Why Financial Time-Series Prediction Is Hard
Financial time-series forecasting has long resisted general-purpose models. Traditional approaches (ARIMA, LSTM) treat price as a continuous scalar, but markets exhibit three stubborn properties:
- High noise — most price movement is noise rather than signal.
- Non-stationarity — distributions drift over time; yesterday's edge vanishes tomorrow.
- Heavy tails — extreme events occur far more often than a Gaussian assumption predicts.
- NLP GPTs treat text tokens as language.
- VLMs discretize image patches into tokens.
- Speech models discretize waveforms into tokens.
- Kronos discretizes candlesticks into tokens.
- Native multimodal scaling laws — different modalities need their own tokenizers; Kronos's financial tokenizer reinforces this.
- Möbius RoPE and positional encoding — position encoding matters for temporal data; Kronos's context-length choices (2048 for mini, 512 for others) reflect explicit assumptions about temporal dependency depth.
- Editing the construction plan, not the blueprint — innovation sits in the data representation layer, not the model architecture.
- Project: https://github.com/shiyu-coder/Kronos
- Paper: https://arxiv.org/abs/2508.02739
- Live Demo: https://shiyu-coder.github.io/Kronos-demo/
- Models on HuggingFace:
NeoQuasar/Kronos-{mini,small,base} - License: open source
- Venue: AAAI 2026
General time-series foundation models (TSFMs) trained on cleaner domains (electricity load, traffic flow) tend to underperform on markets because they never learn the financial "dialect."
Kronos: A Two-Stage Framework
Kronos (GitHub Trending +217 stars/day, accepted at AAAI 2026, arXiv: 2508.02739) is the first open-source foundation model pretrained exclusively for financial markets. Its central innovation is a two-stage pipeline.
Stage 1 — Hierarchical Tokenization of OHLCV
OHLCV (open, high, low, close, volume) is continuous and multi-variate. Rather than uniform binning—which collapses a 1% move and a 5% move into the same bucket and loses multi-scale structure—Kronos applies hierarchical quantization: coarse tokens capture trends, fine tokens capture intra-bar volatility. The tokenizer encodes multiple temporal scales simultaneously, mirroring how human technicians read daily, weekly, and monthly charts side by side.
Stage 2 — Autoregressive Transformer Pretraining
On top of the hierarchical discrete tokens, a standard autoregressive Transformer is pretrained, GPT-style. The architecture is off the shelf; the innovation is the new *vocabulary*—the way continuous financial data is translated into tokens.
Data and Model Family
Pretraining data spans candlesticks from 45+ global exchanges, exposing the model to diverse market microstructures rather than a single national "dialect."
| Model | Parameters | Context Length | Open Source | |---|---|---|---| | Kronos-mini | 4.1M | 2048 | ✅ | | Kronos-small | 24.7M | 512 | ✅ | | Kronos-base | 102.3M | 512 | ✅ | | Kronos-large | 499.2M | 512 | ❌ |
The mini variant is striking: 4.1M parameters with 2k context is tiny by GPT norms, suggesting the "language" of financial series has a smaller vocabulary, simpler grammar, and shorter dependencies than natural language.
The Deeper Trend — "X as a Language"
Kronos exemplifies a widening pattern in modern AI: any signal that can be discretized into tokens becomes a candidate for Transformer pretraining.
The common thread is discretize, then pretrain. Transformer's real strength is not natural language per se—it is sequence modeling over any well-designed token space.
Cross-Paper Convergence
Kronos echoes several contemporary threads: