Kronos: market‑native foundation model

Kronos is an open-source artificial intelligence model built for market charts, not words, and its authors say it beats earlier time-series systems on several finance tasks. (arxiv.org) Market charts package each interval into five fields — open, high, low, close and volume — a format traders call OHLCV or K-lines. Kronos turns those numeric bars into discrete tokens, then trains an autoregressive Transformer to predict what comes next, using the same next-token setup popularized in language models. (arxiv.org) The paper says the model was pre-trained on more than 12 billion K-line records from 45 global exchanges. The public GitHub repository describes Kronos as a family of decoder-only models and says code, pretrained weights and fine-tuning scripts have been released. (arxiv.org) (github.com) The authors report zero-shot results, meaning the model was tested on new tasks without task-specific retraining. On their benchmarks, they say Kronos improved price-series forecasting Rank Information Coefficient by 93 percent over the leading time-series foundation model and 87 percent over the best non-pretrained baseline, while cutting mean absolute error in volatility forecasting by 9 percent. (arxiv.org) That pitch lands as developers keep adapting foundation-model ideas to structured data such as sensor logs, sales histories and prices, where values arrive as sequences over time rather than sentences. Financial data has been a harder target because prices are noisy, regimes shift, and small errors can compound quickly in forecasting. (arxiv.org 1) (arxiv.org 2) Kronos also targets synthetic data generation, where a model creates realistic-looking market sequences for testing and simulation. The paper says it improved generative fidelity for synthetic K-line sequences by 22 percent on the authors’ evaluation. (arxiv.org) The project surfaced publicly in stages: the GitHub repository was available before the paper, the arXiv preprint was posted on August 2, 2025, and the repository later noted the work had been accepted to the Association for the Advancement of Artificial Intelligence 2026 conference. (github.com) (arxiv.org) The paper and repository frame Kronos as the first open-source foundation model focused on financial candlesticks, but the field is getting crowded. Other recent research has pushed foundation models into market microstructure and trade-flow data, using finer-grained transaction records rather than chart bars. (github.com) (arxiv.org 1) (arxiv.org 2) The next test is whether outside researchers can reproduce the gains on live, changing markets rather than static benchmarks. For now, Kronos gives quants and developers a public model, public code and a clear claim: market data may be trainable as its own language. (github.com) (arxiv.org)