Kronos

Open-source financial foundation model trained on 45+ global exchanges

Kronos is an open-source financial time-series foundation model published at AAAI 2026 by researchers from Shanghai Jiao Tong University and Fudan University. It is trained on historical OHLCV (Open, High, Low, Close, Volume) candlestick data from 45+ global stock exchanges, covering US equities, A-shares, Hong Kong stocks, and international markets. Unlike most financial ML models that require exchange-specific fine-tuning, Kronos uses a universal tokenizer that converts candlestick patterns into discrete tokens, enabling zero-shot forecasting on unseen assets. The architecture is an autoregressive transformer available in three scales: 4.1M, 24.7M, and 102.3M parameters. Kronos is trained with a hybrid objective that combines next-token prediction (for pattern learning) and contrastive learning (for distinguishing market regimes like trending vs. mean-reverting). All three model sizes are available on HuggingFace, and the repository includes a live BTC/USDT 24-hour forecast demo served as a Gradio app. Kronos reached 6,486 GitHub stars in its first trending week, driven by interest from quantitative finance communities on Reddit and Twitter. While the academic paper carefully avoids strong trading performance claims (noting Sharpe ratios rather than absolute returns), the community reception has focused on its potential as a base model for fine-tuning on specific asset classes — similar to how LLaMA is used as a base for specialized language models.

Panel Reviews

The Builder

Developer Perspective

Ship

“Clean HuggingFace release with all three model sizes, clear tokenization docs, and a working Gradio demo is exactly how academic code should be shipped. The AAAI peer review adds credibility. As a base model for quantitative feature extraction (not necessarily direct trading signals), this is worth evaluating.”

The Skeptic

Reality Check

Skip

“Financial forecasting models are notoriously data-mined. The paper's backtests look good, but they always do before live trading. Markets are adversarial — anything broadly publicized gets arbed away. The BTC/USDT demo is a marketing piece, not a trading signal. Test on out-of-sample data before trusting anything here.”

The Futurist

Big Picture

Ship

“A universal tokenizer for financial candlestick data could be as important as the BPE tokenizer was for NLP. Once you can represent market data as discrete tokens, the entire LLM architecture toolkit becomes applicable to financial time series. This is early-stage but directionally important.”

The Creator

Content & Design

Skip

“Not a creator tool by any stretch — but the visualization work in the paper's figures is genuinely well-designed. The candlestick-to-token visualization makes a technically complex concept legible. If you're building fintech UX, there's inspiration in how they communicate model uncertainty.”