Kronos: The First Foundation Model That Reads Candlestick Charts
Financial quants have a new toy. Kronos is a foundation model pre-trained exclusively on the "language" of financial markets — K-lines, the candlestick patterns showing Open, High, Low, Close, and Volume. Pre-trained on 12 billion K-line records from 45 global exchanges. Now trending on GitHub with 600+ stars per day.
The key insight: financial time series aren't like weather data or web traffic. Candlestick patterns encode a specific visual and statistical language that generic time-series models completely miss. Kronos introduces a custom tokenizer that discretizes continuous market data into token sequences, preserving both price dynamics and trade activity patterns. Then it trains a decoder-only transformer on this tokenized market language using autoregressive prediction.
The results are striking. On benchmark datasets, Kronos improves price forecasting RankIC by 93% over the leading time-series foundation model, and 87% over the best non-pre-trained baseline. Volatility forecasting gets 9% lower MAE. Synthetic K-line generation improves fidelity by 22%. These aren't incremental improvements — they're massive jumps.
Accepted at AAAI 2026, which gives it academic credibility. The family includes multiple model sizes, and the pre-trained weights are publicly available. If you're building trading agents or financial analysis tools, Kronos is the first purpose-built foundation model you can actually use as a backbone.
The bigger picture: we're seeing foundation models specialize. Not everything needs a general-purpose LLM. Sometimes the right answer is a model that speaks one language — candlestick charts — and speaks it really well.
https://github.com/shiyu-coder/Kronos
← Back to all articles
The key insight: financial time series aren't like weather data or web traffic. Candlestick patterns encode a specific visual and statistical language that generic time-series models completely miss. Kronos introduces a custom tokenizer that discretizes continuous market data into token sequences, preserving both price dynamics and trade activity patterns. Then it trains a decoder-only transformer on this tokenized market language using autoregressive prediction.
The results are striking. On benchmark datasets, Kronos improves price forecasting RankIC by 93% over the leading time-series foundation model, and 87% over the best non-pre-trained baseline. Volatility forecasting gets 9% lower MAE. Synthetic K-line generation improves fidelity by 22%. These aren't incremental improvements — they're massive jumps.
Accepted at AAAI 2026, which gives it academic credibility. The family includes multiple model sizes, and the pre-trained weights are publicly available. If you're building trading agents or financial analysis tools, Kronos is the first purpose-built foundation model you can actually use as a backbone.
The bigger picture: we're seeing foundation models specialize. Not everything needs a general-purpose LLM. Sometimes the right answer is a model that speaks one language — candlestick charts — and speaks it really well.
https://github.com/shiyu-coder/Kronos
Comments