随着汽车金融下沉市场的拓展与多源异构数据的爆发，传统信用评分模型在兼顾预测精度与特定场景泛化能力时遭遇瓶颈。本文提出一种基于多模态混合专家模型（Multimodal Mixture of Experts, MMoE）的深度风控框架。该框架依托企业级AI中台，通过动态门控网络（Gating Network）将借款人的结构化征信、非结构化文本语义及动态行为特征智能路由至专属专家网络。基于 LendingClub 公开数据集的实证研究（有效映射汽车金融多模态场景）表明，MMoE 模型在 AUC 与 KS 指标上显著优于 LightGBM 等主流基准模型，且其期望校准误差（ECE）降至 0.015。研究证实，门控路由机制不仅提升了长尾人群的逾期预测准度，更为深度学习在金融领域的应用提供了宏观可解释性视角。本研究为金融机构构建高并发、易扩展的下一代智能风控底座提供了系统性的工程路径与理论支撑。

This paper provides a comprehensive analysis of excess returns specific to corporate bonds. We construct a measure of excess returns that uses synthetic Treasury securities with identical cash flows as benchmarks, thereby fully removing interest rate effects and isolating the component of returns specific to corporate bonds. Using a monthly sample from 2002 to 2024, we find that, in addition to being lower on average, the corporate-bond-specific excess return differs significantly in the cross section from both the standard excess return based on T-bills and the duration-adjusted return. We further examine the effects of a broad set of bond-level characteristics and systematic risk factors on bond excess returns. Together, these findings provide a foundational benchmark for future research on corporate bond returns.

大语言模型等生成式人工智能技术正深刻重构企业的业务模式与创新路径。本文以2018—2024年中国A股上市公司为研究对象，依托企业在线招聘数据与职业层面的大语言模型暴露度，构建了企业人工智能技术暴露度指标，实证检验了企业核心业务面临的技术冲击对其实质性创新的影响及其作用机制。研究发现，企业人工智能技术暴露度能够显著促进企业实质性创新，该结论在经过一系列内生性与稳健性检验后依然成立。机制分析表明，面对外生技术冲击，企业主要通过推动人力资本结构升级与提升研发投入强度，将技术红利转化为实质性的创新产出。异质性分析表明，在技术吸收能力较强、转型动机强烈以及具备长期积累优势的企业中，人工智能技术暴露度对企业实质性创新的促进作用更为显著。本文从外生技术冲击的视角，为理解人工智能重塑企业创新活动的内在机理提供了经验证据，对引导企业妥善应对技术更迭、优化要素配置进而实现高质量发展提供了现实参考。

We document a striking anomaly in international capital flows that we term "sanction arbitrage": U.S. investors exploited the 2014 sanctions on Russia by significantly increasing holdings in Russian equities while Rest-of-World (ROW) investors fled. We rationalize this behavior through a simple game-theoretic model where the sanctioning government faces a trade-off between geopolitical objectives and domestic welfare, effectively creating a protective shield for domestic investors and driving out ROW investors. Empirically, we confirm that pre-sanction U.S flows negatively predicted subsequent sanction designations. Consequently, U.S. investors internalized this protection to act as opportunistic buyers, absorbing fire-sale assets from exiting foreign investors and capturing significant excess returns from Russian stock holdings. These findings reveal that "smart" sanctions designed to preserve market access can inadvertently generate wealth transfers from foreign to domestic agents.

We study asset demand and price formation in a two-asset rational expectations equilibrium with narrow thinking, where traders imperfectly coordinate decisions across assets under non-nested price information. When the price of one asset increases, cross-asset inference from prices reduces expected demand for the other asset, which feeds back into the demand response for the original asset. Narrow thinking weakens internal coordination and amplifies reliance on price-based inference. As a result, more severe narrow thinking leads to higher own-price elasticities. The model delivers sharp implications for market liquidity and price informativeness in the presence of bounded rationality.

Serving on the official unsecured creditors' committee (UCC) of a bankrupt firm provides hedge funds with access to material nonpublic information (MNPI), which can facilitate their informed trading across firms and asset markets. We find that hedge funds increase equity turnover and execute more large trades in the quarters following UCC membership. In contrast, hedge funds do not exhibit such trading behavior after accessing public information about bankrupt firms or holding the bankrupt firm's debt without committee involvement. Importantly, these large trades often target firms with close economic ties to the bankrupt entity. Returns from these MNPI-driven trades are substantial.

Can fully agentic AI nowcast stock returns? We deploy a state-of-the-art Large Language Model to evaluate the attractiveness of each Russell 1000 stock each trading day, starting in April 2025 when AI web interfaces enabled real-time search. Our data contribution is unique along three dimensions. First, the nowcasting framework is completely out-of-sample and free of look-ahead bias by construction: predictions are collected at the current edge of time, ensuring the AI has no knowledge of future outcomes. Second, this temporal design is irreproducible once the information environment passes. Third, our framework is fully agentic: we do not feed the model curated news or disclosures; it autonomously searches the web, filters sources, and synthesises information into quantitative predictions. We find that AI possesses genuine stock-selection ability, but that its predictive power is concentrated in identifying future winners. A daily value-weighted portfolio of the 20 highestranked stocks earns a Fama-French five-factor plus momentum alpha of 19.4 basis points and an annualised Sharpe ratio of 2.68 over April 2025–March 2026. The same portfolio accumulates roughly 49.0% cumulative return, versus 21.2% for the Russell 1000 benchmark. The strategy is economically implementable: the average bid-ask spread of the daily Top-20 portfolio is 1.79 basis points, less than 10% of gross daily alpha. However, the signal remains asymmetric. Bottom-ranked portfolios generally exhibit alphas close to zero, while the strongest predictive content sits in the extreme top ranks. Delayed-entry tests further show that predictability does not vanish after a single day; rather, the signal remains positive over a broad window of subsequent entry dates, consistent with slow information diffusion rather than a fleeting overnight anomaly.

This paper studies how mutual fund managers update their beliefs as macroeconomic conditions change. Using regulator-mandated reports from Chinese mutual funds, we measure the intensity of belief updating from year-over-year changes in stated outlooks and decompose those updates into macro and micro themes. We show that belief updating is state-contingent: funds with more intensive belief updating shift their narratives toward macro (micro) topics during recessions (expansions) and concurrently reduce (increase) procyclical stock exposures and on-site company visits. This state-contingent belief updating predicts superior performance when matched to prevailing economic conditions, with macro-oriented updates paying off mainly for high-updating funds in recessions and micro-oriented updates paying off more broadly in expansions. Investors recognize this signal of skill, allocating greater flows to these funds, especially when past returns are less informative. Finally, belief updating is stronger for younger managers and for funds from newer, smaller families, consistent with signaling under career and competitive pressures.

This paper introduces reinforcement learning to examine the effect of trading on noise in a dynamic limit order market equilibrium. It shows that intensive noise liquidity provision (consumption) increases speculators' liquidity consumption (provision), improving (reducing) market liquidity. Channeled by uninformed chasing and informed aggressive liquidity provision, the increasing noise liquidity provision and consumption, respectively, improve price efficiency, generating a U-shaped price efficiency to the noise trading uncertainty on liquidity provision and consumption. Associated with a hump-shaped (U-shaped) profitability for the informed (uninformed) at a U-shaped noise trading cost in the noise trading uncertainty, this implies that, at increasing noise trading cost, intensive noise liquidity provision improves market liquidity, price efficiency, order profitability of informed traders, and reduces the loss, even makes profit, for uninformed traders.

We document a novel "news extrapolation" behavior among investors, which distorts the market reaction to corporate news. Specifically, investors tend to extrapolate the value of past news in the immediate reaction to the newly arrived news. News extrapolation generates a biased price reaction to news, which is completely reversed afterwards. Furthermore, the tendency of news extrapolation is related to the recency, consistency, and value uncertainty of news. Investors extrapolate not only from news of the same category but also from news of different categories. By analyzing the trading behavior and sentiment of different investor groups, we find that retail investors tend to be news extrapolators, while institutional investors trade against the news extrapolators.