Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Network Models of Risk Propagation

Abstract. Risks propagate when the consequences of a disturbance at one entity change exposures, resources, operations, or decisions at other entities connected by financial, contractual, productive, informational, or physical relationships. The scientific task is to reconstruct those mechanisms from an explicit network representation and to establish when local disturbances remain local, are absorbed, or produce nonlinear cascades. This monograph integrates mathematical network theory with insurer and reinsurer balance sheets, operational dependencies, cyber accumulation, infrastructure failures, and macroeconomic production linkages. It distinguishes direct losses, transmitted obligations, common-cause shocks, correlated outcomes, and second-round adjustments; specifies directed, weighted, multiplex, temporal, and contingent-link representations; derives linear resolvent and threshold results under stated assumptions; compares fixed-point clearing, exposure-network contagion, reinsurance contract nonlinearities, and interdependent infrastructure models; and presents reproducible hypothetical examples involving feedback, a counterparty default chain, and a shared digital provider. The analysis shows why density, centrality, and diversification have conditional and sometimes nonmonotonic effects, while graph topology alone cannot identify loss severity or legal liability. Particular attention is given to aggregation without double counting, observability of hidden edges, sensitivity to recovery rates and thresholds, and the translation from gross economic disruption to insured claims and capital demands. Original contributions consist of the cross-domain analytical synthesis, an explicit reconciliation of propagation and contractual loss accounting, and transparent illustrative calculations. No independent new theory, proprietary network model, or empirically estimated insurance loss parameter is proposed. The conclusion establishes a disciplined, evidence-sensitive approach to applying existing network models to internationally diverse insurance and risk systems.

Keywords: Network models; risk propagation; financial contagion; cascade failures; systemic risk; counterparty credit; reinsurance networks; excess-of-loss contracts; cyber accumulation; cloud service concentration; critical infrastructure; production networks; input-output linkages; directed graphs; spectral radius; threshold models; multiplex networks; network observability; contagion stress testing; insurance aggregation; ORSA; operational resilience; risk science.

Introduction

Interconnected systems redistribute the consequences of events. A flood can halt a specialized supplier, interrupt downstream production, generate property and business-interruption claims, impair the liquidity of an insurer, and change the reinsurance payments on which several cedents depend. A disruption at a cloud provider can simultaneously affect enterprises that share a service while generating many contractual claims from one incident. A financial institution's inability to honor a debt can reduce the recoverable assets of another institution, producing a further default when losses exceed available capital. These events illustrate different mechanisms: propagation along connections, synchronized loss from a shared source, and changes in liability arising from risk-transfer contracts. A useful scientific account specifies their joint operation without treating them as identical.

Network models provide a language for representing who depends on whom, how much is at stake, which directions events can travel, and what happens when thresholds are crossed. The early formal foundation of simultaneous payment consistency in Eisenberg and Noe (2001) addressed clearing among interdependent debtors. The broader literature developed models of threshold diffusion, contagion, cross-holdings, market liquidity, production-network transmission, and interdependent infrastructure. An important general result from Acemoglu, Ozdaglar, and Tahbaz-Salehi (2015) is that richer connectivity can absorb relatively small shocks while spreading sufficiently severe ones. The same graph therefore admits contrasting risk outcomes when shock magnitude and the loss-absorption regime change.

For insurance, the research object extends beyond a graph of financial institutions. An insurer is simultaneously a collector of premiums, a holder of investment assets, a payer of indemnity or fixed benefits, a purchaser and seller of reinsurance, an employer, a purchaser of essential services, and sometimes a counterparty to derivatives. An insured enterprise occupies still more networks: suppliers, customers, power, telecommunications, cloud software, physical premises, logistics, and legal claims. Each relation has a distinct loss function and may introduce delays or recovery mechanisms. A directed edge representing an indemnity contract cannot be substituted mechanically for an edge representing electricity supply.

Three research questions organize this investigation. First, which properties of connections and shocks determine whether losses are absorbed, transmitted, or amplified? Second, what does the network model need to represent before its output can be interpreted as insured loss or capital demand? Third, how can empirical observation, explicit counterexamples, and stress testing distinguish robust deductions from results that depend on hidden edges or disputed parameters? These questions are pursued through source-based historical and mathematical analysis, algebraic derivations, and labeled hypothetical examples. Assertions about regulatory obligations are limited to the stated supervisory framework and jurisdiction.

The methodology proceeds from definitions to mechanisms, then to model classes, worked calculations, institutional applications, and validation. Sources include peer-reviewed journal articles and authoritative publications by international supervisors, regulators, and standard setters; a publisher or institutional abstract is identified as such when detailed article text was not independently reviewed. No external microdata set is claimed to have been acquired. Mathematical examples are independent demonstrations of established model classes, rather than estimates of historical crises or insurance event losses.

Distinguishing a network outcome from a risk measure

Let a shock be an exogenous change in a specified state variable, such as asset value, operating capacity, or supplier availability. Let a propagation rule map the shock and current node states into later states. Let an exposure rule transform those states into financial consequences for the decision maker, and let a contract rule transform economic damage into a covered payment. Finally, let a risk measure summarize the distribution of those payments under a chosen horizon. The four operations can be studied in sequence, although actual systems contain feedback between them.

A path through a network establishes a potential channel. It does not by itself establish a payment default, an insurance claim, or the probability of catastrophic loss. Financial institutions may continue paying after suffering a mark-to-market decline; enterprises may substitute suppliers; a policy may exclude or limit a contingent-business-interruption event; and a reinsurer may pay only above attachment. Interpretation requires all of those missing conditions to be specified. This separation is central to the monograph's analytical synthesis and prevents the conflation of network reachability with realized loss.

Network Nodes and Exposures

Research object and historical foundations

A network is a collection of entities and relations defined over a common analytical horizon. In mathematical notation, a graph is G = (V,E), where V is the set of nodes and E the set of ordered or unordered relations. A directed edge from j to i means that the relevant effect of j can reach i under the chosen edge semantics. A weighted network associates a nonnegative quantity or transformation with that edge. The units must be declared: monetary liability, dependency fraction, probability of infection, percentage of business capacity, or fraction of unmet demand. Networks with different units cannot be superposed by adding their adjacency matrices.

The approach to network cascades predates current systemic-risk practice. In Watts (2002), a threshold-diffusion process shows that a small initiating event may trigger widespread adoption or failure in specified regions of network connectivity. The paper is a model of agents' threshold responses on random graphs, not a universal formula for an insurer's claims. In Newman (2002), epidemic processes on contact networks yield results sensitive to degree distributions, transmissibility, and structural correlations. Their value for risk science is methodological: they demonstrate how a specified contact and state-transition mechanism generates nonlinear aggregate outcomes. Their literal epidemiological parameters cannot simply be relabeled as insured dollars.

Financial network research adds balance-sheet accounting. Gai and Kapadia (2010) analyzed how network configuration, shocks, and asset-market liquidity interact. Their robust-yet-fragile observation differentiates a low probability of widespread contagion from its high possible impact. Haldane and May (2011) examined analogies to ecological and epidemiological systems for bank stability. These analogies provide candidate mechanisms and warning signs, while institution-specific accounting remains necessary.

A second strand begins with production linkages rather than default. Acemoglu and colleagues (2012) establish that the effect of local productivity shocks on aggregate economic volatility depends on input-output structure and the importance of indirect suppliers. The later Production Networks primer (2019) reviews these mechanisms and their empirical interpretation. Insurance risk analysis must distinguish a production-network reduction in value added from covered claims, which are governed by contract, deductibles, waiting periods, limits, and proof of loss.

Formal graph representations and loss ontology

For n nodes, define an adjacency indicator B_ij equal to 1 if the relevant direction j → i exists, and 0 otherwise. A weighted matrix A records the intensity or contractual magnitude of the relation; all calculations below specify whether A_ij denotes a transmission coefficient or a monetary exposure. A path j → k → i represents a possible second-round effect; the existence of a path does not establish that the relevant shock traverses every edge. Transmission probabilities, thresholds, and failure states determine the realized path.

A complete exposure record may include node attributes (capital, liquidity, physical assets, geographic coordinates, insurance policies, critical operations, sector), edge attributes (counterparty name, gross claim, collateral, attachment, seniority, recovery, payment deadline, service-level arrangements), and event attributes (hazard, origin, duration, severity, remediation, correlation with other hazards). For a financial contract, claim direction must be distinguished from cash-payment direction. A receivable on j is an asset of i and a liability of j; default by j can reduce the value of i's asset. For a service dependency, failure of j can impair i's throughput even where no debt is owed between the two.

The observable loss categories also differ. Gross economic loss can comprise destroyed capital, lost output, additional expense, and opportunity cost. Contractual insured loss is the payable amount determined by policy conditions. Reinsurance recoverable is contingent on the ceded loss, contractual wording, and reinsurer performance. Insurer net loss is computed after recoveries actually available and according to an explicit balance-sheet boundary. A network that represents all four categories as one generic loss weight risks both causal error and double counting.

Direction, layers, time, and contractual contingencies

A directed exposure graph is asymmetric: failure of a cloud supplier may harm a customer even if that customer's failure has little effect on the cloud supplier. A bipartite graph separates firms from service providers, banks from assets, or insurers from reinsurers. A multiplex network represents several edges between the same entities—credit, ownership, outsourcing, equity price, insurance, and payment dependencies—each assigned its own layer-specific rule. A temporal network represents contracting and failure at times t, t+1, ...; an annual aggregate graph can erase sequence, grace periods, liquidity buffers, and recovery delays.

In a multilayer system, physical, financial, informational, and contractual links form distinct but connected networks. A storm may simultaneously damage a power substation and disable digital access to claims systems. A delayed reinsurance recovery may be immaterial to ultimate net profit but critical to one-week liquidity. A high-degree service provider may become a single point of operational dependence while remaining a low-debt counterparty. In practice, the model must decide which state variables transmit across layers and which controls can decouple them.

In the interdependent-network model of Buldyrev and colleagues (2010), one network's loss of functional nodes can disable dependent nodes in another. The study illustrates abrupt transitions in a particular stylized percolation framework. It does not establish that every real infrastructure or every insurer exhibits the same critical threshold. Where operators deploy redundancy, local storage, islanding, manual controls, or alternate service arrangements, the dependency function changes, and so can the apparent cascade threshold.

Propagation, common cause, and correlation

Contagion means that a change at one node produces a further change at another through a specified transmission relation. Common-cause dependence means several nodes suffer losses from the same underlying event, regardless of whether either node transmits loss to the other. Observed statistical correlation can reflect either mechanism, both mechanisms, or an unobserved third factor. A matrix of loss correlations therefore does not identify the direction or financial content of a network. A pair of insured firms hit by a regional outage may have strongly correlated business-interruption claims but no edge between them.

Let Z be a shared environment such as a hazard event or vendor outage, and X_i losses on different policies. The identity Cov(X_i,X_j) = E[Cov(X_i,X_j | Z)] + Cov(E[X_i|Z],E[X_j|Z]) distinguishes within-environment dependence from covariance induced by variation in the common environment, when second moments exist. Even when claims are conditionally independent given Z, the second term can generate large marginal dependence. That decomposition is an actuarial warning against claiming to have discovered contagion from correlated claims alone. Establishing contagion requires evidence about sequence, edges, and counterfactual behavior.

Shock Transmission

Linear propagation and the resolvent

Consider a one-period, nonnegative linear transmission approximation. Let s be an n-dimensional vector of direct shock magnitudes, measured in consistent units, and let h be the vector of total transmitted effects under all modeled rounds. Define A_ij as the amount of incremental effect on i per unit effect at j. Suppose the system obeys h = s + Ah, without any capacity cap, bankruptcy threshold, endogenous policy response, or double-counting adjustment. If the spectral radius ρ(A) is less than 1, the unique solution is h = (I−A)^(−1)s. The Neumann-series identity gives h = s + As + A²s + A³s + ..., with the kth matrix power representing the contribution of paths of length k.

The derivation follows by repeatedly substituting the right side into the equation: h = s + A(s+Ah) = s + As + A²h, and so forth. For a finite matrix with ρ(A) < 1, the series converges. For nonnegative A, each term is componentwise nonnegative when s is nonnegative. The result makes indirect paths visible and identifies the reason a seemingly modest direct shock can have a larger final effect. It is an algebraic consequence of a chosen model, not a measured coefficient of real-world financial distress.

For any chosen output functional wᵀh, the marginal effect of a small change in s is wᵀ(I−A)^(−1), so the same source may matter differently depending on the insurer, economy, or population whose loss is measured. A node with few direct links can be influential because its effects pass through highly consequential downstream nodes. Conversely, a large adjacency weight may imply only a constrained and temporary operational consequence when the omitted mechanisms of substitution or recovery are included. Measurement units and the choice of w are therefore part of the result.

The ρ(A) < 1 condition is sufficient for this unlimited linear feedback representation to converge. A failure of that condition does not prove that a real insured system suffers infinite economic loss. Monetary obligations, insurance limits, emergency shutdowns, resource constraints, and finite facilities generally bound real transactions. If ρ(A) ≥ 1, this particular uncapped infinite-round linear expansion lacks the stated convergence property and needs a different state-transition rule or limiting treatment. A nonlinear capped model may still have finite fixed points, and those points can depend on initialization or selection criteria.

Nonlinear thresholds and network clearing

Many failures are discontinuous. A debtor can continue to pay after a minor asset loss but default when resources fall below amounts due. A factory can tolerate a partial input delay but close after its inventory runs out. A policy triggers when a peril meets conditions or a loss exceeds a deductible; a reinsurance layer pays only beyond attachment. Threshold models encode such discontinuities directly. One simple default rule is d_i(t+1) = 1 if accumulated impairment at i exceeds its loss-absorbing buffer K_i, and 0 otherwise, with d_i(t+1) constrained to remain 1 after an irreversible default in that scenario. The accumulated impairment can depend on direct shocks and exposures to already defaulted counterparties.

Suppose default at j causes creditor i to lose (1−r_j)L_ji, where L_ji is creditor i's nominal claim on j and r_j is an assumed recovery fraction. A simplified update is

D_i(t+1) = 1{u_i + Σ_j (1−r_j)L_ji D_j(t) > K_i},

where u_i is a primary impairment and the indicator is one when the condition holds. This is a stylized binary-loss threshold model. Recoveries and buffers are treated as exogenous, there is no seniority conflict, and losses are not redistributed dynamically through simultaneous payment clearing. It can illustrate a cascade, but its final default set should not be presented as the solution to the more general Eisenberg–Noe payment problem.

Watts's global cascade analysis considers threshold activation under a different rule in which the fraction of active neighbors matters. It is useful to compare mechanisms: a fixed monetary exposure threshold, a fraction-of-neighbors threshold, and a capacity-constrained flow network can react differently to the same graph. The correct formulation follows from institutional and physical facts. A tiny loss from one very large debtor may be more damaging than simultaneous failures of several immaterial counterparties; a count-based rule could misrank these scenarios.

When contracts are mutually dependent, solving sequentially can also be wrong. If a reinsurer's outflows depend on cedents' payments and cedents' ability to pay depends on reinsurer recoveries, the liability structure can involve fixed points with piecewise contractual functions. The detailed treatment in Cascading Losses in Reinsurance Networks shows why contract nonlinearities and cycles merit explicit numerical and mathematical study. For each model class, existence, uniqueness, convergence of the algorithm, and economic interpretation must be examined separately.

Temporal cascades and interdependent systems

A model with a single simultaneous clearing time gives different information from a simulation of a multi-day service outage or a multi-year insurance loss. Let x_i(t) describe available operational capacity, a_i(t) recoverable financial assets, c_i(t) liquidity, and q_i(t) covered claims falling due. A state transition may depend on exogenous event z(t), upstream capacities, substitute suppliers, insurer payments, and intervention rules. One implementation could use x(t+1) = F(x(t),z(t),u(t)) and a separate cash-accounting update c(t+1) = c(t) + inflows(t) − outflows(t). Each function must be explicitly parameterized. The displayed equations are notation for established dynamic modeling choices and do not constitute an independently proposed theory.

Time changes the consequences of correlation. A three-hour outage may fit within a business continuity plan and a contractual waiting period; a three-day outage may exhaust inventory and disrupt an insured's customers; a three-week outage may produce lost contracts or changed market shares. The same graph and initiating event can generate different claims as duration increases. Reinsurance payments can follow substantially after claims settlement, so liquidity effects need a time-indexed cash model even where ultimate contractual recoveries are expected.

In interdependent infrastructure, a network edge may activate only after a period of degraded service. Buldyrev et al. (2010) study failure propagation across mutually dependent networks, while physical systems introduce restore-and-repair pathways absent from a purely irreversible binary deletion process. Operational risk science therefore needs to model recovery and resilience alongside failure. A cascading loss count alone can overstate lasting damage if rapid substitutes are available, and understate disruption if a single disabled service blocks restoration of another layer.

Amplification, fragility, and risk indicators

Connectivity effects are conditional on shock size and node strength. Dense connections can distribute one default's impact over many well-capitalized counterparties; they can also give a severe shock more channels through which to move. Acemoglu et al. (2015) establish a phase-transition result in a particular financial-network setting. That finding can be understood as a threshold interaction between local loss-absorbing capacity and network reachability. It cannot be reduced to the blanket proposition that a more connected system is always safer or always more dangerous.

A linear amplification measure such as ||(I−A)^(−1)s|| / ||s|| depends on the chosen norm, units and initial vector. A centrality score based on A may be informative for some s but poorly predictive for a rare shock that affects unrelated nodes. For a probabilistic model, useful reported statistics include probability of any secondary failure, expected secondary loss conditional on the initiating scenario, distribution of final affected nodes, expected net insurer payments, 99th or 99.5th percentile loss, and required liquidity at specified dates. Capital risk and operational service continuity must be treated as distinct outputs even when their causes interact.

Glasserman and Young (2015) evaluate contagion bounds with comparatively limited knowledge of detailed interbank network structure. Their approach provides a valuable methodological alternative to pretending that a fully observed graph is always available. Structural uncertainty can be characterized by ranges or classes of compatible networks, with bounds on losses, instead of quietly imputing a single undocumented matrix. The review in Glasserman and Young (2016) further cautions that leverage, bankruptcy costs, short-term funding and common exposures shape contagion alongside topology.

Threshold Cascades

Worked linear chain and feedback example

Consider three hypothetical operational nodes A, B and C. Their state variables measure monetary-equivalent business disruptions under a common analytical convention; they are not necessarily insured claim payments. A shock of 100 units first reaches A. B suffers 0.4 units of disruption for each unit at A, and C suffers 0.5 units for each unit at B. There are no other links. With ordering (A,B,C), the nonzero transmission entries are A_BA = 0.4 and A_CB = 0.5. Consequently the direct-and-indirect shock vector is h = (100,40,20). The total 160 is the sum of separate node effects, which may include overlapping foregone transactions; it cannot automatically be added and called 160 of net economic damage. If the nodes have separate non-overlapping capacity-loss measures, one can then define a compatible aggregate metric.

Introduce a feedback dependency from B to A with coefficient 0.2, leaving the other links unchanged. The equations become h_A = 100 + 0.2 h_B, h_B = 0.4 h_A, and h_C = 0.5 h_B. Substituting yields h_A = 100 + 0.08 h_A, so h_A = 100/0.92 ≈ 108.6957, h_B ≈ 43.4783, and h_C ≈ 21.7391. The feedback loop contributes the geometric multiplier 1/(1−0.08), which converges because the loop gain is below one. The three values are illustrative and reproduce directly from the equations; no operational field data or losses of an actual insurer are implied.

This example explains how an indirect supplier feedback can magnify a local shock. It also exposes a model boundary. If a 100-unit initial disruption represents all capacity at A, the uncapped 108.6957 would exceed that capacity and become physically impossible. Saturating losses, restoration, and substitution would require revised equations. Likewise, if A and B record the two sides of the same interrupted sale, their sum measures affected business-unit exposures rather than society's net destroyed output. These are accounting questions, not algebraic errors.

Worked counterparty default chain

Consider another hypothetical network with companies A, B and C. A suffers an exogenous asset impairment of 20 units and, under its assumed pre-shock equity and liability structure, defaults. B holds a claim on A of 30 units; recovery in this simplified scenario is 50%, so B's direct impairment on that claim is 15. B has a loss-absorption buffer of 12, and therefore crosses its stipulated default threshold. C holds a claim on B of 20 units; under the scenario's separate 50% assumed recovery on B, C loses 10. C's buffer is 8, so C too crosses its threshold. The propagation sequence is A → B → C.

All contractual claims and recovery fractions are hypothetical. The binary threshold model is internally consistent with the stated inequalities: 15 > 12 and 10 > 8. It does not endogenously calculate the two 50% recoveries from a simultaneous payment network, so no exact post-default balance sheet or aggregate creditor payout is claimed. The 20 original asset impairment and the 15 and 10 creditor impairments should not simply be counted as 45 of independent final economic loss. The latter are claims revaluations and redistribution of credit losses within a defined financial perimeter; consolidated wealth and bankruptcy costs require a separate accounting exercise.

Now change only B's capital buffer from 12 to 18 while leaving the same assumed recovery on A. B absorbs the loss of 15 and does not default under the binary rule. Consequently the stipulated secondary default-induced 10-unit loss at C does not occur. Network topology is unchanged, but the cascade stops. This makes precise the importance of buffers, recoveries and thresholds, and illustrates why a degree-only ranking misses the system's state-dependent stability. It also shows how counterparty stress calculations should be tied to actual liabilities and group capital rather than borrowed wholesale from an epidemiological graph.

Conditional connectivity and topology comparisons

Compare a concentrated star graph with a diversified graph. A central institution owing a very large fraction of liabilities to one weak creditor can expose that creditor to a severe loss on default. If the same total obligation is divided among many creditors with adequate buffers, the loss to each may remain below its individual default threshold. This is a diversification benefit. A contrasting shock that simultaneously disables the center can nevertheless affect more nodes in the diversified topology. The direction of the net result depends on external asset distribution, debt seniority, and threshold allocation; an unweighted graph alone cannot determine it.

In Financial Networks and Contagion (2014), diversification and integration have distinct nonmonotonic effects. More links may introduce new paths; greater spread of exposure may reduce each link's severity. In Systemic Risk and Stability in Financial Networks (2015), the magnitude of the original shock determines which connectivity mechanism dominates in the modeled setting. For research interpretation, the lesson is to hold aggregate claims, capital, and shock severity constant when testing one topological change, or explicitly explain which variables necessarily vary along with the topology.

A topology comparison also needs care with counterfactual construction. Rewiring an obligation without changing the identities and capacities of the creditors can reveal a pure network effect. Moving obligations to an institution with a different asset base changes both topology and node vulnerability. Introducing a common catastrophe risk to every counterparty can dominate distinctions in graph structure. Comparative research therefore reports an exposure-preserving scenario, a node-capital sensitivity, and a common-shock sensitivity rather than treating one network diagram as evidence of a universal ranking.

Reinsurance layer illustration and scenario design

Suppose an insurer faces an illustrative covered event loss of 70 monetary units. It purchased a simple per-event excess-of-loss treaty attaching at 50 with a limit of 100. Assuming the loss is fully eligible, the nominal reinsurance recovery is R(70) = min{max(70−50,0),100} = 20. The insurer's gross covered payment obligation remains 70; if the reinsurer pays in full, its retained ultimate amount is 50, subject to premium and other accounting terms. If only 50% of the 20 recoverable is actually collected because of a specified reinsurer credit event, the insurer's ultimate borne amount rises to 60. The 10 difference is credit impairment of the expected recovery, not a second hurricane or a new primary claim.

The contract is nonlinear. For eligible event losses of 40 and 50, R(x)=0. At 70 it pays 20; at 150 it pays 100; and at 200 it remains 100 after exhaustion. Therefore a change in gross loss of 10 can create zero, ten, or less than ten additional recovery depending on the layer. Treating the 70-unit scenario as a homogeneous proportional ceded fraction and extrapolating it to all event sizes misstates both tail retention and counterparty dependency. Klages-Mundt and Minca (2020) examine the significance of this kind of reinsurance contract nonlinearity in network contagion.

Real treaties also introduce hours clauses, reinstatement premiums, multi-event definitions, aggregate limits, exclusions, claims development, collateral, retrocession, contract priority and dispute resolution. A scientific model selects and encodes the provisions relevant to the asserted conclusion. No generic reinsurance-network simulation can determine an insurer's actual recoverable without contract data. The worked calculation is deliberately narrow so that every result follows visibly from stated terms.

From deterministic cascades to probability distributions

A deterministic cascade scenario reports what happens conditional on its inputs. Insurance capital and pricing require the frequency and severity distribution of such scenarios, including uncertainty about hazard intensity, network membership, and state-dependent recovery. Let S be a random exogenous event, Θ the structural and contractual parameters, and L(S,Θ) the insurer's net covered claim after defined recoveries and limits. Then the risk distribution derives from uncertainty in S and Θ and the model mapping. Assigning a return period or a 99.5th-percentile number requires a justified probability model for those inputs. One cannot infer a plausible annual probability merely because a modeled cascade appears severe.

The research distinction is particularly important for low-frequency, high-severity events. A scenario that is extreme in conditional financial effect may be unlikely, poorly identified, or structurally incompatible with certain counterparty behaviors. Conversely, a network that has not experienced a severe historic cascade may still possess the necessary configuration. The robust-yet-fragile network results of Gai and Kapadia explain why stress periods should be treated as distinct from routine experience rather than extrapolating stability in calm samples.

Shared Cyber Dependencies

Digital service networks and insured loss

Cyber loss aggregation is often driven by common operating dependencies rather than by direct exchange of risk-bearing liabilities. An insurer may write policies for thousands of companies that all authenticate through one identity service, store data with one cloud provider, distribute software through one update mechanism, or route payments through one processor. In such a portfolio, a single upstream event can disable policyholders that otherwise have distinct industries, geographies, and balance sheets. The relevant network is a many-to-many dependency map: insured organization → operational service → underlying technology provider → shared infrastructure. The economic claim network is then attached through the policy coverage terms.

A compromise that spreads laterally from one enterprise to another represents one propagation mechanism. A central provider outage that simultaneously harms several customers is a common-cause mechanism. A software update delivered to many systems is a shared distribution channel; the mechanism may be propagation of malicious code, common configuration error, or a defective update. These mechanisms can coexist within a particular incident, but they imply different likelihoods, mitigations, event definitions, exclusions, and forensic evidence. A graph model that uses the same undirected edge for all of them discards material causal information.

Biener, Eling, and Wirfs (2015) examine the historical conditions for cyber-risk insurability, including difficulty estimating losses and dependence. Their dataset and period cannot be treated as a census of contemporary global cloud concentration, yet the underlying risk-science questions persist: what is the event unit, what losses are covered, and which joint failures remain plausible? The Geneva Association's cyber accumulation study further discusses the need to represent systemic cyber events, accumulation and risk transfer together. These sources justify explicit scrutiny of concentration; they do not justify a single worldwide percentage for cyber loss correlations.

A common-service probability counterexample

Consider four hypothetical insured businesses. Each suffers six monetary units of eligible loss if a designated software-as-a-service provider has an annual outage of a stipulated duration; otherwise its loss from this cause is zero. The provider fails with probability 0.02 for this illustrative year, and all four businesses depend on that single provider. The total portfolio loss attributable to the specified provider event is therefore 24 with probability 0.02 and zero with probability 0.98. Expected loss is 24 × 0.02 = 0.48, and variance is 24² × 0.02 × 0.98 = 11.2896.

Now construct a comparison portfolio in which the four six-unit losses are triggered by four independent providers, each with the same stipulated 0.02 outage probability. Expected total loss remains 4 × 6 × 0.02 = 0.48. Variance becomes 4 × 6² × 0.02 × 0.98 = 2.8224. Under the lower-quantile convention VaR_0.99(X) = inf{x: P(X ≤ x) ≥ 0.99}, the fully shared-provider portfolio has VaR_0.99 = 24, because the probability of zero loss is 0.98. For the independent-provider comparison the probability of no outage is 0.98⁴ = 0.92236816, and the probability of at most one outage is 0.98⁴ + 4(0.02)(0.98³) = 0.99766352; consequently VaR_0.99 = 6. The common provider quadruples variance relative to the independent comparison and raises the 99th-percentile outcome from 6 to 24, while expected loss is unchanged.

These values are mathematical scenarios, not cyber-frequency estimates, loss predictions, or quotations from insurance-industry statistics. The calculation assumes binary annual events, identical payout severity, no deductibles or sublimits, and no additional causes of loss. A real cloud outage may harm customers unevenly, be mitigated by fallback systems, produce uncovered losses, or occur together with other vendors' failures. Those considerations alter the result's inputs rather than its arithmetic. The example isolates why underwriting an apparently diversified portfolio by business sector can conceal an operationally concentrated exposure.

Frequency models, epidemic processes, and common causes

A contagious process and a common-cause outage should not be assigned the same event-generation equation. In a simple contact process, a node may transmit infection or compromise to neighboring nodes along a directed contact network. In a common-vendor model, one exogenous provider state activates many downstream losses simultaneously, conditional on the duration of lost service and each customer's business continuity provisions. The two mechanisms can be nested: a compromised vendor introduces malicious artifacts at many organizations, after which lateral spread proceeds over their distinct contact networks.

In percolation-based spreading research, a transmissibility parameter combines with a graph's excess-degree structure to influence the possibility of large outbreaks. Newman (2002) derives epidemic results for specified network and transmissibility assumptions. Its threshold logic is informative when constructing a stylized cyber-spread model, but biological contact and cyber compromise differ in detection, adversarial adaptation, access controls, credential reuse and intervention. A direct transfer of disease-transmission rates to enterprise cyber incidents would have no empirical foundation. The governing task is to translate mechanisms, not to import coefficients.

Different cyber events require different event windows. A ransomware campaign may reach organizations over weeks; a dependency failure may have a single start time but heterogeneous restoration times; an exploited software library can create latent vulnerabilities that are monetized later. The number of exposed software installations is therefore not the same as the number of affected firms, and the number of affected firms is not the same as the number of covered claims. A valid loss map carries distinct variables for technical exposure, compromise, service impact, contractual attachment and monetary settlement.

Dependency layers and a loss map

For an insurer's policyholder i, let D_i denote its dependence on service vendors, T_v an outage or compromise state at vendor v, M_i the resulting measurable operational effect, and C_i the policy's claims function. A transparent general specification is L_i = C_i(M_i(D_i,T),W_i), where W_i contains relevant contract terms such as waiting period, retention, limit and exclusions. Portfolio gross insured loss is Σ_i L_i. This notation simply separates established modeling objects; its form supplies no calibrated probability law. If the insurer has reinsurance, the net amount is derived by applying each treaty to the legally defined eligible event set, rather than subtracting a fixed percentage from the entire gross loss.

The definition of event and related claims is particularly important. If two insured customers suffer interruption from the same third-party outage, their underlying lost revenue may be separate insured claims. If the provider also faces liability claims for the same customers, the claim graph contains several legal obligations linked to partly overlapping business losses. Contract scope, double-insurance rules, subrogation and recovery determine what is payable. A simplistic sum of all declared losses across insurers, insureds and suppliers can count several financial claims on the same underlying economic disruption.

The Lloyd's Cloud Down analysis offers an explicitly modeled stress scenario for cloud-service interruption and connected economic losses. Its scenario estimates are useful evidence about a class of hypothetical events, not records of observed claims resulting from one actual incident. Scenario output should be quoted only with assumptions, valuation perimeter and event window. Where those inputs are unavailable, the robust inference is that cloud dependency permits correlated interruption; precise portfolio capital remains to be estimated from a documented model.

Concentration can migrate across apparent substitutes

Two named cloud vendors do not necessarily constitute independent alternatives. They may rely on the same physical connectivity, software dependencies, identity services, domain-name resolution, electric grid, regional facilities or outsourced maintenance provider. Conversely, applications nominally hosted on one platform may have different failure domains because of regional architecture and active failover. Exposure aggregation must follow the actual dependency graph down to the level that matters for the initiating hazard. Vendor identity alone is an incomplete concentration unit.

The FSB third-party risk toolkit (2023) addresses the financial system's dependencies on external service providers and the need for oversight and coordination. The toolkit is addressed to financial institutions and authorities; it does not supply the policy-level cyber severity model an insurer needs. The NIST supply-chain risk guidance and NIST CSF 2.0 provide operational practices and governance categories that can be mapped to node vulnerabilities and control assumptions. Their programmatic categories should not be treated as probability estimates or actuarial rate tables.

Cyber-risk controls enter the network model as causal modifiers. Service segmentation reduces pathways for compromise; independently validated recovery environments can reduce outage duration; supplier exit plans can lower switching times; multi-factor authentication can reduce certain credential attacks; offline backups can reduce loss severity conditional on compromise. Each proposed modifier requires evidence appropriate to the specific organization and attack mechanism. A questionnaire answer that a control exists is not equivalent to an observed reduction in the frequency of insured claims.

Regulatory and supervisory applicability

European Union Regulation (EU) 2022/2554, DORA, applicable from January 17, 2025, creates a financial-sector operational-resilience framework that addresses ICT risk, incidents, testing, third-party services and concentration. Its covered entities and exemptions require reference to the regulation and subsequent applicable technical measures; it is not a globally uniform contract law. EIOPA's DORA overview summarizes the supervisory significance for insurance and pensions within the European regime. DORA can support a documented approach to ICT supply-chain risk while leaving actuarial calibration, insurance policy coverage and national contract interpretation to their separate sources.

The IAIS operational-resilience application paper supplies more recent international insurance-supervision context for critical operations and dependencies. IAIS supervisory guidance and the legally applicable rules of a particular jurisdiction are distinct types of authority. A sound multinational analysis labels the relevant year and jurisdiction, identifies whether a provision is binding, and reports how operational dependencies affect insurer or policyholder exposure instead of asserting a universal regulatory formula.

Contingent claims and reinsurance counterparty feedback

Cyber claims can connect operational and financial networks over time. A large shared-service outage can trigger claims under multiple insurers; each insurer may seek reinsurance recoveries from overlapping counterparties. If a common reinsurer faces multiple requests simultaneously, the original operational shock can become a credit, claims-settlement and liquidity shock. The liability of the reinsurer remains contract-specific. Claims from one underlying incident may attach differently to occurrence, aggregate and cyber-specific treaties. Recovery rates and collateral can change after the incident, and even solvent reinsurers may reimburse after primary claim dates.

This layered pathway is analytically more informative than assigning every cyber policyholder a direct link to every reinsurer. Underwriter records reveal insured assets and coverage; vendor records reveal service dependence; event forensics reveal affected systems; treaty schedules reveal ceded protection; and financial statements reveal recovery and liquidity capacity. Each layer has its own evidence quality, update frequency and confidentiality limitations. A portfolio risk metric is defensible only if the transformations between layers are specified and independently checked where possible.

Observability and Model Limits

What counts as a measured edge

Network graphs are sometimes drawn with more precision than the underlying data support. A measured obligation edge may originate in a legal contract with amount, seniority, maturity and governing law; a vendor dependency edge may originate in a service inventory or architecture map; an inferred contact edge may be estimated from incident traces; a common-factor edge may express statistical association rather than a bilateral relationship. These edges should be reported with different evidentiary designations and with the time as of which they are known. Incomplete records can produce high apparent centrality at one well-documented node and invisibility at a poorly documented one.

Observed co-losses cannot by themselves identify propagation. Two insurers may suffer simultaneous losses because they insure the same region; because a reinsurer defaulted after a catastrophe; because both selected similar investment assets; or because one insurer's financial distress impaired the other. The data may exhibit the same positive correlation in all four mechanisms, while targeted prevention differs sharply. A statistical covariance matrix describes joint variation; a causal network needs evidence of direction, timing and the counterfactual effect of changing a dependency.

One way to structure empirical identification is to distinguish three hypotheses: common exposure creates co-loss; a direct obligation transmits a counterparty impairment; or a shared operational dependency transmits service failure. Evidence can come from legal obligations, physical flow records, detailed incident chronology, contractual clauses, intervention dates and natural experiments. Every identification claim should explicitly address reverse causality, omitted common factors, selection into the observed network and reporting bias. Where such evidence is unavailable, scenario analysis with clear assumptions is appropriate; a causal conclusion is not.

Incomplete financial network observation

Public insurer reporting often aggregates counterparty information by broad classes rather than disclosing every ceded exposure, retrocession link, collateral arrangement and inter-company claim. A fully observed graph is rarely available outside specialized supervisory datasets or privileged market records. Aggregation can conceal dominant bilateral claims or important maturity mismatches. One must resist a reconstructed graph that appears complete merely because each node has been assigned one plausible link.

Glasserman and Young (2015) develop an approach to contagion likelihood that does not rely on assuming perfect information about detailed networks. This supports bounding exercises with available exposures, buffers and leverage rather than arbitrarily choosing one convenient adjacency matrix. Glasserman and Young (2016) review several channels of contagion and distinguish direct claims from other amplification pathways. The practical research procedure is to mark an edge as observed, inferred or stress-imposed; carry its uncertainty into scenario outcomes; and report the range of plausible cascade losses conditional on that information.

One transparent analysis can hold marginal exposure totals fixed while testing extreme allocations of the bilateral matrix. Another can sample matrices compatible with known insurer balance-sheet constraints, clearly stating that the sample represents structural assumptions rather than empirical measured network frequencies. If network reconstructions imply contradictory balance-sheet totals, that is a data and model validation failure. System stability conclusions should be withheld until the mismatch is corrected.

Risk metrics and denominator discipline

For a random portfolio net insured loss L over a specified year, the mean E[L], standard deviation, VaR at specified probability, expected shortfall, probability of exhausting capital and probability of failing a dated cash payment answer distinct questions. Their denominators and boundaries must be specified. Counting nodes that fail, counting policies that make claims and counting monetary units of claims are not interchangeable. A network can be resilient under the failed-node measure yet generate a material loss through degradation of one highly insured node.

Risk metrics can reverse model rankings. In the four-policy cyber example, expected loss is the same under common and independent supplier states, while variance and the 99th-percentile loss differ. A manager ranking only expected losses would treat the cases as equivalent. A manager ranking only connected component size might reach yet another answer. The appropriate collection of metrics follows the insurance decision: pricing, underwriting limits, capital, collateral, liquidity, accumulation control or operational continuity. A single centrality scalar is not a complete risk profile.

The systemic risk metrics in Adrian and Brunnermeier's CoVaR analysis and Acharya et al.'s systemic expected shortfall study concern conditional tail vulnerability and systemwide capital stress, primarily in a financial-institution context. They need not identify the same thing as a propagation path in a legal obligation network. Systemic risk contribution may be associated with common exposures even without one institution's failure mechanically causing another. The scientific responsibility is to preserve that distinction while explaining when indicators complement structural stress testing.

Model selection and comparative stress design

Models should be selected according to the mechanism and evidence, not by mathematical fashion. A linear network model is useful for incremental service disruption inside bounded ranges. An Eisenberg–Noe clearing model is useful for internally consistent proportional repayment under its defined legal and accounting assumptions. A threshold cascade highlights capital or capacity discontinuities. An epidemic process captures a specified transmission mechanism. A network input-output model describes dependence among producers under its own production assumptions. A multi-layer cyber-insurance simulation can connect several of these, but each bridge between layers must be calibrated or acknowledged as a scenario premise.

The literature itself helps check competing mechanisms. Eisenberg and Noe (2001) provide a consistent payment-clearing framework. Watts (2002) studies threshold-based cascades on random networks. Acemoglu et al. (2012) demonstrate the effect of production-network heterogeneity on aggregate fluctuations, and Carvalho and Tahbaz-Salehi (2019) survey production-network structures. Each addresses a different estimand and set of mechanisms. An insurer applying one model to all financial, cyber and physical propagation without adapting the state variables would make the interpretation unreliable.

A disciplined stress program varies initiating hazard, exposure allocation, network topology, node resilience, contract terms, capital, liquidity, counterparty recoveries and recovery time separately where possible. A useful parameter grid records which change is controlled and which parameters must move jointly for accounting consistency. Report sample-size and convergence diagnostics when Monte Carlo simulation is used. For rare cascading tails, naive sampling may observe too few failures to estimate a high quantile reliably; importance sampling or targeted stress constructions can be explored, with their design assumptions disclosed. A larger number of simulated scenarios does not fix a misspecified transmission graph.

Validation, falsification and out-of-sample testing

Validation begins with accounting and mathematical identities: row and column sums of contractual matrices, nonnegative recoveries, legal priority, finite liability bounds, capacity floors, event timing and prevention of duplicated claims. Sensitivity analyses then ask whether outcomes depend on a small number of undocumented links or arbitrary thresholds. Empirical validation compares predictions with held-out incidents or operational disruptions, but the investigator must state which event types and severities the observations cover.

A model can match ordinary claims frequencies while missing extreme shared-service events, because the frequency sample is uninformative about rare tail dependence. A model can reproduce historical bank defaults while misrepresenting insurers whose contracts, capital regulation and reinsurance structures are different. Backtesting must therefore combine event-level outcome checks with process checks: were the exposures known at the relevant date; were policies in force; were the contract terms captured; did settlement occur inside the forecast horizon; and were the reported historical losses revised afterward? An apparent forecast error may come from versioned exposure data rather than the hazard distribution.

For forensic reproducibility, maintain a versioned parameter table, source URLs and retrieval dates, exposure metadata, the precise network direction convention, code version, random seed where applicable, scenario weights and derivations. If confidentiality prevents releasing raw policyholder or counterparty data, an aggregated synthetic example can demonstrate mathematical behavior but should be labeled illustrative. The 3-node feedback, 3-node default chain, reinsurance layer and 4-policy common-vendor examples in this manuscript are fully specified on that basis. They require no privileged insurer data and are reproducible from their stated assumptions.

System boundaries and economic double counting

The accounting boundary should be decided before aggregation. Economic output lost by a factory, the insured indemnity paid to the factory, the reinsurance recovery obtained by the insurer and the reinsurer's payment are separate cash flows associated with partly overlapping consequences of one disruption. Adding all four as if independent national economic damage is incorrect. Likewise, a creditor's impairment and a debtor's unpaid liability are two views of one obligation; bankruptcy costs and destruction of production capacity may be additional losses, but require separate measurement.

If analysis seeks total direct insured claims across policies, each policy contract must be treated as a distinct payment obligation, and overlap or contribution provisions must be respected. If analysis seeks consolidated insurer net loss, payments and recoveries between entities inside the consolidation perimeter require elimination as appropriate. If analysis seeks social welfare loss, transfers between parties and actual resource destruction are treated separately. Confusion among these three ledgers can change a catastrophe accumulation estimate by orders of magnitude without any change in the physical hazard.

This issue also affects financial stability. A reinsurer's payment may reduce one insurer's loss while increasing the reinsurer's loss, leaving the system's combined underlying catastrophe liability largely unchanged; the transfer can nevertheless distribute the loss toward better-capitalized entities and improve claims-paying continuity. A reinsurance default can reverse that redistribution and cause liquidity strain. Interpreting networks as graphs of linked legal obligations helps track these changes while keeping aggregate welfare accounting distinct.

Insurance supervisory implications

The IAIS ICP and ComFrame framework provides internationally oriented principles for supervision, risk management and capital-related practice. Its online text and adoption history matter because individual standards evolve, and adoption into binding domestic law varies. The framework supports assessing material risk concentrations and outsourced functions, but it cannot be read as requiring one specific network measure or graph topology. The IAIS Holistic Framework addresses the monitoring and mitigation of insurance-related systemic risk, including cross-sectoral channels. It provides context for collecting network-relevant evidence, without establishing that any identified interconnectedness will create a cascade.

At the regional level, DORA's ICT third-party risk provisions make digital dependencies an explicit governance and oversight concern for in-scope European financial entities. At the international level, the FSB third-party risk toolkit emphasizes managing third-party risks and identifying concentration and potential systemic effects. These instruments differ in legal character and institutional audience. Their shared practical relevance lies in the need for an accurate dependency inventory, governance accountability, incident reporting, resilience planning and proportionate stress assessment.

Risk reporting for an insurance group should connect network outputs to underwriting concentration, reinsurance counterparty quality, investment exposures, liquidity resources, operational outsourcing and solvency planning. A quantified expected loss may belong in pricing; a plausible low-frequency cascade may belong in capital scenarios and ORSA-style risk assessments; a critical-service failure may require a continuity test. The level of technical precision in any report should match what was actually measured and verified. A glossy network visualization is a weak substitute for a balance-sheet-constrained loss mapping and an independently reviewed assumption set.

New evidence and counterevidence that would change the assessment

The most consequential missing information often includes detailed treaty attachment and aggregate limits, collateral status after a catastrophe, actual vendor-to-vendor dependencies, time-resolved insurer payment flows and observed interruption-duration distributions. New verified data on any of these could materially change tail estimates and reverse a centrality ranking. Strong counterevidence might show that a purported shared provider dependence is segregated across independent failure domains, or that supposedly independent vendors share a hidden infrastructure bottleneck. Neither result invalidates network modeling; both improve the factual graph.

A finding that an apparent cascade was actually several simultaneous responses to a common external shock would change causal attribution. A finding that multiple creditors' impairments were all versions of one already accounted-for loss would change aggregate economic damage. A finding that rapid contractual recoveries occur before cash shortages become binding would change a liquidity cascade, even if ultimate asset impairment remains. These distinctions should be treated as testable research questions rather than residual uncertainties buried under an aggregate risk number.

Original analytical synthesis

The evidence supports five connected conclusions. First, graph structure is informative only together with a defined state variable and edge meaning: contractual exposures, supplier dependence, causal transmission and statistical correlation cannot be interchanged. Second, amplification depends on the shock domain and resilience of the nodes; a link can diversify one small loss and transmit a severe common failure. Third, insurance-network risk must pass through legal policy and treaty mappings before a graph perturbation becomes a claim or net insurer liability. Fourth, a common operational vendor creates a form of concentrated loss dependence even across otherwise diversified insured businesses. Fifth, the principal obstacle to reliable large-scale network estimates is often observability of topology, contract terms and event timing, not the absence of sophisticated mathematical machinery.

These conclusions are an analytical synthesis of established research and the transparent examples in this work. The linear multiplier and the four-policy vendor comparison are elementary derivations under fully disclosed assumptions; the default and reinsurance examples are conceptual applications of existing models. They are not asserted as new scientific theories, new systemic-risk indices or discoveries of previously unknown empirical facts. Their scientific contribution within the dossier is to clarify what mathematical inferences can legitimately pass from network topology to insured claims and financial capacity, and where extra evidence is indispensable.

Conclusion

Network models describe risk propagation by making relations among exposures, counterparties, services and institutions explicit. The underlying relations determine the appropriate mathematics: bilateral obligations require financial clearing and recovery assumptions; physical production and infrastructure dependencies require capacity and time; contagion requires a transmission rule; shared digital suppliers require a joint-event model; and insurance payments require contract-by-contract mapping to eligible claims. A single incident can connect several of these layers, but its loss must be accounted for consistently at each layer.

The central scientific result is that network connectivity has no unconditional sign as a determinant of stability. It can distribute shocks across strong nodes or enable failures to reach additional entities, depending on the initiating shock, the distribution of claims, capitalization, contractual thresholds, alternative capacity and common-cause exposures. Linear multiplier stability, threshold cascade behavior and clearing consistency are each conditional statements tied to a particular model. The worked scenarios here demonstrate those conditions explicitly without presenting their illustrative numbers as empirical market estimates.

For insurance and risk science, the practical standard is to define the loss perimeter and decision horizon; identify exposure and contract relations; model common shocks separately from subsequent transmission; retain nonlinear policy and reinsurance terms; examine buffer, liquidity and recovery timing; test plausible alternative networks; and disclose the source and uncertainty of every critical edge. Expected loss, tail loss, cash shortfall, service downtime and aggregate economic damage should be reported as different quantities. Regulatory sources supply governance and supervisory context, while jurisdiction-specific applicability and actual contract provisions govern legal consequences.

The original synthesis is therefore a discipline of interpretation: mathematical propagation, insured liability and system-level economic loss are related through evidence-dependent mappings whose assumptions must remain visible. The dossier offers an integrated scholarly account, reproducible examples and a structured validation agenda. It creates no new proprietary theory or doctrine and attributes no publication date to this unpublished source manuscript.

References