or granularities and finetune predictors for downstream tasks.
Model zoo and capabilities
Kronos releases several pre-trained sizes to balance compute and performance:
- Kronos-mini (≈4.1M params, context 2048) — tokenizer Kronos-Tokenizer-2k
- Kronos-small (≈24.7M params, context 512) — tokenizer Kronos-Tokenizer-base
- Kronos-base (≈102.3M params, context 512) — tokenizer Kronos-Tokenizer-base
- Kronos-large (≈499.2M params, context 512) — (not open-sourced at time of README snapshot)
Capabilities include probabilistic forecasting (temperature / top_p sampling, sample aggregation), batch prediction utilities, and a predictor class that handles normalization/denormalization, truncation to model context length, and generation of OHLCV forecasts.
Getting started & tooling
Kronos provides:
- Ready-to-use Python classes: KronosTokenizer, Kronos, and KronosPredictor.
- Example scripts for single-series and batch predictions (including plotting examples).
- Finetuning pipelines and training scripts (tokenizer and predictor) designed to run with torchrun for multi-GPU training.
- A Qlib-based example showing data preprocessing, finetuning, and a simple backtest to illustrate how to turn model forecasts into trading signals.
- Live demo showcasing a BTC/USDT 24-hour forecast: the project hosts an interactive demo page.
Typical usage flow:
- Load tokenizer and model from the Hugging Face Hub.
- Instantiate KronosPredictor with device and max_context.
- Prepare historical K-line DataFrame (open, high, low, close, optional volume/amount) and timestamps.
- Call predictor.predict or predict_batch with sampling options to produce forecasts.
Practical notes & considerations
- Context length limits: Kronos-small/base have a max_context of 512 tokens; Kronos-mini supports longer context. Inputs longer than max_context will be truncated by the predictor.
- Forecast semantics: outputs are forecasts for OHLC(V) and may be used as raw signals. For production quantitative strategies, outputs should be post-processed (risk-neutralization, portfolio optimization) and backtested with realistic transaction costs and slippage modeling.
- Fine-tuning: the repo includes scripts to finetune the tokenizer and predictor on custom datasets. The README demonstrates a pipeline using Microsoft Qlib, but is explicit that this is a demo and not a production trading system.
Research & licensing
- The project provides a citation entry and links to an arXiv preprint (paper related to Kronos) and notes acceptance to AAAI 2026 (per README news items).
- Licensed under the MIT License.
Where to find models and demo
- GitHub repository: the code, examples, and finetune scripts are in the repo.
- Hugging Face Hub: tokenizers and model checkpoints (e.g., NeoQuasar/Kronos-mini, NeoQuasar/Kronos-small, NeoQuasar/Kronos-base) are published for easy download.
- Live demo: an interactive demo visualizes forecasts for an example asset.
Overall, Kronos is positioned as a domain-specialized foundation model stack for financial time series research and applied forecasting, offering both pre-trained artifacts and an end-to-end pipeline for adaptation and evaluation.
