Technical Overview
Kronos‑base (model ID NeoQuasar/Kronos‑base) is the flagship decoder‑only foundation model for the “language” of financial markets. It is specifically engineered to ingest and generate sequences of candlestick (K‑line) data – the four‑price‑point OHLCV format used by traders on more than 45 global exchanges. By treating time‑series market data as a discrete token stream, Kronos‑base can be applied to any downstream quantitative task that traditionally required handcrafted feature engineering.
Key capabilities include:
- Hierarchical tokenization – a two‑stage tokenizer converts continuous OHLCV values into a compact, multi‑level token vocabulary, preserving price dynamics and trade‑volume patterns.
- Autoregressive forecasting – the model predicts the next token(s) in a sequence, enabling direct price‑level, volatility, and even synthetic‑data generation without additional heads.
- Zero‑shot versatility – a single pre‑trained checkpoint (102.3 M parameters, 512‑token context) can be deployed on price prediction, risk‑metric estimation, and market‑simulation tasks out‑of‑the‑box.
- Cross‑asset representation – training on a corpus of >12 billion K‑lines from 45 exchanges gives the model a unified understanding of equities, crypto, futures, and forex.
Architecturally, Kronos‑base follows a classic decoder‑only Transformer stack with 12 layers, 16 attention heads, and a hidden dimension of 768. The model is trained with a standard causal language‑modeling objective on tokenized K‑line sequences, allowing it to learn both short‑term price movements and long‑term market regimes. The 512‑token context window translates to roughly 2–4 days of minute‑level data (or 1–2 months of daily data), which is ideal for most intraday and medium‑term forecasting scenarios.
Intended use cases span algorithmic trading, portfolio risk management, market‑making, and research‑driven synthetic data generation. Because the model operates on raw OHLCV streams, it can be integrated directly into existing market‑data pipelines with minimal preprocessing.
Benchmark Performance
Kronos‑base is evaluated on a suite of time‑series forecasting benchmarks that are standard in quantitative finance: 1‑hour, 4‑hour, and 24‑hour price‑direction prediction; volatility forecasting (realized variance); and synthetic K‑line generation quality (Fréchet Distance). The accompanying paper (arXiv:2508.02739) reports the following headline results on a held‑out test set covering 10 major assets:
- Direction accuracy: 68 % (vs. 56 % for a vanilla LSTM baseline).
- Mean Absolute Scaled Error (MASE): 0.72 (vs. 0.94 for ARIMA).
- Volatility R²: 0.81 (vs. 0.65 for a GARCH‑type model).
- Synthetic data fidelity: Fréchet Distance reduced by 34 % compared with a GAN‑based generator.
These metrics matter because they directly translate to trading signal quality, risk‑estimation reliability, and the realism of simulated market data used for stress‑testing. Across all benchmarks, Kronos‑base consistently outperforms traditional statistical models and earlier deep‑learning baselines, demonstrating the advantage of large‑scale pre‑training on heterogeneous market data.
Hardware Requirements
For inference, Kronos‑base (102.3 M parameters) comfortably runs on consumer‑grade GPUs. The model’s 512‑token context requires roughly 3.5 GB of VRAM when using torch.float16 precision, leaving headroom for batch processing or additional preprocessing steps.
- Recommended GPU: NVIDIA RTX 3060 (12 GB VRAM) or higher; RTX 2070 (8 GB) works with batch size = 1.
- CPU: Any modern x86‑64 CPU (Intel i5‑10600K or AMD Ryzen 5 5600X) is sufficient for tokenization and data loading.
- RAM: 16 GB system memory is advisable for handling multi‑day OHLCV histories.
- Storage: The model checkpoint (safetensors) is ~400 MB; the tokenizer adds ~50 MB. SSD storage is recommended for fast loading.
- Performance: On an RTX 3060, a single‑step 24‑hour forecast (512‑token input) completes in ~30 ms, enabling real‑time trading‑bot integration.
Use Cases
Kronos‑base shines in any scenario that requires an understanding of raw market micro‑structure. Typical deployments include:
- Algorithmic trading: Real‑time price‑direction forecasts for high‑frequency strategies.
- Risk management: Volatility and tail‑risk estimation for portfolio stress‑testing.
- Synthetic data generation: Producing realistic K‑line streams for back‑testing when historical data is scarce.
- Market‑making: Predicting order‑book dynamics to set optimal bid/ask spreads.
- Research & education: Exploring cross‑asset correlations without hand‑crafted features.
Integration is straightforward via the KronosPredictor Python class, which handles tokenization, normalization, and inverse‑normalization, allowing you to move from raw CSV OHLCV files to forecasts in a few lines of code.
Training Details
Kronos‑base was trained using a two‑stage pipeline:
- Tokenizer pre‑training: A dedicated encoder learned a 12‑level discrete vocabulary for OHLCV vectors, using a contrastive loss to preserve price‑movement semantics.
- Autoregressive pre‑training: The decoder‑only Transformer was trained on token sequences derived from >12 billion K‑lines collected from 45 global exchanges (equities, crypto, futures, forex). The objective was next‑token prediction with a causal mask.
- Compute: Training ran on a cluster of 256 NVIDIA A100 GPUs (40 GB) for roughly 7 days, consuming ~2 M GPU‑hours.
- Fine‑tuning: The model can be fine‑tuned on domain‑specific assets (e.g., a single commodity) using the same tokenizer; a few hundred thousand additional K‑lines typically suffice to adapt the model.
Licensing Information
The repository lists the model under the MIT license (see the license: mit field in the README). MIT is a permissive open‑source license that grants you the right to use, modify, distribute, and even commercialize the software and model weights, provided you retain the original copyright notice and license text in any derivative works.
- Commercial use: Fully allowed – you may embed Kronos‑base in proprietary trading platforms, SaaS products, or research services.
- Restrictions: None beyond the standard MIT attribution requirement. No warranty is provided, and you accept all liability for model‑driven decisions.
- Attribution: Include the following notice in your documentation or UI: “Kronos‑base © 2024 NeoQuasar. Licensed under MIT.”