Investment Rating - The report does not explicitly provide an investment rating for the industry or specific companies Core Insights - The rapid evolution of large language models (LLMs) like DeepSeek-R1 has garnered significant attention from investors, particularly in their application to quantitative strategies in industry rotation, style rotation, and market timing tasks [1][2] - DeepSeek-R1 has demonstrated stable excess returns in industry rotation tasks, outperforming the equal-weighted industry return by 22.3% since 2024 [4] - The report emphasizes the unique advantages of LLMs in industry allocation tasks compared to stock selection and market timing, as industry performance is more influenced by macro policies and industry conditions [4] Summary by Sections Large Language Models and Alternative Data - LLMs are built on deep learning techniques and can process vast amounts of unstructured text data, making them suitable for extracting investment signals from alternative data sources [11][12] - The growth of alternative data is significant, with projections indicating that global data volume will reach 175ZB by 2025, highlighting the potential for LLMs to analyze non-traditional data types [13] Applications of LLMs in Investment - LLMs can assist quantitative analysts in efficiently constructing code, significantly reducing the development cycle from strategy logic to implementation [2][29] - They can also help analysts quickly find and summarize recent research, forming expert knowledge bases for specific inquiries [2][29] - The integration of LLMs with alternative data sources, such as news and research reports, enhances their effectiveness in strategy development [2][29] Performance of DeepSeek-R1 - The 671 billion parameter version of DeepSeek-R1 has shown superior performance in various tasks, particularly in industry rotation, with a stable excess return of 22.3% since 2024 [4] - The model's performance in size rotation strategies has a win rate of 54.33%, yielding an excess return of over 12% [4] - Market timing strategies have also yielded an excess return of approximately 18% since 2024, although with slightly less stability [4] Limitations of LLMs - Despite their rapid development, LLMs face limitations such as knowledge hallucination, randomness, memory constraints, and data leakage, which can impact the reliability of quantitative strategies [5] - The report highlights the need for caution regarding the accuracy of outputs generated by LLMs, particularly in high-stakes investment contexts [5][43] Future Trends - The report discusses the ongoing evolution of LLMs, emphasizing the importance of cross-modal capabilities and the integration of various data types to enhance their application in investment strategies [26][27]
中金：大模型系列（1）：DeepSeek-R1量化策略实测