时间：2026年03月24日（星期二）下午 13:00
地点：会计学院 207 室
主题1：IPO问询主题与企业违规——基于大语言模型的发现
主讲人：李艳茹 中央财经大学
摘要：在全面实行注册制改革的背景下，IPO审核问询作为监管机构与拟上市企业信息交互的重要载体，其信息价值日益凸显。本文以科创板为例，运用大语言模型对38123个IPO审核问询问题进行主题信息挖掘，探讨IPO问询主题与企业上市后违规行为之间的关系。研究结果表明：（1）IPO问询主题相较于传统的财务数据和公司治理数据，能为企业违规预测提供显著的增量信息，预测指标AUC值平均提升约17.01%。（2）在不同问询主题中，收到较多核心技术、治理与独立性以及财会信息相关问询的企业，其上市后一年的违规概率较高；并且当问询表现出更强的负面语调时，其风险预警作用进一步增强。（3）机制分析发现，IPO问询主题能够发挥风险预警作用，主要源于监管机构的信息穿透优势与企业的组织惯性特征。（4）从上市后表现来看，技术主题问询较少的企业具有更为优异的持有期收益率，而财会信息类问询较多的企业在上市后更容易出现业绩变脸。本研究创新性地运用大语言模型进行信息挖掘，探讨IPO审核问询主题的风险预警价值，为监管机构优化审核策略以及投资者增强风险识别能力提供了实践启示。
主题2：Connected Shareholders at the Gate: Investment Banking Ties and Mutual Fund Voting
主讲人：张志宇 香港城市大学
摘要：This paper examines how investment banking ties influence the voting behavior of affiliated mutual funds in shareholder-sponsored proposals. We posit and document a clear alignment, where mutual funds tend to support the management of their parent bank's corporate clients. While it reveals a potential conflict of interest, we show that such alignment can yield positive governance outcomes. Specifically, by siding with management, affiliated funds help empower managers to effectively reject frivolous and often distracting shareholder proposals, particularly those submitted by “gadfly” individual activists. Further evidence indicates that pro-management voting by connected funds is associated with favorable market reactions and a higher likelihood of future business engagements, underscoring the strategic dimensions of voting behavior within financial conglomerates.
主题3：A Transparent Deep Learning Framework for Technical Analysis on Chart Patterns
主讲人：陈玮政 香港科技大学（广州）
摘要：We introduce a novel deep-learning-based framework to extract chart patterns indicating future return direction for technical analysis. The extraction is accomplished through the cooperation of a deep neural network and a human, in which the neural network explores possible patterns and the human performs definition and selection by a designed algorithm. The extracted patterns are then tested in the out-of-sample testing period. Empirical results demonstrate that the extracted chart patterns are strong technical indicators during the out-of-sample future period. Using the market index as the baseline, the extracted bullish patterns indicate higher subsequent returns on average, while the bearish patterns indicate lower ones. Also, the portfolio constructed from stocks exhibiting bullish chart patterns achieves a higher cumulative return compared with the market portfolio. Moreover, the proposed framework provides transparency. By treating our neural network as a pattern miner rather than a block-box predictor, we avoid the black-box inference process in conventional deep learning-based technical analysis applications. Meanwhile, the extracted chart patterns, in the form of specific geometry, are viewable by users.
