RESEARCH DOSSIER
Risk as Hazard, Exposure, and Vulnerability
Abstract. Risk emerges through the interaction of uncertain hazardous processes, the people and assets exposed to them, the susceptibility and adaptive capacities of those systems, and the consequences recognized by a specified valuation or welfare criterion. This monograph examines how those components acquire causal, probabilistic, economic, and actuarial meaning. It compares the terminology of the Intergovernmental Panel on Climate Change (IPCC), the United Nations Office for Disaster Risk Reduction (UNDRR), and international risk-management standards with scenario analysis, probabilistic catastrophe modeling, vulnerability research, and insurance accounting of loss. The analysis distinguishes event occurrence, local intensity, exposed value, conditional damage, realized economic loss, and contractual insured loss. It derives a transparent conditional expectation representation, tests simplified risk products against counterexamples, works through numerical illustrations, and analyzes interdependence, dynamics, uncertainty, social vulnerability, and insurance protection gaps. The principal finding is that hazard, exposure, and vulnerability provide an indispensable causal vocabulary, while quantitative conclusions depend on joint distributions, time horizons, units of exposure, damage mechanisms, institutional conditions, and coverage terms. The research contributes a documented comparative synthesis and original analytical demonstrations rather than a new named theory or forecasting model.
Keywords: Risk theory; hazard; exposure; vulnerability; susceptibility; capacity; catastrophe modeling; conditional loss; insurance loss; disaster risk; probabilistic risk assessment; loss exceedance; social vulnerability; risk financing; nonstationarity; adaptation.
Introduction
The study of risk requires an explicit answer to several connected questions: what can happen, where and to whom it can happen, what determines the severity of its effects, and how those effects are evaluated. A flood can traverse an unoccupied floodplain without producing property claims. An occupied riverbank can remain undamaged in a season without major inundation. The same flood depth can destroy one building and leave another usable because elevations, materials, maintenance, and protective measures differ. A building can sustain substantial economic damage while its insurer pays only a fraction of the loss because of deductibles, exclusions, and limits. Each statement concerns a distinct stage of the risk process.
The conventional hazard–exposure–vulnerability vocabulary identifies these stages and their interaction. The IPCC's AR6 Working Group II glossary treats climate-related risk as a dynamic interaction among climate hazards, exposure, and vulnerability, influenced by responses and changing conditions. The UNDRR's definition of disaster risk also incorporates capacity and specifies a period within which potential losses arise. Both perspectives direct attention beyond the physical event toward the system that may be harmed.
In risk management more broadly, ISO 31000:2018 provides guidance for handling uncertainty concerning organizational objectives. That management concept extends to opportunities and objectives that exceed disaster-loss analysis. This monograph therefore fixes its principal quantitative object as an uncertain adverse consequence over a stated observation horizon, while explicitly identifying whenever the more inclusive management concept applies. A definition suitable for disaster damage does not automatically constitute the complete definition of financial, organizational, ethical, or strategic risk.
Kaplan and Garrick's (1981) quantitative scenario formulation provides a complementary structure: scenarios, their likelihoods, and their consequences. A scenario connects a physical or social initiating process with a specified state of exposed elements, vulnerability mechanisms, and outcomes. The scenario formulation identifies the object whose probability is calculated. The hazard–exposure–vulnerability analysis explains why a given scenario has the consequences it does. Their combination is particularly useful to insurance, where ground-up losses must then be translated into insured claims.
This dossier uses comparative conceptual analysis, conditional probability, transparent arithmetic examples, institutional comparison, and critical review of primary standards and peer-reviewed research. The numerical illustrations use explicitly stipulated assumptions; they are not empirical estimates for a location, hazard, insurer, or portfolio. The treatment spans natural and anthropogenic hazards when a physical or operational loss mechanism can be specified. It emphasizes catastrophe and property risk as analytically tractable examples while testing the limits of that emphasis for human well-being, social vulnerability, and indirect economic loss. The organization follows six substantive chapters approved for this dossier's research architecture.
Risk Events and World States
Definitions and analytical boundaries
A hazard denotes a potentially damaging process or phenomenon. The UNDRR hazard terminology encompasses natural and human-induced processes and describes hazards through location, intensity, frequency, and probability. An event is a particular manifestation of such a process, with a location, duration, and intensity footprint. A flood hazard is the possibility of inundation; a flood event is a realized spatial-temporal inundation episode. Hazard maps that show only the chance of an intensity threshold being exceeded describe a specific feature of hazard, not a complete loss distribution.
Exposure concerns the presence of persons, infrastructure, activities, or economic and environmental interests in settings that can be affected. The UNDRR exposure definition emphasizes tangible assets and people in hazard-prone locations. Climate assessments also include livelihoods, ecosystems, and cultural assets. An exposure inventory can be measured in people, structures, floor area, replacement value, business revenue, ecological services, or an insured limit; these measures are related but not interchangeable. The phrase exposed value must always specify what is valued and in which units.
Vulnerability describes susceptibility to harm. The UNDRR vulnerability definition explicitly includes physical, social, economic, and environmental conditions. For buildings, vulnerability can be represented by a conditional distribution of damage given hazard intensity and structural attributes. For a household, it may include livelihood dependence, access to transportation, financing, social support, and recovery capacity. A single numerical vulnerability factor often conceals materially different mechanisms.
Capacity denotes resources and institutions available to prevent harm, respond to an event, and recover. The UNDRR capacity definition includes infrastructure, knowledge, social relations, management, and organization. In a quantitative model, some capacities reduce physical damage before an event, while others affect emergency consequences, duration of business interruption, or recovery costs afterward. Treating all capacities as an undifferentiated subtraction from vulnerability obscures their temporal placement and can count the same protection twice.
A rigorous risk proposition consequently states its object: the probability of death, expected building repair cost, annual portfolio claim amount, probability of service interruption, or loss of ecosystem function. Each object requires its own consequence measure and causal scope. The phrase greater risk is incomplete until the loss measure, horizon, population, and reference state are specified.
Scenario structures and probability
The Kaplan–Garrick triplets are useful precisely because physical and financial uncertainty can be attached to explicit scenarios. Let a scenario s describe a hazard event, exposed system, and relevant environmental and institutional circumstances; let p(s) be its probability under an identified model; and let x(s) be its consequence. In a finite exhaustive partition of mutually exclusive scenarios, expected loss is E[L] = Σ_s p(s)x(s). Each scenario must have a clearly defined temporal basis. A collection of overlapping event descriptions is not such a partition: simply summing its probabilities or losses creates double counting.
A hypothetical urban storm illustrates the distinction. Possible states might be no damaging inundation, shallow flooding, deep flooding without electricity failure, and deep flooding with prolonged electricity failure. These scenarios can be made mutually exclusive and exhaustive for the specified event mechanism. Their conditional losses will depend on floor elevation, occupancy, contents, utility dependence, and available recovery resources. Merely ranking storm intensity does not determine scenario loss because infrastructure outcomes and social conditions remain uncertain.
Probabilities may be conditional on an event occurring or unconditional for a calendar year; changing the denominator changes the interpretation. If a hazard event occurs with annual probability q, and damage has conditional probability r given the event, the annual probability of this damage mechanism equals q r under the specified conditional structure. If there are multiple event types, their losses are added in expectation only after duplicate events, overlapping categories, and dependence are handled consistently.
Scenario completeness is an epistemic obligation, not an empirical guarantee. Missing hazards, omitted dependence, flawed exposure inventories, and unmeasured recovery needs can make a modeled risk distribution materially narrower than the possible consequence set. IEC 31010:2019 addresses the selection, application, verification, and validation of assessment techniques; a credible choice of technique begins with the question, evidence, and decision that it is intended to support.
Causal states and compound events
Risk changes with the state of the world at the time a hazard is realized. Let S_t denote the relevant state of assets, location, construction, infrastructure, wealth, health, regulation, and preparedness at time t. A hazard event E_t and this state jointly determine a conditional distribution of consequences. Formally, P(L ≤ l | E_t, S_t) can vary even when the hazard intensity is held fixed. This formulation separates the stochastic realization of a hazardous process from the changing response of an exposed system.
The IPCC's chapter on key risk concepts explicitly addresses compound, cascading, and aggregated climate risks. A flood can disable power, which impedes pumps, which increases inundation and extends business interruption. The event consequences arise through a network of conditional mechanisms. The UNDRR's Global Assessment Report 2022 similarly emphasizes how contemporary systemic exposure and connected infrastructures can transmit disruption beyond the original physical footprint.
The distinction between coincidence and causation is consequential. Two severe storms occurring in one season may simply be statistically dependent; a fire following electrical damage from a storm is a causal cascade; a regional telecommunications outage that impairs multiple firms is a common-cause mechanism. The hazard–exposure–vulnerability vocabulary applies to each, but the stochastic structure differs. For compound risk, the event model must specify joint occurrence or conditional sequence rather than multiply independent probabilities by default.
Hazard and Exposure
Hazard fields, frequency, and intensity
A hazard is described adequately for loss analysis only when its spatial, temporal, and intensity dimensions match the vulnerability mechanisms under consideration. An earthquake intensity map can indicate local shaking intensity; a coastal flood map may represent depth and duration; a wind model may describe gust speed and orientation. A damage relation calibrated to flood depth cannot be applied directly to a storm's recurrence interval. Hazard intensity is an input to a conditional consequence function, whereas occurrence probability governs the frequency or likelihood with which those inputs arise.
A continuous-time event count N(T) may be modeled by a Poisson process with constant rate λ solely when its assumptions are justified. Under that assumption, P[N(T) ≥ 1] = 1 − exp(−λT). If λ = 0.01 events per year, the probability of at least one event in a year is approximately 0.995%, and over thirty years it is approximately 25.9%. The distinction between a 1% annual exceedance probability and a strict mean waiting time of 100 years is significant; event clustering, changing conditions, and threshold definitions can invalidate a simple return-period reading. These calculations establish mathematical implications of a model, not measured hazard frequencies.
Intensity–frequency curves require special interpretation. A 1% annual probability of exceeding a particular flood depth does not describe the probability of any flood, and a return level is not necessarily the expected intensity of a flood that occurs. Furthermore, the hazard distribution and the loss distribution differ because the latter includes exposure and vulnerability. The OpenQuake classical risk workflow illustrates the need to combine site hazard curves with loss-relevant vulnerability functions and exposure inventories.
Exposure is geographical only in part. A port may be directly exposed to storm surge while its customers are indirectly exposed to supply interruptions. A property portfolio concentrated near one shoreline shares a hazard footprint; an insurer with exposures across two distant regions can nevertheless share a single reinsurance tower or operational dependency. For a narrowly defined direct physical-damage model, exposure can be indexed by asset coordinates. For system-level consequences, the relevant exposure domain includes operational and economic relationships.
Geographical and temporal exposure
Suppose a flood polygon overlaps fifty buildings. A building count is an exposure measure; the sum of replacement costs is another. Neither includes vulnerability until floor elevation, construction, flood duration, and damage functions enter the calculation. An address-level geocoding error can assign a building to the wrong side of a levee, causing a discrete loss-model error despite only a small positional displacement. Spatial aggregation can conceal extreme concentrations inside a census unit or coastal parcel. Exposure datasets therefore need stated location precision, valuation date, occupancy definition, taxonomy, and reconciliation controls.
A stock of exposed capital also evolves. New construction raises replacement value; stronger building codes may lower physical vulnerability for new construction; inflation raises nominal repair cost without changing physical hazard. Let V_i(t) denote the value of asset i and A_i(t) its location, occupancy, and structural attributes. An expected-loss forecast conditional on today's inventory is not the same as an expected-loss forecast using a future inventory. Annual risk requires a time-indexed joint specification of event occurrence and the evolution of those assets.
The IPCC climate-risk glossary explicitly notes that hazards, exposure, and vulnerability can vary through time and space. This is an instruction for empirical attribution. An observed increase in insured storm claims may reflect more structures, higher replacement costs, changing policy conditions, altered claim practices, physical hazard changes, or several interacting causes. Relative trends in insured loss cannot, by themselves, isolate changes in meteorological hazard.
Exposure migration also produces selection effects. Households may relocate out of flood-prone areas after a damaging event; commercial value may accumulate behind flood defenses; insurers may restrict coverage or increase deductibles in high-risk locations. The subsequent insured exposure set is endogenous to perceived hazard, loss experience, and institutional responses. Comparisons of observed claims across years must therefore distinguish changes in risk from changes in which risks are observed and insured.
Attribution, resolution, and evidence
Bouwer (2011), in a review of historical disaster losses, finds that exposure and asset value have been major drivers of recorded increases in weather-disaster costs in the historical literature he reviewed. That finding belongs to its data period, methods, and publication date; it is not a universal claim that physical climate hazards have remained unchanged or that climate change cannot affect present or future loss. The more general methodological point remains: raw economic loss trends mix physical, socioeconomic, and institutional changes.
The same separation is apparent in Swiss Re Institute's 2026 natural catastrophe assessment, which reports USD 107 billion in insured natural catastrophe losses in 2025 and emphasizes continued exposure growth and adaptation needs. This total is an observed insurance-sector aggregate under Swiss Re's reporting conventions, not a hazard-frequency statistic or the total economic damage worldwide. A year of comparatively modest realized claims can coexist with material long-horizon tail exposure.
To attribute a loss change to a particular determinant, a study must state its counterfactual. Holding exposure and vulnerability constant while changing the hazard distribution answers one question; holding the hazard fixed while allowing urban development answers another. Counterfactual reconstructions are model-based and subject to data limitations. Causal separation is impossible where available data cannot identify distinct effects, or where hazard, land use, protection, and prices coevolve without credible controls. Reporting such limits constitutes an essential part of scientific interpretation.
Vulnerability and Susceptibility
Physical fragility and economic damage
Physical vulnerability describes the conditional consequences of hazard intensity for a specified object or class. A fragility function F_d(h,x) = P(D ≥ d | H = h, X = x) yields the probability of exceeding damage state d conditional on hazard intensity h and characteristics x. A vulnerability or loss-ratio function may instead specify m(h,x) = E[R | H = h, X = x], where R is the fraction of a defined value lost. The functions are distinct: damage-state probabilities do not determine a monetary loss distribution without additional state-to-loss mappings, and an expected loss ratio does not specify the distribution of extreme damage.
The OpenQuake risk documentation distinguishes fragility-based damage calculations from vulnerability-based loss calculations. The event-based loss calculation methodology combines event-specific ground-motion intensity, asset vulnerability, and stochastic loss ratios, including optional dependence assumptions. These are concrete examples of conditional physical-to-financial modeling. Their earthquake-specific parameters cannot be moved without calibration to flood, wildfire, cyber, or health hazards.
Consider two fictional properties, each with a replacement value of USD 1 million, facing a specified event with annual probability 2%. Under the stipulated event, the first has a deterministic 20% damage ratio and the second a 60% ratio; assume zero loss otherwise and only one event state per year. Their respective annual expected ground-up losses are 0.02 × 1,000,000 × 0.20 = USD 4,000 and 0.02 × 1,000,000 × 0.60 = USD 12,000. Hazard probability and exposed value are equal, but conditional vulnerability changes expected loss by a factor of three. These figures are teaching examples, not empirical vulnerability calibrations or premium indications.
Suppose a mitigation measure reduces the second property's event-conditional loss ratio from 60% to 36%, with event frequency, exposed value, and all else held fixed. Its expected annual ground-up loss becomes USD 7,200, a reduction of USD 4,800 or 40% from the prior expectation. This result follows from the deliberately linear, single-event example. The percentage reduction could differ for a measure that prevents damage only below a threshold, affects downtime rather than repair cost, or changes the event process itself.
Loss-ratio models need careful denominators. An expected repair cost equal to 30% of building replacement value is not necessarily 30% of market price or 30% of the total insured amount. Building and contents losses can have different vulnerability curves. A facility's business-interruption loss may exceed direct repair cost without violating any physical damage relation, because operational revenue and restoration time follow separate conditional mechanisms.
Social vulnerability and coping capacity
Physical susceptibility alone cannot represent vulnerability across people and institutions. Turner et al. (2003), in their framework for vulnerability analysis, relate vulnerability to the coupled human-environment system, exposure, sensitivity, and adaptive capacity. Cutter, Boruff, and Shirley (2003), in social vulnerability research, construct a geographically differentiated index of socioeconomic factors affecting vulnerability in the United States. Their empirical index depends on its study period, geographic units, and variable selection; it is not a universal monetary conversion of social conditions into a stable building damage percentage.
For two households in comparable flood-affected dwellings, identical property damage can translate into very different welfare and recovery effects. One may have liquid savings, safe transportation, temporary housing, health coverage, reliable communications, and insurance; another may face disability-related evacuation barriers, insecure tenancy, lost wages, and limited access to finance. These differences concern the capacity to avoid and absorb harm, the duration of disruption, and the welfare implications of a given economic loss. A model of insured building claims will record only the effects that enter covered claims.
The UNDRR disaster risk assessment guidance includes health, environmental, social, and economic dimensions and the effectiveness of coping capacity. The scientific question is how to represent those dimensions without falsely claiming a single natural metric. Household displacement, mortality, income interruption, service availability, and cultural losses may need a multidimensional consequence vector rather than a single dollar outcome. Monetary conversion is a normative valuation choice that must be disclosed and justified.
Socioeconomic indices can support geographical screening and prioritization, yet aggregation can disguise within-area inequality. The validity of an index depends on construct definition, weighting, missing data, spatial scale, and the population to which it is transferred. An index that predicts inability to recover may not reliably predict the probability that a specific bridge collapses; the two questions engage different vulnerability mechanisms. Good assessment preserves these distinctions.
Heterogeneity, dependence, and mitigation
Heterogeneity matters even when mean damage ratios coincide. Two categories can both have an expected event loss ratio of 20%, yet one may almost always suffer 20% damage while another suffers complete loss in 20% of comparable events. Their expected property losses coincide at the same exposed value, but deductibles, limits, repair planning, solvency risk, and distributional welfare effects differ. A mean vulnerability curve cannot replace a full damage distribution when tail outcomes govern the decision.
Damage uncertainties are also correlated. In the same storm, common wind-direction error, unnoticed construction defects, shared power failure, or the same model-calibration bias can affect many assets together. Independence assumptions may leave expected aggregate loss approximately unchanged while understating extreme portfolio loss. The event-based OpenQuake documentation explicitly distinguishes independent and correlated stochastic loss-ratio sampling, which demonstrates why dependence belongs in vulnerability modeling rather than solely in hazard-event modeling.
Mitigation can influence more than one determinant. A seawall may reduce local flood intensity under ordinary events, change which land attracts development, alter emergency behavior, and affect residual extreme failure scenarios. A building retrofit primarily modifies conditional damage; relocation primarily modifies exposure; early warning can change injury and interruption outcomes through preparedness. Each intervention should be evaluated under an explicit baseline and counterfactual event distribution. Counting the same effect as reduced hazard and again as reduced vulnerability exaggerates the benefit.
Vulnerability evidence must be hazard- and asset-specific, documented in units, and tested outside the construction data when possible. Historical claims may offer observed repair costs but can be truncated by deductibles, limits, uninsured damage, and claim reporting. Engineering experiments offer physical information but may fail to represent workmanship, maintenance, or socioeconomic recovery. Combining these sources requires a mapping between the quantity observed and the parameter inferred.
Valuation of Consequences
Damage, loss, welfare, and financial accounting
Damage is a physical or functional change; loss is an assessed consequence under a selected measure. Repair cost can exceed pre-event market value in some circumstances, and an asset's absence from a commercial property register does not imply absence of welfare or ecological value. The analyst must specify direct physical damage, debris removal, temporary relocation, business interruption, mortality or morbidity, infrastructure service disruption, and indirect supply-chain effects as separate categories before adding quantities that share a compatible unit and accounting boundary.
An exposed asset may carry replacement cost V_r, accounting book value V_b, market value V_m, and insured sum S. Each is valid for a different question. A repair-loss model often uses replacement cost and damage ratios; a financial statement may use accounting rules and impairment; a policy may contain actual-cash-value provisions, deductibles, sublimits, or coinsurance. Treating these valuations as interchangeable silently changes the modeled loss. Exchange rates, inflation, rebuilding-cost inflation, and valuation dates must be standardized before comparative totals are interpreted.
Consider a fictional facility with direct damage of USD 500,000 and six weeks of interruption. Suppose the lost contribution margin after avoided expenses is USD 150,000, emergency relocation costs are USD 30,000, and USD 20,000 of repair cost is already included in the direct damage estimate. Economic loss under a specified owner-accounting boundary is 500,000 + 150,000 + 30,000 = USD 680,000; adding the overlapping USD 20,000 again would be double counting. The calculation excludes impacts on employees, suppliers, health, and the community. A broader social assessment would require further categories with their own attribution and valuation rules.
The IPCC Working Group II risk assessment recognizes climate impacts and responses across coupled human and ecological systems. Such consequences cannot be represented fully by insured monetary loss. Insurance figures are invaluable evidence for covered financial burden, but they are a filtered observation of wider social and ecological consequences.
Expected loss, exceedance, and the horizon
Expected annual loss is the probability-weighted mean of loss over a defined year, given a specified model and exposure state. For mutually exclusive hypothetical annual states with probabilities 0.90, 0.08, and 0.02 and losses USD 0, USD 100,000, and USD 1,200,000, respectively, E[L] = 0.90(0) + 0.08(100,000) + 0.02(1,200,000) = USD 32,000. The probability that annual loss exceeds USD 100,000 is 2%, because the USD 100,000 state equals rather than exceeds the threshold; the probability of loss at least USD 100,000 is 10%. The choice of strict or inclusive exceedance convention matters when the distribution contains discrete masses.
A loss-exceedance curve records P(L > l) or, under a declared convention, P(L ≥ l) against threshold l. It conveys features of severity that expected loss alone cannot show. A purported 100-year loss estimate must specify whether it refers to an occurrence exceedance level, an annual aggregate exceedance level, and which annual exceedance convention is used. The labels cannot be read directly as a promise that one event occurs at a fixed interval. OpenQuake's classical risk outputs demonstrate site-level loss curves; calculating portfolio aggregates additionally requires dependence information.
For an illustrative stationary independent annual model with annual exceedance probability 1%, the chance of one or more exceedances in thirty years is 1 − 0.99^30 ≈ 26.03%. If the annual probabilities change with exposure, protection, or climate conditions, the corresponding probability becomes 1 − ∏_(t=1)^30 (1 − p_t) only under conditional independence of annual exceedance events. Under interannual dependence, the joint probability of no exceedance must be estimated instead of multiplying annual complements.
Risk evaluation needs multiple distributional summaries. Expected loss informs average burden and some pricing questions; high quantiles inform severe but finite thresholds; tail conditional expectations illuminate losses beyond those thresholds; scenario analysis clarifies plausible mechanisms. The full distribution may be inadequately observed, and different measures answer different decisions. A single number should not be represented as the entire content of risk.
Dependency and model uncertainty
Suppose two fictional properties each incur USD 1 million if damaged, with annual damage probability 1%. Their expected aggregate annual loss is USD 20,000 whether their damage indicators are independent or perfectly correlated. Under independence, the probability that both incur damage is 0.01^2 = 0.01%, and the probability that at least one incurs damage is 1 − 0.99^2 = 1.99%. Under perfect positive dependence, both damage together with probability 1%, while the probability of at least one damaged asset is 1%. The probability of a USD 2 million annual loss thus differs by a factor of 100 between those two models despite identical marginal hazard exposure and expected aggregate loss.
Dependence can originate in the hazard field, vulnerability residuals, supply-chain relationships, or insurance contract structure. It cannot be inferred automatically from portfolio geography alone. A model that aggregates independently calculated per-asset loss curves may fail to reproduce an aggregate exceedance curve where common shocks dominate. The OpenQuake probabilistic risk methodology makes correlation treatment an explicit model choice.
Epistemic uncertainty concerns incomplete knowledge of hazard rates, spatial exposure, vulnerability parameters, economic valuations, dependence, and model structure. Aleatory variation concerns events and outcomes represented as random under a fixed model. The distinction is useful for evidence disclosure, though in practical Bayesian modeling uncertainty about parameters is expressed probabilistically as well. A confidence interval or posterior range for expected annual loss is a statement about estimation and assumptions; it must not be confused with the probability distribution of one year's realized losses.
Sensitivity analysis should vary material assumptions, including exposure growth, damage functions, event clustering, policy terms, and inflation. A model can be internally correct and still deliver misleading conclusions if its input inventory omits large classes of buildings or if historical observations systematically underrecord small disasters. The UNDRR GAR2022 systemic-risk account supports assessment of interacting and omitted drivers rather than treating model outputs as complete descriptions of social risk.
Insurance Applications
From ground-up damage to contractual claims
A catastrophe model produces a distribution of physical and economic losses under defined coverage-relevant assumptions; an insurance model applies the actual financial terms. Let G be eligible ground-up loss after contractual definitions of insured property and covered peril are satisfied. For a simplified single occurrence with deductible d and payment cap u, insurer payment is I(G) = min(max(G − d, 0), u). This function excludes coinsurance, participation, exclusions, reinstatements, aggregate deductibles, expenses, and claims-development effects. Their inclusion requires further contract-specific operations and careful ordering.
For the fictional loss G = USD 800,000, deductible d = USD 100,000, and maximum policy payment u = USD 500,000, the insurer pays min(700,000, 500,000) = USD 500,000, leaving USD 300,000 of the eligible gross loss to the insured or other parties. This example specifies u as a cap on insurer payment after the deductible. If a contract instead defines a total damage limit before a deductible, the correct calculation may differ. Identical hazard, exposure, and vulnerability can therefore yield different insured losses across insurance contracts.
Oasis Loss Modelling Framework's financial module documentation distinguishes ground-up, insured, and net-of-reinsurance loss perspectives. These are separate transformation stages. An exposure measured by insured limits may not equal the economic value available to be damaged, and a treaty's ceded recovery depends on attachment, exhaustion, share, reinstatement, and aggregation provisions. Policy and treaty operations should be applied to event-level or period-level loss arrays as specified in the contracts, not indiscriminately to averages.
The distinction affects evidence. Insurer claims data reflect reporting behavior and policy conditions as well as hazardous events and economic damage. A reduction in reported claims following tighter deductibles does not establish lower physical vulnerability. Increasing take-up can raise insured loss while leaving expected economic loss unchanged. Exposure-based pricing and policy analyses should preserve both ground-up and insured measures.
Underwriting, pricing, and accumulation
An insurance underwriting model asks which covered event and consequence distribution belongs to the proposed portfolio over the policy term. It may use hazard data, geocoded values, building taxonomies, vulnerability curves, claim frequency and severity, policy wordings, and contractual limits. Expected claims provide one input to technical pricing, but administrative expense, capital, risk margin, reinsurance, taxes, legal constraints, and competitive conditions remain separate from the pure expected-loss component.
Loss occurrence models and premium models use different exposure measures. Policy-years may be appropriate for claim frequency; replacement cost may determine severity sensitivity; insured limits bound some payouts but are not equivalent to hazard-prone asset value; population at risk is relevant to human consequences. Combining observations from different insurance products without harmonizing these denominators can create spurious comparisons in claim incidence and severity.
Accumulation risk arises when multiple covered properties or insureds share hazard events or correlated vulnerability. A portfolio with individually low expected annual losses can generate severe aggregate claims when geographically concentrated or dependent on shared infrastructure. A proper portfolio analysis retains event identifiers and joint losses. Applying a uniform diversification credit without testing geographical, physical, and operational dependence understates potential insurer capital and reinsurance needs.
The World Bank's disaster risk financing work describes the role of insurance alongside reserves, contingent financing, and other instruments across different loss layers. The economic rationale is that financing instruments differ in cost, timing, loss-trigger requirements, and access. These instruments affect the allocation and financing of losses; they do not by themselves change the physical occurrence of hazards or the underlying damage process, although associated incentives and investments may influence behavior over time.
Insurance design, adaptation, and risk financing
Indemnity insurance commonly links payment to measured eligible losses; parametric insurance links payment to a predefined observable index or event parameter. Parametric design can support rapid liquidity but introduces basis risk: the trigger may fail to pay despite substantial actual damage, or may pay when actual damage is limited. The magnitude of basis risk depends on correlations between the index, local hazard realization, exposed values, and actual vulnerability. Hazard data alone cannot guarantee alignment with beneficiary needs.
An adaptation program has different effects depending on its mechanism. Strengthening a roof alters damage conditional on wind intensity; flood zoning influences future settlement and thus exposure; improving forecasting may reduce injury and business interruption through response capacity; insurance can accelerate recovery through liquidity. Complementarity is possible, yet moral hazard or adverse land-use incentives may offset part of the intended benefit. Evaluation therefore asks both how the intervention changes the loss distribution and how it changes the behavior generating exposure.
Swiss Re's 2025 loss analysis issued in 2026 describes USD 107 billion of insured natural catastrophe losses for 2025, with recurrent losses from wildfire and storm risk and sustained attention to the protection gap. Aggregate insured totals do not establish the distribution of coverage across income groups or countries; social protection analysis requires household and jurisdiction-specific data. A protection gap may be measured as observed economic losses minus insured losses for a defined event set, or as a prospective modeled protection need; those two measures must not be conflated.
In regulatory and public-policy analysis, affordability, access, contract clarity, adaptation incentives, and fiscal capacity matter alongside solvency. Price signals can guide mitigation where households can act on them, but vulnerable households with constrained finance may be unable to modify their exposure. Combining risk reduction, transparent risk financing, and targeted protective institutions is therefore a substantive policy problem, not a mechanically solved optimization within hazard curves.
Conceptual Synthesis
Conditional loss representation
The preceding sources support a precise accounting interpretation. Hazard, exposure, vulnerability, and consequences are distinct analytic components whose interaction generates a loss distribution. A simplified multiplicative expression such as Risk = Hazard × Exposure × Vulnerability can provide a qualitative teaching heuristic when terms and units are defined. As a general quantitative identity it is incomplete. Hazard may be a probability, an event rate, or an intensity field; exposure may be people, dollars, or assets; vulnerability may be a conditional mean ratio, a probability of damage, or a multidimensional social condition. Their product has no stable interpretation without compatible definitions, a scenario structure, and assumptions about dependence.
For a discrete set of mutually exclusive event states e with probabilities p_e, exposed assets i, states x_i, values V_i, and random event-conditional loss ratios R_i(e,x_i), the following illustrative identity holds whenever the expectations exist and the assets' losses are additively defined:
E[L] = Σ_e p_e Σ_i E[V_i R_i(e,x_i) | e].
Here V_i may itself be random and correlated with R_i; it must remain inside conditional expectation unless conditional independence or determinism justifies factoring it out. The event-state partition can include a no-loss event. This identity is an application of total expectation and linearity, not a new authored mathematical theory. It describes expected additive gross losses, not the entire joint loss distribution or nonadditive social damages.
For a fixed positive V_i, the simplified term V_i E[R_i | e] is available if V_i is deterministic at the scenario's valuation date. For a changing exposure stock, V_i(t) and R_i(e,t) must be indexed by time and relevant states. An event-rate integral may be used for expected loss under a specified point process, but the rate measure, double-event handling, loss attribution, and observation window must be explicitly defined. It is scientifically preferable to retain the discrete conditional identity than to imply unjustified continuous-time precision.
For contract payments, replace each eligible loss by its policy transformation and calculate E[I(G)] at the correct aggregation level. Because deductible and limit functions are nonlinear, generally E[I(G)] differs from I(E[G]). For example, two equally likely values G = USD 0 and G = USD 200,000 with deductible USD 100,000 and no binding limit give E[I(G)] = USD 50,000, while I(E[G]) = I(USD 100,000) = USD 0. Any practice that applies a deductible to average modeled loss instead of the underlying loss distribution can therefore severely misstate expected claims.
Comparative cases and identification
A coastal flood assessment can decompose the analysis into an inundation intensity footprint, parcel-level buildings and occupants, hazard-conditional damage and displacement responses, and repair/recovery losses. Model validation must compare flood maps against credible observations, inventory values against property records, vulnerability curves against building-specific damage evidence, and insurance payment estimates against contract terms. A reduction in observed claims after sea defenses must be evaluated alongside potential additional development behind the defenses and rare overtopping scenarios.
An earthquake portfolio assessment requires seismic-source and ground-motion representation, geocoded asset taxonomies, structural fragility and economic vulnerability, correlated event outcomes, and policy loss layers. In this setting, the OpenQuake classical calculator makes distinctions between hazard curves, exposure models, and vulnerability functions operational. Its event-based counterpart addresses joint event losses and relevant correlation choices. A model of expected per-building damage that lacks common event structure is inadequate for aggregate extreme-loss pricing.
A wildfire assessment requires ignition and spread-related hazards, exposed structures and environmental systems, construction and defensible-space susceptibility, evacuation capacity, and recovery conditions. A house's rebuilding costs can be monetized; ecosystem destruction, health impacts, and community displacement require additional measures. Insurance policies may respond to smoke, fire, or interruption under different terms. A single wildfire footprint cannot support all these consequences without distinct models.
A severe-weather risk assessment for agriculture must specify crops, planting stage, the time-varying location and condition of productive assets, weather intensity, yield vulnerability, commodity prices, and coverage triggers. A policy-year average hectare count can miss the seasonal concentration of crop exposure. An index insurance contract may pay under adverse rainfall even when individual yield loss is low, demonstrating a further distinction between physical hazard and financial payment.
Across these examples, identification requires a clear contrast. To evaluate hazard change, hold a counterfactual exposure and vulnerability state fixed or model their interaction explicitly. To evaluate mitigation, compare outcome distributions under the intervention and a defensible baseline. To evaluate insurance penetration, separate covered loss from ground-up economic damage. To evaluate social resilience, measure recovery and distributional effects with units suited to welfare. A causal claim becomes defensible through a specified counterfactual, evidence of comparability, and an account of unmeasured confounding.
Evidence discipline and research implications
The sources examined cover different layers of knowledge. UNDRR and IPCC provide internationally used terminology and assessed conceptual relations. Kaplan and Garrick (1981) contribute a probabilistic scenario framework. Turner et al. (2003) and Cutter et al. (2003) establish important approaches to human-environment vulnerability and measurable social differences. Bouwer (2011) provides a historically bounded review of disaster-loss attribution. OpenQuake and Oasis document implemented computational distinctions. World Bank and Swiss Re materials illuminate financing and empirical insurance-sector practice. These sources complement one another; none alone verifies a complete claim about all hazards, assets, social consequences, and insurance markets.
Scientific verification proceeds from the precise material proposition to the appropriate evidence type. A standard proves what a standard specifies, not whether a quantitative hazard model predicts well in an untested region. A software manual proves a modeled workflow and its stated limitations, not independently measured predictive accuracy. A catastrophe loss estimate documents a particular organization's scope and methodology; it cannot by itself determine the causal contributions of climate, urban development, or insurance selection. Historical correlation cannot demonstrate the efficacy of a hypothetical retrofit without a credible causal research design.
Model uncertainty is decision-dependent. A portfolio capital question is sensitive to joint tail outcomes and reinsurance terms; community protection is sensitive to who bears uninsured losses and the time required to restore services; public land-use planning is sensitive to exposure growth and unintended concentration of value. Accordingly, a scientifically useful assessment documents scenario scope, spatial resolution, time horizon, values and units, vulnerability provenance, dependence assumptions, uncertainty intervals, coverage rules, and validation evidence in a form that permits independent reconstruction.
The comparative synthesis yields several bounded analytical conclusions. First, the probability of a hazard and the probability of a monetary claim refer to distinct events, connected by exposure, damage, and coverage conditions. Second, equal average losses can conceal very different tail risks, as the dependence and vulnerability examples demonstrate. Third, social vulnerability changes both the nature of harm and recovery, so a property claim distribution is an incomplete proxy for disaster welfare. Fourth, the effect of intervention must be attributed to the component and causal pathway it changes, while accounting for induced changes in exposure. Fifth, time-varying exposures and contracts make simple historical comparisons hazardous unless the analysis normalizes and models the relevant determinants.
These conclusions are original analytic integrations and demonstrations of established theory and evidence. They do not constitute a separate new doctrine, named model, or assertion of priority. The research question remains empirical whenever application requires parameter estimation: the distributions, mechanisms, and interventions must be measured anew for each location, hazard, system, and jurisdiction.
Conclusion
Risk as hazard, exposure, and vulnerability is a productive analytical formulation because it directs investigation toward the event process, the location and value of interests at stake, and the mechanisms through which an event becomes harm. The complete object of risk assessment is a distribution of possible consequences over a specified horizon, with explicit measures, populations, and institutional boundaries. A hazard map, exposure inventory, or vulnerability index contributes to that object and answers only its corresponding part.
The comparison of IPCC, UNDRR, ISO, scenario analysis, physical catastrophe modeling, and insurance methods establishes a compatible layered interpretation. Hazard determines which event intensities can arise and how often; exposure identifies who and what may be affected; vulnerability determines conditional damage and broader consequences; capacity changes prevention, response, and recovery; financial and policy terms allocate and transform those consequences. These relations are dynamic and may be mutually influential.
The numerical analyses establish three important limits of shortcut reasoning. Identical hazard probability and exposed value can generate markedly different losses through vulnerability. Identical expected portfolio loss can coexist with radically different aggregate-tail probabilities because of dependence. And deductibles and limits applied to a distribution cannot generally be applied to its mean. Each result follows from elementary probability under explicitly stated assumptions and transfers directly to the interpretation of more complicated models.
Physical, financial, and social measures deserve independent treatment. Insured loss reflects coverage and claims conventions; economic damage includes consequences outside insured payment; social and ecological harm may require distinct outcome metrics. Public and private risk financing can improve recovery and distribute financial burdens, while durable risk reduction depends on changes in hazard-related processes, exposure, vulnerability, and capacity.
Reliable empirical application requires event and loss definitions, geospatial and temporal resolution, valuation consistency, vulnerability validation, dependence analysis, documented uncertainty, clear counterfactuals, and access to the underlying evidence. Risk estimates should be interpreted with their conditions intact. This monograph supplies the conceptual and mathematical discipline for such work while reserving hazard-specific estimates, product pricing, and jurisdictional conclusions for appropriately documented research.
The general conclusion is that risk is best understood by tracing a complete causal and probabilistic passage from hazardous possibility to contingent consequence, and from consequence to any institutional arrangement that measures, finances, or absorbs it. Its determinants are scientifically intelligible only when their interactions are specified; its practical governance depends on maintaining that clarity across models, data, policy, and insurance contracts.
References
Bouwer, L. M. (2011). Have disaster losses increased due to anthropogenic climate change? Bulletin of the American Meteorological Society, 92(1), 39–46. https://doi.org/10.1175/2010BAMS3092.1
Cutter, S. L., Boruff, B. J., & Shirley, W. L. (2003). Social vulnerability to environmental hazards. Social Science Quarterly, 84(2), 242–261. https://doi.org/10.1111/1540-6237.8402002
Global Earthquake Model Foundation. (n.d.). Classical damage and risk. OpenQuake Engine documentation. https://docs.openquake.org/oq-engine/master/manual/user-guide/workflows/classical-risk.html
Global Earthquake Model Foundation. (n.d.). Probabilistic event-based risk calculator. OpenQuake Engine documentation. https://docs.openquake.org/oq-engine/manual/latest/underlying-science/event-based-risk-calculator.html
International Electrotechnical Commission. (2019). IEC 31010:2019: Risk management — Risk assessment techniques. https://www.iso.org/standard/72140.html
International Organization for Standardization. (2018). ISO 31000:2018: Risk management — Guidelines (2nd ed.; reviewed and confirmed in 2023). https://www.iso.org/standard/31000
Intergovernmental Panel on Climate Change. (2022a). Chapter 1: Point of departure and key concepts. In Climate Change 2022: Impacts, Adaptation and Vulnerability, Working Group II contribution to the Sixth Assessment Report. Cambridge University Press. https://www.ipcc.ch/report/ar6/wg2/chapter/chapter-1/
Intergovernmental Panel on Climate Change. (2022b). Annex II: Glossary. In Climate Change 2022: Impacts, Adaptation and Vulnerability, Working Group II contribution to the Sixth Assessment Report. Cambridge University Press. https://www.ipcc.ch/report/ar6/wg2/chapter/annex-ii/
Kaplan, S., & Garrick, B. J. (1981). On the quantitative definition of risk. Risk Analysis, 1(1), 11–27. https://doi.org/10.1111/j.1539-6924.1981.tb01350.x
Oasis Loss Modelling Framework. (n.d.). Financial module. Oasis LMF documentation, version 2.5.9. https://oasislmf.github.io/2.5.9/oasislmf/explanation/financial-module.html
Swiss Re Institute. (2026). Sigma 1/2026: Natural catastrophes in 2025: The persistent rise of wildfire and storm risk (web summary, March 16, 2026). https://www.swissre.com/institute/research/sigma-research/sigma-2026-01-natcat-2025-wildfire-storm-risk.html
Turner, B. L., II, Kasperson, R. E., Matson, P. A., McCarthy, J. J., Corell, R. W., Christensen, L., Eckley, N., Kasperson, J. X., Luers, A., Martello, M. L., Polsky, C., Pulsipher, A., & Schiller, A. (2003). A framework for vulnerability analysis in sustainability science. Proceedings of the National Academy of Sciences, 100(14), 8074–8079. https://doi.org/10.1073/pnas.1231335100
United Nations Office for Disaster Risk Reduction. (2017a). Sendai Framework terminology: Capacity. https://www.undrr.org/terminology/capacity
United Nations Office for Disaster Risk Reduction. (2017b). Sendai Framework terminology: Disaster risk. https://www.undrr.org/terminology/disaster-risk
United Nations Office for Disaster Risk Reduction. (2017c). Sendai Framework terminology: Disaster risk assessment. https://www.undrr.org/terminology/disaster-risk-assessment
United Nations Office for Disaster Risk Reduction. (2017d). Sendai Framework terminology: Exposure. https://www.undrr.org/terminology/exposure
United Nations Office for Disaster Risk Reduction. (2017e). Sendai Framework terminology: Hazard. https://www.undrr.org/terminology/hazard
United Nations Office for Disaster Risk Reduction. (2017f). Sendai Framework terminology: Vulnerability. https://www.undrr.org/terminology/vulnerability
United Nations Office for Disaster Risk Reduction. (2022). Global Assessment Report on Disaster Risk Reduction 2022: Our World at Risk. https://www.undrr.org/gar/gar2022-our-world-risk-gar
World Bank Group. (n.d.). Disaster risk finance and insurance. https://www.worldbank.org/ext/en/topic/financial-sector/disaster-risk-finance-and-insurance