金融研报AI分析

国泰海通 · 宏观聚焦|扩内需、促转型、稳市场——7月中央政治局会议前瞻

预计7月中央政治局会议将把扩大内需置于更突出位置,财政从克制转向“追赶”,加快存量落地、增量结构性加力,六张网等重大基建将获得重点资金支持,货币延续宽松但更强调结构优化与传导效率提升;消费端以服务消费为战略重心,投资端推进基础设施网络建设,产业政策聚焦科技自强与能源转型双轮驱动,推动前沿技术攻关、伦理治理以及可再生能源扩量提质与可靠替代,同时全力维护资本市场平稳运行,形成投资者友好、逆周期调控与高质量发展的合力。该框架对下半年投资格局将产生系统性影响,尤其在基建、科技创新与新能源等领域的资金投放与政策协同方面具有较强的指引意义 [page::1][page::2][page::3]

Skill Substitution, Expectations, and the Business Cycle

本文基于1995–2018年德国六百万以上高中文凭毕业生的行政数据,研究毕业时劳动力市场状况对后续技能投资的影响 [page::0]。 研究发现总体上大学入学随经济周期呈正相关,且对学术型大学的入学影响尤为明显 [page::13]。 在长期 attainment 方面,对学术型大学的效应为负向(-1.04pp),对应用型大学为正向(+0.40pp),总体上呈现不同学位类型的替代效应 [page::13]。 大规模调查数据支持“预期回报”作为通道:经济下行会降低对高等教育的相对回报预期,而对学徒制等外部选项的回报预期相对稳定 [page::22]。 研究还发现,预期回报的变化主要来自全国性冲击而非地方冲击,并且它们直接影响毕业生的 enrollment intentions [page::22][page::24]。 在制度层面,研究强调德国职业教育体系的作用,以及宏观与行为层面的信念更新在教育选择中的关键作用 [page::27]。

Industry Dynamics with Cartels: The Case of the Container Shipping Industry

本文建立静态需求-边际成本模型并结合动态进入/退出与船舶投资的博弈,量化明示卡特尔对价格、产能分配以及行业形成过程的双向影响 [page::3]。估计结果显示,卡特尔价格楔在1973–1979年为 \$1,807/TEU、1980–1983年为 \$998/TEU,相当于各路平均运价的约30%至70% [page::3]。在移除卡特尔的情景下,1973–1983期消费者剩余提升约2–4%、生产者剩余下降约17–27%,净社会福利在三大市场呈现显著的异质性趋势,源自动态投资与产能成本的权衡 [page::28]。内部配额规则的改变会使不同市场的动态福利排序不同,容量比例分配并非在所有市场均为最优选择 [page::33]。结论强调监管者应同时评估价格效应与产能内部分配规则对长期福利的影响,而非仅关注短期价格变动 [page::33]。

Generative AI Availability, Grades, and Student Satisfaction at a Large University

本研究以美国大型公立大学(2015–2025,156,135名学生、87,936门课程)为样本,利用课程大纲与行政数据,采用差分中差分设计检验ChatGPT发布后GenAI可用性对成绩与学习满意度的影响,聚焦GenAI易感性与评估结构之间的关系以及COVID-19对结果的混淆。结果显示,GenAI可用性对总体成绩及低先前表现学生的影响并不显著;自我理解的影响也不显著,只有在将疫情效应视为短暂性时,学习兴趣才呈现小幅提升。将疫情效应与GenAI效应分离后,结论变得更为保守,即GenAI尚未显著削弱成绩信号或满意度相关的课程评估功能。 [page::1][page::24][page::21]

PAI-ECON-CLAUDE: A GATED HUMAN-IN-THE-LOOP MULTI-AGENT ARCHITECTURE FOR AI-ASSISTED ECONOMIC THEORY DEVELOPMENT

This paper proposes pAI-Econ-claude, a gated, human-in-the-loop multi-agent architecture for AI-assisted economic theory development, integrating inspectable intermediate records and targeted quality gates. [page::0] The design emphasizes interception at visible decision points, non-certifying gates, and irreversible human checkpoints to improve auditability without replacing formal verification. [page::0] In a matched-task evaluation across five economic-theory challenges, the gated workflow outperformed ungated baselines on four tasks, with substantial reductions in reported failure severity and higher usefulness. [page::0] A notable success occurred when a reality check rejected a false market-structure premise and a subsequent proof review corrected a welfare claim, while Task 1 shows limits where extra structure may compress key mechanisms. [page::0] The results support the claim that bounded gated oversight enhances interpretability and auditability of AI-assisted theory, with human judgment concentrated where irreversibility is highest. [page::0] The code and workflow are publicly available at the linked GitHub repository. [page::0]

Group boarding for airplanes: benchmarking static policies and optimizing dynamic assignment with deep reinforcement learning

This paper introduces a dynamic boarding group assignment framework for commercial aircraft, formulating the problem as a finite-horizon MDP and solving it with proximal policy optimization (PPO) using a CNN-based state encoder. The model integrates seat location, carry-on luggage declarations, and travel-companion constraints. In simulations across six layouts, the dynamic policy consistently outperforms strong static benchmarks (back-to-front with optimized group sizes) on both total boarding time and average individual boarding time, achieving up to 9.8% and 22.8% improvements respectively. The policy generalizes to varying load factors, companion distributions, and luggage loads, and a Pareto frontier is reported by sweeping the objective weight. These results suggest online dynamic grouping at check-in can meaningfully reduce turnaround time while improving passenger experience [page::6][page::7].

The Evolution of Digital Search: From Blue Links to Delegated Decision-Making

Digital search is undergoing a fundamental transformation from a human-driven process of discovery to an agent-mediated system of delegated decision-making. In an AI-native world, users articulate goals in natural language, agents interpret these intents, and outcomes are delivered as recommendations or executed decisions. This shift redefines search from a visible list of links to an embedded system component that governs information access, trust, incentives, and market structure. The paper identifies four structural shifts—ranked lists to delegated decisions, attention economy to preference economy, evaluating performance to verifying delegation, and retrospective study to prospective research design—and argues that the future of search will be determined by open, transparent, and competitive agentic systems that regulate how decisions are made and markets operate. Early experimental evidence from agent-mediated marketplaces and theoretical work suggest that small design choices in information access, option surface, and action execution have first-order effects on efficiency, competition, and welfare for both consumers and firms. The authors emphasize the need for agent-native infrastructure, auditable signals of quality, and governance mechanisms to sustain trust and competition over time [page::0]

Electricity demand has not become more price-responsive despite ninety years of technological change∗

本研究对1934–2024年的4720条 own-price 弹性估计进行了元分析,覆盖462项研究,采用六层识别梯子与出版偏差校正等方法进行综合评估。核心发现包括:第一,最佳识别设计的研究给出较朴素研究更小的短期弹性,短期约0.09(设计基估计),综合回归约0.16;第二,弹性随调整时间延长而显著上升,长期约0.38,但这一梯度在几十年内基本稳定;第三,九十年数据中未观测到弹性的上升趋势,且在科技密集情景(如时段定价)下,总用电消费的价格弹性仍偏低。价格本身并未使总用电量更具价格弹性,灵活性更多来自启用技术、契约设计与项目规划等因素 [page::2][page::3]。

Portfolio Optimization under Dynamic Rebalancing via Topological Data Analysis and News Sentiments

本研究将拓扑数据分析(TDA)与新闻情感结合,融合多源特征(RSI、SO、MACD及FinBERT情感分数)构建资产的四维点云表征,通过PD与PL的拓扑总结来计算资产之间的非线性相似性,并在聚类筛选后以带交易成本约束的动态均值-方差(DRMV)框架进行滚动窗口再平衡。情感信息提升了资产表征与聚类质量,显著提升了在样本内外的回报与风险调整性,同时通过保留机制降低轮换成本并增强组合稳定性。实证结果基于标普500成分股,在3/5/10日再平衡设定下,TDA基准的组合相较于相关性与欧氏距离方法及若干基准策略,呈现更高的平均收益、夏普及STARR等风险调整指标,且在冲突期等市场动荡时具有更好的鲁棒性。 [page::0][page::2]

【计算机】超节点产品集中亮相WAIC,国产大模型加速突围——计算机行业周报

本报告聚焦计算机行业在AI算力、超节点与国产大模型领域的最新进展,梳理 WAIC 技术亮点与落地潜力,评估龙头企业在芯模协同、算力投资与端侧应用等维度的竞争力,并给出行业中性评级及头部标的关注要点,同时提示宏观、供应链及技术迭代等风险因素。 [page::0][page::1]

Accelerating fossil gas independence in Europe

本研究通过高时空分辨的欧洲能源系统模型在严格的化石气供应上限下对各部门进行耦合优化,发现降低进口依赖在经济上具有可行性:通过替代发电、低温工业热和建筑热,欧洲有望在2030–35年前将天然气消费减半至约200 bcm/年,与欧洲生产能力相匹配。但独立供应并不能完全保护消费者免受全球天然气价格波动影响,因为 gas 在边际电价形成中的作用仍然显著,需辅以其他政策工具来对冲价格冲击 [page::0]

The Human-AI Substitution Principle: When will you be replaced by AI in your organization?

本研究提出人类–AI任务分配(HAT)模型,将层级组织中的任务分配、成本、风险与组织深度统一建模,揭示AI替代人力的“风险调整成本”阈值条件,并给出替代、混合结构及组织扁平化等结构性后果的明确推论。核心发现包括:当AI在风险调整后的成本低于人类时,AI将替代人力;替代具有阈值性质,可能引发离散的波动式 workforce transitions;中层管理在特定情形下易成为替代的峰值区域;在高合规与声誉风险环境下,AI成本优势可能被风险惩罚抵消,从而维持混合型组织;部署规模的放大可通过培训成本摊销效应加速替代,进而推动组织深度下降与管理跨度的扩张等。这一框架将自动化经济学、组织设计、AI治理与 workforce planning 统一起来,提供可检验的多层次预测与治理工具 [page::4][page::10][page::18]

Execution and Evaluation: A New Occupational Measure and Long-Run Employment Gradients

This paper introduces an occupation-level execution share by scoring 19,265 O*NET task statements under fixed execution and AI-capability rubrics, enabling a direct comparison with AI capability and linking to employment trends. The execution share is shown to be reproducible across model coders and O*NET vintages, distinct from AI-capability and routine-task intensity, and adds only modest incremental power beyond existing evaluation measures. In harmonized panels, employment growth is lower in execution-heavy white-collar occupations across multiple periods, with a secular, between-family gradient rather than a pure AI-era shock; the vintage-valid capability gradient steepens after 2022, though the change is dated and not causally attributed to AI. The results establish a workable measure and a chronology for potential labor-market effects, emphasizing measurement framework and future causal-testing needs rather than asserting a causal AI-driven decline.

Quantifying Sub-Optimality in Routing for Automated Market Makers

We provide a large-scale empirical audit of DEX routing using 2.98 million WETH–USDC swaps on Ethereum. We compare realized routes with optimized benchmarks, finding an average shortfall of 2.02 bps per trade, amounting to about \$24 million on roughly \$120 billion of volume. We introduce three reproducible optimal benchmarks—SCO (activated pools only), FVO (all pools, no gas), and G-FVO (gas-aware)—and a scalable bisection-based routing algorithm to compute these benchmarks. The study highlights two regularities: information timeliness critically affects routing outcomes, and inefficiency is heavy-tailed with a few extreme outliers driving most of the dollar losses. Sandwiching attacks emerge as a concrete channel for routing sub-optimality. The framework provides a rigorous, reproducible basis for timing-aware, gas-aware routing improvements [page::1]

Good Guys With Guns? The Relationship Between Legal Firearm Ownership and Firearm Deaths and Crime in Canada

本研究在加拿大情境下,利用2013-2019年的行政数据,检视合法持枪与枪支相关死亡、谋杀及犯罪之间的关系。结果显示,在研究期内,手枪持有显著增加,但与枪支相关死亡、犯罪之间的整体关系并不强烈,且在不同检验中多为不显著。研究还讨论了未来监管政策(如手枪冻结与 buyback)在现有制度下的潜在效果与局限性,以及对非法枪支与合法枪支之边际作用的差异性。 [page::0]

【山证新兴产业】圣兆药物:布比卡因脂质体获批,国际化布局加速推进

圣兆药物获得布比卡因脂质体注射液注册证,国内获批+海外授权的双轮驱动商业化路径由此开启,且与Torrent的美国市场独家商业化合作意向为全球化布局奠定关键支撑 [page::0][page::1]。全球布比卡因脂质体市场潜力巨大,2025年原研药全球市场规模约5.75亿美元,中国市场预计2030年达到约18亿元,2025-2030年的复合增长率约62.7% [page::2]。公司通过平台化高端制剂路线,已实现5款产品上市并推动产能扩张及国际化扩展,未来盈利能力有望随海外销售与国内放量协同而改善 [page::2][page::3]。同时,定增5.0848亿元用于持续推进管线研发、扩产与国际化布局,相关风险包括研发、核心人员流失与盈利周期的不确定性等 [page::3]。

新三板公司分析 圣兆药物:布比卡因脂质体获批,国际化布局加速推进

圣兆药物在高端制剂领域实现关键里程碑,布比卡因脂质体获批并推进中美双报与海外授权,开启国内放量与海外市场的双轮驱动模式 [page::0]。受益于全球市场空间扩张与国际化布局,未来产能扩张和商业化落地将推动估值重估,但也伴随研发、治理等风险提示 [page::1]。

中金 • REITs | REITs二季报点评:经营符合预期,持续关注企稳改善信号

本轮REITs二季报呈现业绩分化与分项信号共存的格局,数据中心与保租房的现金流确定性较高,产业园与物流仓储则受区域供给与租约结构分化影响更大,后续仍需关注出租率修复与租金跌幅收敛的动向,以及路网引流对高速板块的潜在支撑。结合86家REITs的披露口径,二季度收入完成度约24%,可供分配完成度约27%,板块间差异显著,数据中心/保租房等核心资产具备相对防御性,而产业园/物流等周期性板块仍需耐心观察其基本面拐点。整体结论提示:在估值回调背景下,优先配置现金流稳健的资产,持续关注企稳信号与区域性供需变化对估值的推动作用 [page::1][page::2][page::8][page::9][page::13][page::20]]

中金 | 高频监控:科技调整期间,公募“固收+”减配了吗?

本报告聚焦2026年7月科技板块调整期,分析公募“固收+”在科技敞口上的配置演变及其背后逻辑,揭示二级债基在科技敞口上的轮动特征与风控取向 [page::0]。 通过八因子净值回归与二次规划约束,反推出二级债基的高频仓位变化,60日窗口的R²中位数约0.77、20日窗口约0.80,说明模型解释力较强并能较好映射实际操作约束 [page::4]。 自2024年9月以来,二级债基持续“加科技、减转债”调仓,2026年6月末科技仓位达到约7.96%(在股票市值中的占比约48%),转债仓位降至约6.42% [page::4]。 进入2026年7月,科技板块调整叠加,二级债基科技仓位呈现“先小幅上升后快速下降”的特征,与向金融地产、制造和医药等板块的加仓轮动相呼应 [page::5]。 此外,2026Q2“固收+”规模持续攀升,二级债基仍为主要增量来源,新发产品与净申购明显,表明市场在寻求弹性收益的同时对科技敞口保持关注,回看历史数据与当季披露均支持上述结论 [page::1]。 7月初至7月中旬,科创50回撤约-22.3%,二级债基相关区间也承压,显示在科技阶段性调整中,减仓动作逐步开展以控制风险 [page::2]

中金:AI的金融时刻

本报告围绕AI资本开支从现金流向债务的转折点,估算未来五年全球外部融资需求约3.5万亿美元,梳理融资渠道及容量分布(投资级债券、公开市场股权、杠杆融资、证券化、私募资本等),并讨论偿债可持续性、再融资压力及对金融体系的传导风险,指出银行和私募资本等在新融资周期中的机会与风险,强调利润率、折旧期限等关键变量对“回本”可能性的决定性作用 [page::1][page::4][page::5]