Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Cyber Ransomware Loss Development and Recovery Lags

Abstract. Ransomware insurance losses evolve across the technical recovery of systems, restoration of productive activity, contractual assessment of insured consequences, cash disbursement, claim reopening, and external recovery. This monograph examines the distinct histories of gross paid and case incurred amounts, reimbursement lags, and recoveries after extortion incidents. Its source base comprises original cyber policy specimens, official recovery guidance, empirical research, supervisory and insurer reporting, public incident documentation, and established actuarial literature. Documentary comparison identifies differences in discovery triggers, restoration windows, waiting periods, and policy limits. Separate records of gross loss, realized recoveries, operational impairment, closure, and subsequent development permit a consistent treatment of claims whose financial consequences persist after service resumption. Explicitly synthetic examples examine interruption valuation, gross and net transaction reconciliation, reopening, development factors, and censored time to first payment. Public data do not identify a representative international insured ransomware development distribution. The contribution is a source-critical integration of contract and actuarial measurement with recovery processes, including implications for research data, model validation, and financial governance.

Keywords: Ransomware; cyber insurance; claims development; paid loss; case incurred; outstanding claims; claims reopening; restoration lag; business interruption; reimbursement; salvage and subrogation; recovery timing; cryptocurrency seizure; actuarial reserving; censoring; liquidity; contract wording; international risk science.

Introduction

Ransomware insurance losses develop through a sequence of operational, contractual, accounting, and legal events. The time at which malicious activity becomes visible, the time at which productive operations resume, the time at which losses are quantified, and the time at which cash is paid or recovered are different observations of a common incident. An actuarial evaluation that aligns these observations solely by the date of the ransom demand loses information about the process that generates insured severity.

The object of this investigation is the temporal development of affirmative cyber insurance claims arising from ransomware incidents. It includes extortion-related expenses, system and data recovery, business interruption, dependent interruption when covered, and later claim components attributable to the same originating event. It follows gross direct payments, case outstanding estimates, reopened claims, recoveries from external sources, and the insured's reimbursement. It situates these mechanisms within the parent Science, Cyber Insurance, while leaving the comprehensive reserving science for all cyber perils and the optimization of ransom negotiation to their respective research subjects.

The research answers three questions. Which observations establish the temporal and monetary development of a ransomware claim? Through which mechanisms do technical, economic, contractual, and administrative lags alter insured amounts and payment timing? How can recoveries, reopenings, and observed claims be compared without confounding exposure, maturity, coverage, currency, and selection?

The method combines documentary comparison of specimen insurance contracts, official technical guidance, original empirical scholarship, jurisdiction-specific public incident records, insurer-reported experience, and actuarial reasoning. Sources are chosen for their relationship to the questions rather than weighted as if each supplied an independent observation of the same universe. The evidence is bounded by its origin: a contract specimen establishes available wording; a corporate filing establishes the issuer's disclosed experience; a complaint system describes reports to that system; a case study reveals possible sequences; an insurer portfolio summary characterizes the selected portfolio. The analysis is a structured scholarly synthesis, not a systematic review with a preregistered exhaustive search, a portfolio calibration, or a legal opinion.

The cutoff is November 18, 2026. Numerical examples constructed for this monograph use artificial data and nominal monetary units except where a primary document explicitly reports an actual currency. The study does not infer undisclosed indemnity, consent, sanctions status, or recovery allocation from public incidents. Scientific conclusions concern the relationships that measurement and contract structure require, with empirical magnitude left conditional on the available population and data.

Conceptual and Historical Foundations

From malicious extortion to insured consequences

Ransomware describes a malicious cyber event involving coercive denial of access or exploitation of confidentiality, often by encrypting or removing information and demanding a payment. The European Union Agency for Cybersecurity's 2022 ransomware landscape release describes an evolving combination of interference with assets and extortion. An incident may therefore generate costs even when the victim refuses a demand, succeeds in restoring its systems, or experiences no material interruption. A cyber extortion coverage grant may also encompass events beyond the narrower technical classification.

A demand is neither an invoice for a covered expense nor a measure of economic severity. A paid demand is one transaction. The originating event can create independent expenditures and economic losses through investigation, containment, reconstruction, compromised information, impaired production, alternative service arrangements, and subsequent liability. Some consequences remain with the insured, some are contractually transferred, and others fall on customers, suppliers, employees, creditors, or the public.

Connolly, Wall, Lang, and Oddson (2020), in their empirical ransomware study, evaluated 55 cases in 50 organizations in the United Kingdom and North America and treated interruption and recovery time as distinct severity dimensions. Their reported associations with security posture and targeted attacks are evidence from a historical observational sample. The study's qualitative and quantitative structure helps identify mechanisms; it does not establish a universal recovery-time distribution or a causal return on a specific control.

Early ransomware research placed considerable emphasis on encrypted files and the restoration of access. Contemporary business recovery includes identity infrastructure, validated production, dependencies, reconciliation of interrupted transactions, and the treatment of potentially stolen information. The financial implication is that the discovery of an exfiltration layer can increase incurred estimates after encryption has stopped. In the 2026 cyber claims summary, Coalition reports that 70% of ransomware events included both encryption and exfiltration and that 86% of affected businesses rejected the demand in its 2025 experience. Its population includes more than 100,000 policyholders across cyber perils; that figure is not the ransomware-event denominator. Neither percentage determines the probability of a particular claim's closure.

Distinct units and quantities

An incident is an episode in the observed technical and economic world. An insurance claim is an assertion of entitlement under a specified contract, recorded according to the insurer's administrative system. One incident may generate multiple claims, multiple coverage components, or claims across tiers and participating carriers. One claim may combine covered consequences that arise at different times. A recovery is a transaction or right that offsets a relevant cost under a specified accounting and legal treatment. The record therefore requires stable identifiers at incident, policy, insured entity, claim, coverage component, payment, and external-recovery levels, with explicit links among them.

Economic loss is the reduction in resources or economic value attributed to an incident under an explicit counterfactual and horizon. Covered loss is the portion satisfying applicable coverage terms before or after contractual adjustments as defined. Indemnity is the obligation to fund the eligible amount under the contract. Paid loss is the cash transaction recorded by the insurer under an agreed convention; a claimant may receive value through payment directly to a vendor. Case outstanding is the estimated remaining amount on reported claims; case incurred equals cumulative paid plus the corresponding case outstanding estimate. Ultimate direct loss is the final amount on the claims under the specified gross convention. Expenses, salvage, subrogation, and reinsurance must be separately labeled.

A cyber insurance claim can have zero paid indemnity and positive adjustment expenses. An incident can result in a large economic loss and a small covered amount. A claim described as closed with zero indemnity may be below a retention, outside the coverage grant, withdrawn, disputed, or resolved without payment. Such outcomes have different interpretations and should not be combined under a single administrative label if the research concerns coverage or settlement quality.

The clocks of loss development

At minimum, the following event times belong in a longitudinal record: estimated compromise onset; first operational impairment; discovery; insurer notice; first case estimate; initial covered expense; first insurer disbursement; first usable service; secure production restoration; financial reconstruction; contractually recognized restoration endpoint; partial and final payments; closure; reopening; external asset recovery; recognized insurer recovery; and final legal settlement where relevant. Each timestamp needs a definition, observation source, precision, and an indication of whether it was actually observed or estimated.

There is no general total ordering of all these dates. A vendor can receive an insurer payment while productive operations remain disrupted. Discovery may precede an outage or follow the onset of compromise. A claim may close before a liability demand emerges. A court can resolve a subrogation action long after the underlying indemnity is paid. In a data set, negative differences between genuinely comparable clocks sometimes reveal erroneous time-zone normalization or a date-field substitution; in other cases they describe a valid sequence, such as reimbursement before final restoration.

The useful lag is always a difference between specified endpoints. A reporting lag measures notice from the relevant discovery or loss occurrence. An operational disruption lag measures the onset of impaired operations to a defined functionality threshold. A reimbursement lag measures the insured's eligible expenditure to reimbursement. A settlement lag measures a specified notice or valuation origin to a specified settlement outcome. A recovery lag measures the originating payment or right to realized external recovery. Calendar days, business days, and hour-level contractual waiting periods require separate units. A single median of these different intervals has no stable meaning.

Scope of international comparison

The incident mechanism is international, whereas policy interpretation, permissible payments, claims handling, and insurance reporting remain jurisdiction-dependent. The United States examples below establish specific contract wording and federal enforcement documentation. United Kingdom evidence concerns victim interviews and national survey respondents. The European agency release covers a stated cross-regional observation window. None supplies an unrestricted global insured cohort. International comparison proceeds by matching definitions and denominators and by treating legal differences as explanatory variables, rather than imposing an observed local pattern on the whole market.

Claim Development

Contractual gateways and coverage components

Insurance contracts transform an incident into a set of contingent payment obligations. To reproduce a claim calculation, the investigator needs the issued wording, declarations, endorsements, policy period, identity of insureds, applicable trigger, causation standard, exclusions, and the conditions governing notice, consent, expenses, and valuation. An insurer's portfolio may contain multiple forms and negotiated amendments. A product brochure cannot replace the issued contract for this purpose.

The AXIS cyber specimen (2020), first-party grants A-F, illustrates how a single extortion incident can invoke separate clauses for response expenses, interruption, service-provider interruption, data restoration, equipment replacement, and extortion. The business interruption provisions require a measurable disruption satisfying the stated temporal and coverage conditions. The form also provides a 180-day ceiling and a reasonable-speed restoration test for the relevant restoration period. The extortion reimbursement language operates as a distinct grant. Coverage is further conditioned by the declarations, retention, limits, and other policy terms. These observations describe the specimen rather than an issued contract.

The AmTrustCyber sample wording (2021) organizes direct loss around separately described insured events and includes a proof-of-loss expense grant. Its declarations supply the business interruption waiting period and the restoration period, with separate entries for supplier interruption. It distinguishes direct loss discovery from claims made against an insured. Its aggregate may be depleted by defense costs under the wording. The two specimens establish contractual variation: an assumed fixed 180-day window, universal consent rule, or uniform order of retentions and limits cannot be applied to all claims.

A data analyst should reproduce the exact operative layer for every claim component. An amount that begins as a reasonable estimate of system restoration may be reclassified when invoices distinguish ordinary replacement from eligible reconstruction. A customer refund could be a component of another grant or outside coverage entirely. Losses may compete for a common limit. Under such conditions components cannot be treated as independently capped variables and summed without checking the aggregate and contractual ordering.

Financial reconstruction of interruption

Business interruption assessment begins with the productive function that was impaired and the loss measure adopted in the contract. Gross revenue decline is rarely the complete measure of insured income loss. The analyst must reconstruct the relevant counterfactual, allow for saved variable expenses, assess continuing contractual costs, examine mitigation and substitute production, and account for delay rather than destruction of sales. Net profit, gross earnings, continuing operating expenses, and extra expense are contract-defined quantities; their precise combination must follow the applicable grant.

At a daily resolution, a simplified contribution-margin loss is the expected contribution in the absence of the event minus realized contribution, adjusted for verified displaced or recaptured production. If the expected contribution already includes normal operating costs, adding the same continuing expense again duplicates loss. If alternative service expense is included as a separate covered mitigation cost, the contribution comparison must use the actual realized production associated with that expense.

The counterfactual must reflect ordinary business variation. Seasonal demand, a scheduled maintenance outage, prior financial deterioration, unrelated customer departures, or a simultaneous macroeconomic shock can explain part of the difference between forecast and actual turnover. The period preceding an incident may contain atypical growth or contraction. Use several plausible baselines where the precise counterfactual cannot be identified, reconcile them against contemporaneous budgets, orders, comparable locations, and independently observed market conditions, and distinguish a sensitivity range from a statistical confidence interval.

A company can recover transactions without recovering their original timing or value. In a billing service, for example, production may be restored while submitted claims or invoices remain in a backlog. The eventual catch-up may reduce lost gross receipts without eliminating financing costs, overtime, contractual service credits, or cancelled business. Contractual recoverability of each component needs independent assessment.

Hypothetical contract calculation

Consider a synthetic insured with normal daily contribution of 20 monetary units. Production is completely interrupted on days 1-5, then operates at half normal contribution on days 6-15, and resumes on day 16. Assume a one-day time deductible, no later lost sales or catch-up adjustment, and a contract that counts eligible interruption only from day 2. Eligible contribution loss is therefore 4 times 20 plus 10 times 10, or 180 units. Assume incident response costs of 35, eligible system restoration of 80, and separately eligible mitigation expense of 25, with no ransom. In this example only, the contract applies an interruption sublimit of 150, then a single retention of 20, then an overall limit of 250, and permits all other amounts without further restriction.

Component Eligible amount before limits Amount after component sublimit
Business interruption 180 150
Incident response 35 35
System restoration 80 80
Extra expense 25 25
Total 320 290

The resulting indemnity is min(250, 290 - 20) = 250. This calculation reveals the distinction between economically impaired operations and the contractually payable share. Under the assumed time deductible, the excluded first day of lost contribution is 20, and the aggregate and interruption sublimit further restrict coverage. The numerical order of operations was explicitly assumed; it is not an interpretation of either cited specimen.

A longer interruption can increase underlying covered-category loss without increasing an already exhausted indemnity limit. It can still increase insurer adjustment work, the insured's residual burden, and potential third-party obligations. If a second coverage layer becomes involved, the exhaustion timestamp and allocation of earlier payments affect that layer's development history. Analyses that use only final indemnity conceal this economic and contractual saturation.

Paid, outstanding, and incurred movements

Friedland's (2009) unpaid-claims study note explains the importance of consistent paid, reported, ultimate, and outstanding claim definitions. In this study the term gross refers to direct insurer loss before external recoveries and reinsurance, with specified expenses included only when declared. Denote cumulative gross payments at valuation v by P(v) and case outstanding by O(v). Case incurred I(v) equals P(v) + O(v). The incremental change in incurred over a period equals incremental payment plus the change in the outstanding estimate. Ultimate loss U is the final gross settlement amount within that unit's defined claim scope; estimated unpaid gross liability at v is an estimate of U minus P(v). At a portfolio level the estimate includes provision for relevant claims that are unreported and for further development beyond recorded case estimates.

An incurred increase can occur without payment: a forensic report can reveal extra affected systems, a forensic accountant can establish a longer loss period, or a previously disputed clause can be resolved. Equally, a payment can have no effect on incurred if outstanding falls by the same amount. Incurred decreases can reflect revised loss estimates, exhausted coverage, a settlement below the prior case estimate, or correction of a duplicate invoice. The gross ledger should distinguish adjustments and reversals from separately booked external recoveries.

Closure is an administrative state. A closed claim with no case outstanding at v has a point-in-time estimate of zero outstanding; it does not establish the eventual absence of reopened payments. For empirical development, retain all historical valuations. Rewriting earlier data with the latest estimate turns historical forecasting into retrospective reconstruction and creates look-ahead bias.

Synthetic reopened claim ledger

The following separate hypothetical claim has no connection to the preceding contract example. Amounts are cumulative and use arbitrary nominal units. Recovery is direct external cash allocated to the insurer; there is no reinsurance. The claim closure at day 90 is followed by a supplementary allowance at day 150. The final recovery occurs after the second closure.

Day from notice Gross paid Case outstanding Case incurred External recovery Net cash paid State
0 0 200 200 0 0 Open
7 40 210 250 0 40 Open
30 160 110 270 0 160 Open
90 240 0 240 0 240 Closed
150 240 60 300 0 240 Reopened
210 290 0 290 0 290 Closed
300 290 0 290 45 245 Recovery realized

The initial incurred estimate understates the eventual gross amount by 90. At closure on day 90 it understates that amount by 50. Reopening increases incurred by 60 despite no disbursement on that day. Subsequent payment of 50 and release of the unused 10 reduce incurred to 290. A realized external recovery of 45 reduces net cash outflow to 245 but does not erase the original gross cost. A data warehouse that observes day 90 only, or nets the recovery backward into the date of theft, will substantially distort both emergence and maturity. In this synthetic claim net paid remains high until the recovery date, a liquidity feature invisible in ultimate net loss.

Payout Lags

Restoration as a sequence of productive states

Bartock and colleagues (2016), in NIST's Guide for Cybersecurity Event Recovery, make the prioritization of resources and realistic recovery testing central to resilience planning. In insurance research, the relevant application is the mapping from recovery activities to restoration of economically productive services. A recorded backup success or a written recovery-time objective is planning evidence. It does not establish the actual date of safe restoration or the amount of covered interruption.

The recovery chain may require independent authentication infrastructure, validated backups, replacement resources, clean production environments, security assurance, application integration, reconciliation of transactions, and confirmation of service availability by counterparties. These tasks can run concurrently when independent but have sequencing constraints where a downstream function cannot be used before an upstream function becomes trustworthy. A technical restoration milestone should specify both the service and the functional threshold. A status message that server capacity is available does not establish the reliability of the business process that depends on it.

The Home Office and Ipsos UK (2025) describe this distinction in their ransomware victim research. The 39 qualitative interviews were undertaken during February-April 2023; most attacks occurred in 2019-2022. Section 3.3 records experiences spanning days to months and distinguishes initial server reinstatement from time spent restoring and reconciling information. That is qualitative evidence of heterogeneity and multiple clocks. It is insufficient to derive population quantiles or the effect of an insurance intervention.

Productive capacity can be represented in a research record as a time-varying, defined service fraction q(t) between zero and one. This is an ordinary operational measure, not a proposed scientific theory. For a constant baseline daily contribution m, the simplified interruption amount before contractual adjustments is the sum of m times (1 - q(t)) across days. A more realistic assessment lets the baseline vary and records recaptured sales, saved expenses, and mitigation consistently. The area under the impairment curve matters more for this calculation than the first date at which any application comes online. If a low-volume service has a high economic dependency weight, a count of restored devices can be a misleading proxy.

Consider two artificial organizations with identical ten-day security recovery dates and a baseline contribution of 20 units per day. The first remains at zero capacity for five days and at 80% capacity for the next five. Its modeled contribution loss is 5 times 20 plus 5 times 4 = 120. The second remains at 50% capacity for ten days and loses 100 units. Ordering these incidents solely by time to full recovery yields a tie, although the impairment areas differ. A waiting period, a coverage cap, or variation in demand could change the indemnity ordering again.

Administrative and contractual determinants

Payment timing depends on where the payment sits in the claim sequence. Emergency forensic costs may be evidenced by an engagement letter, an approved purchase order, or an invoice. A claim for lost business contribution may require monthly accounts, a counterfactual model, documentation of mitigation, and agreement over the eligible period. The evidence-filing and determination processes are different. Timely partial or vendor payments can reduce the insured's financing burden before remaining components are settled, where the contractual and legal arrangements allow them.

The analyst should separate at least five intervals: recognition to notice, notice to triage, incurred expense to documented submission, submission to determination, and determination to disbursement. Each has a different institutional mechanism. A claim file with a late initial payment might reflect late notice, absence of a qualifying payment, missing documentation, a legitimate unresolved coverage dispute, a disputed amount, or avoidable administrative delay. Cash flow by itself cannot identify which of these explanations holds.

The insured's reimbursement horizon may be longer than the insurer's payment interval measured from receipt of a complete submission. If a vendor was paid on day 5, documentation was submitted on day 45, and reimbursement occurred on day 55, the insured financed the eligible cost for 50 days while the insurer's submission-to-payment interval was 10 days. An analysis of service quality must report both and identify who controlled the intervening steps. Direct billing to an insurer changes the insured's cash outflow without necessarily changing the restoration process.

Payment for a ransom, where lawfully permitted and covered, raises separate compliance questions. The U.S. Office of Foreign Assets Control's September 2021 updated ransomware advisory warns that payment facilitation by insurers and incident response firms can create sanctions exposure and discourages ransom payment. It is explanatory guidance, not itself binding law; applicable sanctions, parties, licenses, and legal conditions must be established for the relevant date and jurisdiction. Payment speed therefore cannot be treated as an unconditional objective. A coverage or reserve estimate must not presume reimbursement of a prohibited transfer.

Illustrative event record

UnitedHealth Group's 2024 Form 10-K reports direct response costs of $2.2 billion and estimated Optum Insight business disruption impacts of $867 million for the year following the Change Healthcare cyberattack. Management anticipated additional, smaller impacts in 2025 while rebuilding transaction volumes. These figures are corporate disclosure categories, not sums of cyber insurer indemnity. The reported direct costs include aspects of medical expenditure and liquidity support, among other response activities. They cannot be mapped mechanically to first-party cyber grants, nor should separately reported medical-cost components be added to the disclosed response total without reconciling overlaps.

The case illustrates the difference between restoring a service and reversing losses already imposed upon organizations connected to it. A claims clearinghouse can return to service while customers continue resolving transaction delays or cash shortages. For a dependent business interruption claim, the relevant causal chain concerns the policyholder's own affected function and the expressly covered provider or dependency, not the headline cost to the upstream company. The case gives no complete insurer payment or recovery schedule, and it is used solely as evidence of possible prolonged and heterogeneous economic consequences.

Financing and severity dependence

A delayed reimbursement changes the timing of available resources even when the final indemnity is fixed. For a cash payment A made at t and discount rate r, the ordinary present value under annual compounding is A/(1+r)^(t/365). An analyst evaluating insurer liability must use a prescribed valuation basis rather than an arbitrary policyholder borrowing rate. The insured's liquidity problem is distinct: the entity faces the cost of financing expenses until funds become available. In a purely illustrative simple-interest calculation, financing 100,000 currency units for 90 days rather than 30 at an annual rate of 12% adds about 1,973 currency units of financing cost (100,000 times 0.12 times 60/365). Such cost is not automatically an insured loss.

Delay may be associated with severity because severe incidents require more work, evidence, and negotiations. It may also increase severity if delays in eligible recovery activities prolong disruption. The causal effect of settlement speed on ultimate insured losses is therefore not identified by a correlation between final loss and time to payment. To investigate causation, distinguish predetermined variation, timing of insurer intervention, changing operational exposure, and the terms that determine coverage. Compare only claims eligible for similar interventions, and avoid conditioning on post-intervention quantities that are themselves consequences of the treatment.

Vendor capacity adds a portfolio dimension. Several insured entities affected by a common technology incident can require similar specialist services at the same time. A shared capacity constraint can lengthen restoration and evidence-processing times across claims. The association is conditional on exposure to the same resource and the absence of spare capacity or substitutes. It is a mechanism to examine with calendars, vendor allocation records, and correlated incident identifiers, not grounds for estimating a universal cyber catastrophe development factor.

Recoveries and Reopening

Restoration, reimbursement, and external receipts

The term recovery is especially ambiguous in ransomware science. Technical recovery is the reestablishment of secure functionality. Business recovery is a return to a specified productive or financial state. Insured reimbursement is payment by an insurer of a covered loss or expense. Third-party recovery is a receipt or legally enforceable entitlement associated with restitution, salvage, contractual recourse, subrogation, or another external source. Reinsurance recovery stems from a separate risk-transfer contract. These events differ in creditor, debtor, measure, timing, uncertainty, and contractual basis.

For a direct insurer, maintain gross loss payments, allocated expenses, third-party receipts, recovery expenses, and ceded cash transactions as separate fields. Direct net paid can be calculated as gross paid minus realized direct recoveries on the chosen basis. If recovery expenses are included in gross payments, they must not be added again; if separately classified, label the net measure clearly. A final amount net of reinsurance is only comparable with another amount on the same ceded and expense basis. The distinction also matters when a recovered asset belongs wholly or partly to the policyholder rather than the insurer.

Expected future recoveries can be a valuation input under the applicable accounting or solvency approach, but an expectation is not a realized receipt. Separate probability of entitlement, conditional collectible amount, timing, cost, currency, and uncertainty. A speculative confiscation, unresolved warranty claim, or possible reinsurance response is not established cash. Recording it as paid recovery suppresses the observed net exposure and obscures the difference between legal and credit risk.

The AXIS subrogation clause provides a concrete illustration of how a right of recourse can depend on payment and cooperation. The relevant rights, limits, preservation obligations, priority, and allocation need to be established from the operative contract and governing law. An insured loss below retention, an excess of policy limits, or multiple insurers' participation can materially affect allocation. For empirical comparisons, assign the external receipt once to the appropriate parties on the appropriate basis.

Law enforcement seizure and currency

On June 7, 2021, the U.S. Department of Justice's Colonial Pipeline seizure announcement reported seizure of approximately 63.7 bitcoin associated with a ransom of approximately 75 bitcoin in the May 2021 incident, relying on allegations in a supporting affidavit. The distinction between payment and seizure is important. The seized quantity is about 85% of the paid bitcoin quantity. That ratio is not a dollar recovery rate and establishes neither restitution to the victim nor subrogation proceeds to an insurer. The official $2.3 million seizure valuation describes the asset at the relevant reported valuation, not necessarily its value at original payment or distribution.

A cryptocurrency-denominated event requires at least two currency records: the quantity and denomination of the transferred asset, and the relevant fiat valuation with its timestamp and valuation source. Additional fees and conversion spreads may affect incurred costs. A seizure records control by an authority; a forfeiture proceeding or distribution has a separate legal basis and date. A recognized receivable, realized insurer receipt, and later currency conversion should each be recorded in their own ledger fields. This prevents conflating traceability of an asset with enforceable and collectible insurance recovery.

The Colonial Pipeline case demonstrates that recovery may occur after the extortion event and that the transfer and seizure quantities can differ. It does not reveal the conditional probability of recovery among ransomware claims. Cases selected for public disclosure of a successful seizure disproportionately illustrate successful enforcement; claims in which assets are never found or compensation is disputed may never generate a comparable release.

Reopening and long-tail components

A reopened claim is an administrative transition from closed to open under a stated rule. Possible reasons include additional covered invoices, revised interruption calculations, a delayed liability notice, reopened coverage litigation, escalation through an appeal or dispute process, a subsequent recognition of related loss, or the pursuit of an external recovery. These reasons differ in their implications for future gross claim cost. A reopening solely to record subrogation receipts should be distinguished from one creating a new payment obligation.

A claim management system may instead create a separate claim number for a subsequent coverage component. Identical loss processes will then produce different counts of reopened files across insurers. A comparative statistic needs a common incident-based grouping rule and a definition of closure: temporary operational closure, cash closure, final legal closure, or all components resolved. Record the administrative reason code, the linked underlying event, and the valuation histories both before and after the transition.

Reopening rates should be evaluated at a comparable observation horizon after first closure. The denominator consists of eligible initially closed claims with their exposure time or appropriate censoring recorded. If 1,000 claims first closed in a year and 50 subsequently reopen within twelve months, a crude 5% rate can be reported only if all 1,000 have twelve months of follow-up or the missing observation is handled appropriately. A cohort assembled six months before valuation cannot be compared directly with a two-year-old cohort. Recovery-only reactivation should be separately counted.

A subsequent attack is scientifically distinct from the late emergence of consequences of the original compromise. Determination of a new or related event requires event linkage, endpoint and identity evidence, timing, and policy aggregation rules. In anonymized research data, preserve the reason for grouping rather than reconstructing it from claim identifier alone. The aim is to avoid both duplicate ultimate loss and the artificial collapse of genuinely distinct events.

Comparative data definitions

Cash recovery ratios require maturity and comparable eligible amounts. For a cohort of original gross indemnity payments P and realized allocated recoveries R at horizon h, R(h)/P(h) is a cash ratio at that horizon, provided both are measured on the same cohort and basis. A ratio of ultimate recovery to ultimate gross loss is a different quantity. Ratios limited to successfully recovered files condition on success and cannot estimate recovery likelihood. In the presence of zero denominators a ratio is undefined; negative cash entries may represent adjustments rather than economically negative ultimate claim cost.

A risk manager may reasonably wish to measure how rapidly liquidity is restored. The appropriate analysis tracks the insured's outlay and receipts before considering its uninsured burden. An insurer instead needs the gross and direct net payment profile, the timing of ceded cash, and uncertainty in collectibility. A public authority measures stolen assets, seizure, forfeiture, and restitution. Each actor's performance indicators answer a separate question. A reproducible comparison aligns these questions instead of constructing a single unqualified recovery statistic.

Evidence Validation and Sensitivity Analysis

Source populations and limitations

Ransomware incident data are assembled through heterogeneous institutions. ENISA's 2022 official release used 623 observed incidents from the EU, UK, and US in May 2021-June 2022 and reported missing payment status in 94.2% of them. Unknown status is an independent observation category; nonpublication of stolen data does not uniquely identify payment, and public leakage does not reliably establish nonpayment. The source supplies direct evidence of missingness within the mapped sample. A portfolio estimate based on a complete-case subset would require a defensible mechanism for the missing values.

The FBI's 2025 IC3 Annual Report distinguishes recorded ransomware complaint losses from broader costs, ordinarily excluding several categories of interruption and outside remediation. Reporting channels and unreported amounts constrain interpretation. A complaint amount divided by a complaint count is not an insured average severity, and complaints need not uniquely identify indemnifiable incidents. The original report establishes this metric boundary on page 14; this study does not use an exact complaint total.

The UK's Cyber security breaches survey 2025/2026 is a population-oriented survey of organizational experiences. It provides context for observed threats and organizational practices. Self-identified and remembered attacks are conditioned on recognition and response, and the organizations reporting breaches do not constitute a set of cyber insurance claims. The actuarial exposure base instead consists of covered insured entities and relevant coverage periods, with policies, endorsements, and limits accounted for. Survey incidence requires a separate treatment from claim frequency per policy-year.

The National Association of Insurance Commissioners' 2026 cybersecurity market report presents aggregate insurance financial and loss ratio data across coverage segments. This is useful market context but cannot recover the individual ransom, operational, payment, and recovery event histories from those aggregates. A loss ratio combines exposure, premium, claims, valuation, and reporting conventions; it is not an estimate of payout lag or a ransomware recovery curve.

Together with insurer summaries, case studies, original contract comparisons, and technical guidance, these sources establish multiple valid views of the problem. None contains a fully linked, mature, representative international series of ransomware-specific insured claims with all the needed event timestamps. The study therefore makes structural and conditional conclusions supported by the evidence and uses explicit synthetic calculations. It does not publish unsupported market development factors or global recovery estimates.

Development triangles and changing composition

A conventional development triangle indexes cumulative claim amounts by origin cohort and elapsed development period. For a ransomware study, origin should be chosen to answer the research question: occurrence or impairment cohorts for incident emergence, discovery cohorts for response and notification, policy periods for underwriting results, and report cohorts for claim handling. These cohorts cannot be pooled without an explicit mapping. A policy-year analysis of a discovery-triggered grant and an occurrence-year security analysis can describe the same events at different origins.

For positive gross cumulative paid amounts C(i,j), a standard volume-weighted one-step development factor is the sum of C(i,j+1) divided by the sum of C(i,j) across the same cohorts observed at both ages. A sequence of factors projects observed cumulative claims to a chosen later horizon. The procedure assumes that the development relationships used for prediction remain sufficiently applicable to immature cohorts. Mack's (1993) chain ladder uncertainty analysis derives a standard-error formulation under assumptions for stochastic development. Distribution-free in that setting does not remove the need for conditional development and independence assumptions. Correlated incidents, evolving contract structures, calendar-year effects, and shifts in response practices can materially weaken them.

Consider this artificial cumulative gross paid triangle in arbitrary units, with ages one, two, and three measured in equal reporting periods.

Origin cohort Age one Age two Age three
A 100 180 240
B 120 198 unknown
C 140 unknown unknown

The first age-to-age factor is (180+198)/(100+120) = 1.71818. The next factor is 240/180 = 1.33333. If age three is assumed ultimate solely for this example, projected ultimate for B is 264 and for C is approximately 320.73. With A fixed at 240, total projected ultimate is approximately 824.73. Observed cumulative paid across the most recent diagonal totals 578, so the illustrative unpaid estimate equals 246.73. No real calibration, tail estimate, or prediction interval is claimed.

Suppose that recent ransomware cases are paying more slowly and that an independent scenario assumes age-one payment equals 35% of ultimate for cohort C. Its scenario ultimate is 140/0.35 = 400, with unpaid of 260, approximately 79.27 more than the chain-ladder projection for this cohort. This is a conditional scenario, not an estimated paid proportion. If the delay is purely administrative and the ultimate expected severity is unchanged, a greater case-based or independent expected-loss estimate may reveal that the paid triangle alone cannot identify the combined speed and severity changes. Conversely, if coverage limits were reduced for C, borrowing factors from older cohorts could distort ultimate in the opposite direction.

For incurred triangles, late recognition, settlement adjustments, and reopening create nonmonotonicity. For net paid triangles, recoveries can cause negative incremental amounts or zero and negative cumulative amounts. Fitting a positive ratio model blindly to such data may be undefined or misleading. Analyze gross payments and recoveries separately, reconcile them under a common valuation basis, and examine case incurred on the same claims. This preserves economically interpretable emergence and avoids a mechanically large ratio caused by a small net denominator.

Time-to-event and multistate analysis

Time-to-event methods are useful when payment or restoration is unobserved at valuation. The event must be explicit: first positive indemnity, completion of a defined technical milestone, final settlement, first reopening, or realized recovery. For a particular event, define time zero, eligible risk set, competing outcomes, and observation cutoff. An unpaid open claim at cutoff contributes right-censored time where future payment remains possible. A denied or exhausted claim may be a competing outcome rather than a noninformative censoring event. Treating all zero-paid files as if they may eventually pay can overstate an eventual payment probability.

With the product-limit estimator of Kaplan and Meier (1958), survival at a time is the product over observed event times of one minus the number of events divided by the number at risk immediately before that time, under the usual independent censoring conditions. Consider ten hypothetical eligible claims. Three first payments occur at day 7, two claims are censored at day 10, two payments occur at day 14, and one at day 21. Assuming no other departures before these dates, survival of the no-payment state is 0.70 after day 7, 0.70 times 3/5 = 0.42 after day 14, and 0.42 times 2/3 = 0.28 after day 21. The estimated median first-payment time is day 14. The example shows why mean observed payment time among completed cases misses still-unpaid claims. The numbers are illustrative and independent censoring is assumed, not demonstrated.

If payment, denial, withdrawal, or settlement without indemnity are mutually exclusive terminal outcomes for the analysis, estimate cumulative incidence with the other outcomes treated as competing events. If claims can reopen, closure is not absorbing; a multistate or recurrent-event description is needed. Distinguish time since original incident and time since entry into the present state. For recovery outcomes that are recorded only at monthly evaluations, time is interval-censored within the observed interval rather than exactly known at the later reporting date. Choosing endpoint midpoints without sensitivity analysis creates artificial clustering and biased hour-level comparisons. The statistical distinction is established in the competing-risks analysis of Austin, Lee, and Fine (2016); its application to claim outcomes here is methodological, not an inference from clinical observations.

Risk-set selection also arises when a database starts after the incident and includes only claims still open. This is left truncation or length-biased sampling depending on the selection rule. A cross-sectional inventory of unsettled cases oversamples claims with long settlement duration. Longitudinal inception cohorts are preferable for estimating delays; an open-claims inventory is useful for assessing current operational workload but answers a different question.

Identifiability and conditional sensitivity

Suppose observed paid amounts at a maturity h satisfy expected paid approximately equal to exposures times expected ultimate loss per exposure times the paid fraction by h. Many combinations of loss frequency, ultimate severity, and payment speed yield the same observed amount. Even if exposure and event frequency are known, paid dollars alone do not identify severity and payment speed. Case estimates, independent predictions of loss magnitude, transaction dates, mature cohorts, and change-point indicators can constrain the possibilities. They cannot guarantee identification if important coverage or reporting changes are unobserved.

Sensitivity analyses should vary one defined assumption or a coherent set of related assumptions: incident occurrence dating; severity and maturity profiles; coverage mix and limit erosion; case reserve adequacy; reopened claim probability conditional on closure; recovery eligibility and collectible amount; exchange rate; external dependency and claims capacity; and inflation or calendar trends. A scenario for faster technical restoration should change the impairment curve and eligible business interruption calculation consistently rather than merely applying a lower arbitrary ultimate factor. A scenario for faster administrative payment at unchanged ultimate changes liquidity and the paid pattern, not the assumed gross nominal liability.

Backtesting requires preserving forecasts as they existed at the historical valuation. Compare projected ultimate with later observed amounts at a common mature horizon, including reopenings and any unresolved tail. Use signed error, absolute error, calibration across severity and policy groups, and the ability to characterize uncertainty. For timing, compare predicted event-time probabilities with observed appropriately censored outcomes. For recoveries, test both realization probability and allocated amount. Cross-validation randomly mixing old and new valuation snapshots of the same event leaks future information; split by incident or origin cohort and time.

Causal evaluation of controls or insurer assistance requires more than a stratified duration table. Matching or regression adjustment must use preincident or pretreatment variables where appropriate. The largest incidents are often the ones receiving intensive assistance, potentially producing a misleading association between assistance and high severity. Differences in contract terms and selection into coverage may also influence observed outcomes. A robust empirical program would specify the estimand, treatment timing, eligibility, censoring, confounding, and sensitivity to unmeasured factors before interpreting an association as an effect.

Minimum data specification and reproducibility

A reproducible longitudinal record should include the following fields and semantics.

Domain Required information Main validation
Exposure Policy, insured entity, period, form version, limits, retention and coverage-component keys Match effective dates and endorsements
Incident Common incident identifier, affected functions, discovery, impairment onset and evidence precision Preserve uncertainty and related-event rules
Restoration Service threshold, measured status dates, operational dependencies and recaptured production Distinguish technical and economic endpoints
Claim Report time, status history, cause attribution, open/close/reopen reason and valuation timestamp Retain immutable historical snapshots
Amount Incremental and cumulative payments, case outstanding, expense classification and currency Reconcile transaction and valuation ledgers
Recovery Type, legal right, eligibility, gross and allocated amount, receipt time, fees and reinsurance separation No double allocation or speculative cash
Data quality Source lineage, missingness reason, changes in definitions, disclosure restrictions Audit coverage of risk sets and duplicates

Useful exact checks include whether cumulative gross payments equal the sum of signed transaction entries, whether case incurred reconciles at each timestamp, whether realized direct net cash equals gross cash less allocated receipts, whether currencies were converted at declared dates, whether reopened files retain their prior closure history, and whether time units match the policy language. Record missing values as missing, with a reason category, rather than assigning zero recovery or a zero-day delay.

Data from insurers and incident response providers can contain sensitive personal, financial, or security information. Research extracts should use approved pseudonymous identifiers, access controls, controlled event linkage, and disclosure thresholds. The minimum fields above describe research needs without demanding access to confidential attack traces or personally identifying claim documents. Outputs should retain sufficient aggregation to prevent reconstruction of an individual insured's loss while preserving meaningful maturity and coverage comparisons.

Integrated Scientific Synthesis and Research Agenda

The link between operational and financial processes

Ransomware-specific loss development is generated by the interaction of technological disruption, productive dependencies, contractual coverage, evidence acquisition, settlement, and external recovery. These mechanisms are linked but remain measurable at distinct stages. Technical improvement can reduce the underlying gross economic loss if it accelerates verified productive capacity. It may have little effect on an already exhausted coverage layer. Administrative acceleration can change cash timing without changing nominal entitlement. A recovery can change net cash without altering the original operational harm. An actuarial analysis must therefore report which state or quantity a measured improvement concerns.

The contribution of this research is an integrative application of existing actuarial definitions, recovery science, original incident reporting, and comparative contract reading to ransomware-specific development. The case ledger shows that the same claim may have multiple coherent paid, incurred, and net histories. The synthetic triangle shows that a change in paid fraction changes a projection without, by itself, establishing a new severity level. The survival illustration shows that immature observations carry information even when payment has not occurred. These examples are derivations from established methods and accounting identities; no new named theory or universal causal law is asserted.

The most important implication is the preservation of separate observation histories. Loss development estimates are evidence about the ultimate values of defined monetary quantities and the likely times of emergence. They require information about the evolving insurance payment obligation. Recovery performance estimates concern the timing and strength of an independent technical or financial outcome. The operation of a single metric, such as average days to close, can conceal both an improvement in cash delivery and a worsening of reopened severity.

An insurer observing a lower average duration after changing claim procedures should first check whether settlement categories, coverage limits, claim composition, and censoring changed. A lower average among closed claims may simply reflect the faster removal of simple claims while complex ones accumulate. Similarly, an enterprise observing faster restoration after investing in backups should test the restored service threshold, verified data integrity, dependency availability, and economic impact. The operational endpoint used in that analysis should not silently replace the insurance policy's restoration definition.

Actuarial and institutional implications

Underwriting and reserving require different perspectives on the same data. Underwriting studies prospective insured exposure and loss consequences under contract terms. Reserving studies outstanding liabilities as of the valuation instant. A response program's historical association with faster settlement can inform hypotheses for both, but any transfer of results needs the same population, form, eligibility, and event definition. Claims development measured on an occurrence basis may be informative for a discovery-triggered policy only after explicit mapping.

Insurers can study the timeliness of undisputed payment and the progression of complex components separately. Analysts should segment by coverage and contract generation, size and sector, shared dependency group, incident characteristics, and available response resources, subject to data sufficiency. Excessive segmentation produces unstable estimates; pooling across segments with demonstrably different development mechanisms produces structural bias. Hierarchical or credibility methods can help where permitted by the data and validated for the task. The choice of method is conditional on the estimand and maturity, not on the rhetorical importance of ransomware.

Reserving oversight should separate data quality, estimated ultimate, actual gross cash, expected recovery, realized recovery, and solvency or accounting presentation. A model predicting lower net ultimate because of a higher projected recovery must make that assumption transparent, particularly when the recovery is speculative, jurisdiction-specific, or sensitive to shared counterparties. A valuation decision under one accounting standard cannot be imported into a cash study without reconciliation. Likewise, reinsurance terms need their own attachment, aggregation, allocation, reporting, and collection data.

Insured enterprises need clarity about the financial bridge between incurred response expenses and reimbursed eligible amounts. A sound research dataset can quantify the funding gap without assuming that all liquidity effects are covered. For policy wording comparison, present actual definitions and claim examples rather than inferred market averages. Supervisors and researchers can use comparable data fields to examine tail development and concentration, subject to legal disclosure and confidentiality limitations.

A focused research program

A first empirical workstream would build linked incident and claim inception cohorts across several valuation dates. It would estimate discovery-to-notice, notice-to-first-payment, documented-expenditure-to-reimbursement, and restoration thresholds, with competing risks and censoring explicitly handled. Stratification by policy form and actual coverage components would be essential. All cohorts would retain reopened histories and both nominal and currency-normalized amounts.

A second workstream would reconstruct gross and net development profiles by covered component and evaluate the effect of calendar changes in contract terms, ransomware tactics, and response capacity. It would preserve historical forecasts for out-of-time backtesting. Alternative development and expected-loss approaches would be compared at common evaluation horizons, not selected only for their in-sample fit.

A third workstream would join documented recovery rights to realized recoveries and estimate their conditional probability, time, cost, and allocation. Law enforcement seizure cases would be included only where the chain from seizure to distribution and policy allocation can be substantiated. A separate provider-recourse dataset would identify contractual indemnity and legal recovery outcomes. Reinsurance collectibility would remain an analytically separate process.

A fourth workstream would investigate restoration performance using service-level trajectories and evidence of mitigation activities. It could examine whether specific verified controls or response arrangements are associated with shorter economically weighted disruption under comparable exposures. Strong causal claims would require designs addressing selection, intervention timing, and unobserved confounding. Comparative studies should publish the measurement definitions and uncertainty, and protect sensitive incident-level information.

A final workstream would study simultaneous cyber incidents with shared infrastructure or response providers, including correlated development delays. The object is to distinguish common shocks from ordinary heterogeneity in case processing and to validate whether dependence persists after accounting for contract limits and incident grouping. The empirical result, whatever its direction, would be more useful than a blanket assumption about cyber dependence.

Boundaries of the conclusions

The examined documents establish heterogeneity of restoration experiences, coverage terms, data sources, and public recovery records. They establish that multiple monetary and temporal measures must be distinguished. They do not establish internationally representative claim-development factors, precise recovery distributions, an empirical estimate of the value of paying a ransom, or a causal effect of insurance on restoration. The absence of a linked public longitudinal portfolio is a limitation of the available evidence. Rigorous conclusions can be made about accounting identities, contractual examples, and conditions for valid inference while leaving empirical parameters to suitable future studies.

Conclusion

Ransomware losses continue to develop after the immediate technical event through document acquisition, loss measurement, contractual determination, payment, reopening, and possible recovery. The date of service restoration, the date that insured costs cease accruing, and the date that an insurer completes reimbursement describe different processes. The evidence of heterogeneous restoration experience, the contrast between original policy specimens, the corporate record of downstream disruption, and the distinct legal event of asset seizure all support the practical necessity of keeping these temporal and financial records separate.

The first conclusion concerns measurement. A reliable claim history preserves the incident, policy, coverage component, valuation date, and transaction identities. Gross paid plus case outstanding gives case incurred on a consistent basis; gross cash less allocated realized recoveries gives direct net cash. Ultimate uncertainty and reinsurance are separately described. Reopening is an observed state change rather than a retrospective alteration of prior valuations. Comparative recovery statistics must distinguish legal entitlement from a receipt and an administrative closure from final maturity.

The second conclusion concerns mechanism. Secure technical recovery and business recovery shape the loss through the period and depth of impaired productive capacity. Insurance wording determines which part of that harm enters covered liability. Evidence and administrative processes determine how it becomes recognized and paid. External recovery changes the direct net burden through separate legal and financial channels. The timing of any one outcome cannot be inferred reliably from another without the individual causal and contractual record.

The third conclusion concerns statistical inference. Contemporary public complaint totals, survey results, market aggregates, and insurer summaries are valid for their defined populations and measures. They cannot substitute for mature longitudinal insured claims. Conventional development factors, time-to-event analysis, and recovery valuation remain useful conditional on observability, applicable assumptions, stable cohort definitions, and an adequate treatment of changing coverage, censoring, and dependence. Transparent examples in this monograph show these relationships and their sensitivity; they provide no asserted empirical calibration.

The sound scientific treatment of ransomware loss development consequently requires a coherent chronological record of restoration, economic loss, insurance liability, gross payments, external recoveries, and outstanding uncertainty. Such a record allows insurers, researchers, and institutions to distinguish operational resilience, contractual performance, liquidity, ultimate severity, and the remaining protection gap. The central research priority is the development of comparable appropriately governed longitudinal observations at these distinct levels. This conclusion follows from the examined evidence and established scientific methods; no independent new theory of ransomware insurance or universally valid recovery law is claimed.

References

  • Austin, P. C., Lee, D. S., & Fine, J. P. (2016). Introduction to the analysis of survival data in the presence of competing risks. Circulation, 133(6), 601–609. DOI: 10.1161/CIRCULATIONAHA.115.017719. Source

  • AXIS. (2020). AXIS Cyber Insurance Policy. Specimen form AXIS 1012561 0120, pp. 2, 8, 22. Source

  • Bartock, M., Cichonski, J., Souppaya, M., Smith, M., Witte, G., & Scarfone, K. (2016). Guide for Cybersecurity Event Recovery. NIST Special Publication 800-184. National Institute of Standards and Technology. DOI: 10.6028/NIST.SP.800-184. Source

  • Coalition. (2026). 2026 Cyber Claims Report: Global key findings. Public report landing page, experience January 1-December 31, 2025. Source

  • Connolly, L. Y., Wall, D. S., Lang, M., & Oddson, B. (2020). An empirical study of ransomware attacks on organizations: an assessment of severity and salient factors affecting vulnerability. Journal of Cybersecurity, 6(1), tyaa023. DOI: 10.1093/cybsec/tyaa023. Source

  • Department for Science, Innovation and Technology & Home Office. (2026). Cyber security breaches survey 2025/2026. UK official statistics. Source

  • European Union Agency for Cybersecurity. (2022, July 29). Ransomware: Publicly reported incidents are only the tip of the iceberg. Official release of ENISA Threat Landscape for Ransomware Attacks. Source

  • Federal Bureau of Investigation, Internet Crime Complaint Center. (2026). 2025 IC3 Annual Report. Ransomware as Reported to IC3 in 2025, p. 14. Source

  • Friedland, J. F. (2009). Estimating Unpaid Claims Using Basic Techniques. Casualty Actuarial Society, Chapter 1. CAS syllabus study note, pp. 1-416. Source

  • Home Office & Ipsos UK. (2025, January 14). The experiences and impacts of ransomware attacks on individuals and organisations. UK Home Office. Source

  • Kaplan, E. L., & Meier, P. (1958). Nonparametric estimation from incomplete observations. Journal of the American Statistical Association, 53(282), 457–481. DOI: 10.1080/01621459.1958.10501452. Source

  • Mack, T. (1993). Distribution-free calculation of the standard error of chain ladder reserve estimates. ASTIN Bulletin, 23(2), 213-225. DOI: 10.2143/AST.23.2.2005092. Source

  • National Association of Insurance Commissioners. (2026). Report on the Cybersecurity Insurance Market. Section 1, The Cyber Insurance Market: Financials and Data, figure 8. Source

  • Technology Insurance Company, Inc. (2021). AmTrustCyber. Sample policy form CY990001 0321, declarations and pp. 1, 2, 9. Hosted by Victor Insurance. Source

  • U.S. Department of Justice. (2021, June 7). Department of Justice seizes $2.3 million in cryptocurrency paid to the ransomware extortionists Darkside. Office of Public Affairs. Source

  • U.S. Department of the Treasury, Office of Foreign Assets Control. (2021, September 21). Updated advisory on potential sanctions risks for facilitating ransomware payments. Source

  • UnitedHealth Group Incorporated. (2025). Annual Report on Form 10-K for the year ended December 31, 2024. Securities and Exchange Commission, accession 0000731766-25-000063, MD&A. Source