Paper: arXiv 2508.02739
Authors: Yu Shi, Zongliang Fu, Shuo Chen, Bohan Zhao, Wei Xu, Changshui Zhang, Jian Li
Abstract
The success of large-scale pre-training paradigm, exemplified by Large Language Models (LLMs), has inspired the development of Time Series Foundation Models (TSFMs). However, their application to financial candlestick (K-line) data remains limited, often underperforming non-pre-trained architectures. Moreover, existing TSFMs often overlook crucial downstream tasks such as volatility prediction and synthetic data generation. To address these limitations, we propose Kronos, a unified, scalable pre-training framework tailored to financial K-line modeling. Kronos introduces a specialized tokenizer that discretizes continuous market information into token sequences, preserving both price dynamics and trade activity patterns. We pre-train Kronos using an autoregressive objective on a massive, multi-market corpus of over 12 billion K-line records from 45 global exchanges, enabling it to learn nuanced temporal and cross-asset representations. Kronos excels in a zero-shot setting across a diverse set of financial tasks. On benchmark datasets, Kronos boosts price series forecasting RankIC by 93% over the leading TSFM and 87% over the best non-pre-trained baseline. It also achieves a 9% lower MAE in volatility forecasting and a 22% improvement in generative fidelity for synthetic K-line sequences. These results establish Kronos as a robust, versatile foundation model for end-to-end financial time series analysis. Our pre-trained model is publicly available at https://github.com/shiyu-coder/Kronos.
Complexity vs Empirical Score
- Math Complexity: 8.5/10
- Empirical Rigor: 9.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced neural network architectures with specialized quantization and autoregressive modeling (high math), while backing it up with massive-scale pre-training on 12 billion records across 45 exchanges and extensive benchmarking across multiple downstream tasks including forecasting, volatility, and generation (high empirical rigor).
Research Flowchart
flowchart TD A["Research Goal:<br>Develop a robust Time Series Foundation Model<br>for financial K-line data?"] --> B["Methodology:<br>Specialized Tokenizer + Autoregressive Pre-training"] B --> C["Input Data:<br>12B K-line records from 45 global exchanges"] C --> D["Process:<br>Pre-train Kronos Model"] D --> E["Key Findings (Zero-Shot):"] E --> F["+93% RankIC<br>Price Forecasting"] E --> G["-9% MAE<br>Volatility Prediction"] E --> H["+22% Fidelity<br>Synthetic Data Generation"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes Kronos, a unified, scalable pre‑training framework for financial K‑line data that includes a specialized tokenizer and autoregressive Transformer.
Source sentence
we propose Kronos, a unified, scalable pre-training framework tailored to financial K-line modeling. Kronos introduces a specialized tokenizer that discretizes continuous market information into token sequences, preserving both price dynamics and trade activity patterns.
Assumes 5
- The data cleaning pipeline can reliably filter out low‑quality K‑line segments such as abnormal price spikes or prolonged inactivity. data
Source sentence
To guarantee data quality, we develop a streamlined data-cleaning pipeline tailored to the unique characteristics of financial K-line data, which identifies and filters out low-quality segments such as those with abnormal price spikes or prolonged periods of inactivity.
- Trading simulations assume a top‑k/drop‑n portfolio construction with a minimum holding period of 5 days and a limit on daily turnover. execution
Source sentence
... a minimum holding period of 5 days is enforced for all positions.
- The model can learn hierarchical market dynamics via a coarse‑to‑fine token prediction scheme. model
Source sentence
This coarse-to-fine prediction scheme allows Kronos to explicitly model multi-scale market dynamics.
- Cross‑attention uses the embedding of ˆbc_t as query and the history h_t as key and value. model
Source sentence
We use a cross-attention mechanism where the embedding of ˆbc_t acts as the query, and the history h_t provides the key and value.
Show 1 more
- The pre‑training corpus consists exclusively of financial K‑line data, assuming this domain‑specific data is sufficient for learning useful representations. data
Source sentence
To address this fundamental gap in pre‑training, we introduce Kronos, a foundation model built from the ground up on a massive corpus composed exclusively of financial K-line data.
Data
Chinese A-shares · equities · daily
- K‑line dataset market-prices unknown
Source sentence
Our dataset spans over 12 billion observations across 7 sampling frequencies, encompassing a broad spectrum of asset classes drawn from 45 global exchanges.
- Stock data market-prices unknown
Source sentence
To test both in‑distribution and out‑of‑distribution generalization, we use data from nine global stock exchanges.
- Binance cryptocurrency data market-prices public
Source sentence
Cryptocurrency: All spot trading pairs available on the Binance exchange.
- Forex data market-prices unknown
Source sentence
Forex: A comprehensive dataset of over 1,000 foreign exchange pairs.
Qlib datasets
Costs
Unclear - sources disagree
Source sentence (the authors' words)
To manage turnover and trading costs, a maximum of n stocks are bought or sold daily, and a minimum holding period of 5 days is enforced for all positions.
Finds 4
- Kronos improves price series forecasting RankIC by 93% over the leading TSFM and 87% over the best non‑pre‑trained baseline. performance
Source sentence
Kronos excels in a zero-shot setting across a diverse set of financial tasks. On benchmark datasets, Kronos boosts price series forecasting RankIC by 93% over the leading TSFM and 87% over the best non-pre-trained baseline.
- Kronos achieves a 9% lower MAE in volatility forecasting compared to baselines. predictive-accuracy
Source sentence
It also achieves a 9% lower MAE in volatility forecasting.
- Kronos improves generative fidelity for synthetic K‑line sequences by 22%. performance
Source sentence
and a 22% improvement in generative fidelity for synthetic K-line sequences.
- In backtesting simulations, Kronos demonstrates superior investment performance across multiple asset classes. performance
Source sentence
Subfigure (e) presents the investment simulation backtesting results.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 67 related papers). Coverage is partial, so this is neither replication nor refutation.
6 papers in the archive : AllocFlow: End-to-End Portfolio Trading Action Generation …
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (61)
- Language models are few-shot learners (2020)
- Gpt-4 technical report (2023) — link
- Learning transferable visual models from natural language supervision (2021)
- Segment anything (2023)
- TimeGPT-1 (2023) — link
- Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts (2025)
- Japanese candlestick charting techniques: a contemporary guide to the ancient investment techniques of the Far East (2001)
- itransformer: Inverted transformers are effective for time series forecasting (2023) — link
and 53 more
- Units: Building a unified time series model (2024)
- Major Issues in High-Frequency Financial Data Analysis: A Survey of Solutions (2025)
- Addressing the Non-Stationarity and Complexity of Time Series Data for Long-Term Forecasts (2024)
- Chronos: Learning the language of time series (2024) — link
- Totem: Tokenized time series embeddings for general time series analysis (2024) — link
- Neural discrete representation learning (2017)
- Language Model Beats Diffusion–Tokenizer is Key to Visual Generation (2023) — link
- Image and video tokenization with binary spherical quantization (2024) — link
- Scalable Image Tokenization with Index Backpropagation Quantization (2025) — link
- Lag-llama: Towards foundation models for time series forecasting (2023)
- Timer: Generative pre-trained transformers are large time series models (2024) — link
- Moment: A family of open time-series foundation models (2024) — link
- A decoder-only foundation model for time-series forecasting (2024)
- Unified Training of Universal Time Series Forecasting Transformers (2024)
- Bridging continuous and discrete tokens for autoregressive visual generation (2025) — link
- Scaling laws for neural language models (2020) — link
- The curious case of neural text degeneration (2019) — link
- Deep time series models: A comprehensive survey and benchmark (2024) — link
- Diffusion-ts: Interpretable diffusion for general time series generation (2024) — link
- Timevae: A variational auto-encoder for multivariate time series generation (2021) — link
- Time-series generative adversarial networks (2019)
- Timemixer: Decomposable multiscale mixing for time series forecasting (2024) — link
- Timesnet: Temporal 2d-variation modeling for general time series analysis (2022) — link
- Are transformers effective for time series forecasting? (2023)
- Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting (2022)
- Timexer: Empowering transformers for time series forecasting with exogenous variables (2024) — link
- A time series is worth 64 words: Long-term forecasting with transformers (2022) — link
- Non-stationary transformers: Exploring the stationarity in time series forecasting (2022)
- Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation (1982)
- Generalized autoregressive conditional heteroskedasticity (1986)
- Towards Neural Scaling Laws for Time Series Foundation Models (2024) — link
- On layer normalization in the transformer architecture (2020)
- Roformer: Enhanced transformer with rotary position embedding (2024)
- Decoupled weight decay regularization (2017) — link
- Intra-day trading volume patterns of equity markets: A study of US and European stock markets (2008)
- The week-of-the-month effect in stock returns: The evidence from the S&P composite index (1992)
- Sundial: A Family of Highly Capable Time Series Foundation Models (2025) — link
- Root mean square layer normalization (2019)
- PLUTUS: A Well Pre-trained Large Unified Transformer can Unveil Financial Time Series Regularities (2024) — link
- DELPHYNE: A Pre-Trained Model for General and Financial Time Series — → 2506.06288
- MarketGPT: Developing a Pre-trained transformer (GPT) for Modeling Financial Time Series — → 2411.16585
- MarS: a Financial Market Simulation Engine Powered by Generative Foundation Model — → 2409.07486
- Qlib: An ai-oriented quantitative investment platform (2020) — link
- Investor sentiment and the cross-section of stock returns (2006)
- In search of attention (2011)
- Macroeconomic factors do influence aggregate stock returns (2002)
- Trading volume and price reactions to public announcements (1991)
- Financial econometric analysis at ultra-high frequency: Data handling concerns (2006)
- The effectiveness of discretization in forecasting: An empirical study on neural time series models (2020) — link
- Addressing representation collapse in vector quantized models with one linear layer (2024) — link
- The variation of certain speculative prices (1963)
- Cross-correlations between volume change and price change (2009)
- The global economic effects of Trump’s 2025 tariffs (2025)
3 of 61 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.