倍漾量化冯霁：大模型重构量化投研整条生产线

Core Insights - The fourth China Quantitative Investment White Paper Seminar was held, featuring a keynote speech by Feng Ji, founder of Beiyang Quantitative, on "Quantitative Investment in the Era of Large Models" [1] Group 1: Machine Learning in Finance - Beiyang Quantitative emphasizes high turnover and has adopted an "AI-native" approach to asset management from its inception, akin to building a tech company [3] - The core of machine learning is generalization, which allows models trained on historical data to perform well on unseen data, as formalized by Valiant's PAC learning framework [3] - The financial market is not efficient, meaning there is exploitable information beyond current prices, and high-frequency data is particularly suitable for machine learning due to its slower drift [3] Group 2: AI and Quantitative Research - The arrival of large models has rewritten the rules of the game, with a streamlined process for natural language processing (NLP) now consisting of pre-training, supervised fine-tuning, and reinforcement learning [4] - Beiyang has divided its team into two groups: a machine learning group focused on accuracy and a high-performance computing group focused on speed, eliminating traditional factor roles [4] - Shorter trading cycles are more susceptible to AI due to their inefficiencies and stable distributions, while longer cycles present exponentially greater challenges [4] Group 3: Future of AI in Investment - AI-driven research systems have the advantage of planned upgrades, contrasting with traditional research that relies on inspiration; Beiyang has a three-month development schedule for internal capabilities [4]