Introduction
Financial crises have been a recurrent feature of economic history, and understanding their causes and propagation is a central concern in financial economics. In the wake of the 2008 global financial crisis, a rich literature has emerged that uses financial network models to study systemic risk and contagion. These models view the financial system as a network of interconnected balance sheets – banks, financial institutions, and markets linked by interbank loans, derivative contracts, or common asset holdings. This review surveys the literature on financial crises through the lens of network models, emphasizing post-2008 contributions (approximately 80% of our focus) while also covering foundational pre-2008 work (around 20%). We begin by summarizing mainstream explanations for why financial crises occur and key factors affecting financial stability more generally. We then provide an overview and comparison of influential network-based models of financial crises – including seminal contributions by Allen and Gale (1998, 2000), Eisenberg and Noe (2001), Elliott, Golub, and Jackson (2014), Acemoglu, Ozdaglar, and Tahbaz-Salehi (2015), and other relevant models. For each, we highlight the insights offered as well as their strengths and weaknesses, avoiding excessive technical detail but noting any tricky conceptual aspects. Next, we discuss the role of “information sensitivity” – a concept from the post-crisis literature that helps explain panics – and consider how it might be incorporated into network models of crises. Finally, we briefly survey the computational techniques used in financial network modeling (e.g. clearing algorithms, simulations, stress tests), before concluding. The goal is to provide a structured, “Journal of Economic Literature” style overview that is clear, comprehensive, and well-referenced, drawing on authoritative sources from academic journals (e.g. AER, JF), NBER working papers, SSRN and arXiv preprints, and insights from leading scholars in the field.金融危机是经济历史中反复出现的特征，理解其原因和传播是金融经济学的核心关注点。在 2008 年全球金融危机之后，出现了大量文献，利用金融网络模型研究系统性风险和传染。这些模型将金融系统视为一个相互关联的资产负债表网络——银行、金融机构和市场通过银行间贷款、衍生合同或共同资产持有相互连接。本文通过网络模型的视角回顾金融危机的文献，强调 2008 年后贡献（大约 80%的重点），同时也涵盖了 2008 年前的基础性工作（约 20%）。我们首先总结了主流解释金融危机发生的原因以及影响金融稳定的关键因素。然后，我们提供了影响力网络模型的概述和比较，包括 Allen 和 Gale（1998, 2000）、Eisenberg 和 Noe（2001）、Elliott、Golub 和 Jackson（2014）、Acemoglu、Ozdaglar 和 Tahbaz-Salehi（2015）等的开创性贡献，以及其他相关模型。 对于每个模型，我们强调所提供的见解以及它们的优缺点，避免过多的技术细节，但注意任何棘手的概念方面。接下来，我们讨论“信息敏感性”的角色——这是一个来自危机后文献的概念，有助于解释恐慌——并考虑它如何可能被纳入危机的网络模型中。最后，我们简要回顾金融网络建模中使用的计算技术（例如清算算法、模拟、压力测试），然后总结。我们的目标是提供一个结构化的“经济文献期刊”风格的概述，清晰、全面且引用充分，参考来自学术期刊（例如 AER、JF）、NBER 工作论文、SSRN 和 arXiv 预印本的权威来源，以及该领域领先学者的见解。
Mainstream Explanations for Financial Crises
Research on financial crises has identified a set of recurring causes and mechanisms that make the financial system vulnerable. Many theoretical and empirical studies find that financial crises are often preceded by booms – periods of rapid credit expansion and asset price inflation – that eventually turn to busts
. In other words,
unsustainable leverage and asset bubbles tend to build up during good times, sowing the seeds of subsequent distress when expectations reverse. Across different types of crises – banking panics, debt crises, currency crashes, etc. – common patterns emerge. Four features are frequently cited in the literature as prime drivers of major crises
:
对金融危机的研究已经确定了一系列反复出现的原因和机制，使金融系统变得脆弱。许多理论和实证研究发现，金融危机通常是在繁荣时期之前发生的——快速信贷扩张和资产价格膨胀的时期——最终转向萧条IMF.ORG。换句话说，不可持续的杠杆和资产泡沫往往在好时光中积累，当预期逆转时，播下后续困境的种子。在不同类型的危机中——银行恐慌、债务危机、货币崩溃等——出现了共同的模式。文献中经常提到的四个特征被认为是重大危机的主要驱动因素IMF.ORGIMF.ORG：
-
Asset Price Booms and Busts: Large increases in asset values (such as real estate or equities) that prove unsustainable, leading to sharp price corrections. Historical analyses show that many crises follow periods of extraordinary asset appreciation which then collapse, eroding bank balance sheets and net worth
. Reinhart and Rogoff (2009) document that countries with the biggest run-ups in housing and equity prices were among the most vulnerable in the 2008 crisis.资产价格的繁荣与崩溃：资产价值（如房地产或股票）的大幅上涨是不可持续的，导致价格的急剧修正。历史分析表明，许多危机发生在资产异常升值的时期之后，这些升值随后崩溃，侵蚀银行资产负债表和净资产IMF.ORGIMF.ORG。Reinhart 和 Rogoff（2009）记录了在 2008 年危机中，房价和股价上涨幅度最大的国家是最脆弱的国家之一IMF.ORGIMF.ORG。
-
Credit Booms and Excess Leverage: A rapid expansion of credit often accompanies the asset boom, resulting in high leverage and debt burdens
. When the boom turns to bust, highly leveraged borrowers (households, firms, or banks) face distress, and a credit crunch ensues. Many crises can be traced to prolonged credit booms that led to deteriorating lending standards and risk-taking in search of yield. Empirical studies confirm that credit booms are a robust predictor of banking crises.信贷繁荣与过度杠杆：信贷的快速扩张通常伴随着资产繁荣，导致高杠杆和债务负担IMF.ORGIMF.ORG。当繁荣转为萧条时，高杠杆借款人（家庭、企业或银行）面临困境，信贷紧缩随之而来。许多危机可以追溯到长期的信贷繁荣，这导致了贷款标准的恶化和为了追求收益而冒险IMF.ORGIMF.ORG。实证研究证实，信贷繁荣是银行危机的一个强有力的预测指标IMF.ORGIMF.ORG。
-
Deterioration of Loan Quality and Risk Build-up: During the boom, there is often a build-up of marginal or risky loans and investments – for example, subprime mortgages before 2008 – which means the financial system’s asset quality declines
. This creates hidden vulnerabilities (high systemic risk) that become apparent only when defaults rise. In good times, default rates stay low, masking the fragility; but once the cycle turns, losses on these risky assets can mount rapidly, threatening bank solvency.贷款质量恶化与风险积累：在繁荣时期，往往会出现边际或高风险贷款和投资的积累——例如，2008 年前的次级抵押贷款——这意味着金融系统的资产质量下降IMF.ORGIMF.ORG。这会造成隐藏的脆弱性（高系统性风险），只有在违约率上升时才会显现。在经济好时，违约率保持在低位，掩盖了脆弱性；但一旦周期转变，这些高风险资产的损失可能迅速增加，威胁到银行的偿付能力。
-
Regulatory Lapses and Innovation Outpacing Oversight: A failure of regulation and supervision to keep up with financial innovation is a recurrent theme
. Crises often involve new financial instruments or complex practices (e.g. securitization, derivatives) that evade existing rules. In the mid-2000s, for instance, the widespread use of opaque structured products and off-balance-sheet vehicles contributed to excessive risk-taking that regulators did not fully grasp. Weak oversight, inadequate capital buffers, and flawed incentives (e.g. “too-big-to-fail” expectations) can allow systemic risks to grow unchecked. Indeed, regulatory failure is frequently identified as a major cause of crises despite decades of efforts to strengthen oversight.监管失误与创新超越监管：监管和监督未能跟上金融创新的步伐是一个反复出现的主题IMF.ORGIMF.ORG。危机通常涉及新的金融工具或复杂的实践（例如证券化、衍生品），这些工具和实践规避了现有规则。例如，在 2000 年代中期，模糊的结构性产品和表外工具的广泛使用导致了监管者未能完全理解的过度风险承担IMF.ORG。薄弱的监督、不足的资本缓冲和有缺陷的激励（例如“太大而不能倒”预期）可能导致系统性风险失控增长。事实上，尽管经过数十年的努力加强监督，监管失误仍常被认为是危机的主要原因IMF.ORGIMF.ORG。
In addition to these four core factors, scholars have noted other elements that played a role in the Global Financial Crisis of 2007–2009. Notably, this crisis featured new aspects on top of the usual boom-bust dynamics: (1) the proliferation of complex, opaque financial instruments (such as CDOs) that obscured risk; (2) increased interconnectedness of financial markets globally (a network dimension, with the U.S. at the core); (3) extremely high leverage in the banking and shadow-banking sectors; and (4) the central role of the household sector (e.g. subprime mortgage borrowers)
. These factors, combined with the traditional problems listed above, led to the worst financial crisis since the Great Depression
. In summary, mainstream explanations view crises as the outcome of
vulnerabilities accumulated during boom periods – excessive leverage, asset bubbles, risky lending – which, when triggered by some shock or loss of confidence, lead to a rapid unraveling of financial positions. This general narrative is echoed in many analyses of 2008 (e.g. Gorton 2010, Brunnermeier 2009) and earlier crises, and it sets the stage for considering how
financial network structures might amplify or mitigate these dynamics.
除了这四个核心因素，学者们还注意到其他在 2007-2009 年全球金融危机中发挥作用的因素。值得注意的是，这场危机在通常的繁荣-萧条动态之上还出现了新的方面：(1) 复杂、不透明的金融工具（如 CDO）的激增，这些工具掩盖了风险；(2) 全球金融市场的相互关联性增强（网络维度，以美国为核心）；(3) 银行和影子银行部门的极高杠杆；以及(4) 家庭部门的核心角色（例如次贷借款人）IMF.ORGIMF.ORG。这些因素与上述传统问题相结合，导致了自大萧条以来最严重的金融危机IMF.ORG。总之，主流解释将危机视为在繁荣时期积累的脆弱性的结果——过度杠杆、资产泡沫、风险贷款——这些在某种冲击或信心丧失的触发下，导致金融头寸的迅速解体。 这种普遍叙述在对 2008 年（例如 Gorton 2010，Brunnermeier 2009）及早期危机的许多分析中得到了呼应，并为考虑金融网络结构如何可能放大或缓解这些动态奠定了基础。
Key Factors Affecting Financial Stability
Financial stability is determined by a combination of macroeconomic conditions, institutional factors, and importantly, the structure of the financial system itself. Beyond broad credit and asset market cycles, research has identified several specific factors that influence how resilient or fragile a financial system is to shocks. Many of these factors are precisely what network models aim to capture. Key factors include:金融稳定是由宏观经济条件、制度因素以及金融系统本身的结构的组合决定的。除了广泛的信贷和资产市场周期，研究还确定了几个特定因素，这些因素影响金融系统对冲击的韧性或脆弱性。许多这些因素正是网络模型旨在捕捉的关键因素，包括：
-
Interconnectedness and Network Topology: The pattern of linkages among financial institutions – who is exposed to whom – can critically shape systemic stability. Highly interconnected networks can spread shocks widely, but they also allow risks to be shared. There is a complex trade-off: more interbank links can dilute individual shocks through diversification, yet they also create additional channels for contagion
. The topology matters: for example, a “fully connected” network (where everyone is linked to everyone) might absorb small shocks well, whereas a sparsely connected or hub-and-spoke network might localize small shocks but be vulnerable to a big hub failure. Research shows that the relationship between connectivity and contagion is often non-monotonic – increasing connections can first increase stability (better risk-sharing) and then decrease it once a certain threshold is passed. We discuss this trade-off in detail later. In short, who is connected to whom (and how many connections each node has) is a key determinant of systemic risk.互联性和网络拓扑：金融机构之间的链接模式——谁与谁相互暴露——可以在关键上影响系统稳定性。高度互联的网络可以广泛传播冲击，但它们也允许风险共享。这是一个复杂的权衡：更多的银行间链接可以通过多样化来稀释个别冲击，但它们也创造了额外的传染渠道FINANCIALRESEARCH.GOVFINANCIALRESEARCH.GOV。拓扑结构很重要：例如，一个“完全连接”的网络（每个人都与每个人相连）可能很好地吸收小冲击，而一个稀疏连接或中心辐射网络可能会将小冲击局部化，但对大中心的失败则很脆弱。研究表明，连接性与传染性之间的关系往往是非单调的——增加连接最初可以提高稳定性（更好的风险共享），但一旦超过某个阈值后又会降低ARXIV.ORGARXIV.ORG。我们稍后将详细讨论这种权衡。简而言之，谁与谁相连（以及每个节点有多少连接）是系统性风险的关键决定因素。
-
Common Exposures and Asset Correlations: Not all contagion happens through direct counterparty links; indirect contagion via common asset holdings is another channel. If many banks invest in the same assets or sectors, a decline in asset values can hit multiple institutions simultaneously (a form of correlation risk). Such common exposures mean that even in the absence of direct lending links, banks’ fortunes become correlated – a mechanism highlighted in the literature on fire sales and mark-to-market losses. For instance, the failure of Lehman Brothers in 2008 impacted money market funds not because of direct loans, but via losses on commonly held securities. Studies distinguish between contagion through networks of bilateral obligations and contagion through price effects or shared risk factors
. A highly diversified portfolio at the individual bank level can paradoxically increase systemic risk if everyone diversifies in a similar way (creating overlap in portfolios).共同风险暴露和资产相关性：并非所有的传染都是通过直接的对手方联系发生的；通过共同资产持有的间接传染是另一种渠道。如果许多银行投资于相同的资产或行业，资产价值的下降可能会同时影响多家机构（这是一种相关风险）。这种共同风险暴露意味着，即使在没有直接贷款联系的情况下，银行的命运也会变得相关——这一机制在关于火灾销售和按市值计价损失的文献中得到了强调。例如，2008 年雷曼兄弟的倒闭影响了货币市场基金，并不是因为直接贷款，而是通过对共同持有证券的损失。研究区分了通过双边义务网络的传染和通过价格效应或共享风险因素的传染。个别银行层面的高度多元化投资组合如果每个人以相似的方式进行多元化，反而可能增加系统性风险（造成投资组合的重叠）。
-
Leverage and Buffers: The level of leverage (debt relative to equity) in financial institutions is a fundamental driver of stability. High leverage means even small asset losses can wipe out equity and render institutions insolvent. System-wide, if many institutions are highly leveraged, a modest shock can produce a wave of defaults. Conversely, strong capital and liquidity buffers absorb losses and prevent contagion. In network terms, a well-capitalized node is less likely to fail and transmit distress. Empirical evidence from crises consistently shows that banks with thinner capital cushions and greater reliance on short-term funding are more prone to collapse. Thus, network models often incorporate leverage as a parameter that can amplify contagion cascades
.杠杆与缓冲：金融机构的杠杆水平（债务相对于股本）是稳定性的基本驱动因素。高杠杆意味着即使是小的资产损失也可能会抹去股本，使机构破产。在系统层面，如果许多机构的杠杆率很高，适度的冲击可能会引发一波违约。相反，强大的资本和流动性缓冲可以吸收损失并防止传染。从网络的角度来看，资本充足的节点不太可能失败并传递困境。危机的实证证据一致表明，资本缓冲较薄且对短期融资依赖较大的银行更容易崩溃。因此，网络模型通常将杠杆作为一个参数纳入，以放大传染级联。
-
Size and Concentration (Too Big To Fail): The distribution of sizes (or importance) of nodes matters. A system with a few very large, dominant banks (a concentrated system) may be stable day-to-day, but if a huge node fails, the impact is catastrophic. By contrast, a more even distribution (many medium-sized banks) might diffuse risk more, though it could be prone to herding behavior. The failure of a “too big to fail” institution can have disproportionate effects – both through direct exposures and through confidence channels (if a giant can fail, everyone can). Research by Glasserman & Young (2015) finds that heterogeneity in node size can exacerbate contagion, especially if a shock starts at a large, central node
. The presence of a core-periphery network structure (with a core of highly connected big banks and a periphery of smaller ones) also influences stability: the core might withstand small shocks but be vulnerable to a big hit, which then spreads outward.规模与集中度（太大而不能倒闭）：节点的大小（或重要性）分布很重要。一个拥有少数非常大、主导性银行的系统（集中系统）可能在日常运作中是稳定的，但如果一个巨大的节点失败，影响将是灾难性的。相比之下，更均匀的分布（许多中型银行）可能会更好地分散风险，尽管它可能容易出现跟风行为。“太大而不能倒闭”机构的失败可能会产生不成比例的影响——既通过直接暴露，也通过信心渠道（如果一个巨头可以倒闭，所有人都可以）。Glasserman & Young（2015）的研究发现，节点大小的异质性可能加剧传染，特别是当冲击从一个大型中心节点开始时ARXIV.ORGARXIV.ORG。核心-边缘网络结构的存在（核心是高度连接的大型银行，边缘是较小的银行）也会影响稳定性：核心可能承受小冲击，但对大冲击则可能脆弱，随后向外扩散ARXIV.ORGARXIV.ORG。
-
Market Liquidity and Funding Conditions: A more subtle factor is the availability of liquidity during stress. In a crisis, banks may hoard liquidity and withdraw funds from counterparties (a funding run), as in the classic Diamond-Dybvig bank run model. Liquidity dries up precisely when it is most needed, turning illiquidity into insolvency. If the network relies heavily on short-term interbank funding, a loss of confidence can cause cascading funding withdrawals. This dynamic was evident in 2008 in the freeze of the interbank lending market and has been modeled as “network runs” in some frameworks. The ease with which assets can be sold without fire-sale discounts (market liquidity) also matters – if banks forced to sell find no buyers except at steep discounts, one bank’s sale can depress prices for all and inflict losses on others (a price-mediated contagion). Cifuentes, Ferrucci & Shin (2005) incorporate this channel, showing that capital-constrained banks selling illiquid assets can propagate stress to others even absent direct links.市场流动性和融资条件：一个更微妙的因素是在压力期间流动性的可用性。在危机中，银行可能会囤积流动性并从对手方撤回资金（融资挤兑），就像经典的 Diamond-Dybvig 银行挤兑模型所示。流动性在最需要的时候枯竭，将流动性不足转变为破产。如果网络严重依赖短期的银行间融资，信心的丧失可能会导致级联的融资撤回。这种动态在 2008 年银行间借贷市场冻结时显而易见，并在某些框架中被建模为“网络挤兑”。资产在没有火灾销售折扣的情况下被出售的难易程度（市场流动性）也很重要——如果被迫出售的银行发现除了以大幅折扣外没有买家，那么一家银行的出售可能会压低所有银行的价格，并给其他银行带来损失（价格介导的传染）。Cifuentes, Ferrucci & Shin（2005）纳入了这一渠道，表明资本受限的银行出售流动性不足的资产可以在没有直接联系的情况下将压力传播给其他银行。
-
Institutional and Policy Environment: Finally, financial stability is affected by factors like the presence (and credibility) of a lender of last resort, deposit insurance schemes, and the overall regulatory framework. Effective crisis management (e.g. central bank liquidity support or bailouts) can halt a network cascade, while policy failures can worsen panic. These aspects lie somewhat outside static network models, but they form the backdrop – for instance, if authorities are expected to save big banks, that might dampen or delay contagion (while possibly encouraging moral hazard). Conversely, uncertainty about policy can exacerbate information-sensitive runs. Network models typically take the rules of the game as given, but in interpreting results one must remember these broader context factors.制度和政策环境：最后，金融稳定受到最后贷款人存在（及其可信度）、存款保险计划和整体监管框架等因素的影响。有效的危机管理（例如中央银行流动性支持或救助）可以阻止网络级联，而政策失误则可能加剧恐慌。这些方面在静态网络模型之外，但它们构成了背景——例如，如果当局被期望拯救大型银行，这可能会抑制或延迟传染（同时可能鼓励道德风险）。相反，对政策的不确定性可能会加剧信息敏感型挤兑。网络模型通常将游戏规则视为既定，但在解释结果时必须记住这些更广泛的背景因素。
In summary, the stability of the financial system is a multifaceted issue. Interconnections, common risks, leverage, and size structure all interact in complex ways
. A key “takeaway” from both theory and historical crisis analysis is that
no single factor determines stability in isolation – it’s the combination that matters. For example, high connectivity might be benign if banks are well-capitalized and shocks are uncorrelated, but dangerous if banks are thinly capitalized or all exposed to the same asset. This insight motivates the development of network models that can handle
multiple factors simultaneously (connectivity, shock size, heterogeneity, etc.) to understand systemic risk. We now turn to the core of the review: how such financial network models explain crises and what we learn from them.
总之，金融系统的稳定性是一个多方面的问题。相互联系、共同风险、杠杆和规模结构以复杂的方式相互作用FINANCIALRESEARCH.GOVFINANCIALRESEARCH.GOV。理论和历史危机分析的一个关键“收获”是，没有单一因素可以孤立地决定稳定性——组合才是关键。例如，如果银行资本充足且冲击不相关，高连接性可能是良性的，但如果银行资本薄弱或都暴露于同一资产，则可能是危险的。这一见解促使我们开发能够同时处理多个因素（连接性、冲击大小、异质性等）的网络模型，以理解系统性风险。我们现在转向评审的核心：这些金融网络模型如何解释危机，以及我们从中学到了什么。
Financial Network Models of Crises: Overview and Key Models
Network models of financial crises seek to answer the question: how do shocks to one part of the system spread to others via the web of financial linkages? In contrast to traditional models that might treat banks in aggregate, network models make the connections explicit – modeling banks (or other entities) as nodes and their inter-linkages (loans, credits, investments) as edges in a network. This approach gained prominence post-2008, but its intellectual roots trace back to earlier work on bank contagion. Below we discuss several key models in chronological order, highlighting their setup, main findings, and contributions. We start with the foundational pre-2008 models by Allen & Gale (1998/2000) and Eisenberg & Noe (2001), then move to influential post-2008 models by Elliott et al. (2014) and Acemoglu et al. (2015), along with related contributions. Throughout, we compare their assumptions and results.金融危机网络模型试图回答这样一个问题：系统某一部分的冲击是如何通过金融联系的网络传播到其他部分的？与可能将银行视为整体的传统模型不同，网络模型明确了连接——将银行（或其他实体）建模为节点，将它们的相互联系（贷款、信用、投资）建模为网络中的边。这种方法在 2008 年后获得了显著关注，但其知识根源可以追溯到早期关于银行传染的研究。下面我们按时间顺序讨论几个关键模型，突出它们的设置、主要发现和贡献。我们从 Allen & Gale（1998/2000）和 Eisenberg & Noe（2001）的基础性 2008 年前模型开始，然后转向 Elliott 等人（2014）和 Acemoglu 等人（2015）的影响力后 2008 年模型，以及相关贡献。在整个过程中，我们比较它们的假设和结果。
Allen and Gale (1998, 2000) – Interbank Risk-Sharing vs. Contagion艾伦和盖尔（1998, 2000） – 银行间风险分担与传染效应
Franklin Allen and Douglas Gale’s work (published in 2000 in JPE, building on a 1998 working paper) is seminal in analyzing financial contagion through interbank connections. They consider a simple model of banks in different regions that face liquidity demand shocks (à la Diamond-Dybvig). Banks can lend to each other (interbank deposits) to share liquidity risk. Allen and Gale ask: does this risk-sharing arrangement make the system more stable, or can it create fragility? Their model uses a network of four banks in four regions, with interbank lending linking their balance sheets
. Depositors in each region may withdraw early (a liquidity shock). If one region has a surplus of liquidity (fewer withdrawals) and another has a deficit (many withdrawals), interbank loans allow the surplus region’s bank to help the deficit region’s bank meet withdrawals –
mutual insurance against liquidity shocks. In the
baseline case with small shocks, this works well: interbank connections allow each bank to withstand its depositors’ demands by drawing on the network, achieving a first-best allocation
. This finding underscores the
benefit of connectivity: more complete interbank markets can
smooth out idiosyncratic liquidity shocks, enhancing stability.
Franklin Allen 和 Douglas Gale 的研究（2000 年在 JPE 上发表，基于 1998 年的工作论文）在通过银行间连接分析金融传染方面具有开创性。他们考虑了一个简单的模型，其中不同地区的银行面临流动性需求冲击（类似于 Diamond-Dybvig）。银行可以相互借贷（银行间存款）以分担流动性风险。Allen 和 Gale 问道：这种风险分担安排是否使系统更稳定，还是可能导致脆弱性？他们的模型使用了四个地区四家银行的网络，银行间贷款将它们的资产负债表连接起来。每个地区的存款人可能会提前提款（流动性冲击）。如果一个地区有流动性盈余（提款较少），而另一个地区有流动性赤字（提款较多），银行间贷款允许盈余地区的银行帮助赤字地区的银行满足提款需求——对流动性冲击的相互保险。 在小冲击的基线情况下，这种方法效果良好：银行间的联系使每个银行能够通过网络应对存款人的需求，实现最佳配置FINANCIALRESEARCH.GOVFINANCIALRESEARCH.GOV。这一发现强调了连接性的好处：更完善的银行间市场可以平滑特有的流动性冲击，增强稳定性。
However, Allen and Gale also show that the same interbank links can facilitate contagion when a large shock hits. They consider an extreme scenario where one region experiences an unexpectedly high withdrawal demand (a spike in liquidity need that was not anticipated). In that case, the affected bank must liquidate a lot of its long-term assets early (incurring losses), and even with help from others, it may fail to repay all interbank loans. The shortfall on interbank claims then transmits the trouble to other banks. Allen and Gale demonstrate that interbank lending, while allowing risk sharing, can also create fragility – an initially small shock in one place (one bank’s liquidity shortfall) can cascade into a system-wide crisis if it forces multiple banks into insolvency
. In their model, the contagion unfolds as follows: one bank’s inability to meet its interbank obligation causes its creditor bank to incur a loss and possibly default, which in turn hurts that bank’s creditors, and so on – a
domino effect of defaults. Notably, this crisis is driven by fundamentals (an extreme liquidity shock), not just a self-fulfilling panic, so it’s a
rational contagion story
.
然而，艾伦和盖尔也表明，当一个大冲击发生时，相同的银行间联系可以促进传染。他们考虑了一个极端情景，其中一个地区经历了意外的高提款需求（未预料到的流动性需求激增）。在这种情况下，受影响的银行必须提前清算大量长期资产（遭受损失），即使得到其他银行的帮助，它也可能无法偿还所有的银行间贷款。银行间索赔的短缺随后将麻烦传递给其他银行。艾伦和盖尔证明，银行间借贷虽然允许风险分担，但也可能造成脆弱性——一个地方（一个银行的流动性短缺）最初的小冲击如果迫使多家银行破产，就可能级联成系统性危机。在他们的模型中，传染的过程如下：一家银行无法履行其银行间义务导致其债权银行遭受损失并可能违约，这反过来又伤害了该银行的债权人，依此类推——违约的多米诺效应。 值得注意的是，这场危机是由基本面驱动的（极端流动性冲击），而不仅仅是自我实现的恐慌，因此这是一个理性的传染故事FINANCIALRESEARCH.GOVFINANCIALRESEARCH.GOV。
A particularly important insight from Allen & Gale is how the structure of the interbank network affects fragility. They compare different network topologies: a “complete” network where every bank is connected to every other, versus a “ring” network where each bank has just one counterparty (forming a circle). They find the ring network is highly fragile: if one bank in the ring fails, it passes its losses entirely to its one creditor, likely causing that bank to fail, and so on around the ring in a chain reaction
. In a ring, there’s no diversification of risk – each link carries a large exposure – so a single default can travel far. In contrast, in a
fully connected network, the impact of one bank’s failure is
diluted across many counterparties; each other bank only loses a small fraction, so they can absorb the hit
. Thus,
more densely connected networks require larger shocks to trigger widespread defaults
. They also consider an intermediate case of “clustered” networks (groups of banks that are fully connected internally but only one link between groups), finding that a firewall can exist between clusters limiting contagion
. One of the
themes of Allen and Gale’s analysis is the interplay of network topology with shock correlation: if all banks face the same shock (systematic risk), no network can save them; but if shocks are uncorrelated, interbank links are generally stabilizing – except when an unanticipated extreme shock hits one node, at which point a more concentrated network will suffer more
.
来自 Allen & Gale 的一个特别重要的见解是，银行间网络的结构如何影响脆弱性。他们比较了不同的网络拓扑：一个“完全”网络，其中每个银行都与其他所有银行相连，和一个“环”网络，其中每个银行只有一个对手方（形成一个圆圈）。他们发现环形网络非常脆弱：如果环中的一个银行失败，它会将损失完全转嫁给其唯一的债权人，这可能导致该银行也失败，依此类推，形成连锁反应。在环中，没有风险的多样化——每个链接都承载着巨大的风险敞口——因此单一的违约可以传播得很远。 相比之下，在一个完全连接的网络中，一家银行的失败对许多对手方的影响被稀释；其他每家银行只损失一小部分，因此它们可以承受这种打击FINANCIALRESEARCH.GOVFINANCIALRESEARCH.GOV。因此，更密集连接的网络需要更大的冲击才能引发广泛的违约FINANCIALRESEARCH.GOVFINANCIALRESEARCH.GOV。他们还考虑了一种“集群”网络的中间情况（内部完全连接但组间只有一条链接的银行组），发现集群之间可以存在防火墙，限制传染FINANCIALRESEARCH.GOV。艾伦和盖尔分析的一个主题是网络拓扑与冲击相关性的相互作用：如果所有银行面临相同的冲击（系统性风险），没有任何网络可以拯救它们；但如果冲击是无相关的，银行间的链接通常是稳定的——除非一个意外的极端冲击袭击一个节点，此时一个更集中的网络将遭受更大的损失FINANCIALRESEARCH.GOV。
Strengths and Contributions: Allen & Gale (1998/2000) broke new ground by showing explicitly how network structure (pattern of interbank claims) can alter the propagation of bank failures. It provided a theoretical foundation for why “financial contagion” is not just about psychology but can be an equilibrium outcome of optimal contracts in interbank markets
. The model captures a realistic concern: banks lend to each other for efficiency, but those same links can transmit trouble. The clarity of comparing a ring vs. complete network yielded the influential insight of a
robust-yet-fragile trade-off: dense networks share small shocks well but can transmit big shocks widely, whereas sparse networks localize shocks but forego risk-sharing. This trade-off has guided much subsequent work. Another strength is the micro-founded nature – it’s built on an optimally derived liquidity insurance motive, so contagion arises endogenously (no need to assume panic).
优势与贡献：Allen & Gale (1998/2000) 开创性地展示了网络结构（银行间索赔模式）如何改变银行破产的传播。这为“金融传染”不仅仅是心理现象提供了理论基础，而是可以成为银行间市场中最优合同的均衡结果FINANCIALRESEARCH.GOVFINANCIALRESEARCH.GOV。该模型捕捉了一个现实的关注点：银行之间为了效率而相互借贷，但这些相同的联系也可能传播麻烦。比较环形网络与完全网络的清晰性带来了一个有影响力的见解：强韧但脆弱的权衡：密集网络能够很好地共享小冲击，但可以广泛传播大冲击，而稀疏网络则局部化冲击但放弃风险共享。这一权衡指导了后续的许多研究。另一个优势是其微观基础特性——它建立在最优派生的流动性保险动机之上，因此传染是内生产生的（无需假设恐慌）。
Weaknesses and Limitations: The simplicity necessary for tractability is a drawback. The model uses only four banks and very stylized shock scenarios (one special state causes contagion)
. This raises questions about how general the results are for larger, more complex networks. Indeed,
Allen & Gale must keep the network small and symmetric to solve it analytically, meaning real-world complexity (many banks, arbitrary link patterns) is not addressed
. They also focus on
liquidity interbank risk (a bank run style withdrawal shock) rather than solvency shocks or market price effects, so other crisis mechanisms are outside the model. Moreover, the contagion in their baseline is driven by a fundamental shock; the model does not account for
self-fulfilling runs or information contagion (in fact, they contrast their rational-run with Diamond-Dybvig’s panic). Finally, it assumes all interbank debts are senior and must be paid in full or default; no consideration of partial recovery or bankruptcy costs is included. These abstractions, while reasonable for a clean analysis, limit direct quantitative application. Nonetheless, the Allen-Gale framework remains a cornerstone, illustrating that
network configuration itself can be a source of financial fragility or resilience.
弱点和局限性：为了可处理性所需的简单性是一个缺点。该模型仅使用四家银行和非常简化的冲击场景（一个特殊状态导致传染）FINANCIALRESEARCH.GOVFINANCIALRESEARCH.GOV。这引发了关于结果在更大、更复杂网络中的普遍性的问题。实际上，Allen & Gale 必须保持网络小且对称才能进行解析，这意味着现实世界的复杂性（许多银行、任意链接模式）没有得到解决FINANCIALRESEARCH.GOV。他们还专注于流动性银行间风险（银行挤兑式提款冲击），而不是偿付能力冲击或市场价格效应，因此其他危机机制不在模型之内。此外，他们基线中的传染是由基本冲击驱动的；该模型没有考虑自我实现的挤兑或信息传染（实际上，他们将其理性挤兑与 Diamond-Dybvig 的恐慌进行了对比）。最后，它假设所有银行间债务都是优先的，必须全额偿还或违约；没有考虑部分回收或破产成本。这些抽象虽然对于清晰的分析是合理的，但限制了直接的定量应用。 尽管如此，Allen-Gale 框架仍然是一个基石，说明网络配置本身可以是金融脆弱性或韧性的来源。
(Note: Freixas, Parigi, and Rochet (2000) developed a related early model of interbank contagion. They also examine a ring vs. complete network and find that banks in a ring network are more vulnerable to runs because the failure of one bank imposes larger losses on its single neighbor, undermining confidence（注：Freixas、Parigi 和 Rochet（2000）开发了一个相关的早期银行间传染模型。他们还研究了环形网络与完全网络，发现环形网络中的银行更容易受到挤兑，因为一家银行的失败会对其唯一的邻居造成更大的损失，从而削弱信心。）
. That model differed by focusing on depositor flows between regions rather than exogenous liquidity shocks, but it reinforced the message that network structure matters. It also hinted at moral hazard issues – in a complete network, banks may take more risk knowing others will share losses. These early works set the stage for the formal network contagion literature.)该模型的不同之处在于关注地区之间的存款人流动，而不是外生流动性冲击，但它强化了网络结构重要性的观点。它还暗示了道德风险问题——在一个完整的网络中，银行可能会承担更多风险，因为他们知道其他银行会分担损失。这些早期的研究为正式的网络传染文献奠定了基础。
Eisenberg and Noe (2001) – Clearing Payments and Cascading DefaultsEisenberg 和 Noe（2001）– 清算支付与级联违约
Eisenberg and Noe (2001) introduced what is now a standard framework for modeling cascades of defaults in a network of interbank obligations. Unlike Allen and Gale’s focus on liquidity sharing, Eisenberg & Noe consider a static balance-sheet perspective: each financial institution (node) owes certain amounts to others and is owed certain amounts. If one or more nodes suffer a shock to their assets (for example, a sudden loss on loans or investments), they may not be able to fully pay their creditors, causing those creditors to incur losses and possibly default on their own obligations. The key question is: given an initial shock, what is the endogenous pattern of payments and defaults in the network? Eisenberg and Noe provide an algorithm to compute the clearing payment vector, i.e. the actual payments each bank makes on its debts in equilibrium, considering limited liability and priority of claims. This framework is highly influential because it formalizes the notion of counterparty contagion in a solvency sense. As one survey notes, the E&N model “forms the basis of much subsequent work on contagion in financial networks”
.
Eisenberg 和 Noe（2001）引入了现在成为标准的框架，用于建模银行间义务网络中的违约级联。与 Allen 和 Gale 关注流动性共享不同，Eisenberg 和 Noe 考虑的是静态资产负债表视角：每个金融机构（节点）欠其他机构一定金额，并且被其他机构欠一定金额。如果一个或多个节点的资产遭受冲击（例如，贷款或投资的突然损失），它们可能无法完全偿还债权人，从而导致这些债权人遭受损失，并可能违约自己的义务。关键问题是：在初始冲击下，网络中的支付和违约的内生模式是什么？Eisenberg 和 Noe 提供了一种算法来计算清算支付向量，即在均衡状态下每个银行对其债务的实际支付，考虑有限责任和索赔优先权。这个框架具有很高的影响力，因为它在偿付能力的意义上形式化了对手方传染的概念。正如一项调查所指出的，E&N 模型“构成了后续关于金融网络传染的许多工作的基础”FINANCIALRESEARCH.GOV。
Model Setup: Each bank $i$ has some outside assets (like loans to firms, bonds, etc.) and liabilities (debts) which include amounts owed to other banks. These interbank liabilities form a directed network (who owes whom). If all banks can meet their obligations, great – but if a shock hits (say Bank A’s outside assets lose value), Bank A’s equity might go negative. In that case, Bank A defaults and can only pay creditors a fraction of what it owes. Those creditors (other banks) then suffer a loss on their asset side (because the loan to A is not fully repaid), which might push some of them into default, and so on. Eisenberg & Noe show that under reasonable assumptions (debt contracts are senior, pro-rata repayment in default, etc.), one can compute the fixed point of the payment system: basically, find a set of payments that each bank makes such that no bank would pay more than its obligations or less than what can be extracted from it
. They prove that a
clearing equilibrium exists and is essentially unique, and they provide an iterative algorithm (often called the
fictitious default algorithm) to find it
. The algorithm conceptually starts by assuming everyone pays in full, then progressively reduces payments for any bank that would be insolvent, redistributing that shortfall to its creditors, and repeating until consistency is achieved. The result is the set of defaults and partial payments after the dust settles. This process explicitly shows
how payment shortfalls originating at one or more nodes spread through the network, causing a widening series of defaults
.
模型设置：每个银行 $i$ 具有一些外部资产（如对企业的贷款、债券等）和负债（债务），其中包括欠其他银行的金额。这些银行间负债形成一个有向网络（谁欠谁）。如果所有银行都能履行其义务，那就太好了——但如果发生冲击（比如银行 A 的外部资产贬值），银行 A 的股本可能会变为负数。在这种情况下，银行 A 违约，只能向债权人支付其所欠金额的一部分。这些债权人（其他银行）随后在其资产方面遭受损失（因为对 A 的贷款没有完全偿还），这可能会使其中一些银行陷入违约，依此类推。 Eisenberg & Noe 表明，在合理的假设下（债务合同是优先的，违约时按比例偿还等），可以计算支付系统的固定点：基本上，找到每个银行支付的一组款项，使得没有银行会支付超过其义务或少于可以从中提取的款项FINANCIALRESEARCH.GOV。他们证明了清算均衡的存在性，并且基本上是唯一的，并提供了一种迭代算法（通常称为虚构违约算法）来找到它FINANCIALRESEARCH.GOVFINANCIALRESEARCH.GOV。该算法在概念上假设每个人都全额支付，然后逐步减少任何可能破产的银行的支付，将该短缺重新分配给其债权人，并重复这一过程，直到实现一致性。结果是在尘埃落定后的一组违约和部分支付。这个过程明确显示了源自一个或多个节点的支付短缺如何在网络中传播，导致一系列不断扩大的违约FINANCIALRESEARCH.GOV。
Key Insights: One basic insight is that the network of obligations can transmit solvency shocks in a domino-like fashion. If a bank default occurs, its creditors take an immediate loss proportional to their exposure. If those losses exceed a creditor’s capital, that creditor will also default. Thus, a bad enough shock to one bank can theoretically cause a chain of defaults (a cascade) through direct interbank credit links. Eisenberg & Noe’s framework captures this direct default contagion rigorously. They assume all obligations have equal seniority (no complex priority structure), and ignore bankruptcy costs or fire-sale losses, focusing purely on the network of contracts itself as the propagation mechanism
. Under these assumptions, contagion is driven by
loss distribution: essentially, it’s a linear propagation of losses until no further defaults occur.
关键见解：一个基本的见解是，义务网络可以以多米诺骨牌的方式传递偿付能力冲击。如果发生银行违约，其债权人会立即遭受与其风险敞口成比例的损失。如果这些损失超过了债权人的资本，该债权人也将违约。因此，对一家银行的足够严重的冲击理论上可以通过直接的银行间信用链接引发一系列违约（级联）。艾森伯格和诺的框架严格捕捉了这种直接的违约传染。他们假设所有义务具有相等的优先级（没有复杂的优先结构），并忽略破产成本或抛售损失，纯粹关注合同网络本身作为传播机制。在这些假设下，传染是由损失分布驱动的：本质上，这是损失的线性传播，直到不再发生进一步的违约。
A valuable result from their paper is conditions under which the systemic cascade is limited vs. extensive. They note that if banks have enough capital relative to interbank exposures, a single default won’t spread. But if interbank liabilities are large relative to capital, one default can domino. Their framework also allows computing measures like the “impact” of a bank failure (how much total loss it causes in the system) and identifying systemically important nodes. It became the backbone for many later studies and for stress-testing exercises – regulators can apply the E&N algorithm to actual interbank exposure data to simulate what happens if a set of banks fail or suffer shocks. Indeed, this approach is now common in central bank stress tests and was influential in policy discussions on how interbank networks contribute to systemic risk
.
他们论文中的一个重要结果是系统级级联受限与广泛的条件。他们指出，如果银行相对于银行间敞口拥有足够的资本，单一违约不会扩散。但如果银行间负债相对于资本较大，一次违约可能会引发连锁反应。他们的框架还允许计算诸如银行破产的“影响”（它在系统中造成的总损失）以及识别系统重要节点等指标。这成为许多后续研究和压力测试工作的基础——监管机构可以将 E&N 算法应用于实际的银行间敞口数据，以模拟一组银行破产或遭受冲击时会发生的情况。实际上，这种方法现在在中央银行的压力测试中很常见，并在关于银行间网络如何导致系统性风险的政策讨论中产生了影响。
Extensions and Uses: The original E&N model has been extended to include things like bankruptcy costs (if a bank default imposes additional deadweight losses on itself or others), different seniorities (secured vs unsecured lending), and even to model confidence crises (where a change in perceived creditworthiness can trigger a cascade akin to an information-based run)
. For example, one can adapt the framework to a scenario of a
“market run”: if suddenly everyone believes Bank X might default, they might refuse to lend or pull funds, which can be interpreted as an exogenous shock to X’s funding and then analyzed with the same clearing mechanism. In essence, the E&N network provides a baseline for analyzing
pure contagion from counterparty default – often termed
direct contagion. It is complementary to models of liquidity runs or asset fire-sales.
扩展和应用：原始的 E&N 模型已经扩展，包括破产成本（如果银行违约对自身或他人造成额外的无谓损失）、不同的优先级（担保贷款与无担保贷款），甚至用于建模信心危机（当感知的信用 worthiness 变化时，可以触发类似于基于信息的挤兑的级联效应）FINANCIALRESEARCH.GOV。例如，可以将该框架适应于“市场挤兑”的情景：如果突然每个人都认为银行 X 可能违约，他们可能拒绝贷款或撤回资金，这可以被解释为对 X 资金的外生冲击，然后用相同的清算机制进行分析。从本质上讲，E&N 网络提供了一个分析来自对手方违约的纯传染的基线——通常称为直接传染。它与流动性挤兑或资产抛售模型是互补的。
Strengths: Eisenberg & Noe’s model is highly tractable and general. It applies to any sized network (in principle, dozens or hundreds of banks) since it’s computational – one can solve the fixed-point easily with iterative algorithms
. This was a big step forward from earlier small-analytical models. It provides a clear, rigorous way to talk about
systemic risk contribution: how much does one bank’s failure affect others. The framework’s influence is evident – it’s the foundation for many subsequent works (e.g. Acemoglu et al. 2015 build on a similar setup). Another strength is its
simplicity and focus: by abstracting away other complications, it isolates the network effect of interbank credit exposures. It also introduced useful concepts like the
“clearing vector”, now standard in systemic risk analysis. In practical terms, this model has been used on real data (when available) to map actual banking networks and simulate defaults (e.g. Furfine (1999) did an early application for US Fedwire network, and many have followed).
优势：Eisenberg & Noe 的模型具有高度的可处理性和普遍性。它适用于任何规模的网络（原则上，数十或数百家银行），因为它是计算性的——可以通过迭代算法轻松求解固定点FINANCIALRESEARCH.GOV。这是相较于早期的小型分析模型的一大进步。它提供了一种清晰、严谨的方式来讨论系统性风险贡献：一家银行的失败对其他银行的影响有多大。该框架的影响显而易见——它是许多后续工作的基础（例如，Acemoglu 等人 2015 年在类似的设置上进行了研究）。另一个优势是它的简单性和专注性：通过抽象其他复杂因素，它孤立了银行间信用风险暴露的网络效应。它还引入了“清算向量”等有用概念，这在系统性风险分析中已成为标准。从实际角度来看，该模型已在真实数据上（在可用时）用于映射实际银行网络并模拟违约（例如，Furfine（1999）对美国联邦电汇网络进行了早期应用，许多后续研究也遵循了这一方法）。
Weaknesses: The E&N model deliberately omits several important channels of contagion, which can be seen as limitations. It assumes no fire-sale or market price effects – creditors only lose money if a counterparty actually defaults and fails to pay. In reality, even the fear of counterparty trouble can cause market prices to drop (e.g. if Bank A is rumored to be in trouble, Bank B might preemptively sell assets or hoard cash, moving prices). By ignoring bankruptcy costs and liquidity, the model may underestimate the speed and severity of contagion in a real crisis (where panic and illiquidity play a huge role). Additionally, multiple equilibria can arise if one introduces slight non-linearities. Eisenberg & Noe’s neat result of a unique clearing solution relies on proportional repayment and monotonicity; if there are cyclic dependencies and bankruptcy costs, you could get more than one fixed point (a bad equilibrium where everyone freezes vs a good one where they don’t – essentially coordination failure)
. The baseline model cannot capture such
self-fulfilling crises because it’s entirely driven by exogenous shock and balance sheet identities. Another issue is data – the model is only as good as the matrix of interbank liabilities, which is often hard to obtain or incomplete in real-world cases (this has improved with some regulatory reporting, but still an obstacle). Finally, it’s a
static, one-period model: it doesn’t capture dynamics like how banks might react adaptively (e.g. deleveraging, or how a crisis unfolds over time). All banks either default or not in one go; there’s no notion of time except in the algorithm’s iterations. Despite these caveats, Eisenberg & Noe (2001) remains a cornerstone of network contagion modeling, and virtually every subsequent paper on interbank network risk cites it as a point of departure
. It provided the
“wiring diagram” approach to systemic risk, which post-2008 research has enriched with additional ingredients.
弱点：E&N 模型故意忽略了几个重要的传染渠道，这可以被视为局限性。它假设没有火灾销售或市场价格效应——债权人只有在对方实际违约并未支付时才会损失资金。实际上，即使是对对方麻烦的恐惧也会导致市场价格下跌（例如，如果 A 银行被传闻有麻烦，B 银行可能会预先出售资产或囤积现金，从而影响价格）。通过忽视破产成本和流动性，该模型可能低估了在真实危机中传染的速度和严重性（在危机中，恐慌和流动性不足起着巨大的作用）。此外，如果引入轻微的非线性，可能会出现多个均衡。艾森伯格和诺的独特清算解决方案依赖于比例偿还和单调性；如果存在循环依赖和破产成本，可能会出现多个固定点（一个是每个人都冻结的坏均衡，另一个是他们不冻结的好均衡——本质上是协调失败）。基线模型无法捕捉这种自我实现的危机，因为它完全由外生冲击和资产负债表身份驱动。 另一个问题是数据——模型的有效性取决于银行间负债的矩阵，而这在现实世界中往往难以获得或不完整（尽管一些监管报告有所改善，但仍然是一个障碍）。最后，这是一个静态的一期模型：它无法捕捉动态，例如银行可能如何适应性地反应（例如去杠杆化，或危机如何随时间展开）。所有银行要么一次性违约，要么不违约；除了算法的迭代外，没有时间的概念。尽管有这些警告，Eisenberg & Noe（2001）仍然是网络传染建模的基石，几乎所有后续关于银行间网络风险的论文都将其作为出发点FINANCIALRESEARCH.GOVFINANCIALRESEARCH.GOV。它提供了系统性风险的“接线图”方法，2008 年后的研究在此基础上增加了更多内容。
Post-2008 Advances: From Stylized Networks to Complex Systems2008 年后的进展：从简化网络到复杂系统
The global crisis of 2007–09 spurred an explosion of research on financial networks. With fresh evidence of how interconnectedness can propagate distress (e.g. through interbank markets, shadow banking, and common exposures), researchers sought to extend the earlier models and capture new phenomena observed in the crisis. We now discuss two highly influential post-crisis models – Elliott, Golub & Jackson (2014) and Acemoglu, Ozdaglar & Tahbaz-Salehi (2015) – as well as other notable contributions. These models build on foundations like Allen-Gale and Eisenberg-Noe but incorporate additional realism such as arbitrary network structures, distribution of shock sizes, and strategic considerations. A recurring theme is elaborating the trade-off between interconnectivity (integration) and diversification in determining systemic risk. We will see that these models formalize the idea that “density” of the network can be both good and bad, depending on circumstances, and they identify conditions for when the system is robust vs. fragile.2007-09 年的全球危机引发了对金融网络研究的爆炸性增长。随着新证据表明互联性如何传播困境（例如，通过银行间市场、影子银行和共同风险敞口），研究人员试图扩展早期模型，并捕捉在危机中观察到的新现象。我们现在讨论两个在危机后具有高度影响力的模型——Elliott, Golub & Jackson (2014) 和 Acemoglu, Ozdaglar & Tahbaz-Salehi (2015)——以及其他显著贡献。这些模型建立在 Allen-Gale 和 Eisenberg-Noe 等基础之上，但融入了额外的现实主义，例如任意网络结构、冲击规模的分布和战略考虑。一个反复出现的主题是阐述互联性（整合）与多样化之间的权衡，以确定系统性风险。我们将看到，这些模型形式化了“网络的密度”在不同情况下既可以是好事也可以是坏事的观点，并识别出系统稳健与脆弱的条件。
Elliott, Golub, and Jackson (2014) – Cascades of Failures and Network StructureElliott, Golub, 和 Jackson (2014) – 失败的级联与网络结构
Matthew Elliott, Ben Golub, and Matt Jackson developed a general model of cascades of failures in financial networks and analyzed how network structure impacts the scope of contagion. Published in the American Economic Review (2014), their paper “Financial Networks and Contagion” provides a unifying framework that can encompass various types of financial linkages (they mention it can handle scenarios like counterparty defaults, funding shocks, even cross-holdings of equity) and studies the conditions under which a small shock can lead to a large cascade of defaults. A distinctive feature is that they allow for discontinuous changes – i.e. a small change in asset values can suddenly trigger a wave of failures, highlighting the network’s nonlinear amplification.马修·艾略特、本·戈卢布和马特·杰克逊开发了一个关于金融网络中失败级联的通用模型，并分析了网络结构如何影响传染的范围。发表在《美国经济评论》（2014 年）的论文“金融网络与传染”提供了一个统一框架，可以涵盖各种类型的金融联系（他们提到它可以处理诸如对手方违约、资金冲击，甚至股权交叉持有等情景），并研究了在什么条件下小冲击可以导致大规模的违约级联。一个显著的特点是，他们允许不连续的变化——即资产价值的小变化可以突然触发一波失败，突显了网络的非线性放大效应。
Model Basics: They consider a network of financial institutions that have claims on each other. Each institution can fail if the value of its assets falls below a threshold (like its obligations). Importantly, they introduce two parameters to characterize connectivity: “integration” and “diversification”
. In their usage,
integration means how large the interbank exposures are relative to total assets (greater dependence on counterparties), and
diversification means how many counterparties each institution has (number of links). These two aspects capture the intuition that if a bank spreads its exposures across many counterparties (high diversification), each counterparty is a smaller part of its portfolio; if it has few counterparties (low diversification), each one is significant. Meanwhile, high integration means a bank relies heavily on other banks (a large fraction of its assets are claims on others, or equivalently a large fraction of its liabilities owed to others).
模型基础：他们考虑一个金融机构网络，这些机构之间相互有债权。每个机构如果其资产价值低于某个阈值（如其义务）就可能会失败。重要的是，他们引入了两个参数来表征连接性：“整合”和“多样化”AEAWEB.ORGAEAWEB.ORG。在他们的用法中，整合意味着银行间敞口相对于总资产的大小（对交易对手的依赖程度更高），而多样化意味着每个机构拥有多少个交易对手（链接数量）。这两个方面捕捉了这样的直觉：如果一家银行将其敞口分散在许多交易对手之间（高多样化），每个交易对手在其投资组合中所占的比例较小；如果它的交易对手很少（低多样化），每一个都是重要的。同时，高整合意味着一家银行在很大程度上依赖其他银行（其资产的很大一部分是对其他银行的债权，或者等价地，其负债的很大一部分是欠其他银行的）。
Key Findings: Elliott et al. show that integration and diversification have different non-monotonic effects on the extent of cascades
. Specifically, increasing diversification (adding more links per bank) has an ambiguous effect: when diversification is very low (few connections), the network may actually not be fully connected, so a shock might die out simply because it can’t reach many others. As diversification increases from a low level, the network becomes more connected,
permitting cascades to travel further – systemic risk
increases initially
. However, as diversification increases to high levels (each bank has many small exposures), each institution is
better insured against any single counterparty’s failure, so beyond a point, adding more links
reduces systemic risk
. This gives an inverted-U relationship: some intermediate level of connectivity can be worst for stability (enough links to transmit contagion, but not enough to dilute it fully). This result formalized earlier intuitions (like those from simulations by Nier et al. 2007, Gai & Kapadia 2010, and the concepts in Allen-Gale) in a clear theoretical way. It also
distinguishes financial contagion from epidemic contagion: in diseases, more connections
always spread infection more
, but in finance, beyond a threshold, more connections can start to help by sharing risk.
主要发现：Elliott 等人表明，整合和多样化对级联的程度具有不同的非单调效应。具体而言，增加多样化（每个银行增加更多链接）具有模糊的效果：当多样化非常低（连接较少）时，网络可能实际上并未完全连接，因此冲击可能会因为无法到达许多其他地方而消失。随着多样化从低水平增加，网络变得更加连接，允许级联传播得更远——系统性风险最初增加。然而，随着多样化增加到高水平（每个银行有许多小的风险敞口），每个机构对任何单一对手方的失败的保险程度更高，因此在某个点之后，增加更多链接会降低系统性风险。这形成了一个倒 U 型关系：某种中间水平的连接性可能对稳定性最为不利（足够的链接以传播传染，但不足以完全稀释）。这一结果形式化了早期的直觉（如 Nier 等人的模拟结果）。 2007 年，Gai 和 Kapadia 2010，以及 Allen-Gale 中的概念以清晰的理论方式进行阐述。它还区分了金融传染和流行病传染：在疾病中，更多的连接总是会传播感染，但在金融中，超过一个阈值后，更多的连接可以通过分担风险开始提供帮助。
Integration shows a somewhat different trade-off: higher integration means banks are more exposed to each other (so failures hit them harder), but it can also mean they rely less on their own risky projects (if a bank’s assets are largely claims on others, it might have less idiosyncratic risk). Elliott et al. find that increased integration also faces trade-offs – there’s a sweet spot where some interdependence helps (perhaps by mutual insurance) but too much makes everyone sink or swim together
. In their model, they can derive a condition for a
“small shock” to one bank to cause a cascade that wipes out a fraction $\alpha$ of the network. That condition relates to network connectivity and the distribution of weights. They define a measure of network
fragility (something like the largest eigenvalue of a matrix of influence, or a percolation threshold in graph terms). If the network is above a certain connectivity threshold and banks are sufficiently leveraged, then even a tiny shock can cause a big cascade (the network is fragile); below that threshold, even moderate shocks remain contained (the network is robust).
整合显示出一种略有不同的权衡：更高的整合意味着银行之间的相互暴露更大（因此失败对它们的打击更重），但这也可能意味着它们对自身风险项目的依赖减少（如果一家银行的资产主要是对其他银行的索赔，它可能会有更少的特有风险）。Elliott 等人发现，增加整合也面临权衡——存在一个最佳点，在这个点上某种相互依赖是有益的（可能通过相互保险），但过多的相互依赖会使每个人一起沉浮。在他们的模型中，他们可以推导出一个条件，使得对一家银行的“小冲击”导致级联效应，抹去网络中一部分$\alpha$。该条件与网络连通性和权重分布有关。他们定义了一种网络脆弱性的度量（类似于影响矩阵的最大特征值，或图论中的渗透阈值）。如果网络的连通性超过某个阈值且银行的杠杆率足够高，那么即使是微小的冲击也会导致大规模的级联（网络是脆弱的）；低于该阈值，即使是适度的冲击也会保持在可控范围内（网络是稳健的）。
One intuitive illustration they give is: suppose each bank has $n$ counterparties (diversification) and keeps a fraction $\phi$ of its balance sheet in claims on each counterparty. If one bank fails, it causes a $\phi$ loss to each of its $n$ counterparties. Those counterparties will fail if $\phi$ exceeds their capital buffers, etc. So if $n$ is too low, maybe the failed bank only hits a few others but hard (possibly failing them). If $n$ is very high, it hits many others but each very lightly (they survive). At intermediate $n$, you hit many enough others and each loss is sizable enough to topple them – that’s the dangerous region. They formalize this kind of reasoning.他们给出的一个直观例子是：假设每家银行有 $n$ 个对手方（多样化），并且在每个对手方的债权上保持其资产负债表的一部分 $\phi$。如果一家银行倒闭，它会对每个 $n$ 个对手方造成 $\phi$ 的损失。如果 $\phi$ 超过了这些对手方的资本缓冲，他们将会倒闭。因此，如果 $n$ 太低，也许倒闭的银行只会对少数其他银行造成严重影响（可能导致它们倒闭）。如果 $n$ 非常高，它会对许多其他银行造成轻微影响（它们幸存）。在中间的 $n$ 值时，你会对足够多的其他银行造成影响，并且每个损失都足够大以使它们倒闭——这就是危险区域。他们将这种推理形式化。
They also examine core-periphery networks (a few highly connected core banks and many peripheral banks with fewer connections). Interestingly, they note that a core-periphery structure can sometimes amplify cascades, because a failure in the core (especially a large core node) can take down the other core nodes and then the periphery – essentially the worst-case scenario
. Large core banks tend to be robust to small shocks (they have many partners to share loss with), but if a
large shock hits a core bank, it can be catastrophic
. This resonates with what was witnessed in 2008: core institutions (like large investment banks) were seemingly fine with small daily fluctuations but were felled by an outsized shock (the subprime collapse), after which the whole core was at risk.
他们还研究了核心-边缘网络（少数高度连接的核心银行和许多连接较少的边缘银行）。有趣的是，他们指出，核心-边缘结构有时会放大级联效应，因为核心中的失败（尤其是一个大型核心节点）可能会导致其他核心节点和边缘节点的崩溃——本质上是最坏的情况ARXIV.ORGARXIV.ORG。大型核心银行通常对小冲击具有韧性（它们有许多合作伙伴来分担损失），但如果大型冲击袭击核心银行，后果可能是灾难性的ARXIV.ORGARXIV.ORG。这与 2008 年所见的情况相呼应：核心机构（如大型投资银行）在小的日常波动中似乎表现良好，但却被一个过大的冲击（次贷崩溃）击垮，之后整个核心都面临风险。
Strengths: The Elliott, Golub, Jackson model is very flexible. It can encompass different kinds of networks (they even discuss equity cross-ownership as an application). It yields clear, economically interpretable metrics – integration and diversification – that policymakers and researchers can think about when assessing systemic risk
. The result that diversification can be bad at first but good later succinctly captures the
robust-yet-fragile nature of financial networks. The model also provides insights into how targeted interventions might stop cascades (e.g. shoring up a particular bank might break the chain). Moreover, the paper bridges network theory and economics, borrowing from percolation theory (they cite results analogous to those in random graph literature)
. It thereby helped bring the tools of network science into the economic discussion on systemic risk.
优势：Elliott, Golub, Jackson 模型非常灵活。它可以涵盖不同类型的网络（他们甚至讨论了股权交叉持有作为一种应用）。它产生了清晰、经济可解释的指标——整合和多样化——政策制定者和研究人员在评估系统性风险时可以考虑这些指标AEAWEB.ORGAEAWEB.ORG。多样化在初期可能是有害的，但后来可能是有益的这一结果简洁地捕捉了金融网络的稳健而脆弱的特性。该模型还提供了关于如何通过有针对性的干预来阻止级联的见解（例如，支持某个特定银行可能会打破链条）。此外，本文将网络理论与经济学结合起来，借鉴了渗流理论（他们引用了与随机图文献中类似的结果）ARXIV.ORGARXIV.ORG。由此，它帮助将网络科学的工具引入了关于系统性风险的经济讨论。
Weaknesses: One limitation is that the model, to get clear results, makes some simplifying assumptions. For example, it uses a kind of threshold model of default (if assets < liabilities, fail completely) rather than modeling partial losses or continuous responses. In reality, banks might experience distress without full default, and might recapitalize or get rescued, etc. The binary nature of “fail or not” simplifies the cascade analysis but may overstate how abrupt real failures are. Additionally, while the model allows arbitrary network graph structures in principle, the analytical results are most crisp either in highly symmetric cases or asymptotic random networks. Like many network models, they often assume a random network topology or regular structure to derive general insights, which might not capture all the nuances of a specific real network (which could have specific patterns). That said, they do explore core-periphery and such. Another challenge is the assumption that when a node fails, it completely defaults on its obligations (zero recovery). In practice there might be partial recovery, bankruptcy processes, etc., which could dampen or delay the cascade – the model’s worst-case cascade might be somewhat overstated (they do mention one could incorporate fractional recovery, but at cost of simplicity). Finally, their model, like E&N, largely treats the network as exogenous and does not incorporate dynamics of behavior during the crisis – e.g. no one is selling assets or hoarding liquidity in response, it’s a one-shot cascade. Despite these caveats, the paper’s conceptual clarity and generality have made it a benchmark in the literature on network contagion. It is frequently cited alongside Acemoglu et al. (2015) as the leading theoretical work of the post-2008 era on systemic risk.弱点：一个限制是该模型为了获得清晰的结果，做了一些简化假设。例如，它使用了一种违约的阈值模型（如果资产 < 负债，则完全失败），而不是建模部分损失或连续反应。实际上，银行可能会经历困境而不完全违约，并可能进行再资本化或获得救助等。“失败与否”的二元性质简化了级联分析，但可能夸大了真实失败的突然性。此外，虽然该模型原则上允许任意网络图结构，但分析结果在高度对称的情况下或渐近随机网络中最为清晰。像许多网络模型一样，它们通常假设随机网络拓扑或规则结构以推导一般见解，这可能无法捕捉特定真实网络的所有细微差别（可能具有特定模式）。尽管如此，它们确实探讨了核心-边缘等问题。另一个挑战是假设当一个节点失败时，它完全违约其义务（零回收）。 在实践中，可能会出现部分恢复、破产程序等，这可能会抑制或延迟级联——模型的最坏情况级联可能有些被夸大（他们确实提到可以纳入部分恢复，但代价是复杂性）。最后，他们的模型，像 E&N 一样，主要将网络视为外生的，并未纳入危机期间行为的动态——例如，没有人因应对而出售资产或囤积流动性，这是一种一次性级联。尽管有这些警告，本文的概念清晰性和普遍性使其成为网络传染文献中的基准。它常与 Acemoglu 等人（2015）一起被引用，作为 2008 年后系统性风险的主要理论著作。
Acemoglu, Ozdaglar, and Tahbaz-Salehi (2015) – “Robust-Yet-Fragile” Phase Transition in Network ConnectivityAcemoglu, Ozdaglar, 和 Tahbaz-Salehi (2015) – 网络连通性中的“稳健但脆弱”相变
Another major contribution is by Daron Acemoglu and coauthors, published in American Economic Review 2015, titled “Systemic Risk and Stability in Financial Networks.” This model builds directly on the Eisenberg-Noe type network (with obligations and a clearing mechanism) but introduces randomness and asymmetry to study how the architecture of a financial network affects the likelihood of systemic cascades. Their headline result is that there is a form of phase transition in contagion as the network becomes more connected
. In sparse networks, contagion is limited; as interconnections increase, the system at first becomes safer (because risks are shared), but beyond a certain point, adding more links suddenly makes the system much more fragile to large shocks. This formalizes the idea that a financial network can be
“robust-yet-fragile” – robust to small shocks, yet fragile to big ones – depending on the degree of connectivity.
另一个重要贡献是由达龙·阿西莫格鲁及其合著者发表在《美国经济评论》2015 年的文章，标题为“金融网络中的系统性风险与稳定性”。该模型直接基于艾森伯格-诺伊类型的网络（具有义务和清算机制），但引入了随机性和不对称性，以研究金融网络的架构如何影响系统性级联的可能性。他们的主要结果是，随着网络连接性增加，传染性存在一种相变形式。在稀疏网络中，传染性受到限制；随着互联互通的增加，系统最初变得更安全（因为风险被分担），但超过某个点后，增加更多链接会突然使系统对大冲击变得更加脆弱。这一理论化了金融网络可以是“强韧但脆弱”的观点——对小冲击强韧，但对大冲击脆弱——这取决于连接程度。
Model Setup: They consider $n$ banks with a random network of interbank lending. Each bank has some probability of being the initial “distressed” bank (shock hits there). A key innovation is distinguishing between small shocks and large shocks. A small shock is one that by itself would not bankrupt a bank (if it had no connections, it could absorb the loss), whereas a large shock is one that would overwhelm the bank’s capital. They show that when shocks are small, more complete interbank connections help stability – because the shock is distributed among many counterparties, each of whom can handle a little loss
. Essentially, interbank links act as “risk-sharing” for small shocks (just like Allen-Gale argued). However, when shocks are large enough to cause insolvency, then those interbank links turn into
channels of contagion that can take down others. Beyond a certain connectivity, a large shock at one bank will propagate to many others and cause a systemic crisis
.
模型设置：他们考虑$n$家银行，具有随机的银行间借贷网络。每家银行都有一定的概率成为最初的“困境”银行（冲击发生在这里）。一个关键的创新是区分小冲击和大冲击。小冲击是指单独不会使银行破产的冲击（如果没有连接，它可以吸收损失），而大冲击是指会压垮银行资本的冲击。他们表明，当冲击较小时，更完整的银行间连接有助于稳定——因为冲击在许多对手方之间分散，每个对手方都可以承受一些损失NBER.ORGNBER.ORG。实质上，银行间链接在小冲击中充当“风险分担”（就像 Allen-Gale 所论证的那样）。然而，当冲击足够大以导致破产时，这些银行间链接就会变成传播渠道，可能会使其他银行倒闭。在某种连接性阈值之上，一家银行的大冲击将传播到许多其他银行，并导致系统性危机NBER.ORGNBER.ORG。
They demonstrate that as the number of counterparties per bank crosses a threshold (in their analysis, the threshold relates to something like the per-bank degree in the network relative to capital), the probability of a full network cascade jumps from near zero to significant – akin to how in percolation theory, there is a critical point where a giant connected cluster emerges. In less technical terms, when the network is “not too connected,” each bank failure remains an isolated incident or causes only a few others to fail; but when the network is “very connected,” a single big failure can lead to a majority of banks failing. This result provided a theoretical underpinning for policymakers’ concern that an increasingly interconnected banking system can suddenly become very fragile if an out-of-bound shock hits (even though in normal times those connections seem to make it safer by distributing risk). The authors phrase it as: “financial contagion exhibits a form of phase transition as interbank connections increase” – with connectivity, stability improves up to a point, then beyond that point, system stability worsens dramatically
.
他们展示了，当每个银行的对手方数量超过一个阈值时（在他们的分析中，阈值与网络中每个银行的度数相对于资本的关系有关），整个网络级联的概率从接近零跃升到显著——这类似于渗流理论中的临界点，在该点上出现了一个巨大的连通簇。用不那么技术性的术语来说，当网络“连接不太紧密”时，每个银行的失败仍然是孤立事件，或者只导致少数其他银行失败；但当网络“连接非常紧密”时，单个重大失败可能导致大多数银行失败。这个结果为政策制定者对日益相互关联的银行系统在遭遇超出预期的冲击时可能突然变得非常脆弱的担忧提供了理论基础（尽管在正常情况下，这些连接似乎通过分散风险使其更安全）。作者将其表述为：“金融传染在银行间连接增加时表现出一种相变形式”——随着连接性的增加，稳定性在某个点之前改善，但超过该点后，系统稳定性急剧恶化。
They also touch on an interesting inefficiency result: the network that banks would form endogenously through bilateral lending may not be socially optimal
. Banks, when deciding on whom to lend to, consider the effect on their immediate counterparties, but
don’t internalize the externality that their lending has on the overall network’s fragility
. This means the equilibrium network (without regulation) could be either too connected or not connected enough from a social standpoint. In their model, they find tendencies toward
excessive connectivity that increases systemic risk – a rationale for macroprudential regulation to limit certain interbank exposures.
他们还提到了一个有趣的低效结果：银行通过双边借贷内生形成的网络可能不是社会最优的NBER.ORGNBER.ORG。银行在决定借给谁时，会考虑对其直接交易对手的影响，但并没有内化其借贷对整体网络脆弱性的外部性NBER.ORG。这意味着在没有监管的情况下，均衡网络可能在社会角度上要么连接过多，要么连接不足。在他们的模型中，他们发现过度连接的倾向增加了系统性风险——这是限制某些银行间风险敞口的宏观审慎监管的理由。
Strengths: Acemoglu et al. provide a clear articulation of the robust-vs-fragile connectivity trade-off with rigorous probability statements. Their result is often cited as explaining why prior to 2008, the system seemed stable (only small shocks occurred and those were absorbed thanks to connectivity), but when a big shock hit (mass mortgage defaults), the very interconnectedness that had been beneficial turned into a disaster amplifier. The use of a random network and probabilistic shock size allows them to derive a threshold condition (like a “tipping point”) which is very valuable conceptually
. It moves beyond deterministic worst-case scenarios to statements about likelihood of systemic failure. They also incorporate the
Eisenberg-Noe clearing mechanism in their model, so it is grounded in that realistic payment dynamics (banks only default after depleting capital, etc.). Another strength is the insight about
endogenous network formation and network externalities: it highlights that individual institutions won’t account for the systemic impact of their connections, justifying regulatory involvement
. This bridges network theory with incentive analysis.
优势：Acemoglu 等人清晰地阐述了强韧性与脆弱性连接的权衡，并提供了严格的概率陈述。他们的结果常被引用来解释为什么在 2008 年之前，系统似乎是稳定的（只发生了小冲击，并且由于连接性得以吸收），但当大冲击发生时（大规模抵押贷款违约），曾经有利的高度互联性却变成了灾难的放大器。使用随机网络和概率冲击大小使他们能够推导出一个阈值条件（类似于“临界点”），这一概念在理论上非常有价值NBER.ORGNBER.ORG。它超越了确定性的最坏情况场景，转向关于系统性失败可能性的陈述。他们还在模型中纳入了 Eisenberg-Noe 清算机制，因此它基于现实的支付动态（银行只有在耗尽资本后才会违约等）。另一个优势是关于内生网络形成和网络外部性的洞察：它强调个别机构不会考虑其连接的系统性影响，从而证明了监管介入的必要性NBER.ORG。这将网络理论与激励分析联系起来。
Weaknesses: To get the sharp phase transition result, the model relies on symmetry and randomness (e.g. each bank might have the same number of links on average, shocks drawn from certain distributions). Real networks are not purely random; they have structure (core-periphery, large vs small banks, etc.), which can change the nature of contagion. For example, if one bank is much larger, the “phase transition” idea might be less relevant than simply that one hub’s failure is devastating. The model can handle heterogeneity in principle, but the neat analytic results are in a homogeneous setting. Another limitation is the focus on a single initial shock. In reality, multiple institutions can be hit by a common shock (like a broad asset price drop) – though their model could handle that by thinking of a “large shock affecting many banks,” it becomes closer to the common exposure channel rather than pure network contagion. Additionally, similar to others, they don’t model panicky behavior or time dynamics: it’s comparative statics of equilibria. So it doesn’t capture, for instance, a situation where the anticipation of contagion leads banks to pull back lending preemptively (which could either mitigate or worsen the crisis). It’s basically “ex-post” analysis: given a shock, who ends up defaulting after the clearing process. Also, while the inefficiency result is insightful, the model of network formation itself is not fully fleshed out – they mostly argue it in words or a stylized way. So, it’s not a full game-theoretic model of link formation (other papers by e.g. Babus or Cabrales have tackled that). Finally, calibrating such a model to real data is challenging – it’s more qualitative. Nonetheless, the phase transition metaphor (“robust-yet-fragile”) from Acemoglu et al. (2015) has become almost a tagline in the systemic risk literature and is supported by other studies and simulations
.
弱点：为了获得明显的相变结果，该模型依赖于对称性和随机性（例如，每个银行的平均链接数量可能相同，冲击来自某些分布）。真实网络并非完全随机；它们具有结构（核心-边缘、大型与小型银行等），这可能改变传染的性质。例如，如果一家银行的规模远大于其他银行，“相变”概念可能不如简单地认为一个中心的失败是毁灭性的更为相关。该模型原则上可以处理异质性，但整洁的解析结果是在同质环境中。另一个限制是关注单一初始冲击。实际上，多个机构可能会受到共同冲击（如广泛的资产价格下跌）——尽管他们的模型可以通过考虑“影响许多银行的大冲击”来处理这一点，但这更接近于共同暴露渠道而非纯粹的网络传染。此外，类似于其他模型，他们没有建模恐慌行为或时间动态：这只是均衡的比较静态。 因此，它并没有捕捉到，例如，预期传染导致银行提前收缩贷款的情况（这可能会减轻或加剧危机）。这基本上是“事后”分析：在经历冲击后，谁在清算过程中最终违约。此外，尽管低效结果很有见地，但网络形成模型本身并没有完全展开——他们主要是用语言或一种风格化的方式进行论述。因此，这并不是一个完整的链接形成博弈论模型（例如，Babus 或 Cabrales 的其他论文对此进行了探讨）。最后，将这样的模型校准到真实数据上是具有挑战性的——它更具定性。尽管如此，Acemoglu 等人（2015）的相变隐喻（“稳健但脆弱”）几乎已成为系统性风险文献中的一个标语，并得到了其他研究和模拟的支持。
Other Notable Models and Extensions其他显著模型和扩展
In addition to the above four, the literature contains many other models that enrich our understanding of financial network crises:除了上述四个，文献中还有许多其他模型丰富了我们对金融网络危机的理解：
-
Gai and Kapadia (2010, 2011) – These authors, including Bank of England’s Andrew Haldane in the 2011 paper, use simulative and analytic models of random networks to examine contagion probability and impact. They find results consistent with the non-monotonic connectivity story: as the connectivity (average degree) of the interbank network increases, the system initially becomes more resilient, but beyond a threshold, the probability of a large cascade increases sharply (a finding akin to a percolation threshold)
. Gai et al. also emphasize the role of bank size heterogeneity and concentration. Their simulations show that contagion is more severe when the network has highly concentrated exposures – e.g. if a few banks hold a large share of interbank assets, their failure triggers bigger cascades. These papers used a simple insolvency cascade model and were among the first post-2008 works to draw regulators’ attention – indeed, their approach fed into regulatory stress tests (they illustrate how one can plug in network data and run scenarios).Gai 和 Kapadia（2010, 2011）——这些作者，包括英格兰银行的 Andrew Haldane 在 2011 年的论文中，使用随机网络的模拟和分析模型来研究传染概率和影响。他们发现的结果与非单调连接性理论一致：随着银行间网络的连接性（平均度）增加，系统最初变得更加有韧性，但超过某个阈值后，大规模级联的概率急剧增加（这一发现类似于渗透阈值）ARXIV.ORGARXIV.ORG。Gai 等人还强调了银行规模异质性和集中度的作用。他们的模拟显示，当网络具有高度集中暴露时，传染更加严重——例如，如果少数银行持有大量的银行间资产，它们的失败会引发更大的级联ARXIV.ORGARXIV.ORG。这些论文使用了一个简单的破产级联模型，是 2008 年后最早引起监管者注意的研究之一——实际上，他们的方法为监管压力测试提供了依据（他们展示了如何插入网络数据并运行场景）。
-
Cifuentes, Ferrucci, and Shin (2005) – This earlier (pre-crisis) model is often cited for incorporating asset price contagion via fire-sales. They extend an E&N-type network with an important twist: when banks are under pressure to meet capital requirements, they may sell assets, which can depress prices if the market lacks depth. Those lower prices then force other banks to mark down their assets, possibly breaching their capital constraints, causing further sales – a spiral. This mechanism can generate contagion even if banks have no direct lending links; it’s an indirect network effect through common asset markets. Cifuentes et al. show that adding even a few risk-neutral buyers (who can absorb assets) greatly improves stability, underscoring the importance of market liquidity in crises. Their work complemented the network literature by showing that “balance-sheet contagiousness” is not just who owes whom, but also who holds what. In the 2007–09 crisis, this effect was evident in how losses in mortgage securities led to fire sales affecting all banks holding similar paper. Recent network models have started combining node-to-node links and common asset links (a multi-layer network) to capture both channels.Cifuentes, Ferrucci, 和 Shin (2005) – 这个早期（危机前）模型常被引用，因为它通过火售引入了资产价格传染。他们扩展了一个 E&N 类型的网络，并加入了一个重要的转折：当银行面临资本要求的压力时，它们可能会出售资产，如果市场缺乏深度，这可能会压低价格。这些较低的价格迫使其他银行降低其资产的账面价值，可能会突破其资本约束，导致进一步的销售——形成螺旋。这一机制可以产生传染，即使银行之间没有直接的贷款联系；这是一种通过共同资产市场的间接网络效应。Cifuentes 等人表明，即使增加少量风险中性买家（可以吸收资产）也能大大提高稳定性，强调了市场流动性在危机中的重要性。他们的研究补充了网络文献，表明“资产负债表传染性”不仅仅是谁欠谁，还包括谁持有什么。在 2007-09 年危机中，这一效应在抵押证券损失如何导致火售影响所有持有类似证券的银行中显而易见。 最近的网络模型开始结合节点间链接和共同资产链接（多层网络），以捕捉这两种渠道。
-
Haldane and May (2011) – In a notable Nature article, Andrew Haldane (a policymaker) and Robert May (an ecologist) drew an analogy between financial networks and ecological or engineered networks. They argued that complexity (interconnectedness) can increase fragility – borrowing the concept of a robust-yet-fragile network from ecology. While not a formal model, their work helped popularize the network perspective among regulators. They warned that the banking network had become too complex, such that a failure could propagate like an epidemic. One specific idea was that tightly coupled systems have little room for error – a shock anywhere is felt everywhere quickly, leaving no time for reaction. This echoed what models like Acemoglu et al. would formalize. They also suggested that measures of node centrality (like degree or betweenness) from network theory might serve as proxies for a bank’s systemic importance. This spurred research into “systemic risk metrics” (like DebtRank, introduced by Battiston et al. 2012, which is essentially an iterative centrality measure of impact).Haldane 和 May（2011）——在一篇引人注目的《自然》文章中，政策制定者安德鲁·哈尔丹和生态学家罗伯特·梅之间建立了金融网络与生态或工程网络之间的类比。他们认为复杂性（相互关联性）可能会增加脆弱性——借用生态学中强韧但脆弱网络的概念。虽然这不是一个正式模型，但他们的工作帮助在监管者中普及了网络视角。他们警告说，银行网络变得过于复杂，以至于故障可能像流行病一样传播。一个具体的想法是，紧密耦合的系统几乎没有容错空间——任何地方的冲击都会迅速在各处感受到，留给反应的时间几乎没有。这与像 Acemoglu 等人将要形式化的模型相呼应。他们还建议，网络理论中的节点中心性度量（如度数或介数）可能作为银行系统重要性的代理。这激发了对“系统性风险指标”的研究（如 DebtRank，由 Battiston 等人于 2012 年提出，基本上是一种影响的迭代中心性度量）。
-
Common Exposure and Sector Network Models: Some research models banks connected not by direct loans, but by correlated portfolios. For instance, Cont and Wagalath (2013) model stress propagation when banks mark to market common assets – effectively forming a network where nodes are linked if they share asset holdings. This strand finds that high overlap in portfolios can create contagion even if direct interbank links are low. In such models, diversification at the individual level can mean concentration at the system level (since everyone holds a bit of everything, if that “everything” crashes, everyone is hit). So there is a network of overlapping portfolios to consider. Empirically, this has become important in analyzing, say, pension funds or insurers who don’t directly lend to each other but might all be long the same bonds.共同暴露和行业网络模型：一些研究模型表明，银行之间的联系并不是通过直接贷款，而是通过相关的投资组合。例如，Cont 和 Wagalath（2013）模型在银行按市值计量共同资产时的压力传播——有效地形成一个网络，其中节点如果共享资产持有则相互连接。这一研究发现，投资组合的高度重叠可以在直接的银行间联系较低的情况下产生传染效应。在这样的模型中，个体层面的多样化可能意味着系统层面的集中（因为每个人都持有一点所有东西，如果那“所有东西”崩溃，所有人都会受到影响）。因此，需要考虑一个重叠投资组合的网络。从经验上看，这在分析例如养老金基金或保险公司时变得重要，这些机构并不直接相互借贷，但可能都持有相同的债券。
-
Endogenous Network Formation: A few models try to endogenize how the network links form under various incentives, and then see what structures emerge. For example, Babus (2016) explores a setting where banks form interbank links for risk-sharing but face a trade-off with contagion risk; she finds that without coordination, the network formed might cluster too much, and adding a “central clearing counterparty” can help by acting as a hub that reduces contagious links. Another by Bluhm, Krahnen & Leippold (2018) consider how banks might choose counterparties based on credit risk and show that safer banks tend to link with each other (assortative pattern), potentially leaving risky banks more isolated but also more vulnerable if hit (no one to save them). While these are beyond our primary scope, they are notable as they address the strategic architecture of networks, not just exogenous structures.内生网络形成：一些模型试图内生化网络链接在各种激励下是如何形成的，然后观察出现什么结构。例如，Babus（2016）探讨了一个银行为了风险共享而形成银行间链接的环境，但面临着传染风险的权衡；她发现如果没有协调，形成的网络可能会过于聚集，增加一个“中央清算对手方”可以通过充当一个减少传染链接的中心来提供帮助。Bluhm、Krahnen 和 Leippold（2018）则考虑了银行如何根据信用风险选择对手方，并显示出更安全的银行倾向于相互链接（同类联系模式），这可能使得风险较高的银行更加孤立，但如果遭受冲击时也更脆弱（没有人来拯救他们）。虽然这些超出了我们的主要范围，但它们值得注意，因为它们涉及网络的战略架构，而不仅仅是外生结构。
Overall, the post-2008 literature has greatly expanded the scenarios and mechanisms considered. But across most models, a few common insights stand out: (1) Network structure critically shapes systemic outcomes – dense networks are a double-edged sword, and certain structures (like highly centralized networks) concentrate risk. (2) Nonlinearities and tipping points are ubiquitous – systems can appear stable under moderate conditions and then suddenly collapse when conditions breach a threshold (like leverage too high, connectivity too high, or asset values too low). (3) Heterogeneity matters – differences in bank size, connectivity, and asset composition can both exacerbate and mitigate contagion (e.g., a network of identical banks behaves differently from one with a few giants – often the latter is less frequent crises, but more severe when they happen). (4) Wider channels beyond direct defaults (like runs and fire sales) need to be integrated for a full picture. Many current studies are working on merging these layers into unified frameworks.总体而言，2008 年后的文献大大扩展了考虑的情景和机制。但在大多数模型中，有几个共同的见解脱颖而出：（1）网络结构对系统结果具有关键影响——密集网络是把双刃剑，某些结构（如高度集中化的网络）会集中风险。（2）非线性和临界点无处不在——系统在适度条件下看似稳定，但当条件突破阈值（如杠杆过高、连接性过高或资产价值过低）时，可能会突然崩溃。（3）异质性很重要——银行规模、连接性和资产组成的差异既可以加剧也可以缓解传染（例如，一个由相同银行组成的网络与一个有几个巨头的网络表现不同——通常后者发生危机的频率较低，但一旦发生则更为严重）。（4）需要整合直接违约以外的更广泛渠道（如挤兑和火灾销售）以获得全面的视角。许多当前的研究正在致力于将这些层次合并为统一的框架。
Comparative Strengths & Weaknesses Summary: To wrap up this overview of models, it’s useful to tabulate the strengths and weaknesses of the key approaches:比较优势与劣势总结：为了总结这些模型的概述，列出主要方法的优势和劣势是有用的：
- Allen & Gale (2000): Rich microeconomic foundation (liquidity preference, optimal contracts) and shows topology matters; but only handles small networks and specific shock setups, not directly quantitative for large systems.Allen & Gale (2000)：丰富的微观经济基础（流动性偏好，最优合同）并显示拓扑结构的重要性；但仅处理小型网络和特定冲击设置，未直接对大型系统进行定量分析。
- Eisenberg & Noe (2001): Elegant algorithm for clearing and widely applicable to real networks; but focuses solely on direct default contagion, ignoring market dynamics and multiple equilibria issues (unless adapted).Eisenberg & Noe (2001)：优雅的清算算法，广泛适用于真实网络；但仅关注直接违约传染，忽略市场动态和多重均衡问题（除非进行调整）。
- Elliott et al. (2014): General network cascade model capturing integration vs diversification trade-offs, giving clear intuition for non-monotonic effects; however, uses a stylized failure mechanism (threshold default) and doesn’t model time or strategic behavior.Elliott et al. (2014)：一般网络级联模型捕捉整合与多样化的权衡，清晰地直观展示非单调效应；然而，使用了一个简化的失败机制（阈值违约），并未建模时间或战略行为。
- Acemoglu et al. (2015): Bridges network structure with probabilistic systemic risk, yielding a clear “phase transition” insight and incorporating the clearing payment logic; but relies on symmetric random networks for analytical tractability and does not include price-mediated contagion.Acemoglu et al. (2015)：桥接网络结构与概率系统风险，提供了清晰的“相变”洞察，并纳入了清算支付逻辑；但依赖于对称随机网络以便于分析，并未包括价格中介的传染效应。
- Other models: Gai-Kapadia and simulations add realism of distribution and heterogeneity; fire-sale models (Shin 2005, etc.) add another channel of contagion; these often require numerical analysis and can be sensitive to assumptions about market liquidity.其他模型：Gai-Kapadia 和模拟增加了分布和异质性的现实性；火售模型（Shin 2005 等）增加了另一种传播渠道；这些通常需要数值分析，并且可能对市场流动性假设敏感。
No single model captures everything – each focuses on particular channels (liquidity vs solvency, direct links vs common assets, exogenous vs endogenous network). This is an active area, and ongoing work is trying to combine channels (for instance, recent papers examine how **“cycles” in networks – interdependent obligations in loops – can lead to multiple equilibria and coordination failures
, bringing bank-run-like indeterminacy into the Eisenberg-Noe world). In the next section, we discuss one such important aspect that is not well covered by the above models: the role of
information and beliefs in triggering financial crises, often referred to as
information sensitivity of financial instruments, and how one might integrate that into network models.
没有单一模型能够捕捉所有内容——每个模型关注特定的渠道（流动性与偿付能力、直接联系与共同资产、外生网络与内生网络）。这是一个活跃的研究领域，正在进行的工作试图结合这些渠道（例如，最近的论文考察了网络中的“循环”——循环中的相互依赖义务——如何导致多重均衡和协调失败ARXIV.ORGARXIV.ORG，将银行挤兑般的不确定性引入艾森伯格-诺伊世界）。在下一节中，我们将讨论一个上述模型未能很好覆盖的重要方面：信息和信念在触发金融危机中的作用，通常被称为金融工具的信息敏感性，以及如何将其整合到网络模型中。
The Role of “Information Sensitivity” in Financial Crises
One lesson from the 2007–09 crisis is that loss of confidence and sudden demands for information can precipitate a crisis, even in the absence of clear fundamental triggers. Gary Gorton and others have argued that a defining feature of crises is a shift in the information environment of financial markets – previously accepted “safe” assets become questioned, leading to runs. This idea is captured by the concept of information sensitivity of debt. In normal times, certain short-term debts (like overnight repo, commercial paper, interbank deposits) are treated as essentially risk-free and “information-insensitive,” meaning investors/lenders see no need to scrutinize the underlying collateral or counterparty too closely
. The debt is safe enough that nobody spends resources to privately evaluate its value; it’s almost like cash. However, in a crisis, that changes – if people suspect the collateral or solvency behind a debt might be impaired, the debt becomes
“information-sensitive”
. Suddenly everyone wants to know “who is exposed to what?” and “how bad are the losses?”, but obtaining that information is hard and slow. The result is often a panic or freeze: lenders withdraw funding or demand higher haircuts because they fear the worst, and without time or ability to prove otherwise, a self-fulfilling loss of liquidity occurs.
2007-09 年危机的一个教训是，信心的丧失和突如其来的信息需求可以引发危机，即使没有明确的基本触发因素。加里·戈顿和其他人认为，危机的一个定义特征是金融市场信息环境的变化——以前被认为“安全”的资产开始受到质疑，从而导致挤兑。这个想法通过债务的信息敏感性概念得以体现。在正常时期，某些短期债务（如隔夜回购、商业票据、银行间存款）被视为基本上无风险和“信息不敏感”，这意味着投资者/贷款人认为没有必要过于仔细地审查基础抵押品或交易对手。债务足够安全，以至于没有人花费资源私下评估其价值；这几乎就像现金。 然而，在危机中，这种情况会发生变化——如果人们怀疑债务背后的抵押品或偿付能力可能受到损害，债务就变得“信息敏感”。突然间，每个人都想知道“谁暴露于什么风险？”和“损失有多严重？”，但获取这些信息既困难又缓慢。结果往往是恐慌或冻结：贷方撤回资金或要求更高的折扣，因为他们害怕最坏的情况，而在没有时间或能力证明相反的情况下，流动性自我实现的损失就发生了。
Dang, Gorton, and Holmström (2015, 2020) formalize this information view of financial crises
. They argue that the financial system produces lots of short-term debt precisely because it’s normally information-insensitive (which is good for trading and liquidity – no haggling over value). But
when the value of the backing collateral falls enough to raise doubts, some agents have an incentive to acquire private information (e.g. a hedge fund might dig into which mortgage tranches are bad)
. This frays the common knowledge that the debt is safe. Traders then worry that others might know something they don’t (adverse selection), so they pull back – leading to a classic lemons market problem or “run”. In Gorton’s famous analogy, the system is like
“nobody worries about the dry powder keg until someone smells smoke” – then everyone runs, because the once-ignored risk suddenly becomes very relevant. In the 2007–08 crisis, for example, repo lenders stopped rolling over loans because they became uncertain about the value of mortgage-backed collateral and suspected other parties might know more about that value. This caused a run on the shadow banking system.
Dang、Gorton 和 Holmström（2015、2020）正式化了这种关于金融危机的信息视角ECONOMICS.MIT.EDUANNUALREVIEWS.ORG。他们认为，金融系统产生大量短期债务正是因为它通常对信息不敏感（这对交易和流动性是有利的——没有对价值的讨价还价）。但是，当支持抵押品的价值下降到足以引发怀疑时，一些参与者就有动机获取私人信息（例如，某个对冲基金可能会深入研究哪些抵押贷款分层是有问题的）ANNUALREVIEWS.ORGANNUALREVIEWS.ORG。这削弱了债务安全的共同知识。交易者随后担心其他人可能知道他们不知道的事情（逆向选择），因此他们会退缩——导致经典的柠檬市场问题或“挤兑”。在 Gorton 的著名类比中，系统就像“没有人担心干粉桶，直到有人闻到烟”——然后每个人都跑了，因为曾经被忽视的风险突然变得非常相关。例如，在 2007-08 年危机中，回购贷款方停止了贷款的展期，因为他们对抵押贷款支持的抵押品的价值感到不确定，并怀疑其他方可能对该价值知道得更多。 这导致了影子银行系统的挤兑。
So how does this concept relate to network models? Most network models we discussed assume either that shocks are common knowledge or that agents act mechanically (no strategic withholding of funds unless insolvent, etc.). Information sensitivity adds a new layer: it can cause endogenous shocks or amplify shocks due to panic. A bank that is actually solvent might lose funding if creditors believe it might be insolvent, especially if they lack transparency. In network terms, this can create contagion of beliefs or confidence, not just direct losses. For instance, consider an interbank lending network: if Bank A suddenly faces a run by short-term creditors due to rumors of losses, Bank A will try to withdraw any funds it lent to Bank B and C (to shore up liquidity), potentially causing funding stress on B and C – this is a network amplification of an information-driven run. Or if one bank is suspected to have big exposure to some toxic asset, all banks with similar portfolios might fall under suspicion (an “information spillover”).那么这个概念与网络模型有什么关系呢？我们讨论的大多数网络模型假设冲击要么是普遍知识，要么是代理人机械地行动（除非破产，否则不会战略性地 withholding 资金等）。信息敏感性增加了一个新层面：它可以导致内生冲击或因恐慌而放大冲击。一个实际上是有偿付能力的银行，如果债权人认为它可能破产，尤其是在缺乏透明度的情况下，可能会失去资金。从网络的角度来看，这可能会造成信念或信心的传染，而不仅仅是直接损失。例如，考虑一个银行间借贷网络：如果银行 A 因为亏损的谣言而突然面临短期债权人的挤兑，银行 A 将试图撤回它借给银行 B 和 C 的任何资金（以增强流动性），这可能会对 B 和 C 造成资金压力——这是信息驱动的挤兑的网络放大效应。或者，如果某个银行被怀疑对某些有毒资产有重大敞口，所有拥有类似投资组合的银行可能都会受到怀疑（“信息溢出”）。
One way to incorporate information sensitivity into network models is via “crises of confidence” scenarios. Recall, the Eisenberg-Noe framework can be adapted by interpreting a shock as a change in expectations: if everyone suddenly believes a certain bank will default, they might act in ways (not rolling over funding) that ensure that default occurs. The OFR survey we referenced notes that the same network model can be used to model information contagion: “a change in market perceptions about the credit-worthiness of a given institution can have knock-on effects on the perceived credit-worthiness of others”, effectively creating an informational cascade
. Morris and Shin (2008) and Acharya, Gale & Yorulmazer (2011) have global games models where banks withdraw funding based on noisy signals of each other’s health – a theoretical approach to runs in interbank markets. These can be interpreted in a network context: each creditor decides to cut links if their confidence falls below a threshold. The result is a coordination game on the network – possibly yielding multiple equilibria (one where everyone trusts and no one defaults, another where no one trusts and many defaults happen).
Cycles in networks (where banks mutually depend on each other’s payments) can particularly give rise to indeterminacy: recent research shows that if there is a loop of interbank obligations, there can exist an equilibrium where all in the loop default (because each expects the others to default) and another where none do – a stark illustration of how beliefs can matter
. In such cases, the
more pessimistic expectations can be self-fulfilling, leading to a systemic collapse even if fundamentally banks could have survived (this is akin to Diamond-Dybvig runs but within a network of banks).
将信息敏感性纳入网络模型的一种方法是通过“信心危机”场景。回想一下，艾森伯格-诺伊框架可以通过将冲击解释为期望的变化来进行调整：如果每个人突然相信某家银行会违约，他们可能会采取某些行动（不续借资金），从而确保违约发生。我们提到的 OFR 调查指出，同样的网络模型可以用来模拟信息传播：“对某一机构信用 worthiness 的市场看法变化可能会对其他机构的信用 worthiness 产生连锁反应”，有效地创造了信息级联FINANCIALRESEARCH.GOV。莫里斯和申（2008）以及阿查里亚、盖尔和尤鲁尔马泽（2011）有全球博弈模型，其中银行根据彼此健康状况的噪声信号撤回资金——这是对银行间市场挤兑的理论方法。这些可以在网络背景下进行解释：每个债权人决定在其信心低于某个阈值时切断链接。 结果是在网络上的一个协调游戏——可能产生多个均衡（一个是每个人都信任而没有人违约，另一个是没有人信任而发生许多违约）。网络中的循环（银行相互依赖于彼此的支付）特别容易导致不确定性：最近的研究表明，如果存在一个银行间义务的循环，可能会出现一个所有在循环中的人都违约的均衡（因为每个人都期望其他人违约），以及一个没有人违约的均衡——这生动地说明了信念如何重要ARXIV.ORGARXIV.ORG。在这种情况下，更悲观的预期可能会自我实现，即使从根本上讲，银行本可以生存，也会导致系统性崩溃（这类似于 Diamond-Dybvig 的挤兑，但发生在银行网络中）。
Another route to incorporate information sensitivity is to consider “network opacity”. Who knows what in the network? In 2008, a big problem was that banks didn’t know which other banks were exposed to subprime – this uncertainty made them cut off lending to virtually everyone (the interbank market freeze). A network model could include uncertainty about link weights or about who has been shocked, forcing banks to form beliefs and perhaps update them when observing others’ failures. If Bank X fails, others might infer “perhaps Bank X had exposure to Y asset; do I also have exposure or my other counterparties do?” leading them to pull back credit generally – an informational contagion beyond the direct loss from X’s default.另一种纳入信息敏感性的方法是考虑“网络不透明性”。网络中谁知道什么？在 2008 年，一个大问题是银行不知道其他哪些银行暴露于次级贷款——这种不确定性使他们几乎切断了对所有人的贷款（银行间市场冻结）。网络模型可以包括关于链接权重的不确定性或关于谁受到冲击的不确定性，迫使银行形成信念，并在观察到其他银行的失败时可能更新这些信念。如果银行 X 失败，其他银行可能推断“也许银行 X 对 Y 资产有暴露；我是否也有暴露，或者我的其他对手方是否有？”这会导致他们普遍收缩信贷——一种超越 X 违约直接损失的信息传染。
Information-sensitive debt and network stability: Essentially, when debt is information-insensitive, the network operates in a smooth manner (no sudden jumps, only actual losses matter). When it becomes sensitive, network ties can break endogenously (credit lines withdrawn, collateral demanded) at the slightest hint of trouble. This can make the network amplify volatility much more. It explains phenomena like why repo haircuts spiked in 2008 or why interbank lending volumes dried up: everyone wanted to conserve liquidity because they were unsure who is safe. From a modeling perspective, one might incorporate this by having a trigger condition based on asset values where lenders switch from “rollover” to “no-rollover” (like a regime switch). For example, a simple addition to a network model: if a counterparty’s perceived asset value falls by X%, then assume short-term lenders withdraw funding (an edge in the funding network is cut), which itself is a shock to the borrower who now must find liquidity elsewhere, potentially causing it to sell assets or default on other obligations, thereby spreading the stress. This kind of mechanism merges solvency and liquidity contagion and is inherently nonlinear and state-dependent (in good times, those edges stay intact; in bad times, they vanish).信息敏感型债务与网络稳定性：本质上，当债务对信息不敏感时，网络运作平稳（没有突然的跳跃，只有实际损失才重要）。当它变得敏感时，网络联系可能在最轻微的麻烦迹象下内生性地断裂（信用额度被撤回，要求提供担保）。这可能使网络放大波动性。它解释了诸如为什么 2008 年回购交易的折扣率飙升或为什么银行间借贷量枯竭的现象：每个人都想保持流动性，因为他们不确定谁是安全的。从建模的角度来看，可以通过基于资产价值的触发条件来纳入这一点，在该条件下，贷款人从“续贷”转变为“非续贷”（就像一个制度切换）。例如，对网络模型的简单补充：如果交易对手的感知资产价值下降 X%，则假设短期贷款人撤回资金（资金网络中的一条边被切断），这本身对借款人造成冲击，借款人现在必须在其他地方寻找流动性，可能导致其出售资产或违约其他义务，从而传播压力。 这种机制将偿付能力和流动性传染合并在一起，并且本质上是非线性和状态依赖的（在好时光，这些边缘保持完整；在坏时光，它们消失）。
While not many of the classic network models explicitly incorporate information runs (they mostly do ex-post analysis of defaults), the literature is moving that way. For instance, He and Xiong (2012) model debt runs in a network of creditors; Chen et al. (2019) look at how the structure of an interbank network can mitigate or exacerbate a coordination failure among depositors. The general lesson is: Crises often involve a regime shift from an equilibrium where certain connections and markets function smoothly to a regime where they break down due to lack of trust. Network models could incorporate that by allowing multiple equilibria or by simulating the endogenous removal of links when stress passes a threshold (perhaps calibrated to something like LIBOR-OIS spreads blowing out, which is a sign of counterparty fear).虽然许多经典网络模型并未明确纳入信息运行（它们主要进行违约的事后分析），但文献正在朝这个方向发展。例如，He 和 Xiong（2012）在债权人网络中建模债务运行；Chen 等（2019）研究了银行间网络的结构如何缓解或加剧存款人之间的协调失灵。一般的教训是：危机往往涉及从某种连接和市场平稳运作的均衡状态转变为由于缺乏信任而崩溃的状态。网络模型可以通过允许多重均衡或通过模拟当压力超过某个阈值时内生性地移除链接来纳入这一点（可能校准为 LIBOR-OIS 利差扩大，这是一种对手方恐惧的迹象）。
In summary, information sensitivity helps explain why crises accelerate so quickly – not only are institutions losing money, but they are losing the trust of their creditors and counterparties. This creates a vicious cycle: as Gorton (2009) put it, in a crisis “everyone is suddenly trying to become informed,” which is impossible to do in real-time, so instead they flee. A practical implication is that transparency and accurate information dissemination are crucial during crises to prevent worst-case rumors. For network modelers, it means that purely mechanical contagion may underestimate the speed of collapse – real crises can develop faster than fundamentals alone would dictate, due to informational contagion. In network terms, edges can “disappear” when confidence is lost, effectively altering the network topology in the middle of the crisis (often fragmenting it). Capturing this dynamic remains an open challenge, but theoretical frameworks like global games on networks are a promising direction. These would model each link’s continuation as a game where parties decide to maintain or cut a link based on noisy signals of each other’s health. Such models could yield insight into which network structures are more robust to panic (for instance, is a complete network more stable because of mutual reassurance, or less stable because everyone’s fate is tied?). Initial work suggests that network structures that facilitate diversification can also help limit runs (because no single counterparty default is too devastating, reducing panic), but the flip side is if something threatens the whole network’s assets (common shock), a highly interconnected system may create a globally bad outlook triggering widespread runs. This area is at the frontier of current research, aiming to integrate informational cascades with financial cascades.总之，信息敏感性有助于解释为什么危机会如此迅速地加剧——不仅机构在亏损，而且它们失去了债权人和交易对手的信任。这造成了一个恶性循环：正如戈顿（2009）所说，在危机中“每个人都突然想要变得知情”，但在实时情况下这是不可能做到的，因此他们选择逃离。一个实际的启示是，在危机期间，透明度和准确的信息传播至关重要，以防止最坏情况的谣言。对于网络模型师来说，这意味着纯粹的机械传播可能低估了崩溃的速度——真正的危机可能比基本面所能决定的更快发展，这要归因于信息传播。在网络术语中，当信心丧失时，边缘可以“消失”，有效地在危机中改变网络拓扑（通常会使其碎片化）。捕捉这种动态仍然是一个开放的挑战，但像网络上的全球博弈这样的理论框架是一个有前景的方向。这些模型将每个链接的持续性视为一个博弈，各方根据彼此健康状况的噪声信号决定维持或切断链接。 这样的模型可以提供对哪些网络结构在恐慌中更具韧性的洞察（例如，完整网络是否因为相互安慰而更稳定，还是因为每个人的命运相互关联而不太稳定？）。初步研究表明，促进多样化的网络结构也可以帮助限制挤兑（因为没有单一的交易对手违约会造成过于严重的后果，从而减少恐慌），但反过来，如果某种情况威胁到整个网络的资产（共同冲击），高度互联的系统可能会导致全球性的糟糕前景，从而触发广泛的挤兑。这个领域处于当前研究的前沿，旨在将信息级联与金融级联结合起来。
Computational Techniques in Financial Network Modeling
Because financial network models often involve potentially large systems of banks with complex interactions, a variety of computational methods have been developed to analyze them. Here we give a brief overview of the main techniques used to study these models and compute outcomes like cascade sizes, default probabilities, and systemic risk measures.由于金融网络模型通常涉及潜在的大型银行系统及其复杂的相互作用，因此开发了多种计算方法来分析它们。在这里，我们简要概述了用于研究这些模型的主要技术，以及计算级联规模、违约概率和系统性风险度量等结果的方法。
-
Fixed-Point Algorithms for Clearing Equilibria: As introduced by Eisenberg & Noe (2001), many network contagion problems boil down to solving a system of equations or inequalities that represent the equilibrium payments or asset values after defaults. These are typically approached via iterative algorithms that converge to a fixed point. The fictitious default algorithm is a prime example: one starts by assuming no one defaults, then iteratively imposes defaults of any banks whose assets are insufficient, adjusting creditor claims, until a consistent set of defaults/payments is reached
. This is essentially a propagation algorithm on the network: in each iteration, the “shock” of a default propagates to neighbors, potentially causing new defaults, and so on. Such algorithms are guaranteed to converge under certain monotonicity conditions and are efficient even for large networks (they amount to linear scans through the network until stability). They have been implemented by central banks to compute contagion impacts in actual interbank networks (requiring data on bilateral exposures). These methods can handle networks with hundreds of banks in fractions of a second, making them suitable for stress-testing. However, if more complex features like bankruptcy costs or multiple asset classes are added, the fixed-point might need more sophisticated solvers (possibly linear programming or iterative nonlinear solvers).清算均衡的固定点算法：正如 Eisenberg 和 Noe（2001）所介绍的，许多网络传染问题归结为解决一组方程或不等式，这些方程或不等式表示违约后的均衡支付或资产价值。通常通过迭代算法来处理这些问题，这些算法收敛到一个固定点。虚构违约算法就是一个典型的例子：首先假设没有人违约，然后迭代地对任何资产不足的银行施加违约，调整债权人索赔，直到达到一致的违约/支付集合FINANCIALRESEARCH.GOV。这本质上是网络上的传播算法：在每次迭代中，违约的“冲击”传播到邻居，可能导致新的违约，依此类推。这类算法在某些单调性条件下保证收敛，并且即使对于大型网络也很高效（它们相当于在网络中进行线性扫描，直到稳定）。中央银行已实施这些算法，以计算实际银行间网络中的传染影响（需要双边敞口的数据）。 这些方法可以在几分之一秒内处理数百家银行的网络，使其适合压力测试。然而，如果添加了更复杂的特征，如破产成本或多种资产类别，固定点可能需要更复杂的求解器（可能是线性规划或迭代非线性求解器）。
-
Monte Carlo Simulation and Agent-Based Modeling: Many network models, especially those involving random networks or stochastic processes (like Gai-Kapadia, or those with behavioral rules), rely on simulation. Researchers will simulate a large number of random network realizations and shock scenarios to estimate the distribution of outcomes – e.g., the probability that a shock to one bank causes $k$ other banks to fail. By varying parameters (connectivity, capital, etc.), one maps out regions of stability vs fragility. For instance, Gai, Haldane & Kapadia (2011) run simulations of contagion on random graphs to see how outcomes depend on network concentration
. Elliott et al. (2014) also use simulations (in a supplemental appendix) to illustrate their theoretical results and to examine cases (like core-periphery networks) that are analytically complex. Agent-based models go further by specifying how individual banks behave (e.g., if a neighbor defaults, reallocate assets, etc.) and then simulating time-step evolution. These can capture dynamic feedback (like asset fire sales over time). For example, the Bank of England and others have built agent-based stress test models where banks respond to shocks according to certain rules (sell assets if capital falls below X, call in loans if liquidity falls below Y, etc.) and see if the system reaches a stable state or a widespread failure. The advantage of simulation is flexibility – one can include many real-world details (multiple layers of the network: interbank, derivatives, repo; various behavior rules; heterogeneity in every parameter). The disadvantage is that it may be hard to derive general insights or guarantee you’ve found worst-case scenarios, etc. But simulation has been heavily used in policy: regulators often simulate network-based stress scenarios because a full analytical solution is infeasible. For instance, the Basel Committee (BCBS 2015) developed a simulation framework for contagion analysis in global banks.蒙特卡洛模拟和基于代理的建模：许多网络模型，特别是那些涉及随机网络或随机过程（如 Gai-Kapadia，或具有行为规则的模型），依赖于模拟。研究人员将模拟大量随机网络实现和冲击场景，以估计结果的分布——例如，冲击一个银行导致$k$个其他银行倒闭的概率。通过改变参数（连通性、资本等），可以绘制出稳定性与脆弱性的区域。例如，Gai、Haldane 和 Kapadia（2011）在随机图上运行传染模拟，以查看结果如何依赖于网络集中度ARXIV.ORGARXIV.ORG。Elliott 等人（2014）也使用模拟（在补充附录中）来说明他们的理论结果，并检查一些案例（如核心-边缘网络），这些案例在分析上是复杂的ARXIV.ORGARXIV.ORG。基于代理的模型更进一步，具体说明个别银行的行为（例如，如果邻居违约，重新分配资产等），然后模拟时间步演变。这些模型可以捕捉动态反馈（如资产随时间的抛售）。 例如，英格兰银行等机构建立了基于代理的压力测试模型，其中银行根据某些规则对冲击做出反应（如果资本低于 X 则出售资产，如果流动性低于 Y 则收回贷款等），并观察系统是否达到稳定状态或发生广泛失败。模拟的优点是灵活性——可以包含许多现实世界的细节（网络的多个层次：银行间、衍生品、回购；各种行为规则；每个参数的异质性）。缺点是可能很难得出一般性的见解或保证找到最坏情况等。但模拟在政策中被广泛使用：监管机构通常模拟基于网络的压力情景，因为完整的分析解决方案是不可行的ARXIV.ORG。例如，巴塞尔委员会（BCBS 2015）为全球银行的传染分析开发了一个模拟框架ARXIV.ORGARXIV.ORG。
-
Percolation and Graph Theory Analysis: Some methods borrow from graph theory to compute things like the percolation threshold (when a giant component of defaults emerges) or use network centrality measures to identify important nodes. Researchers compute metrics like degree distribution, clustering, or centrality of each node (like eigenvector centrality, which is related to systemic impact). An interesting finding is that traditional centrality (like a bank being highly connected) isn’t always a good predictor of systemic risk contribution
– because a highly connected bank might be well diversified (so less likely to fail) whereas a less connected but very large exposure to one other could be more dangerous. So more tailored measures like “financial centrality” have been proposed (Glasserman & Young 2015 define measures based on how much contagion a node can cause or absorb). DebtRank (Battiston et al. 2012) is an algorithm inspired by Google’s PageRank that iteratively calculates how much “centrality” or impact a node has in terms of distress propagation; it’s been applied to banking networks to rank systemically important institutions. These graph-based methods are computationally light (they involve matrix operations or simple iterative procedures).渗透和图论分析：一些方法借鉴了图论来计算诸如渗透阈值（当出现巨大的违约组件时）或使用网络中心性度量来识别重要节点。研究人员计算每个节点的度分布、聚类或中心性等指标（如与系统性影响相关的特征向量中心性）。一个有趣的发现是，传统的中心性（如银行高度连接）并不总是系统性风险贡献的良好预测指标FINANCIALRESEARCH.GOVFINANCIALRESEARCH.GOV——因为一个高度连接的银行可能是良好分散的（因此不太可能失败），而一个连接较少但对另一个有非常大敞口的银行可能更危险。因此，提出了更具针对性的度量，如“金融中心性”（Glasserman & Young 2015 根据节点能够引起或吸收多少传染来定义度量）。DebtRank（Battiston et al. 2012）是一个受谷歌 PageRank 启发的算法，它迭代计算一个节点在困境传播中的“中心性”或影响力；它已被应用于银行网络，以对系统重要机构进行排名。 这些基于图的方法计算量小（它们涉及矩阵运算或简单的迭代过程）。
-
Solving Optimization / Equilibrium with Constraints: Some network problems can be cast as optimization. For example, finding the clearing payments can be seen as solving a linear programming problem: maximize total payments subject to each bank not paying more than it owes or more than it can afford. Similarly, some models of optimal contagion or social planner’s problem lead to convex programs. Off-the-shelf solvers can handle quite large systems if formulated well. In other cases, researchers use equation-based solvers (like coding the equilibrium conditions into MATLAB or Python and using root-finding). For global games or multiple equilibrium analysis, one might have to compute best and worst-case equilibria, which can involve checking many combinations of defaults (which is combinatorially large). There, clever methods are needed – e.g., using monotonicity to narrow possibilities, or focusing on symmetric equilibria for tractability.带约束的优化/均衡求解：一些网络问题可以被视为优化问题。例如，寻找清算支付可以看作是解决一个线性规划问题：在每个银行的支付不超过其欠款或其承受能力的前提下，最大化总支付。同样，一些最优传染模型或社会规划者问题会导致凸规划。如果公式化得当，现成的求解器可以处理相当大的系统。在其他情况下，研究人员使用基于方程的求解器（如将均衡条件编码到 MATLAB 或 Python 中并使用根查找）。对于全球博弈或多重均衡分析，可能需要计算最佳和最坏情况的均衡，这可能涉及检查许多违约组合（这在组合上是巨大的）。在这种情况下，需要巧妙的方法——例如，利用单调性来缩小可能性，或专注于对称均衡以提高可处理性。
-
Empirical Network Analysis: On the empirical side, constructing the network itself can be a computational task. Often data is partial and must be reconstructed using algorithms (e.g., inferring bilateral exposures from aggregated balance sheet data using maximum entropy or other heuristics). Once a network is constructed, researchers compute various stress scenarios algorithmically. They might also perform statistical analyses on networks – like regression of systemic impact on network metrics – which involves computational graph libraries. In recent years, with better data, actual interbank networks of entire countries have been mapped (e.g., Italy, Mexico, Brazil – central banks have published studies) and computational techniques applied to them to simulate historical crisis scenarios and see if the model would have flagged vulnerability.实证网络分析：在实证方面，构建网络本身可能是一项计算任务。数据通常是不完整的，必须使用算法进行重建（例如，使用最大熵或其他启发式方法从汇总的资产负债表数据推断双边敞口）。一旦构建了网络，研究人员就会以算法方式计算各种压力情景。他们还可能对网络进行统计分析——例如，系统影响与网络指标的回归——这涉及计算图形库。近年来，随着数据的改善，整个国家的实际银行间网络已经被绘制出来（例如，意大利、墨西哥、巴西——中央银行已发布研究），并对其应用计算技术以模拟历史危机情景，看看模型是否会标记出脆弱性。
-
Regulatory Stress-Test Platforms: Institutions like the IMF and central banks have developed fairly elaborate models (often not public in detail) that integrate network contagion, market contagion, and feedback loops. These typically run a large number of simulations, and some use parallel computing due to the complexity (especially if doing macro-financial agent-based simulations). However, the core network contagion calculation (who defaults on whom) is usually done with something like the E&N algorithm or its extensions, which is quite fast even for many banks. The heavy part is often Monte Carlo over many scenarios or adding macro feedback.监管压力测试平台：国际货币基金组织和中央银行等机构开发了相当复杂的模型（通常不公开详细信息），这些模型整合了网络传染、市场传染和反馈循环。这些模型通常运行大量模拟，并且由于复杂性（尤其是在进行宏观金融基于代理的模拟时），有些使用并行计算。然而，核心的网络传染计算（谁违约谁）通常使用类似 E&N 算法或其扩展的方式进行，即使对于许多银行来说，这也相当快速。繁重的部分通常是在许多情景下进行蒙特卡洛模拟或添加宏观反馈。
In summary, computational techniques in this field range from analytical algorithms (fixed-point solvers) to brute-force simulations. The choice depends on the question: if one wants a clear threshold result, one might analyze a simplified system with graph theory tools; if one wants realistic loss estimates for a specific crisis, one loads the actual network data and simulates. Increasingly, regulators use a mix: they derive insight from simple models, then “war-game” scenarios on their detailed models to inform policy. From a research perspective, a challenge is ensuring that computational findings (like from an agent-based model) are robust and not artifacts of particular assumptions – hence the value of tying them back to simpler theoretical results. But the overall takeaway is that modern computational power and data availability have made it feasible to study large financial networks explicitly, which was not possible in the Allen & Gale era. As a result, the field has benefited from cross-pollination with network science and the development of specialized tools for systemic risk analysis.总之，该领域的计算技术从分析算法（固定点求解器）到暴力模拟不等。选择取决于问题：如果想要一个明确的阈值结果，可以使用图论工具分析简化系统；如果想要特定危机的现实损失估计，则需要加载实际网络数据并进行模拟。监管机构越来越多地使用混合方法：他们从简单模型中获取洞察，然后在详细模型上进行“战争游戏”场景以指导政策。从研究的角度来看，一个挑战是确保计算结果（如来自基于代理的模型）是稳健的，而不是特定假设的伪影——因此，将其与更简单的理论结果联系起来的价值。但总体而言，现代计算能力和数据可用性使得明确研究大型金融网络成为可能，这在艾伦和盖尔时代是不可行的。因此，该领域受益于与网络科学的交叉传播以及针对系统性风险分析的专门工具的发展。
Concluding Remarks
The literature on financial crises and networks has evolved rapidly, especially after 2008, blending theoretical rigor with practical insights. We now appreciate that financial contagion is not a monolithic process – it can arise from direct interbank defaults, fire-sale spillovers, runs and freezes driven by information, or often a mix of all these. Network models have helped formalize many of these channels, demonstrating how the structure of financial linkages (who is connected to whom) can often be as important as (or amplify) traditional risk factors like leverage or asset quality. A consistent theme is the delicate balance between risk-sharing and risk-spreading: the very connections that make the system safer in normal times can become conduits for disaster in extreme times
. This has informed policy debates on issues like whether to limit interbank exposures or to centrally clear derivatives (reducing the complexity of the network).
关于金融危机和网络的文献发展迅速，特别是在 2008 年之后，理论严谨性与实践洞察相结合。我们现在认识到，金融传染并不是一个单一的过程——它可以源于直接的银行间违约、火售溢出、由信息驱动的挤兑和冻结，或通常是这些因素的混合。网络模型帮助形式化了许多这些渠道，展示了金融联系的结构（谁与谁相连）往往与传统风险因素（如杠杆或资产质量）同样重要（或加剧）。一个一致的主题是风险共享与风险传播之间的微妙平衡：在正常时期使系统更安全的连接在极端时期可能成为灾难的通道NBER.ORGAEAWEB.ORG。这为关于是否限制银行间风险敞口或集中清算衍生品（减少网络复杂性）等问题的政策辩论提供了参考。
We have surveyed both pre-crisis foundational models and post-crisis advances. The early works of Allen & Gale and Eisenberg & Noe laid the groundwork, showing that network topology and clearing mechanisms matter for stability. The post-2008 research – exemplified by Elliott et al. and Acemoglu et al., among others – expanded our understanding with richer networks and new concepts like non-monotonic contagion and phase transitions in connectivity. Alongside, the heightened awareness of information sensitivity in crises has highlighted a frontier for future modeling: integrating agent beliefs and asymmetric information into network contagion models, to capture runs and freezes that are not purely fundamentals-driven
.
我们调查了危机前的基础模型和危机后的进展。Allen & Gale 和 Eisenberg & Noe 的早期工作奠定了基础，表明网络拓扑和清算机制对稳定性的重要性。2008 年后的研究——以 Elliott 等人和 Acemoglu 等人为例——通过更丰富的网络和非单调传染、连接性相变等新概念扩展了我们的理解。同时，对危机中信息敏感性的高度关注突显了未来建模的一个前沿：将代理人的信念和不对称信息整合到网络传染模型中，以捕捉那些并非完全由基本面驱动的奔跑和冻结现象。
In terms of methodology, the field stands out for its interdisciplinary approach: borrowing tools from mathematics (fixed-point algorithms, percolation theory), computer science (network algorithms), and economics (game theory, contract theory) to tackle a complex problem. The result is a much more nuanced understanding of financial stability. We know now, for instance, that simply adding up the size of banks misses the systemic importance that comes from network position – a relatively small bank could trigger a crisis if it’s entangled in the right (or wrong) way. This has led to new regulatory measures of systemic risk (the Basel “GSIB” framework includes criteria for interconnectedness, for example).在方法论方面，该领域以其跨学科的方法而突出：借用数学（不动点算法、渗流理论）、计算机科学（网络算法）和经济学（博弈论、契约理论）中的工具来解决复杂问题。结果是对金融稳定性有了更细致的理解。例如，我们现在知道，仅仅将银行的规模相加并不能反映出来自网络位置的系统重要性——一个相对较小的银行如果以正确（或错误）的方式纠缠在一起，可能会引发危机。这导致了新的系统性风险监管措施（例如，巴塞尔“GSIB”框架包括互联互通的标准）。
Yet, there are certainly open questions. One is how to design networks that are robust – e.g., can we restructure connections or implement policies (like central counterparties or contingent contracts) that preserve the benefits of integration without the fragility? Some research suggests structuring the network in modular clusters (loosely connected) might contain crises, at the cost of some efficiency
. Another question is how
multiple layers of the network interact – banks are linked through lending, derivatives, collateral, etc., and stress can hop from one layer to another (a default on a derivative can cause a liquidity need in the funding market, etc.). Modeling these multiplex networks is challenging but crucial. Computational advances are allowing some progress here. Finally, the element of
behavior and learning in crises – how agents form expectations and change behavior in the midst of a collapse – remains hard to pin down, though it’s clearly important. Bridging network models with insights from behavioral finance or game theory (to capture strategic runs) is a promising avenue.
然而，确实存在一些未解的问题。其中一个是如何设计稳健的网络——例如，我们能否重组连接或实施政策（如中央对手方或或有合同），以在不增加脆弱性的情况下保留整合的好处？一些研究表明，将网络结构化为模块化集群（松散连接）可能会抑制危机，但会牺牲一些效率FINANCIALRESEARCH.GOV。另一个问题是网络的多个层次如何相互作用——银行通过贷款、衍生品、抵押品等相互关联，压力可以从一个层次跳跃到另一个层次（衍生品的违约可能导致资金市场的流动性需求等）。对这些多层网络进行建模具有挑战性，但至关重要。计算技术的进步正在推动这一领域的一些进展。最后，危机中的行为和学习因素——代理人在崩溃中如何形成预期和改变行为——仍然难以确定，尽管这显然很重要。将网络模型与行为金融或博弈论的见解结合起来（以捕捉战略性挤兑）是一个有前景的方向。
In conclusion, financial network models have become an indispensable part of the toolkit for understanding crises. They have brought forth a clearer realization that “the whole is more than the sum of the parts” – the architecture of financial relationships can create systemic properties not evident from individual institutions alone. By combining these models with empirical data and lessons from history, economists and policymakers are better equipped to identify vulnerabilities in the financial system and devise strategies to mitigate systemic risk. While we may never be able to predict or prevent all crises, the network perspective has undoubtedly improved our map of the financial system’s fault lines, helping us move toward a more secure and stable financial environment
.
总之，金融网络模型已成为理解危机的工具箱中不可或缺的一部分。它们使我们更加清晰地认识到“整体大于部分之和”——金融关系的架构可以创造出单个机构所无法显现的系统性特征。通过将这些模型与实证数据和历史教训相结合，经济学家和政策制定者能够更好地识别金融系统中的脆弱性，并制定减轻系统性风险的策略。虽然我们可能永远无法预测或防止所有危机，但网络视角无疑改善了我们对金融系统断层线的认知，帮助我们朝着更安全、更稳定的金融环境迈进。FINANCIALRESEARCH.GOVARXIV.ORG。
Sources: This review drew on numerous authoritative sources, including survey papers and working papers from organizations like the IMF, BIS, and OFR, as well as top-journal publications. Key references include Claessens & Kose (2013) for crisis causes
, Allen & Gale (2000)
, Eisenberg & Noe (2001)
, Elliott, Golub & Jackson (2014)
, Acemoglu, Ozdaglar & Tahbaz-Salehi (2015)
, Gai, Haldane & Kapadia (2011)
, and Gorton & Holmström’s work on information sensitivity
, among others. These and additional citations embedded above provide detailed support and examples for the points discussed.
来源：本综述参考了众多权威来源，包括国际货币基金组织（IMF）、国际清算银行（BIS）和金融研究办公室（OFR）的调查论文和工作论文，以及顶级期刊的出版物。关键参考文献包括 Claessens & Kose（2013）关于危机原因的研究IMF.ORGIMF.ORG，Allen & Gale（2000）FINANCIALRESEARCH.GOVFINANCIALRESEARCH.GOV，Eisenberg & Noe（2001）FINANCIALRESEARCH.GOV，Elliott, Golub & Jackson（2014）AEAWEB.ORG，Acemoglu, Ozdaglar & Tahbaz-Salehi（2015）NBER.ORG，Gai, Haldane & Kapadia（2011）ARXIV.ORG，以及 Gorton & Holmström 关于信息敏感性的研究ANNUALREVIEWS.ORGANNUALREVIEWS.ORG 等。这些及上述附加引用为讨论的要点提供了详细的支持和示例。