Kronos: market‑native foundation model
Kronos was presented as an open‑source foundation model trained on K‑line (OHLCV) data across 45+ exchanges, tokenizing time‑series for forecasting and volatility prediction and claiming up to 93% improvement in some tasks. The project frames market data as model input tokens to support forecasting and quant research workflows. (x.com)
Financial market charts are just time series — numbers laid out in order — and Kronos packages those sequences into tokens so a transformer model can learn patterns across exchanges. The project’s authors released it as open source and say it was trained on candlestick data from more than 45 global venues. (arxiv.org) (github.com) A candlestick, also called a K-line, is a compact record of one trading interval: open, high, low, close, and volume. Kronos turns those five fields into discrete symbols with a two-stage tokenizer, then feeds the symbol stream into a decoder-only transformer, the same broad model family used for next-token prediction in language systems. (arxiv.org) (huggingface.co) The paper says Kronos was trained on more than 12 billion K-line records and evaluated on forecasting, volatility prediction, and synthetic market-data generation. On the authors’ benchmarks, it improved price-series forecasting Rank Information Coefficient by 93% over the leading time-series foundation model, cut volatility mean absolute error by 9%, and improved synthetic-sequence fidelity by 22%. (arxiv.org) (nips.cc) The release lands as finance researchers are testing whether “foundation models” can do for market data what large language models did for text: one pre-trained system, then fine-tuning for narrower tasks. Kronos is aimed at a harder setting than many general time-series models because market data is noisy, non-stationary, and shaped by many exchanges trading around the clock. (arxiv.org) (github.com) The code repository says Kronos is the first open-source foundation model built specifically for financial candlesticks, and the public release includes model weights, examples, tests, and a web interface. A Hugging Face model card for Kronos-base lists the model under time-series forecasting with an MIT license. (github.com) (huggingface.co) The authors on the August 2, 2025 arXiv paper are Yu Shi, Zongliang Fu, Shuo Chen, Bohan Zhao, Wei Xu, Changshui Zhang, and Jian Li. A NeurIPS 2025 virtual poster page also lists the work and repeats the benchmark claims from the paper abstract. (huggingface.co) (nips.cc) The pitch is practical: instead of training separate models for each asset or task, a quant team could start from one pre-trained market model and adapt it to forecasting or risk work. The repository also links a live demo showing probabilistic forecasts for Bitcoin against Tether, with a mean path and uncertainty band rather than a single point estimate. (github.com) (shiyu-coder.github.io) The open question is whether those benchmark gains hold up outside the authors’ test setup and after trading costs, slippage, and regime shifts. For now, Kronos is less a trading system than a public bet that market data can be modeled as a language — and that enough quants will try to prove or disprove it in the open. (arxiv.org) (github.com)
