Source-linked AI summary
Network models of financial systemic risk: A review
Fabio Caccioli, Paolo Barucca, Teruyoshi Kobayashi
TL;DR
Financial systemic risk arises from interactions among interconnected institutions, yet realistic network structures and contagion mechanisms require models beyond stylized economic representations. This review synthesizes network approaches to bilateral defaults, overlapping portfolios, and empirical interbank structure, highlighting both systemic amplification mechanisms and important scope limitations.
Problem
The paper addresses how interconnected financial institutions can transform links that diversify risk into channels for systemic breakdown.
Method
The paper reviews network models of financial systemic risk, including default cascades, overlapping portfolios, leverage targeting, and empirical interbank-network structure.
Results
The review finds that network structure and propagation mechanisms shape systemic outcomes, including instability through diversification, amplification beyond initial losses, and limitations of core-periphery interpretations.
Takeaways & Limitations
Network-based analysis provides a framework for studying feedback between individual institutions and collective financial-system behavior.
Takeaways & Limitations
The review is not exhaustive and does not cover the prediction and control of systemic risk.
Abstract
from arXiv · showhide
The global financial system can be represented as a large complex network in which banks, hedge funds and other financial institutions are interconnected to each other through visible and invisible financial linkages. Recently, a lot of attention has been paid to the understanding of the mechanisms that can lead to a breakdown of this network. This can happen when the existing financial links turn from being a means of risk diversification to channels for the propagation of risk across financial institutions. In this review article, we summarize recent developments in the modeling of financial systemic risk. We focus in particular on network approaches, such as models of default cascades due to bilateral exposures or to overlapping portfolios, and we also report on recent findings on the empirical structure of interbank networks. The current review provides a landscape of the newly arising interdisciplinary field lying at the intersection of several disciplines, such as network science, physics, engineering, economics, and ecology.
1 Introduction
Financial systemic risk is studied as an emergent property of interconnected financial institutions whose interactions create feedback between individual and collective dynamics. This review surveys network-based models of contagion, clearing, distress propagation, overlapping portfolios, and empirical interbank structure.
- Financial networks: Financial institutions interact through liabilities, cross-asset holdings, and return correlations, forming complex webs of financial linkages.These interactions motivate representing systemic risk as a network phenomenon.
- Financial networks: Systemic risk concerns the possible breakdown of the financial system emerging from interactions among market participants.Participants also react to the aggregate market dynamics they collectively create, producing a micro-macro feedback loop.
- Review scope: The review focuses on network models developed outside traditional economics because they directly represent feedback between micro- and macroscopic phenomena.Network structure is presented as essential for linking micro events to collective outcomes.
- Review scope: Financial networks can be multiplex, bipartite, core-periphery, or time-varying rather than simple random or star graphs.The relevant structure depends on the financial linkage being modeled.
- Topics covered: The review covers bilateral default cascades, distress propagation before default, overlapping-portfolio contagion, and empirical interbank network structure.It also discusses clearing algorithms needed to allocate failed borrowers’ remaining assets among creditors.
2 Clearing algorithms
Clearing algorithms determine how interconnected institutions distribute payments when liabilities cannot be fully honored. The Eisenberg-Noe framework formulates this as a fixed-point problem and, under suitable assumptions, yields a unique clearing vector that exposes default cascades and systemic effects.
- Clearing framework: Clearing is central because mutually entangled transactions underpin liquidity problems, failed payments, losses, and insolvencies.The framework applies to bilateral over-the-counter contracts, central clearing, and market-mediated transactions.
- Eisenberg-Noe model: The Eisenberg-Noe model represents bilateral obligations with a liabilities matrix and determines payments subject to limited liabilities and absolute priority.Banks repay as much as possible and cannot retain cash while interbank liabilities remain unpaid.
- Eisenberg-Noe model: The financial system is represented as F = (L, e), combining a non-negative liabilities matrix with operating cash flows.An extra bank can represent external liabilities in the liabilities matrix.
- Scope and limitations: The framework computes clearing payments, contagion dynamics, and conditions for unique solutions, but excludes several realistic financial features.Excluded features include multiple seniority levels, different maturities, stochastic cash flows, correlations, and common asset holdings.
- Clearing solution: The payment vector p records each bank’s actual repayment and satisfies 0 ≤ p ≤ p̄; under natural financial assumptions, the clearing solution is unique.The fixed-point equations are piece-wise linear, monotone, bounded, and continuous, supporting existence and uniqueness.
- Convergence: An iterative algorithm alternates between identifying defaulted banks and updating payments toward a fixed point until convergence.Under regularity, the converged payment vector is the only solution; strictly positive equity commonly satisfies the condition.
- Scope and limitations: Deterministic clearing models propagate losses only after actual insolvency, limiting their ability to represent distress contagion before default.The review identifies DebtRank and stochastic external-asset values as extensions addressing this limitation.
3.1 The Gai-Kapadia model
The Gai-Kapadia model adapts threshold cascades to interbank default contagion and estimates cascade sizes using tree-based and generating-function approaches. These approaches identify cascade conditions, with the second-order condition matching simulations better than the first-order condition.
- Threshold cascade formulation: The Gai-Kapadia model interprets node activation in threshold cascades as bank default caused by sufficiently many defaulted counterparties.It builds on the Watts model, extending threshold contagion to directed interbank lending relationships.
- Tree-based approximation: The tree-based approximation computes the final active fraction from a fixed point for neighbor activation probability under a locally tree-like network assumption.The response function activates a node when the fraction of active neighbors exceeds threshold R.
- Tree-based approximation: The tree-based method predicts the final cascade size very accurately despite its simplifying assumptions.Figure 1 compares the analytical cascade size with simulated averages over 1000 runs for a Poissonian degree distribution.
- Cascade conditions: The first-order cascade condition is based on whether the recursion derivative near q = 0 exceeds one, but simulations show it is not very accurate across (R, z).The second-order condition improves the approximation of the cascade region.
- Cascade conditions: The second-order cascade condition well matches the cascade region predicted by numerical simulation.It is derived by expanding the recursion equation around q = 0 to second order.
- Generating function approach: The generating-function approach characterizes vulnerable nodes and vulnerable-cluster sizes, and its cascade condition is equivalent to the tree-based condition when R > 0 and ρ0 →0.The equivalence requires a sufficiently small seed fraction and F(0) = 0.
- Generating function approach: The vulnerable-cluster size is not identical to the average cascade size, so the correspondence between the generating-function and tree-based approaches is not rigorous.The paper notes this distinction explicitly when relating the two methods.
3.2 Extensions of the threshold cascade model
Extensions relax the Gai-Kapadia model’s assumptions about loan weights, network topology, and asset structure. They introduce heterogeneous exposures, empirical network features, multiplex layers, and overlapping-portfolio contagion, while analytical treatment remains limited for multiplex systems.
- Model assumptions: The simplest Gai-Kapadia model assumes evenly distributed loans, an Erdős-Rényi random graph, and no risk from external assets.Later models relax these assumptions to make cascade dynamics more realistic.
- Heterogeneous exposures: With heterogeneous loan weights, default depends on losses relative to total interbank assets rather than only the fraction of defaulted borrowers.The fraction-based condition is recovered when edge weights are identical.
- Heterogeneous exposures: For heterogeneous edge weights, the number of defaulted banks alone is not informative, making the standard mean-field approximation inappropriate.Numerical simulations and alternative approximation methods are used for more general environments.
- Non-Erdős-Rényi networks: Empirical interbank networks differ from Erdős-Rényi graphs, exhibiting fat-tailed degree distributions and potentially non-tree-like structure.Configuration models can also have substantial clustering when degree distributions are fat-tailed.
- Non-Erdős-Rényi networks: Disassortativity is observed in real-world financial networks, where highly connected banks tend to trade with low-degree banks.Studies find that assortativity of financial linkages strongly affects systemic risk.
- Multiplex networks: Multiplex models represent different asset types or seniority levels as separate layers of interbank connections.Examples include long- and short-term assets, foreign exchange exposures, and derivatives.
- Multiplex networks: Analytical models of contagion in multiplex financial networks remain scarce, leaving this structure relatively underdeveloped.The review describes multiplex financial networks as a relatively premature research area.
- External assets and overlapping portfolios: Overlapping portfolios add an asset-price contagion channel in which devaluation simultaneously affects banks holding common or correlated assets.Asset liquidation can depress prices further and transmit losses to other banks, potentially accelerating interbank contagion.
4 Distress propagation due to credit quality deterioration
Network distress can propagate before default when credit deterioration reduces expected payments, and DebtRank models this through active-bank transmission dynamics. Extensions show that network structure, nonlinear rules, and liquidation-related effects can amplify systemic losses, while some diversification processes may increase instability.
- Counterparty contagion can be practically limited in isolation, yet interbank networks may amplify distress when combined with fire sales or overlapping portfolios.The review cites theoretical and empirical challenges to default contagion while emphasizing amplification through additional contagion channels.
- Credit deterioration can transmit losses before default by reducing the expected cash flow and marked-to-market value of interbank assets.A loss at bank j raises its default probability, lowering the expected value of bank i’s exposure to j.
- 4.1 DebtRank: DebtRank tracks equity losses over discrete time as active banks transmit distress once to creditors in proportion to their distress and relative interbank exposure.Its leverage matrix uses exposure relative to the receiving bank’s equity, while inactive banks no longer transmit further losses.
- 4.2 Extensions: A largest interbank-leverage eigenvalue above one indicates that network shocks are amplified and can lead to bank defaults.The modified dynamics use the initial exogenous shock to analyze stability under small perturbations.
- 4.2 Extensions: Nonlinear distress rules reveal different amplification regimes, while increasing diversification can move an initially stable network toward instability through cyclical exposure structures.The nonlinear family interpolates between threshold and linear DebtRank rules; a separate analysis links instability to cycles emerging during diversification.
- 4.2 Extensions: Second-round DebtRank effects and third-round leverage-targeting effects can dominate direct losses, so stress tests omitting network effects may underestimate systemic risk.The reviewed stress-testing framework applies an exogenous shock, propagates distress, and models further losses from common-asset liquidation.
- 4.2 Extensions: Taxation policies that account for interbank contracts’ systemic impact can promote stability without reducing interbank lending volume.The reviewed policy studies use DebtRank as a systemic-risk measurement tool.
5 Overlapping portfolios and price mediated contagion
Overlapping portfolios create contagion when one bank’s liquidation devalues commonly held assets, imposing losses and potentially triggering further liquidations. Review models analyze this mechanism through bipartite bank–asset networks, market impact, leverage responses, and cascade conditions.
- Mechanism: Common asset holdings transmit stress when one bank’s liquidation lowers asset prices and causes losses to other banks.Further liquidations can devalue additional assets and propagate losses across several banks.
- Model structure: Overlapping-portfolio models represent banks and assets as two node types, with links indicating bank investments and positions recorded in Q.Model dynamics require rules for bank responses to losses and asset responses to liquidation.
- Empirical analysis: 7,846 commercial banks and 13 asset classes were used in a 2007 US stress test, which found abrupt survival transitions and predictive power for failures between 2008 and 2011.The analysis identified commercial real estate loan devaluation as responsible for commercial-bank failures during the subprime crisis.
- Threshold dynamics: A branching-process approximation uses a transfer matrix, whose largest eigenvalue assesses stability and global-cascade conditions in large bipartite networks.For bipartite Erdős–Rényi networks, average diversification has a non-monotonic relation with global-cascade probability, while sufficiently low leverage is always stable.
- Leverage targeting: Banks’ preemptive liquidation to restore leverage significantly widens the parameter region where global cascades can occur.Deleveraging lowers asset prices, creates mark-to-market losses, and can require further deleveraging.
- Leverage targeting: Second-order portfolio overlaps connect institutions even when many pairs have zero direct overlap, so tests omitting second-round losses can underestimate systemic risk.
6 Empirical structure of interbank networks
Empirical interbank-network research studies topology, estimation methods, and time variation because connectivity affects how risks spread. Evidence challenges a universal static core-periphery description: daily Italian networks are often bipartite or community-structured and exhibit regular dynamical patterns.
- Objectives and data: Interbank-network studies measure topology to understand how local risks may spread through financial linkages.When bilateral data are unavailable, network structure must be estimated from aggregate balance-sheet information.
- Network estimation: Maximum entropy estimation distributes total lending evenly across possible borrowers but may produce networks denser than the actual network.This density mismatch may over- or underestimate default-contagion risk.
- Network structure: A core-periphery network has densely interconnected core nodes, while peripheral nodes connect to the core but not to one another.Stochastic block models can represent core-periphery, modular, and bipartite structures.
- Core-periphery vs. bipartite structure: Empirical core-periphery findings can reflect heterogeneous degree distributions or multiple cores rather than a single pure core-periphery structure.Italian networks in 2007 appeared to contain two core-like groups of Italian and foreign banks.
- Why daily scale?: More than 86% of Italian interbank transactions were overnight during 2000–2015, making temporal granularity important for representing contemporaneous relationships.Aggregating edges formed on different days can include relationships that did not coexist.
- Core-periphery vs. bipartite structure: At daily resolution, e-MID networks are characterized by bipartite or broader community structures rather than core-periphery structure; aggregation increases core-periphery detections.Bipartivity in daily interbank networks increased over the past decade.
- Daily network dynamics: N ∝ M^1.5 governs daily Italian market activity across network structures and sizes, alongside power-law transaction durations and tent-shaped weight growth.The relationship uses N and M as the numbers of active banks and edges, respectively.
7 Discussion
The review covers mechanisms of network-based systemic risk but not the full progression from modeling to forecasting and control. It identifies prediction and control as important areas for further work.
- The review is not exhaustive and specifically omits prediction and control of systemic risk.
- Most reviewed studies remain at the first stage of understanding and modeling mechanisms rather than forecasting or controlling systemic risk.The paper frames scientific maturation as progressing from modeling, to forecasting, to control.
- Further studies are needed at every stage to deepen understanding of financial-network complexity and reduce systemic risk.A growing number of researchers are simulating possible regulatory policy tools for controllability.