Kronos: market‑data foundation model
Kronos is an open‑source foundation model trained on billions of K‑line records from more than 45 exchanges, aimed at forecasting and volatility prediction and capable of producing synthetic market data. Social reports claim Kronos improves some forecasting metrics by up to 93% and is being positioned for quant research and signal generation. (x.com)
Financial markets generate streams of price bars — open, high, low, close, and volume — and Kronos is a new open-source model built to learn patterns from that data. (arxiv.org) Those price bars are often called candlesticks or K-lines, and traders use them as compressed snapshots of how an asset moved over a minute, hour, or day. The Kronos paper says the model was pre-trained on more than 12 billion K-line records from 45 global exchanges. (arxiv.org) The paper, posted to arXiv on August 2, 2025, describes Kronos as a “foundation model” for financial time series, meaning one large model is trained first and then adapted to tasks such as forecasting returns, estimating volatility, and generating synthetic sequences. The code repository says the project is open source under the Massachusetts Institute of Technology license. (arxiv.org) (github.com) To make market data fit a language-model style system, the authors convert continuous price and volume moves into discrete tokens — the same basic idea as turning words into IDs before training a chatbot. The paper says that tokenizer is designed to preserve both price dynamics and trading-activity patterns. (arxiv.org) The headline numbers circulating on social media come from the paper’s benchmark section. The authors report a 93 percent gain in Rank Information Coefficient for price-series forecasting over a leading time-series foundation model, a 9 percent lower mean absolute error in volatility forecasting, and a 22 percent improvement in generative fidelity for synthetic K-line sequences. (arxiv.org) Those figures are relative improvements on the authors’ chosen benchmarks, not a claim that markets became 93 percent predictable. In finance research, even small gains can disappear when models face new regimes, transaction costs, and live trading constraints. (arxiv.org 1) (arxiv.org 2) That caution is part of a larger debate around time-series foundation models. A 2024 survey on the field said these models promise reusable representations across tasks, but also flagged challenges including distribution shift, data heterogeneity, and evaluation standards. (arxiv.org) Kronos enters that debate with unusually broad market coverage for an open model. The GitHub repository says it includes pretrained checkpoints, examples, fine-tuning code, and a web interface, which lowers the barrier for researchers and developers who want to test it on equities, futures, foreign exchange, or crypto-style candlestick data. (github.com) The immediate use case is research, not proof of a profitable trading system. The paper presents forecasting and generation results, but it does not claim that Kronos by itself delivers live, net-of-cost trading returns across markets. (arxiv.org) So the story is less that a model has solved markets than that a finance-specific base model is now public, large-scale, and easy to inspect. Whether Kronos becomes a standard tool will depend on what outside researchers find when they test it beyond the paper’s own benchmarks. (github.com) (arxiv.org)
