Resilience to Contagion in Financial Networks

Hamed Amini; Rama Cont; Andreea Minca

Source-linked AI summary

Resilience to Contagion in Financial Networks

Hamed Amini, Rama Cont, Andreea Minca

arXiv:1112.5687v1Dec 24, 2011q-fin.RMmath.PR

TL;DR

The paper asks how insolvency cascades depend on heterogeneous financial-network structure and node characteristics. It analyzes random weighted directed networks with prescribed degrees and weights, derives asymptotic contagion and resilience measures, and finds that these measures predict large cascades in realistic-sized simulations.

Problem
The paper addresses how network topology, heterogeneous exposure weights, and local institution characteristics determine whether small defaults spread widely or die out.
Method
The authors analyze random weighted directed networks with arbitrary degree sequences and prescribed node characteristics, modeling balance-sheet insolvency cascades and related cash-flow mechanisms.
Results
The paper derives an asymptotic default fraction and resilience measure; simulations on realistic-sized networks show that resilience predicts when large cascades emerge.
Takeaways & Limitations
Institutions most associated with network instability have high connectivity and a large fraction of contagious links, motivating targeted capital or liquidity requirements.
Takeaways & Limitations
The main results are formulated for balance-sheet insolvency and rely on network settings that represent heterogeneous degrees and exposure weights through specified assumptions.

Abstract

from arXiv · show

Propagation of balance-sheet or cash-flow insolvency across financial institutions may be modeled as a cascade process on a network representing their mutual exposures. We derive rigorous asymptotic results for the magnitude of contagion in a large financial network and give an analytical expression for the asymptotic fraction of defaults, in terms of network characteristics. Our results extend previous studies on contagion in random graphs to inhomogeneous directed graphs with a given degree sequence and arbitrary distribution of weights. We introduce a criterion for the resilience of a large financial network to the insolvency of a small group of financial institutions and quantify how contagion amplifies small shocks to the network. Our results emphasize the role played by "contagious links" and show that institutions which contribute most to network instability in case of default have both large connectivity and a large fraction of contagious links. The asymptotic results show good agreement with simulations for networks with realistic sizes.

1 Introduction

The paper studies how network topology, node characteristics, and weighted directed exposures determine whether small defaults die out or generate systemic contagion. It develops rigorous asymptotic tools and a resilience measure intended to identify and mitigate systemically important institutions.

Research problem: The central question is whether defaults at a small number of nodes propagate widely or whether contagion quickly dies out, depending on network topology.
Research gap: Prior work often used stylized networks or heuristic mean-field approximations, while financial networks exhibit heterogeneous degrees and heavy-tailed exposure weights.
Approach: The paper analyzes random weighted directed networks with arbitrary degree sequences and prescribed local node characteristics relevant to balance-sheet or cash-flow cascades.
Contributions: It derives rigorous asymptotic results and an analytical expression for the fraction of defaults, extending contagion results from homogeneous undirected graphs to heterogeneous weighted directed networks.
Regulatory relevance: The resilience measure connects network structure and local characteristics to the amplification of small initial shocks and can support decentralized stress testing and capital-adequacy assessment.

2 A network model of default contagion

The model represents financial institutions as nodes in a weighted directed exposure network, with capital and recovery assumptions determining default propagation. It then embeds prescribed networks in a random ensemble to analyze cascades probabilistically.

Network representation: A weighted directed graph represents institutions and their mutual exposures, with edge weights measuring monetary exposure between counterparties.
Balance-sheet characteristics: Each institution’s capital is its loss-absorption capacity, and the capital ratio is defined relative to interbank assets rather than total assets.
Default cascade: Initially insolvent institutions form the starting set, and a counterparty defaults when the loss after recovery exceeds its available capital.
Default cascade: The cascade iterates until no additional institutions become insolvent, with the final fraction of defaults measuring contagion magnitude.
Assumptions: The framework assumes constant recovery rates and formulates the main results for balance-sheet insolvency, while noting analogous applications to cash-flow insolvency.
Random network model: Empirical heterogeneity motivates a random network ensemble that preserves prescribed degree sequences, exposure weights, and capital ratios while randomizing graph structure.

3 Asymptotic results

The paper derives asymptotic results for default cascades in large random financial networks and expresses the final default fraction through network structure and node characteristics. It also identifies conditions for widespread contagion and introduces a resilience measure for small initial shocks.

Setup: The cascade analysis studies the final fraction of defaults generated by initially insolvent institutions in a sequence of random financial networks.The network size grows with n, and the cascade begins from nodes whose capital ratio is zero.
Node characteristics: Default thresholds measure how many counterparty defaults each node can tolerate, given the order in which counterparties default.The threshold framework incorporates exposures, capital ratios, degree classes, and contagious links.
Limiting cascade: The limiting function I(π) gives the expected fraction of counterparty defaults after one cascade iteration when a randomly chosen edge ends in default with probability π.The smallest fixed point π* represents the eventual probability that a randomly selected edge ends at a defaulted node.
Network resilience: The resilience indicator measures whether a sufficiently small initial default fraction produces a negligible final cascade, while its sign and threshold characterize amplification of small shocks.The resilience function is presented as a simple computable indicator of a global network property.
Propagation condition: If contagion does not spread to nodes with threshold 1, it does not spread at all.This identifies threshold-1 nodes as a necessary stage for propagation in the cascade.

4 Numerical results on finite networks

Finite-network simulations broadly support the asymptotic contagion results, including agreement between theoretical and simulated amplification. Resilience depends on connectivity together with contagious-link structure, not average connectivity alone.

4.2 Relevance of asymptotics: The simulations use a 10,000-node scale-free network with Pareto-distributed degrees and exposures calibrated to empirical Brazilian interbank-network properties.The out-degree, in-degree, and exposure tail exponents are 2.19, 1.98, and 2.61, respectively.
4.2 Relevance of asymptotics: Starting from defaults affecting 0.1% of nodes, theoretical amplification agrees well with simulated amplification as the minimal capital ratio varies.Amplification rises dramatically when the minimal capital ratio falls below a critical value.
4.2 Relevance of asymptotics: A single default’s amplification is compared against the defaulting node’s in-degree, with particularly good agreement for nodes having large in-degrees.Large in-degree nodes can produce large amplification when network susceptibility is high.
4.4 Average connectivity and contagion: The resilience threshold can be computed from connectivity and the fraction of contagious links under the model’s minimal-capital-ratio assumption.All nodes are assigned the same minimal capital ratio in the worst-case finite-network analysis.
4.4 Average connectivity and contagion: Across three equal-average-connectivity networks, empirical-degree networks are resilient with high probability, but resilience is not monotonic in average connectivity.The comparison includes heterogeneous-weight scale-free, equal-weight scale-free, and equal-weight Erdős-Rényi networks.
4.4 Average connectivity and contagion: Assessing resilience requires the degree distribution and contagious-link subgraph, rather than aggregate connectivity measures such as average degree or link count.The model attributes non-monotonicity to a trade-off between risk-sharing and contagion.

5 Conclusions

The paper derives asymptotic contagion results for heterogeneous weighted directed financial networks and validates them on realistic finite networks. It identifies highly connected institutions with many contagious links as important sources of instability and motivates capital requirements based on contagious exposures.

5 Conclusions: The paper extends contagion analysis from homogeneous undirected graphs to heterogeneous, weighted directed networks with prescribed degree sequences and node features.It derives an asymptotic expression for default-cascade size.
5 Conclusions: Simulations on a large realistic network corroborate the asymptotic results and show that the resilience measure predicts the onset of large cascades as capital ratios vary.The sample network has empirical properties matching a real interbank network, such as Brazil’s.
5 Conclusions: Institutions most likely to act as contagion hubs are highly connected and have a large fraction of contagious links.The paper suggests setting minimum capital requirements with respect to contagious exposures.
5 Conclusions: The framework also applies to over-the-counter derivatives, where intermediaries may depend on critical receivables to meet payment obligations.These critical receivables are defined analogously to contagious links.
5 Conclusions: The results cover settings where connectivities and weights are exchangeable random variables with arbitrary correlation structures.This supports analysis when the exposure sequence is unobserved and only its distribution is modeled.

A Appendix: Proofs

The appendix proves the asymptotic results by replacing the financial network with a related weighted configuration model and representing contagion as a sequential Markov-chain process.

A Appendix: Proofs: A weighted configuration model is constructed to have the same asymptotic behavior as the random financial network.The proof uses coupling to relate the model’s default cluster to the original contagion process.
A Appendix: Proofs: The default cluster is constructed sequentially, allowing the contagion process to be described by a Markov chain.The proof begins for financial networks satisfying the stated assumptions.

A.1 Link with the configuration model

The configuration model replaces a prescribed-degree directed network with a uniformly matched directed multigraph, enabling contagion analysis through independence properties. Conditioning on simplicity transfers high-probability results back to the original random network.

Scope: The approach extends prescribed-degree random-graph contagion results to inhomogeneous weighted random directed graphs with arbitrary degree sequences.The configuration model is presented as a standard tool for random graphs with prescribed degree sequences.
Transfer to the financial network: The related multigraph has the same degrees and exposures as the financial network and, conditional on simplicity, the same law as the original random network.This permits analysis on the multigraph before translating results to the simple network.
Configuration model: The configuration model matches incoming and outgoing half-edges uniformly to form a directed random multigraph with the prescribed degree sequence.Each matching produces a directed edge from the node owning the outgoing half-edge to the node owning the incoming half-edge.
Transfer to the financial network: Any property holding with high probability for the configuration multigraph also holds for the original network when the multigraph is simple with probability bounded away from zero.The cited degree-sequence condition is that the relevant sum of squared degrees is O(n).
Configuration model: The configuration model is particularly suited to contagion because half-edge matching can be restricted to incoming half-edges entering defaulted nodes.A uniform matching can be generated sequentially by selecting an incoming half-edge and then choosing a uniformly random unmatched outgoing half-edge.

A.2 Coupling

The coupling constructs a weighted configuration model whose contagion process is easier to analyze while preserving the original process’s final defaults. A threshold-based unweighted skeleton then represents the same default structure in law.

Construction: The construction generates a weighted random graph by randomly permuting each node’s outgoing half-edges, initializing all half-edges as black, and processing contagion interactions.The algorithm begins with the prescribed network degrees and edge weights.
Construction: At each interaction, the algorithm selects an incoming black half-edge associated with a defaulted node and connects it to a probabilistically selected outgoing half-edge.The selected edge is colored red, and its weight determines whether the counterparty defaults based on remaining capital.
Construction: The remaining half-edges are matched uniformly after contagion-relevant interactions have been processed.This final matching preserves the random matching structure for unprocessed edges.
Coupling result: The constructed random graph has the same distribution as the weighted configuration model, and its final default set equals the contagion process’s final default set.This coupling makes the contagion process amenable to analysis without changing the defaults it produces.
Threshold representation: The threshold-based graph has the same law as the unweighted skeleton of the weighted coupled graph.For each node, the threshold is determined from its degree and the randomized ordering of counterparties.
Threshold representation: The normalized expected number of nodes in each degree-and-threshold class converges to its limiting degree distribution times threshold probability.Specifically, E[Nn(j, k, θ)/n] converges to µ(j, k)p(j, k, θ).

A.3 A Markov chain description of contagion dynamics

Replacing default rounds with successive bilateral interactions yields a Markov chain with at most one new default per step. Its state tracks solvent and defaulted banks by degree, threshold, and exposure to defaulted neighbors.

Markov-chain representation: Each interaction matches one incoming edge with one outgoing edge and can produce at most one default, while preserving the final default set of the round-based process.This representation provides a simpler Markov-chain description of contagion.
State variables: The Markov-chain state partitions nodes by solvency, degree, default threshold, and number of defaulted neighbors.It includes counts of solvent banks, defaulted banks, total defaults, and incoming edges belonging to defaulted banks.
Transitions: After t interactions, the number of black outgoing half-edges is mn − t.One outgoing edge is colored red at every step, starting from mn black outgoing half-edges.
Cascade termination: The total number of defaults at cascade termination is Dn(Tn), the cardinality of the final defaulted-node set.The Markov chain remains mathematically definable beyond Tn in some cases, but then no longer represents contagion dynamics.
Transitions: At each step, the partner of a defaulted node’s incoming edge is either already defaulted, remains solvent, or defaults after reaching its threshold.The transition probabilities depend on the partner’s degree, threshold, and number of previously deleted outgoing edges.
Cascade termination: For t ≤ Tn, the Markov-chain trajectory is close with high probability to the solution of the associated deterministic differential equations.This establishes the bridge from the stochastic contagion process to its large-network approximation.

A.4 A law of large numbers for the contagion process

The large-network analysis approximates the contagion Markov chain by differential equations and establishes conditions under which the normalized stochastic variables follow their deterministic trajectory. The differential-equation system has a unique solution constructed through finite-dimensional restrictions.

Existence and uniqueness: The differential-equation system has a unique solution in the specified domain for admissible initial conditions.The solution extends to points arbitrarily close to the boundary of the domain.
Existence and uniqueness: The infinite system is defined through finite systems restricted to degree indices satisfying j ∧ k < K.Standard ordinary-differential-equation results establish existence and uniqueness for each finite restriction.
Differential-equation approximation: The central approximation replaces the contagion Markov chain with a system of differential equations in the large-network limit.The approximation follows Wormald’s differential-equation method.
Approximation conditions: The approximation is controlled until a stopping time at which the normalized time-and-state vector leaves a prescribed domain.The stopping time is defined using the distance of the deterministic trajectory from the domain boundary.
Approximation conditions: The law-of-large-numbers result requires bounded variables, an approximate conditional trend, and a sufficiently small error parameter.Under these conditions, the stochastic process remains close to the differential-equation solution with high probability.
Application to contagion: The contagion variables are rescaled by n and compared with the differential-equation solution over the interval before the stopping time.The application defines normalized trajectories using the functions sj,k,θ,l and the corresponding Markov-chain variables.

A.5 Proof of Theorem 3.6

The proof uses differential-equation approximations and stopping-time analysis to characterize the asymptotic default cascade. It distinguishes between complete contagion and termination near a stable fixed point.

Proof strategy: The proof controls higher-order terms in the relevant infinite sums before applying the differential-equation approximation.This step is needed because the number of variables depends on n.
Differential-equation approximation: The cascade variables are approximated by solutions to a system of ordinary differential equations within a bounded domain Uε.The approximation follows from boundedness, Lipschitz, and initial-condition convergence conditions.
Complete contagion: When I(π) > π for every π < 1, the stopping-time analysis yields |Dn(Tn)| = n − op(n).Thus, asymptotically, nearly all network vertices default.
Stable fixed point: When π* < 1 is a stable fixed point of I, δ−(τ) becomes negative immediately after τ* = λπ*, constraining the cascade stopping time.The proof selects ε using the negative minimum of δ− and obtains Tn/n = τ* + O(ε) + op(1).

A.6 Proof of Theorem 3.9

The proof reduces contagious links to a directed configuration-model graph and applies giant-component results under the paper’s degree assumptions. It establishes when the contagious-link skeleton contains a giant strongly connected component.

Assumptions: The proof shows that the supplementary degree conditions used in earlier work can be dropped for the sufficient giant-component condition needed here.The restricted Assumption 3.1 is sufficient for this purpose, although stronger assumptions yield more structural detail.
Contagious-link skeleton: The contagious-link skeleton is analyzed through a rewiring construction that preserves half-edges while removing noncontagious links.New degree-(1,0) nodes are then removed because they cannot belong to a strongly connected component.
Degree sequence: The rewired graph has the same relevant degree properties as the original network, including the unchanged total number of edges and average degree.The proof relates out-going edges after rewiring to contagious out-going edges before rewiring.
Giant component criterion: With high probability, the contagious-link skeleton contains a giant strongly connected component when the stated degree-sequence criterion holds.This conclusion follows by applying Lemma A.11 to the rewired configuration-model graph.