统计与数据科学学院知识库

  在当前全球金融市场复杂性和联动性日益增强的背景下，利用关联网络测度银行系统流动性风险已成为金融研究的重要方向。2008年国际金融危机揭示了金融机构间关联网络在风险传染中的关键作用。传统流动性风险测度方法通常局限于单一银行视角，未能深入探讨银行系统内部的复杂网络关联和风险传播路径，因而难以适应复杂的金融系统。近年来，金融网络理论的兴起为流动性风险研究提供了新的视角，但现有研究在网络结构特性与风险动态演化的结合方面仍有待加强。特别是针对我国复杂的银行网络系统，关于流动性风险的测度、传导和监测预警方法的研究相对不足。为此，本文从银行关联网络的视角出发，借鉴金融网络分析技术、混频模型与机器学习方法，构建符合我国实际国情的商业银行业流动性风险测度、传导和预警框架，以期弥补国内在流动性风险生成机制和风险传导研究上的不足，更好地探析我国商业银行业流动性风险的生成和传导规律，并提高监管效率。

  首先，根据我国商业银行业“短存长贷”的业务特性，本文设计了包含同业拆借和共同贷款网络的银行间关联网络模型。通过对银行节点和关联网络拓扑特征的系统刻画，揭示了银行间直接、间接资金的联系特征，并深入分析了网络结构对商业银行业流动性风险的影响。结果显示，由于同业拆借与共同贷款网络存在交互作用，会显著影响流动性风险的传播路径，如果忽视两者之间联动效应，可能会低估银行系统的流动性风险；银行业系统重要性银行因在银行业间关联网络中处于核心地位，因此容易成为银行业间流动性风险传染的核心节点，而对于边缘中小银行而言，由于对流动性管理能力较弱且更为依赖短期融资，因而容易成为流动性风险扩散的薄弱环节。

  其次，由于本文所构建的银行业流动性风险模型中，数据具有多频率的特点，因此提出了改进的混频动态因子模型（MF-DFM），结合了月、季、年等不同频率的动态数据，构建了商业银行业流动性风险综合指数。该指数不仅能够全面捕捉流动性风险的时间动态特征，还能从关联网络视角定量揭示流动性风险在不同时间尺度下的演化规律。研究结果表明流动性风险指数的动态变化与重大经济事件之间存在密切联系，也验证了混频动态因子模型在捕捉流动性风险动态变化的有效性，并凸显了关联网络视角在揭示风险传播机制中的重要作用。

  再次，除了银行体系内部的关联网络引发的风险传导，银行业外部市场对流动性风险的传导路径也存在不可忽视的影响作用。为了进一步探讨外部市场对流动性风险的动态传导机制，本研究引入时变参数随机波动率向量自回归（TVP-SV-VAR）模型，深入分析了外汇市场、房地产市场、股票市场和债券市场四个外部经济变量对商业银行业流动性风险的影响机制及时变动态特征。研究发现各外部市场通过不同渠道对商业银行业流动性风险产生显著的时变影响，外汇市场的汇率波动、房地产市场的景气度变化、股票市场的价格波动以及债券市场的收益率波动在不同时间段对银行流动性风险表现出复杂的传导路径和动态特征。

  最后，在流动性风险的早期预警方面，本文将商业银行业流动性风险指数与机器学习方法结合，构建了基于关联网络的流动性风险预警模型。通过时间序列分解和模型集成技术提高了对短期流动性风险预测的精准度。与此同时，本文构建马尔可夫区制转移模型识别流动性风险状态，建立多模型对比预警系统。研究发现随机森林模型在测试集上的表现最佳，预测准确率和召回率均达到较高水平，因此被认为是商业银行业流动性风险预警的最优模型。此外，采用SHAP模型可解释技术识别出推动流动性风险上升的关键驱动因素，为防范流动性风险提供了技术支持和科学参考。

  依据上述研究内容和结论，本文的贡献主要体现在：一是提出基于关联网络的流动性风险测度和传导分析框架，同时结合同业拆借网络和共同贷款网络，弥补了传统研究中对银行间复杂关联结构的刻画不足；二是针对金融数据的多频率特性，提出混频动态因子模型，构建商业银行业流动性风险指数，提升了风险测度的准确度和动态捕捉的精度；三是结合机器学习方法开发流动性风险早期预警模型，并通过模型解释技术识别关键驱动因素，为政策制定者和银行管理者提供科学的决策依据。

    In the context of the increasing complexity and interconnectedness of global financial markets, utilizing correlation network measures to assess liquidity risk in the banking system has become an important direction in financial research. The 2008 international financial crisis revealed the critical role of inter-institutional correlation networks in risk transmission. Traditional methods of measuring liquidity risk are often limited to a single bank perspective and fail to deeply explore the complex network connections and risk propagation pathways within the banking system, making them inadequate for the complexities of the financial system. In recent years, the rise of financial network theory has provided new perspectives for liquidity risk research; however, existing studies still need to strengthen the integration of network structural characteristics with the dynamic evolution of risk. Particularly concerning China's complex banking network system, research on the measurement, transmission, and monitoring and early warning methods of liquidity risk is relatively insufficient. Therefore, this paper starts from the perspective of banking correlation networks, drawing on financial network analysis techniques, mixed-frequency models, and machine learning methods, to construct a liquidity risk measurement, transmission, and early warning framework for the commercial banking sector that aligns with China's actual national conditions. This aims to address the shortcomings in domestic research on the mechanisms of liquidity risk generation and risk transmission, better explore the generation and transmission patterns of liquidity risk in China's commercial banking sector, and enhance regulatory efficiency.

    First, based on the "short-term deposits and long-term loans" business characteristics of our country's commercial banking sector, this research designs an interbank association network model that includes interbank lending and joint loan networks. Through a systematic characterization of the bank nodes and the topological features of the association network, it reveals the characteristics of direct and indirect funding connections between banks and deeply analyzes the impact of network structure on liquidity risk in the commercial banking sector. The results show that the interaction between interbank lending and joint loan networks significantly affects the propagation path of liquidity risk. Ignoring the linkage effects between the two may lead to an underestimation of the liquidity risk within the banking system. Systemically important banks in the banking sector, due to their core position in the interbank association network, are more likely to become the core nodes for the transmission of liquidity risk among banks. In contrast, smaller banks on the periphery, which have weaker liquidity management capabilities and rely more on short-term financing, are prone to becoming weak links in the diffusion of liquidity risk.

      Second, due to the multi-frequency characteristics of the data in the liquidity risk model constructed for the banking industry in this research, an improved Mixed-Frequency Dynamic Factor Model (MF-DFM) has been proposed. This model integrates dynamic data of varying frequencies, such as monthly, quarterly, and annually, to construct a comprehensive index of liquidity risk for commercial banks. This index not only captures the temporal dynamics of liquidity risk comprehensively but also quantitatively reveals the evolution patterns of liquidity risk across different time scales from the perspective of the associated network. The research findings indicate a close relationship between the dynamic changes in the liquidity risk index and significant economic events, thereby validating the effectiveness of the Mixed-Frequency Dynamic Factor Model in capturing the dynamic changes of liquidity risk. Additionally, it highlights the important role of the associated network perspective in revealing the mechanisms of risk transmission.

      Third, in addition to the risk transmission caused by the interconnected networks within the banking system, the external market also has a significant impact on the transmission pathways of liquidity risk. To further explore the dynamic transmission mechanism of external markets on liquidity risk, this study introduces a Time-Varying Parameter Stochastic Volatility Vector Autoregression (TVP-SV-VAR) model, which provides an in-depth analysis of the impact mechanisms and time-varying dynamic characteristics of four external economic variables: the foreign exchange market, the real estate market, the stock market, and the bond market on the liquidity risk of commercial banks. The research finds that each external market exerts significant time-varying effects on the liquidity risk of commercial banks through different channels. Fluctuations in exchange rates in the foreign exchange market, changes in the prosperity of the real estate market, price fluctuations in the stock market, and yield fluctuations in the bond market exhibit complex transmission pathways and dynamic characteristics regarding bank liquidity risk over different time periods.

     Finally, in terms of early warning for liquidity risk, this paper combines the liquidity risk index of the commercial banking sector with machine learning methods to construct a liquidity risk early warning model based on associative networks. The accuracy of short-term liquidity risk predictions has been enhanced through time series decomposition and model ensemble techniques. At the same time, this paper develops a Markov regime-switching model to identify liquidity risk states and establishes a multi-model comparative early warning system. The research finds that the random forest model performs best on the test set, achieving high levels of both prediction accuracy and recall, thus being considered the optimal model for liquidity risk early warning in the commercial banking sector. Furthermore, the use of SHAP model interpretability techniques identifies key driving factors that contribute to the rise of liquidity risk, providing technical support and scientific reference for the prevention of liquidity risk.

      Based on the research content and conclusions, the contributions of this research are primarily reflected in the following aspects: First, this research proposes a framework for measuring and analyzing liquidity risk based on an interconnected network, which integrates the interbank lending network and the syndicated loan network. This addresses the inadequacies in depicting the complex interconnections among banks in traditional studies. Second, in response to the multi-frequency characteristics of financial data, this research introduces a MF-DFM to construct a liquidity risk index for commercial banks, thereby enhancing the accuracy of risk measurement and the precision of dynamic capture. Third, this research develops an early warning model for liquidity risk using machine learning methods and employs model interpretation techniques to identify key driving factors, providing scientific decision-making support for policymakers and bank managers.