RESEARCH DOSSIER
Cyber Claim Frequency and Severity Joint Modeling
Abstract. Cyber insurance liabilities arise through a sequence of technical events, economic damage, contractual eligibility, notification, adjustment, and payment. A joint model of claim frequency and insured severity must represent this sequence and the dependence introduced by both underlying risk and the insurance contract. This monograph develops a critical synthesis of collective risk theory, count and severity regression, mixed discrete-continuous dependence models, marked point processes, shared vulnerability exposure, and aggregate loss calculations. Original empirical cyber-event studies, actuarial modeling research, an Australian cyber policy wording, European reporting legislation, supervisory communications, and professional standards are compared according to the populations and outcomes they actually observe. The analysis derives the expected aggregate loss under count-dependent severity, distinguishes event-level and policy-level weighting, demonstrates the nonidentifiability of incident intensity and reporting propensity from observed claims alone, and quantifies the effect of common risk states through transparent synthetic calculations. It also examines deductibles, exhaustion, reporting delays, incomplete monetary records, and the separation of within-policy dependence from cross-policy accumulation. Results establish the conditions under which separate frequency and severity regressions can support expected-loss estimates and the additional information necessary for tail risk, reserving, and reinsurance applications. The contribution is an analytical integration and comparative evaluation of existing methods; no independently calibrated cyber distribution or new named theoretical system is claimed.
Keywords: Cyber insurance; claim frequency; claim severity; collective risk model; conditional severity; marked point processes; exposure; insurance deductibles; reporting selection; copulas; accumulation risk; actuarial validation; claims reserving.
Introduction
The actuarial object of this investigation is the distribution of the amount payable under a set of cyber insurance contracts during a specified risk period, including the relationship between the number of claims and their financial magnitude. The scientific problem begins with the observation process. A cyber incident database may record a technical compromise, a publicly disclosed breach, or a regulatory notification. The claim system records another population, constituted through insurance contract terms, policyholder decisions, and the insurer's handling process. The difference is substantive: the probability of entering a claims database is related to both the magnitude of an incident and the institutional conditions that govern coverage.
A complete account of the joint distribution must therefore distinguish three separate questions. First, what generates the occurrence of adverse events in insured entities? Second, how do the repercussions of these events, when passed through policy terms, become a distribution of insurance payments? Third, how are occurrence and magnitude related, both within an insured risk and across a portfolio? These questions have different denominators, likelihoods, and validation requirements. Identifying a dependence parameter without specifying the population to which it belongs produces a number without a stable actuarial interpretation.
The methodology is a structured narrative and mathematical synthesis of research, rather than a systematic review with an exhaustive search protocol. The evidence corpus is deliberately heterogeneous: original empirical studies, original mathematical research, a survey, a specimen contract, legal provisions, supervisory and insurer communications, and professional standards. Each is used within its own scope. The observed results of published studies are reported as such. All numerical illustrations introduced below are analytical or synthetic, use stated parameters, and require no access to a proprietary claims file. Model equations describe established statistical constructions and logical consequences of their assumptions. The investigation does not estimate current insurance rates, infer a universal relationship between attack frequency and losses, or provide a compliance opinion on a particular insurance contract.
The historical and geographical emphasis of the available empirical material is uneven. U.S.-dominated cyber-event data are considered alongside Australian policy language and European regulatory duties. This composition permits an international comparison of measurement and institutional mechanisms, but not a pooled international estimate of incident or insured-claim frequency. The evidence cutoff is November 18, 2026. The research belongs to Cyber Insurance within Insurance by Bogdanova.
Conceptual and Historical Foundations
The insured loss as a sequence of observations
Let an enterprise be exposed to events that impair confidentiality, integrity, or availability of digital assets and services. The insured system is defined sufficiently broadly to include relevant third-party dependencies, subject to the scope of the coverage. An attack attempt, a successful compromise, a material incident, an economic loss, a report to the insurer, a recognized claim, and a positive payment are distinct possible stages. An event may pass through all of them, stop at an earlier stage, or produce several claim files.
Consider an organization that receives a malicious message. The receipt of that message may be observed in security telemetry. A successful compromise of credentials may subsequently lead to a fraudulent transfer, a network interruption, or no material loss. If it produces a loss, the business may recover the money independently or may notify its insurer. If a particular coverage grant responds, the insurer applies the relevant conditions, deductibles, exclusions, and limits. The resulting payment is a financial consequence of both the underlying event and its contractual treatment.
The same causal origin may subsequently generate privacy litigation. An event-level file may then be associated with first-party restoration expense, business interruption, third-party defense costs, and liability indemnity. These components mature at different speeds. Recording each component as an independent cyber incident would artificially increase frequency and erase the economic relationship among components. Recording everything as one undifferentiated file would obscure the entitlement and development structure needed for pricing particular coverage grants.
The modeling unit must accordingly be declared before estimating any distribution. Useful units include events per organization-year, reported claim files per policy-year, paid claims per coverage-year, and occurrence clusters per portfolio-year. A report may also represent a notice of circumstance with zero payment. No transformation makes these units equivalent merely because they are counted in the same database.
The collective risk model and the meaning of a claim mark
Let N be the count of claims of a specified type during the exposure period. Let Y_j denote the nonnegative ultimate insured amount of the jth claim. The total is
In the classical compound model, N is independent of an identically distributed sequence of mutually independent severities with mean mu and variance sigma squared. If the relevant moments exist, iterated expectation and variance yield
The first identity follows by conditioning on N; the second adds the expected conditional variance to the variance of the conditional mean. With Poisson frequency of parameter lambda, the aggregate variance is lambda times the second raw severity moment. With an overdispersed frequency, the count contribution to variance may be substantially greater even when average frequency remains unchanged.
The collective model is a valuable benchmark because it makes the independence restrictions explicit and permits the investigation of deviations. Awiszus et al. (2023), in their survey of cyber insurance modeling, distinguish multiple sources of dependence and examine how actuarial loss models can be adapted to a rapidly changing risk environment. The benchmark is therefore the starting point for comparing richer models rather than an assertion that independent cyber claims characterize every portfolio.
In a marked point process, the claim mark can encompass amount, event cause, timing, affected service, reporting status, coverage component, and insured identity. Zeller and Scherer (2022) develop a marked point process approach that accommodates different classes of attacks and nonmalicious failures. Their formulation motivates a practical separation of arrival dynamics and loss descriptions while preserving the information linking them.
This monograph uses a mark as the associated set of variables on an event or claim, not necessarily a scalar. Defining marks at the technical incident level makes it possible to simulate insurance outcomes under multiple hypothetical coverage arrangements. Defining marks at the payment level permits direct prediction for existing contracts; transfer to a materially different contract then demands further assumptions.
The historical empirical transition
Early empirical cyber research drew heavily on descriptions of publicly known breaches and associated corporate losses. The study by Romanosky (2016), Examining the costs and causes of cyber incidents, integrates several event categories and distinguishes reported corporate expenses from other dimensions of social damage. Its observations expose a fundamental source of uncertainty: the availability of cost information is substantially more limited than the availability of incident records.
Subsequent research expanded the modeling of event types, data breach sizes, and dependencies. The resulting evolution can be understood as a move from measuring isolated losses toward characterizing the processes connecting losses. A data set of affected records is useful for privacy-event analysis; it supplies neither the number of exposed insured policies nor the ultimate indemnity unless those quantities are independently observed or credibly linked. Policy-level insurance experience and infrastructure exposure maps answer different questions.
Historical comparisons require particular care when the product itself changes. A cyber insurance program that adds contingent business interruption creates a new path from a supplier incident to the policyholder's claim. Growth in insured claims after such a coverage expansion might reflect an increase in coverage rather than in underlying attack activity. A pricing model trained before the expansion must therefore be recalibrated or explicitly translated through the changed coverage terms.
Three scales of statistical dependence
Within-policy count-severity dependence concerns how the expected or complete distribution of amounts varies with the count in the same exposure period. Within-event dependence concerns relationships among financial loss components or among several affected insured entities. Cross-policy dependence concerns the joint distribution of different policyholders' annual totals and may include repeated events, common risk factors, and infrastructure failures.
These scales can occur together, but identification at one scale does not establish dependence at another. A portfolio may have little observable association between each policy's count and its average claim amount yet have a substantial common-event tail. Conversely, a portfolio with strong positive within-policy association may still have near-independent policy totals after conditioning on observable covariates. Each scale must be represented separately if the intended use depends upon it.
The economic role of a joint model is to connect the distribution of individual risks with the distribution of financial obligations. Contractual design determines which consequences from the technical event process become obligations; correlated cyber environments determine how many obligations may arise together. Neither marginal accuracy nor an appealing narrative about attack activity alone establishes aggregate accuracy.
Primary Documents and Scope of Frequency Models
The denominator and the observation window
A count model begins with the exposed population. Let e_it be the fraction of a year that organizational risk i participates in the relevant insurance portfolio during interval t. A Poisson regression can be written
The exposure offset fixes the coefficient of log exposure at one. Its use assumes the expected count scales proportionately with the time interval for an otherwise unchanged risk class. Short-duration policies written in concentrated threat periods may violate this approximation; integrating a time-varying hazard over the actual covered interval is preferable in that situation.
A policy-year offset is appropriate for many insurance counts, whereas endpoints, employee accounts, revenue, and transaction values can characterize the underlying attack surface or economic stakes. Such quantities are normally predictors, unless there is justification for treating them as proportional exposure. A firm can double its endpoints while centralizing security controls, and its expected incidence may not double. Distinguishing offsets from estimated effects prevents imposing the conclusion through model structure.
Exposure data must include risks with zero reported claims. A claim table by construction lacks many zero-count observations. Counting from this table alone gives a frequency conditional on appearance in the claim data, not a population rate. Joining claims to a complete history of policy inception, cancellation, renewal, endorsement, and coverage status is an essential data preparation step.
Calendar-year counts answer questions about reports or payments received in a period. Accident-year counts group events according to their chosen loss-origin date. Policy-year counts group losses attributed to contracts commencing in a period. A claims-made trigger may require still another attribution convention. The choice has consequences for reported frequency and reserves when occurrence, discovery, report, and settlement are separated by long delays.
Overdispersion, heterogeneity, and zero counts
Under a Poisson distribution, variance equals mean. If dispersion remains materially greater after including relevant exposure and covariates, an alternative is the negative binomial with mean m and dispersion k, parameterized so
This can arise from a Poisson count with a gamma-distributed random intensity. The limiting case as k tends to infinity returns the Poisson variance. Overdispersion may reflect unobserved risk classes, common environment, repeated attacks, changing reporting, or inadequate covariates. The choice of a negative binomial distribution accommodates variation without resolving which mechanism generated it.
A zero-inflated count distribution mixes a structural-zero regime and an ordinary count regime. A hurdle model models the probability of at least one claim separately from the positive count distribution. With p equal to that probability and f_+ a normalized distribution on positive integers,
A hurdle is useful when the transition into report or claim handling is analytically distinct from additional claims once a risk has entered that process. A zero-inflated distribution is useful when a well-supported latent or observable class has effectively no opportunity to experience the recorded event. Cyber insurance data with many zeros do not automatically establish the existence of a structural-zero class: low intensities, short observation, and high deductibles can produce identical zero frequencies.
A direct comparative check is to estimate the proportion of positive exposures and the conditional distribution of positive counts separately, then compare fitted zero proportions across exposure bands and time periods. If the model misses the number of risks with at least one claim while accurately predicting total counts, the average may be correct but the dispersion of retained losses and the expected use of claims services can be wrong.
Event intensity and the reporting mechanism
Let incident arrivals form a Poisson process with intensity lambda. Suppose each incident is recorded as a claim independently with probability q that is constant conditional on a defined class. Under independent thinning the claim count is Poisson with intensity lambda q. From a claim-only count likelihood, lambda and q are separately unidentified: lambda=0.4 and q=0.25 imply the same claim intensity 0.1 as lambda=0.2 and q=0.5. These are all hypothetical annual parameters. The observational equivalence is exact under the thinning assumptions.
If reporting depends on gross loss X, the claim intensity for a fixed incident cohort is
In the same model the severity distribution among reports has measure proportional to q(x)dF_X(x). Thus a rise in reporting propensity among smaller losses increases the observed count and may lower observed average severity even if the latent gross-loss distribution is unchanged. The opposite pattern can follow a stricter reporting threshold. The sign of observed count-severity association is jointly determined by latent risk and observation selection.
Consider a hypothetical hundred incidents, seventy with loss $5,000 and thirty with loss $100,000. If all incidents are reported, the mean gross loss is $33,500. If only the thirty large events are reported, the mean recorded loss is $100,000. Underlying incident frequency and gross economic losses are identical. The example illustrates how a population of recorded incidents can display a different frequency-severity relationship solely through a changed observation rule.
When reporting is delayed, an incident that occurs near the valuation date has a lower chance of entering the current report data than an older one. A likelihood for recent accident-period counts must account for the report-delay distribution or restrict inference to sufficiently mature cohorts. Reporting and delay may themselves depend on severity and loss type. A severe breach may be detected immediately, while liability demands arising from the incident may arrive years later. Combining the clocks without distinguishing them biases both frequency trends and estimates of count-severity dependence.
Contractual definition of claim frequency
Primary policy language demonstrates why counting claims requires legal and financial data. The Australian Avant Cyber Insurance policy wording (version 4.0) distinguishes bases of cover across sections, applies a time retention to a specified interruption coverage, and includes aggregation of claims from the same originating cause in Section VI(D). Other clauses set obligations concerning notification and deductibles. The wording is relevant as an example of contractual mechanisms; its actual application requires an issued schedule, endorsements, and governing law.
A single service disruption can trigger multiple forms of economic loss. If a contract treats all related claims as a single claim, the contract-level frequency differs from the number of coverage files created administratively. A model whose observations are individual component entries must reconstruct the contract-level cluster before applying a per-claim deductible or occurrence-based reinsurance attachment.
The boundary between a notification and an insurance claim can also depend on disclosure law. Under Article 33 of the GDPR (European Parliament and Council, 2016), a controller must notify the competent supervisory authority without undue delay and, where feasible, within 72 hours of knowledge of a personal data breach, except where the stipulated low-risk condition applies. The processor has a distinct notification duty to the controller. Article 34 establishes a separate threshold for notice to affected persons. These provisions do not determine coverage under a cyber policy.
For an actuarial model, a change in data protection enforcement, disclosure obligations, or organizational detection processes may affect both the number of known incidents and the financial cost of response. A cross-country comparison should model these institutional changes as observation and cost mechanisms, rather than interpret differences in disclosure counts as differences in intrinsic attack rates.
Risk controls, classification, and causal inference
Security controls can be organized through technical frameworks. The NIST Cybersecurity Framework guidance describes a flexible risk management process and ways to assess current and target states. Such guidance provides a vocabulary for capturing changes in controls, but it supplies no universal insurance frequency multiplier.
A model may include indicators of authentication practices, tested recovery capability, patch timing, external service reliance, and historical incident experience. These covariates should be measured at a specified date. Using a post-claim investigation result as a predictor at policy inception leaks outcome information into training data. Using a security assessment recorded after a major incident may also reverse the apparent direction of causal effect.
To estimate causal effects of an intervention, define the intervention, time zero, treated and comparable untreated populations, relevant confounders, and follow-up. Policyholders that adopt protection after an incident may have elevated baseline risk. A simple association between control adoption and high claims does not establish that the control increases risk. Conversely, favorable selection into a better-controlled insurance portfolio can exaggerate the benefit attributed to an individual control. Prediction and causal evaluation therefore serve different questions and demand different evidence.
Operationally, a data dictionary should preserve versions of risk assessments, policy definitions, and event classifications; provide explicit categories for unknown values; and document transitions between taxonomies. Combining labels across systems without a crosswalk produces artificial shifts in frequency. An initially absent category may appear to surge as soon as a new classifier is deployed, even when the underlying hazard is stable.
Mechanisms and Evidence for Severity Models
Gross damage, covered loss, and indemnity
The economic impact of a cyber incident may comprise direct remediation, restoration of data, interruption of business, fraudulent transfers, extortion, notification, defense, and third-party liabilities. Gross economic loss refers to the covered and uncovered financial consequences within the defined economic perimeter. Covered loss is the amount eligible under the contractual grant before specified financial terms. Insured indemnity is the amount payable after the relevant contractual terms and allocation.
The distinction is particularly important where an empirical source includes alleged, estimated, incomplete, or unaudited corporate costs. Romanosky (2016) documents limitations of reported cost data: monetary information is available for a small subset of observations, and the recorded amount omits some categories of private and social loss. The data inform the investigation of incident repercussions without supplying a full distribution of insurer payments.
The extent of information loss can be observed directly in more recent empirical research. Shevchenko et al. (2023) analyze cyber events and business sectors using an event corpus of 132,126 records spanning 2008–2020. Only 4,667 observations with nonzero financial amounts are included in their positive-loss analysis. The corpus is dominated by U.S. organizations. These properties constrain the interpretation of any distribution fitted to those monetary observations: it is a selected event-loss distribution, not a census of insured severities.
Missing amount, recorded zero, denied indemnity, positive paid amount, policy-limit exhaustion, and an open reserve are distinct states. If missing monetary loss is encoded as numerical zero, the mean is biased downward relative to the corresponding positive population. If all zero outcomes are deleted, a model cannot directly predict the probability of payment. Proper coding preserves the reason an amount is unobserved and the status of the claim at the valuation date.
Choosing a positive severity distribution
For positive ultimate loss X, a gamma GLM with log mean is a tractable model for conditional expected severity. A lognormal model specifies log X as normal with mean parameter eta and variance tau squared; on the original scale,
The exponential of a fitted log-scale predictor is a conditional median when the stated lognormal assumptions hold, not its mean. A model that back-transforms the linear predictor without an appropriate correction underpredicts the mean when log-scale variance is positive.
A gamma distribution can fit moderately right-skewed body losses, while a lognormal distribution provides a wider range of right-tail behavior. Both may be poor descriptions of a highly heterogeneous or severely censored population. Selection should consider predictive calibration at economically meaningful attachment points, not only global fit. A distribution that predicts ordinary repair claims well may still underestimate rare liability or business interruption amounts.
Extreme-value analysis represents the excess of a loss above threshold u. For a generalized Pareto distribution with shape xi and scale beta, its survival function for excess z is
on the support where the expression in parentheses is positive, for nonzero xi; the limit at zero shape is exponential. The mean excess exists when xi is less than one; the variance exists when xi is less than one-half. These are conditions on the uncapped fitted tail under the stated model. Insurance limits can bound indemnity even when a hypothetical ground-up loss distribution has infinite variance.
Zeller and Scherer (2022) use a body-tail severity construction based on separate distributions in their marked-process investigation. Their illustrative parameterization is not an empirical estimate from actual insured claims. The same distinction applies to proposed tail thresholds: the statistical suitability of a splice must be assessed on the observed population and the intended financial layer.
An uncapped severity model can be specified by a body distribution F_B below a threshold and a tail distribution conditional on exceeding it. If pi is the probability of a tail exceedance, the unconditional survival above u is pi times the tail survival. This probability must not be omitted when calculating unconditional expected excess. The splice can be made continuous in its cumulative distribution using normalized body and tail components; equality of densities at the transition is an extra modeling choice, not a necessary condition of a valid distribution.
An important uncertainty is that the threshold and tail shape are usually estimated from few observations. Moving the threshold upward reduces bias relative to a limiting tail approximation but increases sampling variance. The practical response is to report a range of threshold choices, confidence or posterior intervals, and sensitivity of insurance-layer loss rather than attach false precision to a single fitted extreme quantile.
Deductibles, caps, and censored observations
Consider a simplified coverage that pays
where d is an ordinary monetary deductible, ell is a maximum payment, X is the covered ground-up loss, and there is no annual aggregate or coinsurance. Let F denote a continuous ground-up distribution. The mass at zero is F(d), and the mass at the payment limit is 1-F(d+ell). For a payment y strictly between zero and ell, the density is f(y+d).
Using the survival integral for a bounded nonnegative variable gives
This formula incorporates zero payments. If the severity sample contains only positive payments, its mean is instead E[Y]/P(X>d), assuming no other eligibility or reporting selection. An estimator of positive-paid claim frequency must then be paired with this positive-conditional mean. Multiplying incident frequency by positive payment severity without multiplying by the probability of reaching a positive payment overstates expected loss.
Suppose the deductible is $20,000 and the payment cap $100,000. A covered loss of $10,000 yields zero, a $60,000 loss yields $40,000, and a $200,000 loss yields $100,000. Under this hypothetical contract, the capped observation informs the analyst that ground-up covered loss has reached at least $120,000. It does not reveal its precise magnitude.
For a positive-payment-only sample, a correct likelihood for an uncapped observation y is f(y+d)/(1-F(d)); a capped observation contributes (1-F(d+ell))/(1-F(d)). If zero observations are retained, use the unconditional point mass and density. Deductibles can create truncation of the sample, whereas limits create right censoring of ground-up values and a payment atom. These mechanisms should not be conflated.
Actual policies may implement aggregate limits, varying sublimits, time retentions, different treatment of defense expenses, and other insurance. The limits and retention clauses of the Avant specimen provide a concrete example. The simple scalar transform above is therefore an instructional case. It cannot reproduce every insurance contract without adapting its payment function.
Uncertainty in severity development
Cyber claims can involve several amounts at a valuation date. Paid-to-date is cash already disbursed. Case reserve estimates remaining expected amounts for reported claims. Incurred-to-date is often defined as paid plus case reserve, subject to accounting conventions. Ultimate severity is the eventual amount on a specified gross or net basis. It may differ materially from all interim measures.
Training a model on current incurred amounts for open claims and final paid amounts for closed claims mixes forecast and realized outcomes. A recent accident-year cohort will usually contain a different proportion of open claims than an older cohort. A change in valuation conventions can also move reported severity without changing eventual payments. Estimate the claim development process explicitly, or use mature cohorts with an appropriate adjustment and loss of recency.
A severity model can represent time to closure jointly with payments or use a multistate process with open, reopened, settled, and denied claims. Closure may be informative: straightforward restoration cases can settle quickly, whereas contested liability claims often take longer. Treating every open file as a random censoring observation without investigating closure mechanisms can lead to biased inference.
For monetary comparisons across time, define currency conversion dates and a price or cost index suited to the relevant loss component. Inflation in legal services, restoration labor, and interruption costs need not be equal. A common deflator can aid descriptive comparisons but may hide changes in the relative composition of claim amounts. The empirical question is whether residual shifts represent changing risk or changing price and claim mix.
Severity components and aggregation
Let the mark of an incident be a vector of financial loss components, and let T_i denote the contractual payment mapping for policy i. The scalar Y_i is then T_i applied to the mark and the current state of the policy. This mapping can include dependencies explicitly: several expense categories may share one sublimit; restoration expenditure may change the length of interruption; a common legal defense may reduce the loss from multiple demands.
A sum of independent component models is defensible only when conditional independence is plausible and relevant terms can be applied consistently. Modeling business interruption without its duration and recovery relationship can produce erroneous behavior at high quantiles. Modeling litigation independent of affected data and the underlying breach may likewise misrepresent multi-coverage occurrence losses.
The ultimate actuarial target is the amount subject to the intended financial decision. A gross incident model supports alternative product design when loss components and coverage mapping are available. A direct paid claim model is useful when forecasting an established contract and when selected observations are adequate. The choice should reflect which counterfactual changes the model must reliably support.
Comparative Measurement of Dependence
Conditional moments and size weighting
Define mu_n=E[Y_j | N=n] for exchangeable claim severities within a policy with n positive claims. The formula for expected aggregate loss becomes
No within-policy independence assumption is needed for this first-moment identity. If the conditional variance of each severity is v_n and each pair has covariance c_n, the conditional variance of the sum is
The unconditional variance is the expectation of that expression plus the variance of N times mu_N. The covariance term can be material for policies generating multiple files from a single underlying event. Ignoring it leaves the first-moment formula intact but generally distorts uncertainty.
Define A=S/N among policies with positive counts, and set A=0 for zero counts. Then E[S]=E[NA]. A comparison with E[N]E[A] yields
Because A is defined as zero when N=0, a covariance computed across all policies contains the mechanical effect of that convention. To interpret within-positive-policy count-severity dependence, condition on N>0:
The distinction prevents a trivial positive association between claim existence and a positive average amount from being mistaken for count-severity dependence among claim-bearing risks.
There is an equally important weighting qualification. If one samples a claim uniformly from the full population of claims, rather than a policy uniformly, the population mean severity is E[S]/E[N], whenever the ratio exists and E[N]>0. Thus expected count times correctly claim-weighted mean severity gives the correct aggregate mean as an identity, even under dependence. The question is whether a model can predict that mean under changes in risk class, exposure, and conditions, and whether it can predict the distribution of the aggregate. An unconditional claim mean transferred across populations may fail on precisely those dimensions.
For a two-type synthetic population, let half the policy exposures be type A with two claims each of $10,000; the remainder be type B with six claims each of $30,000. Average count is four; the mean of policy-level average severities is $20,000; their product is $80,000. Actual average aggregate loss is $100,000, with $20,000 covariance contribution. But the mean per claim, weighted by claims rather than policies, is $25,000 and the corresponding product with four correctly equals $100,000. Both calculations are valid for their own weighting; interchanging them causes error.
Conditional regression and dependence corrections
Shi, Feng, and Ivantsova (2015) propose dependent frequency-severity modeling with a hurdle and two alternative dependence constructions. The accessible original abstract describes conditional decomposition and a mixed copula approach, demonstrated on automobile insurance. Garrido, Genest, and Schulz (2016) describe conditional severity GLMs with count as a predictor in the severity regression, again illustrated outside cyber insurance. These abstracts support discussion of the model classes, not conclusions about their relative out-of-sample accuracy in cyber insurance.
Consider an instructional joint model with N Poisson of mean m and conditional expected individual severity
Using the probability-generating function E[z^N]=exp(m(z-1)), differentiation gives E[Nz^N]=mz exp(m(z-1)). Taking z=exp(b),
This calculation is derived here from stated assumptions; it is not a new general theory or an estimated cyber dependence adjustment. If b=0, it reduces to m exp(a). If b>0, expected severity increases with count and the correction relative to the conditional n=1 severity is greater than one. If b<0, the correction relative to that comparison mean is less than one. The result changes for a different count law, a different link, or a different basis of comparison.
In a numerical illustration, take m=0.2, a=log(50,000), and b=log(1.5). Individual expected losses at counts one and two are $75,000 and $112,500. The expected aggregate is $15,000 exp(0.1), approximately $16,577.56. A model holding the count-one expected severity fixed yields $15,000 at the same frequency. The approximately $1,577.56 difference is an assumption-driven dependence adjustment, not an empirical loading. One must separately check the plausibility of extrapolation to large n and any policy cap.
The appeal of conditional regression is transparency and predictive tractability. Its disadvantage is that the number of claims is itself an outcome observed after coverage inception. A pricing application cannot simply use a future realized count as a known predictor. It integrates over the predicted count distribution, as in the derivation above. For reserving conditional on claims already observed, a model may instead use the current count and predict remaining events and payments. The information set differs between the two applications.
Copulas with discrete and continuous margins
A copula can couple marginal distributions, allowing different families for frequency and severity. If N is a count and A is a continuous average severity defined for N>0, a joint construction on positive policies can specify C(F_{N+}(n),F_A(a)). A rectangle probability for n and a severity interval (a_1,a_2] is the alternating difference of four copula evaluations at the discrete count boundaries and continuous severity boundaries.
It is important that a count margin has jumps. Unlike the continuous-margin setting, the copula need not be uniquely identified across the entire unit square by the observed joint distribution. A parametric family imposes additional structure through extrapolation. For a continuous A density, a likelihood can use the derivative with respect to its continuous margin, differenced across adjacent count cumulative probabilities. Replacing the count with a continuous density or choosing arbitrary ranks without addressing discreteness can mischaracterize uncertainty.
A Gaussian copula is flexible for ordinary association but, with nondegenerate correlation, lacks asymptotic upper-tail dependence for continuous margins. A t copula with finite degrees of freedom can represent upper- and lower-tail dependence. For a discrete count, the extrapolated copula tail properties should not be presented as directly observed tail coefficients without attention to the available support. At high attachment points the fitted family and tail constraints matter more than a single reported correlation.
A positive-count average is conditional on a risk having claims. It may become more stable with a growing number of independent severity draws, while correlated losses from a common event can weaken that stabilizing effect. A joint distribution for N and A alone does not determine the distribution of individual components or their allocation across exposures. If the intended output includes payment-level deductibles or multi-coverage aggregation, a full claim-level construction may be necessary.
Shared latent risk and portfolio accumulation
A latent environment Z can jointly alter claim frequency and severity. Suppose N_i conditioned on Z is Poisson with rate lambda_i(Z), and the conditional expected amount per claim is mu_i(Z). Under independent severity marks conditional on Z and count,
This exceeds E[lambda_i(Z)]E[mu_i(Z)] whenever their covariance is positive. Even if counts and sizes are independent conditional on the environment, they can be dependent marginally. If S_i and S_j are independent conditional on Z, their unconditional covariance equals the covariance of their conditional expected totals.
Consider a synthetic common-state portfolio of 100 insured firms, each with conditional frequency 0.1 and mean loss $20,000 in a normal state of probability 0.9, or frequency 0.6 and mean loss $100,000 in a stressed state of probability 0.1. For simplicity each conditional severity is deterministic and claims are independent across firms given the shared state.
The expected insurer loss per firm is $7,800. Multiplying the unweighted state-average frequency of 0.15 by the unweighted state-average severity of $28,000 would produce $4,200. The correct claim-weighted mean in this mixture is $52,000. The product of 0.15 and $52,000 again equals $7,800. State mixing changes the weighting even when the model is conditionally independent.
The unconditional portfolio mean is $780,000. The conditional portfolio means are $200,000 and $6,000,000. By the variance decomposition, the shared environment contributes
in squared dollars. The expected conditional variance is $63,600,000,000 because a compound Poisson with deterministic amount y and rate lambda has variance lambda times y squared. The total variance is $3,091,200,000,000. If each firm instead has its own independent state draw with the same probabilities, its marginal mean and variance remain identical but the variance of the 100-firm portfolio is $93,876,000,000. The large difference is attributable to state sharing rather than different marginal risk. This synthetic calculation isolates the diversification issue without pretending to estimate real catastrophe frequencies.
Awiszus et al. (2023) discuss common stochastic risk factors and Cox processes as representations of systematic occurrence dependence. Zeller and Scherer (2024) demonstrate in their accumulation risk study how insufficient information about common vulnerabilities can undermine portfolio modeling. A key inference for data governance is that risk entities and events must retain joinable information about shared technical dependencies.
A latent factor model remains a simplification. It may describe a broad environment that raises intensities across multiple policyholders without identifying a discrete cause. A common-shock model instead generates distinct events such as a service outage and draws a vector of affected firms conditional on the event. A self-exciting process generates a higher short-term arrival intensity after recorded events. These mechanisms can produce similar clustering in historical counts, but they imply different future behavior and portfolio interventions.
Empirical comparison and nonstationarity
The available empirical studies show variation between event types and business sectors. Shevchenko et al. (2023) report in their frequency and severity analysis the lack of a distinct general pattern across the relationships investigated, together with evidence of heavy-tailed losses. Their sample concerns cyber events with available loss information rather than a complete longitudinal insurance-policy panel. The result supports heterogeneity and careful interpretation; it cannot establish the absence of conditional dependence in every insured cohort.
The NAIC (2026) market report's claims discussion describes, based on other carrier reporting, lower frequency and higher average severity for large accounts in 2025 compared with 2024. The approximate reported pairs are 15 claims per 100 policies and $2.2 million, versus 10 per 100 and $4.4 million. If frequencies and means had identical coverage, valuation, and exposure definitions, their products would imply $330,000 and $440,000 per policy respectively. This is an illustrative arithmetic comparison of reported approximate inputs; the prerequisites for an exact actuarial pure-premium comparison are not established by the report summary. Chubb's (2026) official report overview confirms that its study examines claims through 2025 and describes geographical differences in frequency and severity trends.
These published aggregates cannot identify the covariance of count and average severity within a policy. They compare cohort-year summaries and may reflect different coverage terms, portfolio selection, settlement development, claim composition, and threat conditions. An ecological correlation between aggregate rates and costs must not be interpreted as the individual-level dependence coefficient of a joint model.
A longitudinal estimation strategy should capture policy retention, new entrants, control adoption, and contract changes, and should preserve the distinction between stable entity characteristics and period-specific shocks. Repeated observations of the same organization are not equivalent to independent rows. Cross-validation across random rows can leak information about repeated policyholders or common events. Time-blocked, entity-blocked, and event-blocked validation reveals different dimensions of model generalization.
Financial, Legal and Institutional Consequences
Pricing and product design
A technical pure premium equals the expected claim cost on a defined basis over the covered exposure period. It can include an explicit expectation of adjustment expenses if those expenses are part of the chosen target. Expected expenses, cost of capital, risk margin, commissions, taxes, and competitive considerations have separate roles in determining the charged premium. The joint frequency-severity model supplies a probability distribution, not a complete commercial tariff.
For a fixed portfolio and a well-calibrated marginal aggregate expectation, richer dependence may have little impact on expected individual claim cost. It can still change the capital or reinsurance needed to support that cost. A positive association between count and amount within a rating cell may also change the conditional expectation itself when the chosen separate-regression benchmark uses inappropriate severity weights.
A rate change must reflect the exposure and contractual terms of the proposed book. A lower per-claim deductible can increase positive payment frequency and reduce the average severity measured among positive payments. The expected covered loss can rise even while average payment severity falls. A higher limit increases payments in the upper tail and can change the correlation between policy count and amount, particularly when high-frequency risks also have large exposures.
Contract design can thus change observed statistical dependence without changing underlying technology. Capital pricing can respond to shared cloud exposure while expected-loss pricing responds primarily to the risk's own coverage and size. Those two components of the pricing decision benefit from an explicit separation of marginal and cross-policy risk modeling.
A simple annual aggregate illustrates further nonlinearity. Let individual hypothetical amounts after per-claim terms be W_j, and an annual limit A. Then annual insurer loss is min(A,sum W_j). Averaging independent estimates of expected W_j and multiplying by expected count misses any cap effect unless it is already reflected in the observed W_j population. Modeling the ground-up claims and applying annual exhaustion to the entire simulated sequence permits comparison across alternative aggregate limits.
Reserving and reporting development
A joint prediction of ultimate count and ultimate severity supports the estimation of liabilities for reported but unsettled claims and incurred but not reported claims. The latter is a distribution over future reported claim counts and their eventual payments conditional on past incidence periods and known exposure. An ultimate loss estimate must clearly state whether it includes claims handling expenses, whether recoveries are deducted, and the valuation date.
Two separate development factors for claim counts and average severities can be useful descriptive tools, but their product is valid as an ultimate cost forecast only if the target counts and severities are defined and weighted consistently. More complex reporting may require a severity-dependent delay model. A cyber event may be rapidly notified but give rise to additional defense expenses months later; an event-level development path can be more useful than several unlinked claim files.
A robust accounting reconciliation uses multiple ledgers: event occurrences, claim reports, payments, case reserves, recoveries, and coverage allocation. New files are linked to original loss events where known. The reconciliation also records policy limits and amounts remaining after prior payments. Without such linkage, an insurer could inadvertently reserve twice for losses that share a contractual limit.
Synthetic forecasting tests should withhold recent calendar periods and reconstruct what information would have been available at each historical valuation date. A backtest that uses ultimately known event classifications or final settlements to predict earlier estimates is not a valid simulation of reserving practice.
Reinsurance and portfolio risk
Suppose S is the annual gross insurer total and a simplified annual stop-loss cover responds above attachment a up to layer width b. Its ceded payment is min((S-a)_+,b). The expected recovery is the integral of the survival function of S from a to a+b. A model calibrated only at E[S] cannot generally determine this recovery: two distributions with the same expectation can have very different survival in that layer.
An occurrence excess-of-loss cover applies the contract to a defined occurrence total. In cyber insurance, identifying that total may require attribution to shared interruption, software vulnerability, or other defined originating cause according to treaty wording. Without event-level linkage, a model can price each ceded claim as if separate, overstate the number of attachments, or underestimate the total amount accumulating above an occurrence threshold.
Treaty terms also govern limits, reinstatements, time windows, exclusions, and allocation. A mathematical common-event definition should correspond to the legal definition used in the treaty. Risk teams should consider uncertainty about event attribution explicitly when multiple plausible groupings materially alter loss recovery.
Capital questions concern the distribution of the net retained portfolio after primary terms and reinsurance. Value at Risk at probability p is a corresponding generalized quantile. Expected Shortfall can be defined via the average quantile above p,
when finite. This definition remains useful for distributions with atoms, such as loss limits and zero-claim periods. Tail estimation requires finite or otherwise appropriately managed expectations for the intended risk measure; the financial contract can impose bounds even when some underlying economic loss model has infinite moments.
The synthetic common-state example shows why diversification based on individual marginal fits is inadequate without information on shared dependencies. At portfolio level, variation in the expected loss across common states can scale approximately with the square of the number of similarly exposed firms, while conditionally independent variation grows linearly. Exposure growth without diversification of underlying technology can therefore increase risk concentration more rapidly than expected loss.
The EIOPA (2025) financial stability report summary identifies digital interdependencies and accumulation modeling as supervisory concerns. For an insurer, this implies a need to distinguish its own operational cyber exposure from the claims it has undertaken to insure. These risk channels may be related in a scenario, but their losses enter different parts of a solvency assessment and must be recorded without double counting.
Regulatory and professional practice boundaries
The international application of cyber joint models requires recognition of jurisdiction-specific notification rules, insurance law, contractual interpretation, data access, and professional standards. The European Union notification example illustrates the legal determinants of what becomes public or official data. The Australian specimen demonstrates that one policy can apply different triggers and financial terms across coverage parts. Neither source creates a uniform worldwide insurance or disclosure rule.
For actuarial work in the United States, the Actuarial Standards Board's (2016) ASOP No. 23 on data quality addresses the suitability of data for a defined assignment, review and disclosure of known limitations. Its applicability is governed by the scope of that professional standard. The same board's (2019) ASOP No. 56 on modeling addresses intended purpose, model structure, data, assumptions, controls, and output evaluation. The standard is effective from October 2020. These provisions provide a useful documented professional reference for the distinction between a technically executable model and a responsibly supported modeling conclusion.
Statistical discrimination may arise when rating covariates reflect uneven access to security resources or differences in detection and disclosure. An insurer should separately investigate predictive validity, applicable rules governing rating variables, and the consequences of using variables whose values are proxies for other characteristics. The existence of a predictive association does not resolve the legal or ethical suitability of its use.
Model output is most effective when its scope matches the decision. Internal underwriting, regulatory filing, portfolio stress testing, reserving, and public scientific explanation have different standards of confidence, sensitivity, and detail. General actuarial formulas are transferable, but local assumptions and legal conclusions require case-specific evidence.
Integrated Scientific Synthesis and Research Agenda
An integrated data architecture for existing methods
The investigation supports a linked representation of exposures, occurrences, claim files, financial components, and contract state. A policy table identifies the insured entity, exposure interval, applicable coverage, limits, deductibles, and policy version. An event table identifies the underlying occurrence, event type, dates, technical dependency, and confidence in grouping. A claim table records coverage-specific files and stages, including no-payment outcomes. A financial transaction table connects expenses, indemnity, reserves, and recoveries to the correct coverage and event.
The minimum useful linkage is from each claim to its policy and its event where known. Event-attribution confidence should be stored as evidence rather than forced into a single category. An unknown cause should remain unknown; grouping every unknown event into one common event creates artificial accumulation, while splitting a known incident across independent identifiers suppresses accumulation.
Technology-dependency data should be time-stamped. A provider used at policy inception may have been replaced before the loss. Exposure maps should preserve the coverage period relevant to each technology relationship and, where possible, identify shared platforms, products, regions, and services at an economically meaningful resolution. A generic flag indicating cloud use cannot distinguish concentrated reliance on one shared provider from dispersed use of several providers.
A useful structure for an analysis snapshot includes units and data provenance, reporting cutoff, currency, monetary valuation basis, policy terms as of inception or endorsement, and the definition of zeros and missing values. It should also record why the dataset contains each observation. The documentation should be sufficient for a second analyst to regenerate the data set on the same frozen inputs.
Comparing model families by intended output
A benchmark is a separate count GLM and claim-level severity GLM on compatible cohorts. Introduce a positive-count hurdle if zero behavior cannot be explained adequately. Compare a count-dependent conditional severity model if the observations support within-policy dependence. A latent-factor model can share variation across time or risk groups. A copula model is useful when the relevant margins can be estimated and a flexible joint distribution is necessary. A marked event or common-shock model is appropriate when event-level exposure information and contractual allocation are central.
These are existing approaches; their comparison does not establish a new model family. The research question is which dependencies are necessary for a particular decision and which can be supported by the data. More parameters may improve apparent in-sample fit while worsening extrapolation, especially when observations of extreme events are sparse. An overly simple model may omit economically crucial dependencies despite predictable ordinary claims.
For expected-loss pricing, compare observed and predicted ultimate cost per exposure in credible groups. Decompose error in total count, conditional severity, and the composition of risk. For the distribution of the count, consider predictive log likelihood, observed zero proportions, count quantiles, and overdispersion. For severity, compare calibrated conditional quantiles and attachment-point exceedance rates, account for censoring, and examine residuals by claim type and coverage.
For a joint distribution, check the predicted relationship between positive counts and severity within each materially different segment, as well as joint exceedance behavior. Portfolio validation requires occurrence-based cross-sectional checks and aggregated totals at the relevant time scale. A model that accurately matches separate frequency and severity histograms may still omit simultaneous high counts and high losses.
Temporal backtesting should be accompanied by sensitivity to reporting delay, inflation, exposure and contract changes. Rolling windows test performance under evolving threat conditions. Fixed historical backtests isolate reproducibility. Entity- and event-blocked partitions protect against information leakage. The validation protocol depends on the intended deployment: a policy renewal model may legitimately use previous history of the same insured, whereas a new-business model must predict for firms not observed during training.
Reproducible synthetic checks
The following pseudocode reproduces the common-state calculation with explicit units. It is a deterministic analytical check, not a fitted simulation.
from math import exp, log
p = 0.1
n = 100
lam0, lam1 = 0.1, 0.6
y0, y1 = 20000, 100000
mu0, mu1 = lam0*y0, lam1*y1
mean_firm = (1-p)*mu0 + p*mu1
within_firm = (1-p)*lam0*y0**2 + p*lam1*y1**2
between_firm = p*(1-p)*(mu1-mu0)**2
shared_var = n*within_firm + n**2*between_firm
independent_var = n*(within_firm+between_firm)
assert round(mean_firm) == 7800
assert round(shared_var) == 3091200000000
assert round(independent_var) == 93876000000
m, a, b = 0.2, log(50000), log(1.5)
expected = m*exp(a+b)*exp(m*(exp(b)-1))
assert round(expected, 2) == 16577.56
The calculations expose two distinct sources of model error. An unweighted state average may be combined with a count expectation even though states generate different numbers of claims. A model may also estimate the marginal distribution of each insured entity correctly while assigning independent environment draws to entities that in reality share the environment. Correcting the first addresses mean aggregation; correcting the second addresses portfolio risk.
An empirical application of these methods would estimate the probability and effects of shared states rather than assume them. With sparse data, scenario weights may be judgmental. They should be clearly separated from estimated parameters and stressed over a defensible range. Scenario analysis is a complementary source of information about tail exposure, not evidence that the probabilities assigned to scenario events have been statistically identified.
Research priorities and limitations
Priority should be given to longitudinal policy-level data including no-claim exposures and to claims-event linkage with monetary development and coverage detail. The data need observation timing and contract terms to separate changes in insured outcomes from changes in risk. A study of repeated insured entities can estimate the degree to which risk differences persist over time and whether frequency and severity share latent drivers.
Another priority is to improve information on the coverage consequences of infrastructure shocks. A catastrophe model evaluated only on a database of isolated historic incidents cannot establish the distribution of all affected firms in a novel infrastructure outage. Supplementing the record with dependency mapping, stress scenarios, engineering analysis, and careful observation of actual clustered events improves the ability to examine accumulation.
Cross-jurisdictional work requires harmonized event taxonomy and a map of reporting obligations and insurance triggers. Comparable exposed populations and consistent currency/valuation definitions are needed before international loss rates are pooled. Statistical hierarchies may borrow strength across jurisdictions, but they should not erase differences in disclosure law and insurance design.
The present evidence base includes public reports and published research, but no licensed event-level database, insurer policy-year panel, or actual issued schedules. The original articles on generic dependent-regression approaches are used at abstract level only. The two PDF sources were examined through extracted text, with page image retrieval unavailable. Accordingly this dossier uses a limited number of directly supported reported figures, does not rely on graph readings from unavailable page images, and makes no claim of an empirical re-estimation or treatment-effect identification. The synthetic parameters are fully specified and serve methodological illustration.
The resulting scientific conclusions are strongest where they follow from probability identities, observation mechanisms, explicit contracts, and the stated sample properties of empirical studies. The magnitude and direction of residual dependence in any current cyber insurance portfolio remain empirical questions. The same applies to optimal thresholds, catastrophe probabilities, and specific reinsurance prices.
Conclusion
Cyber claim frequency and insured severity are jointly produced by underlying cyber events, organizational responses, insurance coverage, reporting, adjustment, and payment. The unit of observation and the population at risk establish the meaning of these variables. Observed claims generally identify a combination of incident activity and recording mechanisms, rather than the full latent incident process.
The collective risk model supplies a transparent reference point for analyzing dependency. Independent frequency and severity distributions can produce a defensible expected-loss estimate when the relevant conditional assumptions and consistent claim weighting are satisfied. Where average severity varies with count, the expected aggregate is found by integrating the conditional severity over the count distribution. When multiple claims share a common origin, covariance among individual claim amounts also contributes to aggregate uncertainty.
A central conclusion concerns weighting. Expected count multiplied by correctly claim-weighted mean severity equals expected aggregate loss without a general independence requirement. This identity explains why some separate approaches can estimate a mean successfully even in a dependent portfolio. It also explains why multiplying a count expectation by an unweighted mean of policy-level averages, an inappropriate cohort's severity, or a mean conditional on positive payments can produce systematic error. Joint modeling adds the ability to predict distributional changes, coverage transformations, and dependencies relevant to capital and reinsurance.
A second conclusion concerns observation. Reporting thresholds, contract limits, notification duties, and claims settlement timing change measured count and severity. Without distinct records of underlying events and insurance responses, a statistical relationship cannot always be attributed to technological risk. This limitation is an important property of the empirical object itself.
Finally, policy-level and portfolio-level dependencies must be evaluated separately. The common-state calculation establishes how identical marginal distributions can support radically different aggregate risk under different dependence structures. Published cyber-event evidence and supervisory analyses reinforce the material importance of shared vulnerabilities and data limitations. The future development of cyber insurance actuarial science depends on precise definitions, longitudinal exposure data, durable event linkage, contractual fidelity, and validation for the actual financial decision.
References
Zeller, G., and M. Scherer. 2022. A comprehensive model for cyber risk based on marked point processes and its application to insurance. European Actuarial Journal 12: 33–85. doi:10.1007/s13385-021-00290-1. https://link.springer.com/article/10.1007/s13385-021-00290-1
Awiszus, K., T. Knispel, I. Penner, G. Svindland, A. Voß, and S. Weber. 2023. Modeling and pricing cyber insurance: Idiosyncratic, systematic, and systemic risks. European Actuarial Journal 13: 1–53. doi:10.1007/s13385-023-00341-9. https://link.springer.com/article/10.1007/s13385-023-00341-9
Shevchenko, P. V., J. Jang, M. Malavasi, G. W. Peters, G. Sofronov, and S. Trück. 2023. The nature of losses from cyber-related events: risk categories and business sectors. Journal of Cybersecurity 9 (1): tyac016. doi:10.1093/cybsec/tyac016. https://academic.oup.com/cybersecurity/article/9/1/tyac016/7000422
Romanosky, S. 2016. Examining the costs and causes of cyber incidents. Journal of Cybersecurity 2 (2): 121–135. doi:10.1093/cybsec/tyw001. https://academic.oup.com/cybersecurity/article/2/2/121/2525524
Shi, P., X. Feng, and A. Ivantsova. 2015. Dependent frequency–severity modeling of insurance claims. Insurance: Mathematics and Economics 64: 417–428. doi:10.1016/j.insmatheco.2015.07.006. https://ideas.repec.org/a/eee/insuma/v64y2015icp417-428.html
Garrido, J., C. Genest, and J. Schulz. 2016. Generalized linear models for dependent frequency and severity of insurance claims. Insurance: Mathematics and Economics 70: 205–215. doi:10.1016/j.insmatheco.2016.06.006. https://ideas.repec.org/a/eee/insuma/v70y2016icp205-215.html
Zeller, G., and M. Scherer. 2024. Is accumulation risk in cyber methodically underestimated? European Actuarial Journal 14: 711–748. doi:10.1007/s13385-024-00381-9. https://link.springer.com/article/10.1007/s13385-024-00381-9
European Parliament and Council of the European Union. 2016. Regulation (EU) 2016/679 of 27 April 2016 (General Data Protection Regulation). Official Journal of the European Union L 119, 4 May 2016, 1–88. Articles 32–34. https://eur-lex.europa.eu/eli/reg/2016/679/oj/?pid=14006577
Avant. n.d. Avant Cyber Insurance: Policy wording, version 4.0. Sections I and VI, Australian specimen. https://assets.avant.org.au/cdf6134c-01d7-0292-26f5-2f5cf1db96f8/200fe7a3-b253-4957-8797-8be3cc12fd19/Avant%20-%20Cyber%20Insurance%20Policy%20wording.pdf
National Association of Insurance Commissioners (NAIC). 2026. Report on the Cybersecurity Insurance Market. Section 1, p. 10; calendar-year 2025 market context. https://content.naic.org/sites/default/files/2026-report-cybersecurity-insurance-market.pdf
Chubb. 2026. Chubb Cyber Claims Report: Managing the Velocity of Cyber Risk. Official U.S. report landing page. https://www.chubb.com/us-en/business-insurance/products/cyber-insurance/chubb-cyber-claims-landscape-report.html
European Insurance and Occupational Pensions Authority (EIOPA). 2025. EIOPA flags financial stability risks related to private credit, a weakening dollar and global interconnectedness. 15 December 2025. Official summary of December 2025 Financial Stability Report. https://www.eiopa.europa.eu/eiopa-flags-financial-stability-risks-related-private-credit-weakening-dollar-and-global-2025-12-15_en
National Institute of Standards and Technology (NIST). n.d. Cybersecurity Framework: Frequently Asked Questions. CSF 2.0 implementation and risk management guidance. https://www.nist.gov/cyberframework/faqs
Actuarial Standards Board. 2016. Actuarial Standard of Practice No. 23: Data Quality. Revision adopted December 2016. Sections 3.1–3.6. https://www.actuarialstandardsboard.org/asops/data-quality/
Actuarial Standards Board. 2019. Actuarial Standard of Practice No. 56: Modeling. Adopted December 2019; effective October 1, 2020. Sections 1 and 3. https://www.actuarialstandardsboard.org/asops/modeling-3/