文献类型：期刊文献

中文题名：基于大语言模型方法的供应链风险测度与结构分解

英文题名：Measuring and decomposing firm-level supply chain risks:A large language model approach

作者：张宇[1];余典范[1];王超[2]

机构：[1]上海财经大学,上海200433;[2]华东理工大学,上海200237

年份：2026

卷号：37

期号：3

起止页码：27

中文期刊名：财贸研究

外文期刊名：Finance and Trade Research

收录：;北大核心:【北大核心2023】；

基金：国家社会科学基金重大项目“超大规模市场优势与现代化产业体系建设研究”(23&ZD042);中央高校基本科研业务费专项资金资助项目“中国上市公司供应链风险的测度与特征事实——基于大语言模型方法”(CXJJ-2024-428);国家自然科学基金青年项目“数据跨境流动限制对企业全球供应链参与的影响与应对研究”(72503063)。

语种：中文

中文关键词：大语言模型;供应链风险;产业链;供应链

外文关键词：large language model;supply chain risk;industrial chain;supply chain

摘要：准确测度企业层级的供应链风险是完善供应链风险监测预警体系的重要前提。基于2012—2022年A股制造业上市公司年度报告,利用大语言模型方法从367万余句管理层讨论与分析文本中提取供应链关键词,结合关键词共现识别策略测度了企业-年份层级的供应链风险,并通过事件研究、宏观冲击趋势比对和资本市场风险关联分析对指标的合理性进行了多方验证。在此基础上,将供应链风险分解为采购风险、生产风险和销售风险,发现造纸、橡胶等原材料粗加工行业供应链风险主要来自采购端,废弃资源综合利用业等工艺复杂行业供应链风险主要来自生产端,服装、制鞋、计算机通信设备制造等出口优势行业供应链风险主要来自销售端。本文使用的数据和估计方法在一定程度上缓解了使用人工或传统机器学习方法进行关键词选取时可能存在的偏误,对于提高供应链风险监测和预警系统的智能化水平也有积极意义。
Accurately measuring firm-level supply chain risk is a critical prerequisite for improving supply chain risk monitoring and early warning systems.Using the annual reports of A-share manufacturing listed companies from 2012 to 2022,this paper employs a large language model approach to extract supply chain keywords from over 3.67 million sentences in the Management Discussion and Analysis(MD&A)sections,and measures firm-year level supply chain risks through a keyword co-occurrence identification strategy.The validity of the index is verified through multiple dimensions:event study analysis,macro-level supply chain shock trend comparison,and capital market risk correlation analysis.On this basis,we decompose supply chain risks into procurement risk,production risk,and sales risk.The results show that industries such as paper making and rubber processing,which rely heavily on raw materials,primarily face supply chain risks from the procurement side;industries with complex production processes,such as waste resource recycling,mainly encounter risks from the production side;and export-oriented industries,such as textile and apparel,footwear,and computer and communication equipment manufacturing,are predominantly exposed to risks from the sales side.The data and estimation methods utilized in this study help mitigate biases inherent in manual or traditional machine learning approaches for keyword selection,and also contribute to enhancing the intelligence level of supply chain risk monitoring and early warning systems.