证券研究报告 必选消费 看好 AKK 菌产业链投资机会 华泰研究 2026 年 1 月 24 日│中国内地 动态点评 AKK 益生菌——嗜粘蛋白阿克曼氏菌（Akkermansia muciniphila，简称 AKK）是一种独特的肠道共生菌，据上海交通大学医学院附属瑞金医院公众 号数据，AKK 益生菌占肠道菌群的 1%-5%；另据中国食品学报，AKK 益生 菌在改善肥胖、调节血糖血脂及代谢、抗炎等方面有积极作用，并具有潜在 的抗衰研究价值。 中国益生菌市场潜力较大，AKK 菌商业化落地加速 功效明显与技术进步共促中国市场 AKK 菌的成长势头可期 据仙乐健康公众号数据，2024 年全球 AKK 菌市场规模约 23.5 亿元人民币， 预计将在 2031 年快速增长至 44.6 亿元，对应期间 CAGR 为 9.6%，而中国 市场的势头或将更为强劲，预计在 2031 年市场规模有望突破 10 亿元，市 场渗透率约 20%；在中国，AKK 菌相关产品线上销售额从 25 年 1 月的 600 万元冲至 6 月的 2300 万元（来源魔镜洞察数据）。当前 AKK 处于快速发 展阶段，主要的增长驱动因素为较强的功效性与供 ...

证券研究报告 英特尔(INTEL) (INTC US) 制程追赶初见成效,看好 18A 订单落地 华泰研究 2026 年 1 月 23 日 | 美国 年报点评 半导体 英特尔 25Q4 业绩超预期,但 26Q1 指引液软且 18A 无客户更新,盘后股 价下跌超 12%(股价自 12/17 累计上涨 51%,也存在获利了结）。其中 1Q 指引审慎(或受季节性需求下供应约束影响),股价先下跌 6%;业绩会未 对 18A/14A 进展做更多更新,市场随之调整预期。我们认为,短期财报并 非核心，关键在于 Foundry 转型推进节奏,更应关注代工订单进展及制程 领先是否持续兑现。业绩端:Q4 营收 137 亿美元,同比-4.1%,超 VA 预 期（下同）2.1%（全年营收 529 亿美元,同比持平）;Non-GAAP 毛利率 和 EPS 为 37.9%和 0.15 美元，高于预期的 36.3%和 0.08 美元，其中毛利 率高指引 1.4pp。指引端: Q1 营收 117-127 亿美元（预期 125 亿美元）, Non-GAAP 毛利率/EPS 为 34.5%/0.00 美元（预期 36.1%和 0.04 美元）, ...

Investment Rating - The report maintains an "Overweight" rating for the company with a target price of RMB 62.00 [5][4]. Core Insights - The company is expected to achieve a net profit attributable to shareholders of RMB 7.8-9.2 billion for 2025, representing a year-on-year increase of 1093%-1307%. The non-recurring profit is projected to be RMB 6.8-8.2 billion, an increase of 356%-450%, aligning with expectations due to industry demand recovery and market share expansion [1][2]. - The penetration rate of TRB bearings is anticipated to increase significantly, with projections of 50% in 2025, 80% in 2026, and 90% in 2027 for onshore wind turbines. The company is leveraging its technological advantages and increasing production capacity to support this growth [2][3]. - The company is actively expanding into the gearbox bearing market, which has a low domestic production rate. It has completed prototype validations and small-scale supplies to several gearbox manufacturers, with expectations for large-scale supply in 2026 [3]. Summary by Sections Performance Forecast - The company expects to see a continuous increase in TRB main bearing penetration and market share, leading to an upward revision of revenue forecasts for 2026-2027. The projected net profits for 2025, 2026, and 2027 are RMB 8.77 billion, RMB 12.82 billion, and RMB 14.99 billion, respectively, with corresponding EPS of RMB 2.12, RMB 3.10, and RMB 3.62 [4][8]. Valuation - The report assigns a P/E ratio of 20x for 2026, resulting in a target price of RMB 62, reflecting the company's significant first-mover advantage in TRB bearings and leading position in gearbox bearing domestic substitution [4][5].

Quantitative Models and Construction Methods Model: AI CSI 1000 Enhanced Portfolio - **Construction Idea**: The model is based on the full-spectrum fusion factor, which integrates both high-frequency and low-frequency price-volume data using deep learning and multi-task learning techniques[6][7] - **Construction Process**: 1. Train 27 high-frequency factors using a deep learning model to obtain high-frequency deep learning factors 2. Use multi-task learning to extract end-to-end features from low-frequency price-volume data, resulting in low-frequency multi-task factors 3. Combine the high-frequency and low-frequency factors to form the full-spectrum fusion factor[6] - **Evaluation**: The model shows significant excess returns and a high information ratio, indicating strong performance and effective risk management[1][7] - **Backtest Results**: - Annualized excess return: 21.60% - Annualized tracking error: 6.06% - Information ratio (IR): 3.57 - Maximum drawdown of excess return: 7.55% - Calmar ratio of excess return: 2.86[1][7] Model: LLM-FADT Text Stock Selection Strategy - **Construction Idea**: The model enhances the BERT-FADT strategy by incorporating additional interpretations from a large language model (LLM), including new title interpretations, market catalysts, implied meanings, potential risks, and return guidance[2][14][17] - **Construction Process**: 1. Input six types of text into a fine-tuned FinBERT model: original text, new title interpretations, market catalysts, implied meanings, potential risks, and return guidance 2. Convert these texts into text feature vectors 3. Train an XGBoost model using these enriched text features[17] - **Evaluation**: The LLM-FADT strategy is more stable and has smaller excess drawdowns compared to the BERT-FADT strategy, showing better performance in extreme market conditions[2][14][20] - **Backtest Results**: - Annualized return: 30.10% - Annualized excess return: 25.52% - Sharpe ratio: 1.18 - Information ratio (IR): 2.00[2][20][24] Model: AI Industry Rotation Model - **Construction Idea**: The model uses the full-spectrum fusion factor to score 32 primary industries and constructs a weekly rebalancing strategy by equally weighting the top 5 industries[3][38] - **Construction Process**: 1. Score each industry using the full-spectrum fusion factor based on the industry component stocks 2. Select the top 5 industries with the highest scores 3. Equally weight these industries and rebalance weekly[38][43] - **Evaluation**: The model effectively utilizes AI's feature extraction capabilities to capture patterns in multi-frequency price-volume data, complementing top-down strategies[3][38] - **Backtest Results**: - Annualized return: 26.87% - Annualized excess return: 19.02% - Maximum drawdown of excess return: 12.43% - Sharpe ratio of excess return: 1.85[3][41] Model: AI Thematic Index Rotation Model - **Construction Idea**: The model scores 133 thematic indices using the full-spectrum fusion factor and constructs a weekly rebalancing strategy by equally weighting the top 10 thematic indices[4][28] - **Construction Process**: 1. Score each thematic index using the full-spectrum fusion factor based on the index component stocks 2. Select the top 10 thematic indices with the highest scores 3. Equally weight these indices and rebalance weekly[28][31] - **Evaluation**: The model leverages AI to identify and capitalize on trends in thematic indices, providing a diversified and dynamic investment approach[4][28] - **Backtest Results**: - Annualized return: 16.92% - Annualized excess return: 9.37% - Maximum drawdown of excess return: 20.79% - Sharpe ratio of excess return: 0.73[4][30] Model Backtest Performance AI CSI 1000 Enhanced Portfolio - Annualized excess return: 21.60% - Annualized tracking error: 6.06% - Information ratio (IR): 3.57 - Maximum drawdown of excess return: 7.55% - Calmar ratio of excess return: 2.86[1][7] LLM-FADT Text Stock Selection Strategy - Annualized return: 30.10% - Annualized excess return: 25.52% - Sharpe ratio: 1.18 - Information ratio (IR): 2.00[2][20][24] AI Industry Rotation Model - Annualized return: 26.87% - Annualized excess return: 19.02% - Maximum drawdown of excess return: 12.43% - Sharpe ratio of excess return: 1.85[3][41] AI Thematic Index Rotation Model - Annualized return: 16.92% - Annualized excess return: 9.37% - Maximum drawdown of excess return: 20.79% - Sharpe ratio of excess return: 0.73[4][30]

Quantitative Models and Construction Methods 1. Model Name: A-Share Technical Scoring Model - **Model Construction Idea**: The model aims to fully explore technical information to depict market conditions, breaking down the abstract concept of "market state" into five dimensions: price, volume, volatility, trend, and crowding. It generates a comprehensive score ranging from -1 to +1 based on equal-weighted voting of signals from 10 selected indicators across these dimensions[9][14] - **Model Construction Process**: 1. Select 10 effective market observation indicators across the five dimensions[14] 2. Generate long/short timing signals for each indicator individually 3. Aggregate the signals through equal-weighted voting to form a comprehensive score between -1 and +1[9] - **Model Evaluation**: The model provides a straightforward and timely way for investors to observe and understand the market[9] 2. Model Name: Style Timing Model (Small-Cap Crowding) - **Model Construction Idea**: The model uses a crowding-based trend approach to time large-cap and small-cap styles. Crowding is measured by the difference in momentum and trading volume ratios between small-cap and large-cap indices[3][20] - **Model Construction Process**: 1. Calculate the momentum difference between the Wind Micro-Cap Index and the CSI 300 Index across 10/20/30/40/50/60-day windows 2. Compute the trading volume ratio between the two indices over the same windows 3. Derive crowding scores for small-cap and large-cap styles by averaging the highest and lowest quantiles of the above metrics, respectively 4. Combine the momentum and volume scores to obtain the final crowding score. A score above 90% indicates high small-cap crowding, while below 10% indicates high large-cap crowding[25] - **Model Evaluation**: The model effectively captures the dynamics of style crowding and provides actionable insights for timing decisions[20][25] 3. Model Name: Industry Rotation Model (Genetic Programming) - **Model Construction Idea**: The model applies genetic programming to directly extract factors from industry indices' price, volume, and valuation data, without relying on predefined scoring rules. It uses a dual-objective approach to optimize factor monotonicity and top-group performance[28][32][33] - **Model Construction Process**: 1. Use NSGA-II algorithm to optimize two objectives: |IC| (information coefficient) and NDCG@5 (normalized discounted cumulative gain for top 5 groups) 2. Combine weakly collinear factors using a greedy strategy and variance inflation factor to form industry scores 3. Select the top 5 industries with the highest multi-factor scores for equal-weight allocation, rebalancing weekly[32][34] - **Model Evaluation**: The dual-objective genetic programming approach enhances factor diversity and reduces overfitting risks, making it a robust tool for industry rotation[32][34] 4. Model Name: China Domestic All-Weather Enhanced Portfolio - **Model Construction Idea**: The model adopts a macro-factor risk parity framework, emphasizing risk diversification across underlying macro risk sources rather than asset classes. It actively overweights favorable quadrants based on macro momentum[39][42] - **Model Construction Process**: 1. Divide macro risks into four quadrants based on growth and inflation expectations: growth above/below expectations and inflation above/below expectations 2. Construct sub-portfolios within each quadrant using equal-weighted assets, focusing on downside risk 3. Adjust quadrant risk budgets monthly based on macro momentum indicators, which combine buy-side momentum from asset prices and sell-side momentum from economic forecast surprises[42] - **Model Evaluation**: The strategy effectively integrates macroeconomic insights into portfolio construction, achieving enhanced performance through active allocation adjustments[39][42] --- Model Backtesting Results 1. A-Share Technical Scoring Model - Annualized Return: 20.78% - Annualized Volatility: 17.32% - Maximum Drawdown: -23.74% - Sharpe Ratio: 1.20 - Calmar Ratio: 0.88[15] 2. Style Timing Model (Small-Cap Crowding) - Annualized Return: 28.46% - Maximum Drawdown: -32.05% - Sharpe Ratio: 1.19 - Calmar Ratio: 0.89 - YTD Return: 11.85% - Weekly Return: 5.25%[26] 3. Industry Rotation Model (Genetic Programming) - Annualized Return: 32.92% - Annualized Volatility: 17.43% - Maximum Drawdown: -19.63% - Sharpe Ratio: 1.89 - Calmar Ratio: 1.68 - YTD Return: 6.80% - Weekly Return: 3.37%[31] 4. China Domestic All-Weather Enhanced Portfolio - Annualized Return: 11.93% - Annualized Volatility: 6.20% - Maximum Drawdown: -6.30% - Sharpe Ratio: 1.92 - Calmar Ratio: 1.89 - YTD Return: 3.59% - Weekly Return: 1.54%[43] --- Quantitative Factors and Construction Methods 1. Factor Name: Small-Cap Crowding Factor - **Factor Construction Idea**: Measures the crowding level of small-cap style based on momentum and trading volume differences between small-cap and large-cap indices[20][25] - **Factor Construction Process**: 1. Calculate momentum differences and trading volume ratios for multiple time windows 2. Derive crowding scores by averaging the highest and lowest quantiles of these metrics 3. Combine momentum and volume scores to obtain the final crowding score[25] 2. Factor Name: Industry Rotation Factor (Genetic Programming) - **Factor Construction Idea**: Extracts factors from industry indices using genetic programming, optimizing for monotonicity and top-group performance[32][34] - **Factor Construction Process**: 1. Perform cross-sectional regression of standardized daily trading volume against daily price gaps to obtain residuals (Variable A) 2. Identify the trading day with the highest standardized volume in the past 9 days (Variable B) 3. Conduct time-series regression of Variables A and B over the past 50 days to obtain intercepts (Variable C) 4. Compute the covariance of Variable C and standardized monthly opening prices over the past 45 days[38] --- Factor Backtesting Results 1. Small-Cap Crowding Factor - YTD Return: 11.85% - Weekly Return: 5.25%[26] 2. Industry Rotation Factor (Genetic Programming) - Training Set IC: 0.340 - Factor Weight: 18.7% - YTD Return: 6.80% - Weekly Return: 3.37%[31][38]

Investment Rating - The report maintains an "Overweight" rating for the real estate development and service sectors [6] Core Insights - The real estate sector is experiencing a significant reduction in holdings, with public funds and northbound capital reaching new lows in their investment proportions. The market is currently stabilizing, with a focus on recovery in core cities, particularly first-tier cities [1][2] - Recommended investment opportunities include companies with strong credit, urban presence, and product quality, as well as those with robust operational capabilities to manage cash flow during market adjustments [1] - The report highlights a shift in holdings concentration, with Beike rising to the top position among public fund holdings, indicating a narrowing of investor divergence in the sector [3] Summary by Sections Public Fund Holdings - As of Q4 2025, the total market value of public fund holdings in the real estate sector was 38.8 billion yuan, a 31% decrease quarter-on-quarter. The sector's holdings accounted for 0.43% of total stock investments, down 0.19 percentage points [2] - The real estate sector index fell by 8.9%, ranking 30th out of 31 sectors, primarily due to declining fundamentals and some companies hitting new stock price lows [2] Northbound Capital - Northbound capital's total holdings in real estate stocks were 11.5 billion yuan, a 17% decrease quarter-on-quarter, representing 0.45% of total northbound holdings [4] - The top five real estate stocks held by northbound capital include China Merchants Shekou, Poly Developments, and others, with notable increases in holdings for companies with "real estate+" attributes [4] Recommended Companies - Key recommended companies include: - Yuexiu Property (123 HK) with a target price of 7.06 HKD and a "Buy" rating [8] - Longfor Group (960 HK) with a target price of 15.21 HKD and a "Buy" rating [8] - Greentown Service (2869 HK) with a target price of 6.56 HKD and a "Buy" rating [8] - China Overseas Development (688 HK) with a target price of 19.08 HKD and a "Buy" rating [8] - China Merchants Shekou (001979 CH) with a target price of 12.79 CNY and a "Buy" rating [8] - CR Land (1109 HK) with a target price of 36.45 HKD and a "Buy" rating [8] - Others include companies like Greenland China, and Hong Kong local firms benefiting from market recovery [1][8] Performance Insights - Beike's market value increased significantly, reflecting a strong investor interest, while other companies like Poly Developments and China Merchants Shekou saw reductions in their holdings [3][4] - The report emphasizes the importance of operational efficiency and cash flow management for companies navigating the current market challenges [1][3]

证券研究报告 银行 华泰研究 2026 年 1 月 24 日 | 中国内地 动态点评 增持 (维持) | 12 月至今中信银行指数回调 8.4pct,主要受地产舆情、降息预期、资金风 | 沈娟 | 研究员 | | --- | --- | --- | | | SAC No. S0570514040002 | shenjuan@htsc.com | | 格切换影响,指数估值降至 0.65xPB、近 5 年 65%分位,部分优质个股 25E | SFC No. BPN843 | +(86) 755 2395 2763 | | 股息率近6%。险资开门红同比高增,预计后续新增保费对高股息、低波动 | | | | 的优质银行股仍有配置需求:银行开年投放积极,息差降幅明显收窄,盈利 | 贺雅亭 | 研究员 | | | SAC No. S0570524070008 | heyating@htsc.com | | 有望改善,地产链影响或较为可控。南京、宁波、兴业等 8家银行披露 25A | SFC No. BUB018 | +(86) 10 6321 1166 | | 业绩快报，其中６家营收提升、5 家利润改善，我们预计 20 ...

证券研究报告 金工 AI 量化的当下与未来 2026 年 1 月 22 日│中国内地 深度研究 人工智能 100：AI 量化的过去、现在与未来 本文是华泰人工智能系列的第 100 篇研究报告。过往的八年半里，我们亲 历了量化投资行业的这场深刻变革：技术路径上，从早期的机器学习，演进 到深度学习，再到如今以大语言模型为代表的新范式。应用场景上，从早期 的因子合成，拓展至因子挖掘与端到端建模，进而渗透到组合优化、行业轮 动、资产配置、流程管理等投资的各个环节。行业认知上，从最初的质疑与 观望，逐渐转向接纳与尝试，直至今日的全面拥抱。第 100 篇研究，既是 对过往足迹的回顾，也是对未来征途的眺望。 AI 量价端到端策略的演进 在量价研究普遍内卷的当下，端到端建模不仅是效率的提升，亦是一种回归 原始数据的研究范式。我们已实现从日频、周频等低频数据到逐笔成交、 level2 高频数据的全面覆盖，通过引入 GRU 及 Transformer 等架构，模型 得以直接在原始数据空间中学习量价数据间的内在联系。展望未来，全频段 融合或是关键，未来的端到端模型或将致力于打破时间尺度与数据形态的边 界，一方面通过对比学习等技术实 ...

证券研究报告 工业/基础材料 重视顺周期建材均衡配置机会 华泰研究 2026 年 1 月 22 日│中国内地 动态点评 政策持续推进房地产止跌回稳，重视传统与新兴均衡配置机会 1 月 21 日住建部部长倪虹在接受采访时表示，今年将着力稳定房地产市场， 继续因城施策、精准施策，发挥好房地产融资"白名单"制度的作用，支持 房企合理的融资需求，支持居民刚性和改善性住房需求。我们认为积极的房 地产政策仍然有望加速地产止跌回稳，目前地产端数据已较为充分地反映在 建材股股价和估值，上市公司通过国内市占率提升、海外业务拓展和品类扩 张，个别公司已开始逐步体现收入端改善迹象。短期我们建议重视建材板块 传统顺周期与新兴科技成长的均衡配置机会，推荐东方雨虹、中国联塑、兔 宝宝、伟星新材、北新建材。 地产债务化解路径逐步明朗，建材公司减值或有改善 1 月 21 日万科公告"21 万科 02"债券持有人会议高票通过核心展期议案， 以"首付 40%现金兑付+小额持有人保障+明确资产增信"的组合拳，为行 业探索出一条理性协商的债务化解路径。消费建材板块前期信用减值已相对 充分，除业务更偏向工程属性的防水材料外，大部分企业单项减值计提已超 ...

证券研究报告 百龙创园 (605016 CH) 需求高景气支撑 Q4 业绩增长提速 | 华泰研究 | | | 公告点评 | 投资评级(维持): | 买入 | | --- | --- | --- | --- | --- | --- | | 2026 年 | 1 月 | 22 日│中国内地 | 食品 | 目标价(人民币): | 28.00 | | | | | | 吕若晨 | 研究员 | | SAC No. S0570525050002 | lvruochen@htsc.com | | --- | --- | | SFC No. BEE828 | +(86) 755 8249 2388 | 基本数据 | 收盘价 (人民币 截至 1 月 22 日) | 22.48 | | --- | --- | | 市值 (人民币百万) | 9,442 | | 6 个月平均日成交额 (人民币百万) | 116.66 | | 52 周价格范围 (人民币) | 16.17-28.55 | 股价走势图 (6) 15 36 57 78 Jan-25 May-25 Sep-25 Jan-26 (%) 百龙创园 沪深300 资料来源：Wind ...