Kronos

• 1. Product background (2279 chars)

  Project origin and development context:   The full name of Kronos is "Kronos: A Foundation Model for the Language of Financial Markets". It was jointly developed by seven researchers from the Interdisciplinary Information Institute (IIIS) and the Department of Automation of Tsinghua University. The core authors include Yu Shi, Fu Zongliang, Chen Shuo, Zhao Bohan, Xu Wei, Zhang Changshui and Li Jian. The project was first published on arXiv on August 2, 2025, with the paper number arXiv:2508.02739. It then went through rapid iterations: the fine-tuned script was released on August 17, 2025, and a poster was presented at the NeurIPS 2025 Workshop "Recent Advances in Time Series Foundation Models" in November 2025. It was finally officially accepted by AAAI 2026 in March 2026 and published at the AAAI Conference on Artificial Intelligence Vol.40 No.30.   Judging from the statistics of the GitHub warehouse, the project has maintained a very high level of community activity since its release: so far it has received a total of 19.4k Stars, 3.5k Forks, 179 Watchers, received 146 Issues and 25 Pull Requests, and a total of 18 contributors participated in the development. The latest developments displayed on the official GitHub page are dated November 10, 2025. At that time, the project had just been confirmed to be accepted by AAAI 2026, which immediately triggered extensive discussions in the developer community. It is worth mentioning that the project has also been synchronously mirrored to domestic code hosting platforms such as Gitee and GitCode, further expanding its influence among the domestic developer community.   Development team and organizational background:Kronos' R&D team comes from two strong disciplines of Tsinghua University - Interdisciplinary Institute of Information (IIIS, famous for computer science and quantum computing research) and Department of Automation (traditional strengths in control science and engineering). This interdisciplinary combination provides a unique technical perspective for financial time series modeling. Among the authors of the paper, Professor Zhang Changshui is a well-known scholar in the Department of Automation and has profound experience in the fields of machine learning and pattern recognition; Professor Li Jian also has a high academic reputation in related fields. As a joint R&D institution, Microsoft Research Asia provides computing power support and engineering experience for the project. This "academic institution + industrial laboratory" cooperation model is relatively rare in domestic AI open source projects, and also reflects the academic value and commercial potential of the research.

• 2. Core functions and capabilities (1925 chars)

  The core positioning of Kronos is to "understand the K-line sequence of the financial market as a language" and let the model learn the "grammar" and "vocabulary" of the financial market through large-scale pre-training. Different from the general time series basic model (TSFM), Kronos is specially optimized for the high noise, multi-dimensionality and non-stationarity of financial data. It can directly process K-line data of stocks, cryptocurrencies, foreign exchange and other assets and output prediction results.   The model mainly supports three downstream tasks. The first is price sequence prediction, which is the core function. After receiving historical K-line data, the model outputs a prediction of future price trends. It supports 7 time granularities of 1 minute, 5 minutes, 15 minutes, 30 minutes, 1 hour, 4 hours and daily. The output includes the opening price, highest price, lowest price, closing price and transaction volume prediction value at each time point. The second is volatility prediction, which is particularly important for scenarios such as risk management and option pricing. Kronos has shown better performance than the traditional GARCH model and general TSFM on the realized volatility prediction task. The third is synthetic K-line generation, which can generate realistic simulated K-line sequences based on learned market rules, which is of great value for data enhancement and strategy backtesting - the model trained by the researcher on synthetic data performs even better on real data than the baseline model trained on some real data.   In terms of prediction mechanism, Kronos supports two uncertainty quantification methods: temperature sampling (T parameter) and kernel sampling (top_p parameter), allowing users to adjust the "conservative" or "aggressive" degree of prediction according to risk preference. In addition, the model also supports batch prediction mode and implements GPU parallel computing through the predict_batch method, which can efficiently predict multiple assets or multiple time periods at the same time. An online demonstration of BTC/USDT 24-hour real-time prediction is provided in the GitHub repository, allowing users to intuitively experience the prediction effect of the model.

• 3. Technical architecture analysis (2414 chars)

  Kronos uses an innovative two-stage framework to process financial K-line data. In the first stage, a specialized K-line tokenizer (Tokenizer) converts continuous OHLCV six-dimensional data (opening price, highest price, lowest price, closing price, trading volume, trading amount) into discrete Token sequences. The core technology here is Binary Spherical Quantization (BSQ), which projects the embedding of each K-line record onto the hypersphere to generate a 20-bit binary code. In order to capture multi-level market dynamics, each K-line is quantified into two sub-Tokens: coarse-grained sub-Token (10bit) and fine-grained sub-Token (10bit). The former encodes the overall price level, and the latter encodes subtle fluctuations. In the second stage, a Decoder-only autoregressive Transformer is pre-trained on these discrete Tokens, using a causal attention mechanism (Causal Self-Attention) to ensure that predictions can only be based on historical information, and integrating temporal order information through rotational position encoding (RoPE).   The scale of pre-trained data is the biggest advantage of Kronos compared to other TSFMs: the model is trained on more than 12 billion K-line records. The data covers multiple asset classes on more than 45 exchanges around the world, including stocks, cryptocurrencies, futures and foreign exchange. The data time span covers a variety of market cycles. The pre-training task uses a standard autoregressive objective function, allowing the model to learn fine-grained temporal representations and cross-asset representations.Kronos provides four parameter scale model variants to adapt to different computing power needs and performance requirements. Kronos-mini is the most lightweight version, with only 4.1M parameters, but the context length reaches 2048 Tokens. It is suitable for rapid analysis of long-term historical data and only requires about 2GB of video memory to run on consumer-grade GPUs. Kronos-small has 24.7M parameters and a context length of 512. It is currently the most widely used version in the open source community and has achieved a good balance between performance and efficiency. Kronos-base is the standard version, containing 102.3M parameters. It also uses a context length of 512. The performance is more powerful but the computing power requirements are also increased accordingly. Kronos-large is the full version with 499.2M parameters and the strongest performance but it is not yet open source and is for research use only. From the perspective of architectural parameters, Kronos-small uses 8 layers of Transformer, 512 hidden dimensions and 8 attention heads, Kronos-base is expanded to 12 layers, 832 hidden dimensions and 16 heads, and Kronos-large is further expanded to 18 layers, 1664 hidden dimensions and 32 heads.

• 4. Fine-tuning and production deployment (1235 chars)

  Kronos provides complete fine-tuning support, allowing users to customize the training of the model using their own market data, which is considered a key step in applying the model to actual trading strategies. The fine-tuning process is implemented based on the quantitative investment platform Qlib and is divided into four steps: first, use Qlib to prepare data and convert the user's local market data into the format required by the model; second, configure the experimental parameters, including training set/validation set division, learning rate and batch size; then fine-tune the word segmenter to adapt the word segmenter to the data characteristics of a specific market; and finally fine-tune the prediction model, updating the Transformer parameters to learn the rules of specific assets. The GitHub repository provides a complete fine-tuning tutorial using the A-share market as an example. Users can refer to this tutorial to get started quickly. The project document also specifically emphasizes the precautions for production environment deployment: the original prediction signal is not equal to the directly usable Alpha factor. The risk factor needs to be neutralized through the portfolio optimization model, and real factors such as transaction costs and market impacts should be considered. The fine-tuning code itself is only for demonstration purposes and does not constitute a production-level quantitative trading system.

• 5. Pricing and business model (1203 chars)

  Kronos itself is a fully open source MIT-licensed project, and all pre-trained weights are publicly available for download on Hugging Face at no charge. Open source models include Kronos-mini (NeoQuasar/Kronos-mini), Kronos-small (NeoQuasar/Kronos-small) and Kronos-base (NeoQuasar/Kronos-base). Users can load and use them directly through the transformers library of Hugging Face. The official claims that "only 4 lines of code" can be used to complete model calling and prediction. Although Kronos-large is not open source, the research team may plan to provide services through paid API calls, which is not yet clear.   From a business model perspective, Kronos is extremely attractive to researchers and individual quantitative enthusiasts because it eliminates the threshold of massive data and computing power required to train basic financial models from scratch. For commercial institutions, closed-source value-added services can be developed based on Kronos, such as dedicated forecasting services for specific markets or asset classes, SDKs integrated into quantitative trading platforms, or technical consulting to provide model fine-tuning and deployment. Currently known community derivative projects include kronos-dashboard (WebUI visual interface), which is developed and maintained by GitHub user haoyuhu, further lowering the threshold for non-technical users to get started.

• 6. User reviews (3395 chars)

  Positive feedback:   Judging from discussions in the Chinese developer community, the most recognized value of Kronos is its "originality" and "open source integrity." Many Zhihu and CSDN authors pointed out that Kronos is "the first basic model specially built for financial K-line data", filling the gap where there was no open source financial dedicated time series model before. Prior to this, if developers wanted to use deep learning methods for financial prediction, they could only rely on general-purpose TSFM (such as TimesFM, Chronos) or train small models themselves. The former is not ideal for processing financial data, and the latter is limited by the amount of data and computing power.   The high number of Stars on GitHub (19.4k) also reflects the community’s recognition of the project. Some users stated in Issues that Kronos' two-stage framework design is "very elegant" and the idea of ​​BSQ tokenizer is "inspiring", allowing financial time series to be tokenized like text, providing new ideas for subsequent research. In addition, the BTC/USDT real-time prediction demo provided by the project is considered a "very conscientious" function, allowing users to intuitively experience the prediction effect of the model without any installation.   Judging from the experimental results of the paper, Kronos' performance in the A-share backtest has also received certain recognition. The researcher's investment simulation on Shanghai and Shenzhen market data showed that the annualized excess return (AER) and information ratio (IR) of the investment portfolio constructed using Kronos forecast signals were the highest among all tested baselines. However, it should be noted that the backtest results do not represent the actual performance. In actual transactions, factors such as liquidity, slippage, and market impact also need to be considered.   Criticism and Questioning:   Kronos also faces some technical and practical challenges. The first is the dispute about the status of "AAAI 2026 Acceptance": the GitHub page clearly marked AAAI 2026 acceptance, but the inclusion information of the same paper on the OpenReview page was shown as NeurIPS 2025 Workshop Poster. The final verification results showed that the paper was indeed published in AAAI 2026 (DOI: 10.1609/aaai.v40i30.39730, released on March 14, 2026), but its early propagation path passed through the NeurIPS Workshop. This situation of "first published and then officially accepted" at academic conferences and workshops is not uncommon in the rapidly iterative AI field, but it may cause confusion for some users.Secondly, there are technical limitations. Judging from the in-depth analysis of technical bloggers on Medium, some scholars pointed out that "even if a universal model is pre-trained with billions of data points, it may still be inferior to a manually tuned domain-specific architecture on K-line data." The non-stationary nature, high noise and unique market microstructure of financial data make it difficult to directly migrate the "one-size-fits-all" basic model paradigm. The context length of Kronos is limited to 512 Tokens, which may not be enough for scenarios such as high-frequency trading that require longer historical data. In addition, the discretized representation of the word segmenter inevitably loses some fine-grained information. The actual usage rate of coarse-grained Tokens reaches 97.66% and that of fine-grained Tokens is 85.25%, indicating that a certain proportion of codewords are not fully utilized and there may be room for optimization.   At the practical level, some users shared their A-share actual measurement experience on platforms such as Snowball, pointing out that Kronos has certain predictive capabilities for specific stocks, but the effect "varies from stock to stock" and hyperparameters such as sampling parameters need to be adjusted for different targets. Some comments mentioned that forecast signals do not equal actual profits, and that

• 7. Industry evaluation and comparison with competing products (3113 chars)

  Position in the lineage of basic time series models:   The emergence of Kronos fills the gap in financial-specific timing basic models. Judging from the development of time series basic models, between 2023 and 2024, giants such as Google (TimesFM), Amazon (Chronos), Salesforce (Moirai), and Nixtla (TimeGPT) have successively released general time series basic models. These models have demonstrated zero-sample prediction capabilities in their respective benchmark tests, but they are not specifically designed for the financial market—financial data accounts for less than 1% of the training data. The core insight of Kronos is that financial K-line data has a unique OHLCV structure, high noise characteristics and market-specific laws. The general TSFM tokenizer and pre-training target cannot effectively capture these characteristics, so a basic model specially designed for the financial market is needed.   From the perspective of architecture type, Kronos uses Decoder-only Transformer, which is consistent with better-performing timing models such as TimesFM and Lag-Llama. Moirai uses an Encoder-only architecture, which has more advantages in general timing understanding (such as anomaly detection). Chronos' Encoder-Decoder architecture is more balanced in terms of flexibility and multi-tasking capabilities. The uniqueness of Kronos lies in its two-stage framework - first converting continuous financial data into discrete Tokens through BSQ quantification, and then performing autoregressive pre-training at the Token level. This is different from all general TSFMs and is a customized design for financial data.   List of main competing products:   The competing products most directly related to Kronos can be divided into several categories. The first category is general time series basic models, which mainly include TimesFM (Google, 200M parameters, Decoder-only, 100 billion time points pre-training), Chronos (Amazon, Encoder-Decoder, multi-domain probability prediction), Moirai (Salesforce, Encoder-only, multi-variable and anomaly detection) and TimeGPT (Nixtla, Transformer, API call mode, accuracy increased by 60%). The advantages of these models lie in their versatility and cross-domain generalization capabilities, but their specific performance on financial K-line data is not as good as Kronos - the paper shows that Kronos's price prediction RankIC leads TimesFM by 93%.The second category is financial-specific models, including BloombergGPT (Bloomberg, 50 billion parameters, specially trained for financial text, does not process K-line data) and FinGPT (AI4Finance Foundation, open source financial large language model, focusing on financial sentiment analysis and text generation). Kronos complements these models: BloombergGPT and FinGPT handle financial text, and Kronos handles financial price series. The third category is quantitative models from academic and open source communities, such as various tools in the Open Quant.AI and AI4Finance Foundation ecosystems. Kronos, as a "basic model", forms an upstream and downstream relationship with these upper-layer applications.   From a practical application perspective, the performance advantage of Kronos in the subdivided scenario of financial forecasting is clear, but the general TSFM still has advantages in non-financial time series forecasting tasks. Experienced researchers may choose to maintain multiple models at the same time and switch according to the task. Kronos' MIT license and Hugging Face's convenient access make it the current tool of choice for entry and research into financial K-line prediction.

• 8. Industry background and market size (1126 chars)

  The birth of Kronos is closely related to the rapid development of the AI ​​financial market. According to multi-party market research data, the global algorithmic trading market will be approximately US$21.1 billion in 2024 and is expected to grow to US$43 billion in 2030, with a compound annual growth rate of approximately 12.9%; the scale of the AI ​​financial market is expected to grow from US$38.36 billion in 2024 to US$190.33 billion in 2030, with a compound annual growth rate of as high as 30.6%. The AI ​​algorithmic trading platform market will be approximately US$1.26 billion in 2024 and is expected to reach US$3.352 billion by 2031.   These data reflect the strong growth trend of quantitative trading and AI finance, and also provide a broad application space for open source tools like Kronos. The Linux Foundation's 2025 Open Source Financial Services Industry Report shows that 84% of respondents believe that open source is crucial to the development of financial AI, and 56% believe that basic models have the greatest impact on AI development. The emergence of Kronos fits this trend perfectly - it not only lowers the threshold for financial AI research, but also provides commercial institutions with a customizable and auditable underlying tool, which has special value in the highly regulated financial industry.

• 9. Controversies and Risks (2315 chars)

  Academic and engineering challenges:   The primary controversy facing Kronos is the fundamental issue of "forecasting does not equal profitability." Financial markets are highly random, and even technically advanced prediction models cannot guarantee stable returns in real trading. Although the A-share backtest results shown in the Kronos paper are optimistic, there is a risk of overfitting in the backtest - the fact that the model performs well on data from a specific historical period does not mean that this performance can be replicated in the future. More importantly, forecasting signals is only a part of the quantitative trading process. From signals to strategies, it also requires the cooperation of multiple modules such as fund management, risk control, and transaction execution.   The second issue is about the actual usability of "open source large models". Kronos-large (499.2M parameters) is not open source, and users can only use three smaller versions: mini/small/base. For commercial applications seeking maximum performance, this may be a limitation. Even the largest open source version, its 499M parameter Kronos-large is still lightweight compared to general-purpose large models, and there may be an upper limit to its ability to capture complex market patterns.   In addition, Kronos' pre-training data is available as of June 2024. For rapidly changing market environments (such as emergencies and policy changes), the model's generalization ability may be limited. The non-stationary nature of financial markets means that historical laws may not necessarily apply to the future. This is a fundamental challenge faced by all statistical models based on historical data.   Compliance and Ethical Considerations:   At the regulatory level, the use of AI models for trading decisions is strictly regulated in many jurisdictions. As an open source tool, Kronos can be used freely by anyone, but using prediction results directly for real trading may trigger scrutiny from regulatory agencies, especially in the context of high-frequency trading and algorithmic trading that have received a lot of attention. In addition, if commercial services built on Kronos involve misleading publicity (such as claiming that "AI prediction is guaranteed to make money without losing money"), it may face compliance risks.   From a technical ethics perspective, Kronos’ predictive capabilities may be used for market manipulation (such as creating false price signals through fake buying and selling), although this is beyond the design scope of the tool itself. The use of any financial AI tool should be carried out within a legal and compliant framework, and the author of Kronos also clearly stated this in the README.

• 10. Applicable groups and usage suggestions (2450 chars)

  Suitable scenarios:   Kronos is best suited for the following types of users. The first is quantitative finance researchers. Kronos provides a full-process baseline from original K-line data to predicted signals. On this basis, researchers can quickly verify new ideas, compare different model architectures, or use Kronos as the baseline for ablation experiments. The second is AI/ML developers who want to understand or practice "how to apply LLM ideas to time series." Kronos' two-stage framework (word segmenter + autoregressive Transformer) is a good learning case. The third is personal quantification enthusiasts who lack the data and computing power to train models from scratch, but hope to use deep learning methods to assist their investment research. Kronos-mini's low hardware requirements (2GB video memory) and simple API design make it possible for such users to get started. The fourth is financial technology companies, which need to build prototypes to verify concepts or quickly iterate product ideas. Open source Kronos can reduce investment risks in the early stages of research and development.   Suggestions and precautions for use:   For those planning to use Kronos, the following points are worth noting. First of all, do not use the model output directly as the basis for trading decisions. It should be regarded as one of many reference factors and a complete risk management mechanism should be established. Secondly, give priority to starting with Kronos-small or Kronos-mini, and then consider a larger version after becoming familiar with the model behavior. Third, making full use of the Qlib fine-tuning process provided by the project and customizing the model with data from your own target market can often achieve better results than using zero samples. Fourth, follow GitHub Issues and community discussions to learn about other users’ experience and pitfall records, which is more real-time than documentation. Fifth, be sure to read the "Production Recommendations" chapter in the official README to understand the gap and potential risks from predicted signals to actual strategies.   Overview of alternatives:   If Kronos doesn't suit your needs, here are some alternatives to consider. For general time series prediction needs, TimesFM and Chronos are relatively mature open source choices. Although they are not dedicated to finance, they are widely used in energy, transportation and other fields. For scenarios that require longer context and stronger generation capabilities, the Pro version of TimeGPT-2 (Nixtla) provides higher accuracy but requires a paid subscription. For financial text analysis needs, FinGPT and BloombergGPT are more specialized. For highly specialized quantitative trading needs, a comprehensive framework combining traditional econometric models (GARCH, ARIMA, etc.) and machine learning methods may be required.

• 11. Summarize (970 chars)

  Kronos is an important open source milestone in the field of financial AI. It proves that a basic model specially designed for financial K-line data can achieve results far beyond the general time series model. The cooperation background between Tsinghua University and Microsoft Research Asia, the academic recognition of AAAI 2026, the community popularity of 19.4k GitHub Stars, and the complete open source under the MIT license have all made it a tool that cannot be ignored in the current financial quantitative research and development. However, it must be clearly understood that the performance advantages demonstrated by Kronos are mainly in technical indicators (RankIC, MAE, etc.), and converting them into actual investment returns requires a lot of additional work. The complexity and uncertainty of financial markets mean that any prediction model is not a "silver bullet." For researchers and practitioners who really want to make a difference in the quantitative field, Kronos provides a high-quality starting point and a solid baseline, but the final value creation depends on the user's own professional ability and judgment.

No in-depth report yet

• 

  RichardThompsonJr42

  —

  跑了一下Kronos-mini，4M参数在MacBook上确实能跑，但预测准确率也就55%左右，不能说没用但离直接赚钱差得远。作为研究工具还不错，但别抱太高期望，理性看待它的能力边界。这项目值得关注但别冲动去All in。

---

• 

  Clara605

  —

  把K线当语言来训这个思路确实巧妙，BSQ分词器比直接扔数值进模型靠谱多了。但问题在于，市场不是你读懂历史就能预测未来的。Kronos解决了表示问题，但没解决预测问题，这是两码事，很多人都把这两个概念混淆了。

---

• 

  Austin.Brown273

  —

  试了三天Kronos预测A股，RankIC确实比我自己瞎搞的因子强，但回测和实盘是两码事，大家冷静。预测能力和赚钱能力中间隔着一整条量化链路：预测信号→Alpha因子→横截面选股→组合优化→风险控制→交易执行→净收益，每一步都可能衰减。

---

• 

  KMartin_771

  —

  看到有人说和MACD差不多我就笑了，你拿一个120亿条数据预训练的基础模型跟一个几十年前的指标比，本身就是对Kronos的侮辱。但它也确实不是圣杯，千万别all in。作为因子来源很优秀，但别当交易信号直接用。

---

• 

  psnjq

  —

  GitHub 30k star全是量化圈的人点的，说明大家对这个方向很期待。但坦诚讲，现阶段用它做因子筛选比做直接预测靠谱得多。我把它当因子生成器用，加入现有策略后超额收益确实有提升，这就够了，别期待更多。

---

• 

  DorothyMorganSr

  —

  搞了一晚上终于把Kronos挂在Qlib上跑通了，微调A股数据的过程比我想象中简单，finetune目录给了完整pipeline。但README里那些AI生成的注释确实有坑，建议用前仔细读一下文档，别直接照着跑。

---

• 

  Jessica_Thompson

  —

  清华这个团队确实有东西，能把Transformer搬到K线上还入选AAAI顶会。不过零样本没那么神，A股有涨跌停限制，不做微调效果打折扣。建议先用A股数据跑一遍微调流程再做预测，这样效果会好很多。...

---

• 

  烟雨_8

  —

  BTC/USDT的Live Demo我盯了一整天，走势方向判断还行但具体点位偏差挺大。当辅助参考可以全信就完了。demo版模型版本可能滞后，正经研究建议本地部署，跑起来也不难，值得一试。........

---

• 

  JEsch

  —

  Kronos最大的价值可能不是预测股价，而是提供了一个金融领域的开源基础模型。未来基于它的应用层创新才是真正值得期待的东西。这让我想起ChatGPT刚出的时候，大家惊叹方向对了。............

---

• 

  SAlvarez

  —

  我把它当因子生成器用，输出预测方向作为特征加入现有策略，超额收益确实有提升。但单独用它做策略择时夏普比率惨不忍睹。Kronos证明了第一步，但评论区想看最后一步，中间的gap还很大。..........

---

• 

  ZDiaz37

  —

  说实话有点失望，预期拉太高了。跑了一圈发现波动率预测MAE确实低9%，但对我做日内交易帮助有限，我更需要精确的进场时机。也许更适合中低频策略的策略研究吧。........................

---

• 

  Joshua.Thomas_2021

  —

  在V2EX发了贴分享把Kronos接入A股量化工作流的经验，花了一个下午在CPU上跑通了预测。实测下来，对财报前后的走势预测基本失效，基本面事件它完全读不到。......................

---

• 

  JBarnes_8883

  —

  The tokenizer design is the real innovation here. Binary Spherical Quantization for financial data makes so much sense. General TSFMs treat OHLCV like any other time series which is just wrong for financial data patterns.

---

• 

  21ku65o4o6

  —

  Ship or Skip上评分50/50，说明社区也很纠结。技术上确实牛但研究级和生产级之间的鸿沟不是一朝一夕能填平的。期待后续有更多工程化的工作把这个gap补上让更多人能用起来。...........

---

• 

  Dylan_GreenQ

  —

  跑了一下A股数据发现pre-trained权重对美股和crypto效果最好对A股需要微调。毕竟是45个交易所混合训练的，A股的制度特征比如T+1、涨跌停这些模型没有学到建议A股用户一定做微调再用。...

---

• 

  purplesnake184

  —

  I integrated Kronos-mini into my screener at 10% weight alongside nine other signals. The fallback to EMA5/EMA20 momentum when the model fails to load is a nice engineering pattern for production use.

---

• 

  Jesse_Reed_2024

  —

  AAAI接收后会有大量follow-up工作，这个方向才刚起步。现在入局学习是好时机。

---

• 

  Nathan_Mitchell_Pro70

  —

  BTC的实时Demo界面简洁，但正经研究还是得本地部署。作为快速了解的入口不错。

---

• 

  剑客_20

  —

  A股政策市属性无法被纯价格数据捕捉，但这不全是Kronos的问题。

---

• 

  TeresaHughes_Max

  —

  Kronos-large要4GB显存，建议base版就够了，large边际收益递减。

---

• 

  sATURNsWAPcHAVEZ

  —

  量化圈评价两极分化。做研究的觉得是里程碑，做实盘的觉得噱头大于实用。

---

• 

  ACox_2023

  —

  GitHub Issues有人说用Kronos预测财报前后走势基本失效。金融模型不是魔法。

---

• 

  CHOKM0F

  —

  代码质量不错，PyTorch原生实现，BSQ分词器实现值得反复研读。

---

• 

  郑磊

  —

  K线即语言这个类比太妙了。每根K线记录的是市场参与者的贪婪和恐惧。

---

• 

  SHicksQ_476

  —

  跑了一周模拟盘：作为特征工程工具5星，作为交易信号发生器2星。

---

• 

  redsnake107

  —

  不会All in。金融市场是复杂适应系统，Kronos是工具箱里的一把新工具。

---

• 

  David_WrightZ

  —

  MIT协议+AAAI顶会+120亿条数据，比那些割韭菜的项目强太多了。

---

• 

  TAmor

  —

  API风格和HuggingFace一致，上手成本低。但数据清洗花了我一上午。

---

• 

  蝴蝶651

  —

  项目确实良心，MIT协议全开源，四个模型尺寸任选，这点比那些挂羊头卖狗肉的项目强一万倍。

---

• 

  RussellFisherK

  —

  散户友好度确实高，pip install就完事了。但门槛低了用的人多了，alpha也就消失了。

No reviews yet