Most time-series foundation models treat market data as just another numeric signal and drown in its noise. Kronos instead borrows the LLM playbook: it learns the "language" of candlesticks, tokenizing OHLCV bars into a discrete vocabulary and predicting the next bar autoregressively. It is the first open foundation model built specifically for financial K-lines.
Key Capabilities
- Two-stage design. A specialized tokenizer turns continuous, multi-dimensional OHLCV bars into hierarchical discrete tokens, which a decoder-only transformer then models autoregressively. This split is what lets it absorb the high noise that trips up general-purpose time-series models.
- Cross-market scale. Pre-trained on over 12 billion K-line records from 45 global exchanges, so it carries priors across assets and venues rather than overfitting one market.
- One backbone, many tasks. The same model handles multistep price/return forecasting, realized-volatility forecasting, synthetic K-line generation, and trading simulation.
- Reported gains. Versus baselines: ~93% improvement over the leading time-series foundation model and ~87% over the best non-pre-trained baseline on price forecasting, 9% lower MAE on volatility, and 22% better generative fidelity for synthetic data.
Who It's For
Great fit if you are a quant researcher or builder who wants a pre-trained OHLCV backbone instead of training a forecaster from scratch — three of its four sizes ship openly (Kronos-mini at 4.1M params/2048 context, Kronos-small, and Kronos-base at 102.3M). Look elsewhere if you need the strongest variant (Kronos-large, ~499M, is closed-source), or if your edge comes from order-book microstructure, news, or fundamentals: Kronos only sees candlesticks, not order flow or text. Outputs are probabilistic samples controlled by temperature and nucleus sampling, not deterministic trade signals.
