中国科学院大学教授张玉清：大模型开启智能金融新纪元

Core Viewpoint - The financial large models are transitioning towards specialization, lightweight design, and compliance, marking the beginning of a new era in intelligent finance rather than being the endpoint of quantitative trading [1][8]. Group 1: Current State of Quantitative Trading - Quantitative funds have shown relatively strong performance in both returns and risk control compared to fundamental funds, with quantitative trading accounting for over 60% of the U.S. stock market and approximately 20%-30% in the A-share market as of 2023 [4]. - The number of quantitative funds in the A-share market doubled from 2019 to 2022, making up 18% of actively managed public funds [4]. - Despite their strengths, quantitative trading faces challenges such as strategy homogeneity, poor adaptability, narrow information processing, and high R&D costs [4][6]. Group 2: Challenges in Quantitative Trading - A significant issue is the homogeneity of trading strategies, as evidenced by over 70% of quantitative long products underperforming the benchmark index during extreme market conditions in August [4]. - The adaptability of quantitative strategies is limited, particularly in market structures where only a few stocks surge while many others remain stagnant [4]. - Traditional quantitative strategies often rely on outdated financial data and indicators, leading to a lack of unique Alpha returns [4]. - The increasing number of selectable factors complicates strategy development and raises trial-and-error costs [4]. Group 3: Role of Large Models in Quantitative Trading - Large models are set to redefine quantitative trading by shifting from experience-driven to intelligence-driven paradigms, enhancing the ability to process vast amounts of unstructured data and perform logical reasoning [6][8]. - These models can automate information extraction, generate trading signals, and optimize decision-making processes, thereby improving the depth, breadth, and adaptability of trading strategies [6][7]. - The integration of multi-agent systems and multi-source information will empower the entire quantitative trading process, from data collection to risk control [6][7]. Group 4: Practical Applications and Performance - Real-world applications of large models have demonstrated their value, with Chinese models outperforming U.S. models in a recent trading competition, achieving an average of 3.4 trades per day and a single trade profit of $181.53 [8]. - The successful strategies of these models include selective trading, maximizing profits, quick loss-cutting, and patient holding of profitable positions [8]. - However, caution is advised regarding the "hallucination problem" in financial large models, which can lead to significant shifts in market sentiment and trading strategies based on minor adjustments in input [8].