广发证券首席金工分析师 安宁宁 SAC: S0260512020003 anningning@gf.com.cn 广发金工安宁宁陈原文团队 摘要 最近5个交易日，科创50指数涨6.47%，创业板指涨1.96%，大盘价值跌0.34%，大盘成长涨2.48%，上证50涨1.07%，国证2000代表的小盘跌1.27%，电力设 备、有色金属表现靠前，社会服务、综合表现靠后。 广发证券资深金工分析师 张钰东 SAC: S0260522070006 zhangyudong@gf.com.cn | 日 期 | 指数代码 | 指数名称 | | --- | --- | --- | | 20250926 | 950125.CSI | 上证科创板半导体材料设备主题指数 | | 20250926 | 931865.CSI | 中证半导体产业指数 | | 20250926 | 931743.CSI | 中证半导体材料设备主题指数 | | 20250926 | 000685.SH | 上证科创板芯片指数 | | 20250926 | 000682.SH | 上证科创板新一代信息技术指数 | 一、市场涨跌 风险溢价，中证全指静态PE的倒数E ...

近年来，孟买极端降雨频率与强度显著上升，而传统全球预报系统因分辨率不足难以捕捉局地天气特征。为此，印度理工学院孟买分校与马里兰大学合 作，开发了基于卷积神经网络与迁移学习的预测模型，实现了对极端降雨事件的提前预报。 每年 6 月至 9 月，印度孟买进入季风季节。近年来，孟买极端降雨事件频发，平均降雨量较 2019 年前增加了近 40%。这座拥有 1800 万人口的沿海城市 常因暴雨陷入混乱：天气预警滞后往往导致停工停学甚至严重洪灾，季风带来的重大灾害对更精准的本地化天气预测提出了迫切需求。 然而，在热带季风气候下，标准全球天气模型约 25 平方公里的分辨率难以捕捉局部天气系统的细微差异，地形的复杂性也加剧了洪水的空间不确定性。 因此，对季风预测此前仅限于宏观趋势。 为填补城市洪水风险预判对缺口，印度理工学院孟买分校与马里兰大学研究团队合作开发了基于卷积神经网络（CNN）和迁移学习（CNN-TL）的超本地 预测模型，实现了提前数天预测大部分强降雨事件。根据 Science 最新报道，目前，孟买方面已经考虑将其纳入官方预警体系，标志着南亚城市洪水预报 进入了超本地化的新阶段。 相关研究成果以「Hyperlocal ...

Market Performance - The Sci-Tech 50 Index increased by 5.48% over the last five trading days, while the ChiNext Index rose by 2.10%. In contrast, the large-cap value index fell by 0.22%, and the large-cap growth index increased by 2.16% [1] - The performance of sectors showed that electronics and real estate were leading, while comprehensive and banking sectors lagged behind [1] Risk Premium Analysis - The risk premium, measured as the inverse of the static PE of the CSI All Share Index minus the yield of 10-year government bonds, has reached historical extremes. As of October 28, 2022, it was at 4.08%, indicating a market rebound. The latest reading on January 19, 2024, was 4.11%, marking the fifth time since 2016 it exceeded 4% [1] - As of September 12, 2025, the risk premium indicator was at 2.87%, with the two-standard deviation boundary set at 4.76% [1] Valuation Levels - As of September 12, 2025, the CSI All Share Index's TTM PE was at the 78th percentile, while the SSE 50 and CSI 300 were at 72% and 70%, respectively. The ChiNext Index was close to the 48th percentile, indicating a relative median valuation level historically [2] Long-term Market Trends - The Shenzhen 100 Index has historically experienced bear markets every three years, followed by bull markets. The current adjustment, which began in Q1 2021, has shown sufficient time and space for a potential upward cycle [2] Investment Themes - The latest investment themes identified include automotive, communication, artificial intelligence, and chemicals. Specific indices highlighted are the CSI 800 Automotive and Parts Index, CSI All Share Communication Equipment Index, CSI Artificial Intelligence Theme Index, and CSI Sub-segment Chemical Industry Theme Index [2][3] Fund Flow and Trading Activity - Over the last five trading days, ETF inflows totaled 11.6 billion yuan, while margin financing increased by approximately 59.1 billion yuan. The average daily trading volume across both markets was 22,948 billion yuan [2] Market Sentiment - The proportion of stocks above the 200-day moving average indicates market sentiment, with a focus on the long-term trend [12] Financing Balance - The financing balance reflects the overall market leverage and investor sentiment towards equity investments [15]

Core Viewpoint - The article highlights that the concept of Scaling Law was proposed 32 years ago by Bell Labs, not by recent AI advancements, emphasizing the historical significance of this research in machine learning [1][6]. Group 1: Historical Context - The paper titled "Learning Curves: Asymptotic Values and Rate of Convergence" introduced a predictive method for training errors and testing errors converging to the same asymptotic error value as training size increases, following a power-law form [4][6]. - The authors of the 1993 paper included notable figures such as Vladimir Vapnik and Corinna Cortes, who contributed significantly to the field of machine learning [6][25]. Group 2: Methodology and Findings - The research aimed to save computational resources when training classifiers by predicting their performance on larger datasets based on smaller training sets [8][10]. - The study found that as the training set size increases, both training and testing errors converge to a common asymptotic value, denoted as 'a', which typically falls between 0.5 and 1 [10][16]. - The proposed method allows for the estimation of classifier performance on larger datasets without complete training, thus conserving computational resources [10][14]. Group 3: Implications and Applications - The findings indicated that the predictive model was highly accurate for linear classifiers, demonstrating its potential to optimize resource allocation in training models [15][24]. - The research also revealed that the more difficult the task, the higher the asymptotic error and the slower the convergence rate, indicating a relationship between task complexity and learning efficiency [22].

Market Performance - The Sci-Tech 50 Index increased by 7.49% and the ChiNext Index rose by 7.74% over the last five trading days, while the large-cap value index fell by 1.37% [1] - The large-cap growth index gained 5.83%, and the Shanghai 50 Index increased by 1.63%, with the small-cap index represented by the CSI 2000 rising by 0.33% [1] - Communication and non-ferrous metals sectors performed well, while textiles, apparel, and coal sectors lagged [1] Risk Premium Analysis - The static PE of the CSI All Index minus the yield of 10-year government bonds indicates a risk premium, which reached 4.17% on April 26, 2022, and 4.08% on October 28, 2022, leading to a market rebound [1] - As of January 19, 2024, the risk premium indicator was at 4.11%, marking the fifth occurrence since 2016 of exceeding 4% [1] - The indicator as of August 29, 2025, was at 2.92%, with the two-standard deviation boundary set at 4.77% [1] Valuation Levels - As of August 29, 2025, the CSI All Index's P/E TTM percentile was at 78%, while the Shanghai 50 and CSI 300 were at 72% and 70%, respectively [2] - The ChiNext Index was close to 46%, indicating a relatively low valuation level compared to historical averages [2] Technical Analysis - The Deep 100 Index has experienced bear markets every three years, with declines ranging from 40% to 45% [2] - The current adjustment cycle began in Q1 2021, suggesting a potential upward cycle from the bottom [2] Fund Flow and Trading Activity - In the last five trading days, ETF inflows totaled 28.6 billion yuan, and margin financing increased by approximately 96.6 billion yuan [3] - The average daily trading volume across both markets was 29.51 billion yuan [3] AI and Data Analysis - A convolutional neural network (CNN) was utilized to model price and volume data, mapping learned features to industry themes [9] - The latest investment themes include artificial intelligence and related sectors [2]

【免责声明】本文仅代表作者本人观点，与和讯网无关。和讯网站对文中陈述、观点判断保持中立，不对所包含内容 的准确性、可靠性或完整性提供任何明示或暗示的保证。请读者仅作参考，并请自行承担全部责任。邮箱： news_center@staff.hexun.com 隆汇8月28日｜据晚点Auto，理想汽车自研智驾芯片 M100 于今年一季度样片回片，迈过量产前的关键 阶段。随后，M100 在两周内完成功能测试和性能测试，后续通过理想研发人员的压力测试。目前， M100 已经小批量上样车做道路测试。据我们了解，在处理不同类型的计算任务时，M100 表现出特定 的性能特点：如在运行大语言模型(LLM, Large Language Model)的计算任务时，1 颗 M100 所能提供的 有效算力与 2 颗英伟达 Thor-U 大致相当；而在处理卷积神经网络(CNN, Convolutional Neural Network) 相关的传统视觉任务(如图像识别)时，1 颗 M100 所能提供的有效算力可对标 3 颗英伟达 Thor-U。 ...

Core Viewpoint - The article discusses the rising popularity of the AI game "Draw A Fish," which allows users to draw a fish and see it swim in a virtual aquarium, attracting millions of players globally due to its simplicity and interactive features [3][14]. Group 1: Game Mechanics - The game requires players to draw a fish on a canvas, with an AI providing real-time feedback on how fish-like the drawing is, based on a similarity threshold of 60% [5][14]. - Players can name their fish and place it in a shared virtual aquarium, where it swims alongside creations from other users [5][14]. - A leaderboard showcases the highest scores, with the top score reaching 53,245 points for an abstract fish drawing [7][14]. Group 2: User Engagement - The game is designed to be low-barrier, requiring no login or tutorial, which encourages immediate participation [14]. - The AI's feedback mechanism creates a sense of achievement as users improve their drawings, enhancing the overall gaming experience [14]. - The shared aquarium fosters community interaction, allowing users to like or dislike each other's fish, thus creating a social atmosphere [15]. Group 3: Technical Aspects - The game utilizes a convolutional neural network based on the ResNet18 architecture, trained with the Google QuickDraw dataset to classify drawings as "fish" or "not fish" [16][18]. - The model's design includes a lenient recognition approach to enhance user enjoyment and engagement [16]. - Various engineering features, such as early stopping and consistent preprocessing, are implemented to optimize the model's performance [18].

Market Performance - The Sci-Tech 50 Index increased by 13.31% over the last five trading days, while the ChiNext Index rose by 5.85%. The large-cap value index grew by 1.56%, and the large-cap growth index increased by 4.77%. The Shanghai 50 Index and the CSI 2000 Index, representing small caps, saw gains of 3.38% and 3.47%, respectively. The telecommunications and electronics sectors performed well, while real estate and coal sectors lagged behind [1]. Risk Premium Analysis - The static PE of the CSI All Index minus the yield of ten-year government bonds indicates a risk premium. Historical extreme bottoms have shown this data to be at two standard deviations above the mean, with notable instances in 2012, 2018, and 2020. As of January 19, 2024, the indicator reached 4.11%, marking the fifth occurrence since 2016 to exceed 4%. As of August 22, 2025, the indicator stands at 3.03%, with the two standard deviation boundary at 4.77% [1]. Valuation Levels - As of August 22, 2025, the CSI All Index's PE TTM percentile is at 76%. The Shanghai 50 and CSI 300 indices are at 72% and 68%, respectively, while the ChiNext Index is close to 39%. The CSI 500 and CSI 1000 indices are at 58% and 57%, indicating that the ChiNext Index's valuation is relatively low compared to historical averages [2]. Long-term Market Trends - The Shenzhen 100 Index has experienced bear markets approximately every three years, followed by bull markets. The current adjustment cycle began in Q1 2021, suggesting a potential upward cycle from the bottom based on historical patterns [2]. Fund Flow and Trading Activity - In the last five trading days, ETF inflows totaled 24.7 billion yuan, and the margin financing increased by approximately 90.1 billion yuan. The average daily trading volume across both markets was 25.463 billion yuan [3]. AI and Neural Network Analysis - A convolutional neural network (CNN) has been utilized to model price and volume data, mapping learned features to industry themes. The latest focus is on sectors such as telecommunications [8].

Core Viewpoint - The recent market performance shows a significant increase in the ChiNext and Sci-Tech 50 indices, while large-cap value stocks have declined, indicating a shift in investor sentiment towards growth sectors [1][2]. Market Performance - In the last five trading days, the Sci-Tech 50 index rose by 5.53%, the ChiNext index increased by 8.48%, while the large-cap value index fell by 0.76%. The large-cap growth index rose by 3.63%, and the Shanghai 50 index increased by 1.57%. Small-cap stocks represented by the CSI 2000 index rose by 3.86% [1]. - The communication and electronics sectors performed well, while the banking and steel sectors lagged behind [1]. Risk Premium Analysis - The risk premium, measured as the difference between the inverse of the static PE of the CSI All Share Index and the yield of ten-year government bonds, has reached historical extremes. As of October 28, 2022, the risk premium was at 4.08%, indicating a potential market rebound [1]. - The risk premium has exceeded 4% for the fifth time since 2016, with the latest reading on January 19, 2024, at 4.11% [1]. Valuation Levels - As of August 15, 2025, the CSI All Share Index's TTM PE is at the 72nd percentile, with the Shanghai 50 and CSI 300 at 69% and 63%, respectively. The ChiNext index is at a relatively low valuation level of approximately 33% [2]. - The long-term view of the Deep 100 index suggests a cyclical pattern of bear and bull markets every three years, with the current adjustment phase starting in Q1 2021 showing sufficient time and space for a potential upward cycle [2]. Fund Flow and Trading Activity - In the last five trading days, there was an outflow of 10.4 billion yuan from ETFs, while margin financing increased by approximately 41.8 billion yuan. The average daily trading volume across both markets was 20,767 billion yuan [3]. AI and Trend Observation - The use of convolutional neural networks (CNN) for modeling price and volume data has been explored, with the latest focus on mapping learned features to industry themes, particularly in the communication sector [8].

Market Performance - The recent 5 trading days saw the Sci-Tech 50 Index increase by 0.65%, the ChiNext Index by 0.49%, the large-cap value by 1.63%, the large-cap growth by 1.17%, the SSE 50 by 1.27%, and the small-cap represented by the CSI 2000 by 2.74% [1] - The sectors of defense, military, and non-ferrous metals performed well, while pharmaceuticals, biotechnology, and computers lagged behind [1] Risk Premium Analysis - The static PE of the CSI All Index minus the yield of 10-year government bonds indicates a risk premium, which reached 4.17% on April 26, 2022, and 4.08% on October 28, 2022, showing a market rebound [1] - As of January 19, 2024, the risk premium indicator was at 4.11%, marking the fifth time since 2016 it exceeded 4% [1] - The indicator as of August 8, 2025, was at 3.39%, with the two-standard deviation boundary at 4.77% [1] Valuation Levels - As of August 8, 2025, the CSI All Index's PE TTM percentile was at 68%, with the SSE 50 and CSI 300 at 69% and 61% respectively, while the ChiNext Index was close to 25% [2] - The long-term view of the Deep 100 Index shows a technical pattern of bear markets every three years followed by bull markets, with the current adjustment starting in Q1 2021 being substantial [2] Fund Flow and Trading Activity - In the last 5 trading days, ETF inflows amounted to 18.5 billion yuan, and the margin trading increased by approximately 27.8 billion yuan, with an average daily trading volume of 1.6748 trillion yuan [3] Neural Network Trend Observation - A convolutional neural network was utilized to model price and volume data, mapping learned features to industry themes, with a focus on semiconductor materials among the latest configurations [9]