SCIENCE
Risk Theory
Abstract. Risk theory studies how uncertain events acquire consequences for people, organizations, economies, and ecological systems; how knowledge about those events can be represented; and how decisions, institutions, and insurance transform their distribution. Its mathematical foundations include probability, stochastic processes, statistical inference, decision theory, and the mathematics of extremes. Its economic foundations include risk aversion, incentives, information asymmetry, market incompleteness, and the allocation of contingent claims. Its practical foundations include exposure analysis, protection, capital, mutualization, reinsurance, governance, and the interpretation of uncertainty that exceeds the available data. This monograph develops an integrated account of risk while preserving the differences between objective hazards, probability assessments, consequences, perceived risks, and institutional decisions. It explains why expected loss, dispersion, tail severity, dependence, solvency, and social acceptability answer different questions. It derives fundamental results for aggregate insurance losses, diversification, utility, quantile risk measures, and ruin; examines their assumptions; and evaluates the limits of model-based knowledge in climate, cyber, and systemic settings. The central conclusion is that scientifically credible risk analysis requires an explicit account of the event space, exposure, consequences, uncertainty, decision horizon, dependence, and the evidence supporting each inference. Mathematical precision improves decisions when the adequacy of its premises is tested with equal rigor.
Keywords: risk theory; uncertainty; probability; expected utility; ambiguity; insurance; loss distributions; diversification; dependence; tail risk; extreme value theory; coherent risk measures; expected shortfall; model risk; catastrophe; climate risk; systemic risk; risk governance.
The Scientific Object of Risk Theory
Risk becomes a scientific object when a possible change in the world is connected to an identifiable consequence and assessed under conditions of incomplete knowledge. A floodplain may be exposed to flooding, a debtor may default, a patient may require treatment, an insurer may experience claims beyond available capital, and an electrical grid may suffer a cascading outage. These cases differ in their physical mechanisms, legal and economic consequences, repeatability, and accessibility to observation. Risk theory provides methods for describing those differences and establishing what can be inferred from available knowledge.
A useful general analytical question is: What can happen, how plausible is it, what would follow, and what could alter its occurrence or consequences? Kaplan and Garrick (1981) expressed a quantitative risk analysis as a set of scenarios, their probabilities, and consequences. Their contribution helps prevent the frequent compression of multiple distinct events into a single unqualified risk score. The event definition establishes the population of possibilities; the probability specification gives a degree of uncertainty about them; the consequence specification identifies what matters to the affected system. Kaplan and Garrick (1981).
Several analytical levels must be kept explicit. A hazard is a potential source of harm. Exposure describes what lies in the path of the hazard. Vulnerability describes how severely exposed systems may be affected. A probability distribution represents a claim about possible outcomes. An economic loss translates some consequences into a particular unit of account. A choice expresses an agent's preference under uncertainty. A control, insurance policy, or capital buffer changes either the consequences or their allocation. These are related objects, and each must be described on its own terms.
Risk analysis therefore rests simultaneously on ontology, measurement, inference, and valuation. The ontological question asks which events and mechanisms are possible. The measurement question asks how their frequencies, magnitudes, and interdependence can be represented. The inferential question asks how observations warrant claims about unobserved or future events. The evaluative question asks which outcomes matter and for whom. The action question asks which interventions are feasible and how their effects will be tested. Any completed assessment must address all five, at the level appropriate to its decision.
The insurance connection is immediate. An insurance contract transfers specified financial consequences of uncertain events in exchange for a premium, subject to coverage terms, limits, exclusions, and the insurer's capacity to meet claims. The uncertainty of a covered event is necessary to the contractual logic of insurance; uncertainty alone does not establish insurability. Loss aggregation, information, legal enforceability, demand, and capital remain decisive.
Historical Formation of the Discipline
Modern risk theory grew from several interacting traditions rather than a single invention. Commercial navigation, lending, annuities, gambling, mortality records, and maritime insurance made the comparison of uncertain outcomes economically urgent. Mathematical probability developed methods for enumerating chances and studying repeated experiments. Political arithmetic and mortality tables supplied a statistical language for population-level contingencies. Life insurance and public finance turned these methods into institutions with long-dated promises.
Daniel Bernoulli's essay of 1738, known in a 1954 English translation as "Exposition of a New Theory on the Measurement of Risk," challenged the adequacy of expected monetary value for choice and emphasized the utility of wealth. The historical distinction matters: the original essay belongs to the eighteenth century, while the frequently cited English-language Econometrica edition appeared in 1954. Its argument helped establish why the same monetary lottery can have different value to individuals with different wealth or preferences. Bernoulli (1738/1954).
The nineteenth and early twentieth centuries saw the development of increasingly formal probability, statistics, insurance mathematics, and economic theories of uncertainty. Theoretical tools for aggregate claims, premiums, reserves, and insurer failure emerged alongside the institutional expansion of insurance. In 1921 Frank H. Knight examined how measurable risk differs from uncertainty that resists assignment of known probabilities, and how entrepreneurial judgment and profit relate to this distinction. His analysis forms part of a broader theory of imperfect competition and uncertainty, rather than a claim that every contemporary use of the word risk must follow one rigid taxonomy. The 1921 edition is preserved by the Library of Congress. Knight (1921).
The second half of the twentieth century changed the field's reach. Von Neumann and Morgenstern's 1944 work supplied a systematic axiomatic treatment of utility within game theory. Savage's 1954 account developed subjective expected utility and the conditions under which choices reveal personal probabilities. Arrow's 1963 analysis of medical care established how uncertainty and institutional features affect welfare. Pratt's 1964 treatment of risk aversion supplied local curvature measures. Akerlof, Rothschild and Stiglitz, and Pauly developed complementary accounts of asymmetric information and the incentives of insured parties. These works explain why the mere existence of a probability distribution does not ensure a functioning insurance market. Von Neumann and Morgenstern (1944); Savage (1954); Arrow (1963).
By the end of the century, financial risk management had sharpened the formal study of tail losses, dependence, and capital. Artzner, Delbaen, Eber, and Heath (1999) analyzed desirable axioms for monetary risk measures; Embrechts, Klüppelberg, and Mikosch (1997) assembled mathematical methods for extremes in insurance and finance; and Rockafellar and Uryasev (2000) developed tractable optimization using conditional value-at-risk. Contemporary risk research draws increasingly on climate science, network economics, cybersecurity, machine learning, decision science, and public governance. Its intellectual continuity is the attempt to make consequential uncertainty analyzable without confusing the representation with the world.
Meanings of Risk and the Structure of a Risk Statement
There is no single numerical property that exhausts every scientifically legitimate meaning of risk. A variance reports dispersion; a failure probability reports the chance of an event; expected loss reports a probability-weighted average; a tail measure reports the severity of unfavorable outcomes at a chosen confidence level; perceived risk concerns judgment and experience; systemic risk concerns how distress spreads. Each measure establishes a particular aspect of a specified problem.
An operational risk statement should identify at least: the system under study; the relevant event or scenario; the exposure and consequence; the observation and decision horizon; the probability or uncertainty representation; the evidence and assumptions; the intervention being considered; and the responsible decision maker. A statement that a site has "high flood risk" is incomplete if it omits whether high refers to annual probability, expected financial loss, extreme-event severity, life safety, or disruption. A geographical statement without the relevant period, climate regime, defenses, and exposure also lacks a stable interpretation.
Different authoritative domains adopt different definitions. ISO 31000:2018 develops management guidance around risk as the effect of uncertainty on objectives; its scope includes adverse and favorable deviations. The IPCC's Sixth Assessment, Working Group II, treats climate-impact risk in terms of potential adverse consequences for human or ecological systems and the interaction of hazard, exposure, and vulnerability. The standards serve different purposes. One frames organizational objectives; the other examines potential harms in natural and human systems. Neither authorizes transferring an identical numerical indicator into all settings. ISO (2018); IPCC (2022).
A clear classification distinguishes at least physical risk, liability risk, financial-market risk, underwriting risk, credit risk, liquidity risk, operational risk, strategic risk, conduct risk, and systemic risk. Many events inhabit several classifications simultaneously. A hurricane may destroy insured property, impair transportation infrastructure, reduce investment values, increase claims inflation, affect liquidity, and create disputes over policy conditions. Classification should organize evidence and decisions; it should never replace causal investigation.
Probability: Objective, Subjective, and Statistical Interpretations
Probability theory provides a mathematical language for uncertainty. A probability space consists of a sample space (\Omega), a collection of measurable events (\mathcal F), and a probability measure (P) satisfying nonnegativity, normalization, and countable additivity. A loss is a measurable random variable (L:\Omega\to\mathbb R), often taken to be nonnegative for gross insured claims. A distribution (F_L(x)=P(L\le x)) specifies the modeled probabilities of its possible magnitudes. Its expectation (E[L]) summarizes mean loss when integrable; its variance measures dispersion around the mean when the second moment exists.
The mathematics does not by itself say where the probabilities come from. A frequentist interpretation connects probability to stable repeated-event behavior. Subjective approaches express coherent personal or institutional beliefs, potentially informed by evidence. A Bayesian account explicitly updates such beliefs with data. Propensity or physical-chance interpretations focus on mechanisms that make outcomes possible. These interpretive differences matter sharply for unique geopolitical events, climate transitions, new technologies, and catastrophic interactions.
Observed relative frequency (\widehat p=k/n) is a data statistic, not automatically the probability of every future trial. Its inferential significance depends on event comparability, sampling, dependence, changes in exposure, and measurement quality. If the event definition changes, the denominator may cease to represent the same risk. A sample with no observed loss cannot prove zero loss probability. Equally, a confident estimate may conceal a narrow observation window or an untested extrapolation.
The law of large numbers explains why independent and identically distributed integrable observations can have a sample average converging to their common expectation. Its conclusion depends on its assumptions; it does not guarantee that each insurer's future portfolio contains enough comparable exposures, or that a single aggregate catastrophe will average away. The central limit theorem supplies powerful approximations under suitable conditions, yet extreme tails may converge slowly, and some heavy-tailed models lack the moments needed for ordinary normal approximations.
Savage's decision-theoretic framework clarified the role of subjective probabilities in coherent choice, but the evaluation of its axioms is partly normative. A risk model must be clear about whether probabilities are measured frequencies, estimated parameters, judgments, scenario weights, or outputs of a particular simulation. Treating all five as identical would overstate the evidence supporting the estimate. Savage (1954).
Uncertainty, Ambiguity, and the Limits of Forecasting
Knight's distinction draws attention to situations in which the relevant probability law cannot simply be treated as known. Statistical uncertainty arises from finite data, sampling variability, and measurement error. Parameter uncertainty concerns unknown values within a chosen model. Model uncertainty concerns the adequacy of the chosen family of distributions or causal mechanisms. Structural uncertainty arises when the process itself changes. Ambiguity describes uncertainty about which probability assessment should be used. Deep uncertainty can arise when decision makers disagree about plausible models, outcomes, or valuations.
These distinctions have operational consequences. A company may estimate a fire frequency from decades of comparable buildings; it may have only a few observations of novel cyber aggregation pathways; and it may face climate exposure whose future physical distributions depend on changing emissions, adaptation, infrastructure, and land use. Each problem warrants an evidence structure suited to the process.
Ellsberg's 1961 analysis exposed choices that challenge Savage-style subjective expected utility when probabilities are ambiguous. A decision maker can prefer a well-specified risky lottery over an ambiguous one even when no conventional expected-value comparison determines that preference. The lesson is that uncertainty about a distribution can itself affect valuation and behavior. Ellsberg (1961).
Several responses are available: probability intervals, scenario families, sensitivity ranges, minimax and minimax-regret criteria, Bayesian model averaging, robust optimization, and adaptive decision strategies. Each response encodes a decision attitude. A worst-case rule offers protection against a specified set of possibilities but may be excessively conservative if that set includes implausible extremes. Bayesian inference integrates probability beliefs but is sensitive to prior and likelihood specification where data are thin. Scenario analysis illuminates vulnerabilities without automatically assigning frequencies to scenarios.
A responsible analysis records where uncertainty enters, what part of it can be reduced through research, and how much uncertainty remains in a recommendation. When the available evidence cannot discriminate between important alternative mechanisms, additional decimal places in a numerical result cannot supply the missing knowledge.
Decision Theory, Utility, and Risk Aversion
Risk matters because agents must choose among actions whose outcomes vary. In the simplest expected-value calculation, action (a) produces consequence (x_i(a)) with probability (p_i), and its expected monetary value is (\sum_i p_i x_i(a)). Choosing the largest expected value is appropriate only under a decision criterion for which monetary value has the assumed linear significance.
Expected utility replaces money with a utility function. A decision maker maximizes (E[u(W+X_a)]), where (W) is current wealth and (X_a) is the random change produced by action (a). A concave increasing utility function describes risk aversion: Jensen's inequality gives (u(E[Y])\ge E[u(Y)]) whenever the relevant expectations exist. For the same expected monetary return, a risk-averse decision maker may prefer a certain outcome. Bernoulli's utility argument and the axiomatic representation developed by von Neumann and Morgenstern are foundational sources. Bernoulli (1738/1954); von Neumann and Morgenstern (1944).
Pratt (1964) formalized a local measure of absolute risk aversion as (A(w)=-u''(w)/u'(w)), given differentiability and positive marginal utility. For a small zero-mean monetary risk (X), a second-order approximation to the certainty-equivalent risk premium is (\pi\approx \tfrac12 A(w)\operatorname{Var}(X)). This is a local approximation, not a universal insurance pricing formula: finite changes, nonlinear preferences, wealth constraints, and non-small risks require more complete calculations. Pratt (1964).
Insurance can increase utility even when the premium exceeds expected claims, because it exchanges uncertain large losses for a more predictable expenditure. To illustrate, suppose a household faces a small probability of a loss whose payment would severely reduce consumption. A premium covering expected claims plus administrative and capital costs can still be preferable under sufficiently concave utility. The willingness to insure depends on the wealth distribution across states, contract conditions, background risk, trust, and alternative protections.
Risk aversion is not the sole source of insurance demand. Collateral requirements, debt covenants, regulation, household commitments, transactions costs, and fear of ruin also matter. The maximum rational premium depends on the objective: expected utility, liquidity protection, compliance, or the ability to avoid unacceptable harm. Risk theory should state the decision criterion before announcing what is optimal.
Behavioral and Social Approaches
Observed human decisions often differ from the idealized axioms of expected utility. Kahneman and Tversky's 1979 prospect theory describes choices in terms of gains and losses relative to a reference point and introduces decision weights in place of objective probabilities. Their work explains characteristic patterns including the certainty effect and the role of low-probability weighting; it does not mean that every individual or context exhibits one invariant preference curve. Kahneman and Tversky (1979).
Behavioral risk research adds salience, familiarity, controllability, dread, trust, fairness, experience, and social communication. Slovic's 1987 research on risk perception examines how people characterize hazards and why public responses cannot be predicted from mortality frequencies alone. Such research is relevant to take-up of flood coverage, vaccine decisions, cybersecurity behavior, and acceptance of infrastructure risk. Slovic (1987).
The implications for insurance are concrete. A consumer may overreact to a vivid recent disaster but neglect a slow-onset exposure. Complex exclusions may impede informed comparison even when their monetary terms are actuarially defensible. Framing the same deductible as a saving or potential loss may change choice. Limited attention can produce underinsurance even among households facing substantial expected losses. These mechanisms require empirical testing in the relevant population; the theory provides hypotheses, not universal estimates of their magnitude.
Social theory also asks who defines a risk as acceptable and who bears the costs of an intervention. Risk communication must convey magnitude, uncertainty, reference classes, distributional consequences, and the limits of knowledge without pretending that disagreement is always a failure of statistical literacy. Diverse interests and values are part of the problem under examination.
Random Losses and Collective Risk Processes
Insurance losses are often modeled by separating claim frequency from claim severity. Let (N) be the number of claims during a defined period, and let (X_i) be the amount of claim (i). The aggregate gross loss is
[ S=\sum_{i=1}^{N}X_i, ]
with (S=0) when (N=0). In the classical collective-risk model, (N) is independent of independent identically distributed severities (X_i). Under these assumptions,
[ E[S]=E[N]E[X],\qquad \operatorname{Var}(S)=E[N]\operatorname{Var}(X)+\operatorname{Var}(N)(E[X])^2. ]
The variance identity follows from conditioning on (N) and applying the law of total variance. When (N) is Poisson with mean (\lambda), both its expectation and variance equal (\lambda), giving (E[S]=\lambda E[X]) and (\operatorname{Var}(S)=\lambda E[X^2]), provided the moments exist.
This result is exact for the model, not an empirical assertion that real claims always follow a Poisson process. Claim counts may be overdispersed by heterogeneity, weather, common incidents, and contagion. Claim severities may depend on claim counts because a disaster produces both more and larger claims. Expenses, litigation, inflation, and payment development can make observed paid amounts differ from ultimate costs. Model quality is determined by these mechanisms and by validation, not by mathematical convenience alone.
A frequency–severity decomposition improves diagnosis. A rising aggregate claim total may reflect more exposures, a higher event frequency per exposure, more severe claims, changing limits, inflation, claims handling, or a different reporting lag. Studying only aggregate amounts conceals causal alternatives. Exposure measures such as policy-years, insured values, vehicle miles, or building characteristics must correspond to the modeled event.
Compound Poisson processes, renewal processes, Cox or mixed Poisson models, marked point processes, and stochastic severity models give different representations. Model selection should follow the observable event process and the intended use. A reserving problem may need development delays, while a catastrophe capital problem demands spatial and temporal event dependence. These boundaries connect risk theory to actuarial science without conflating the disciplines.
Exposure, Time, and Dynamic Risk
Risk is always relative to a horizon. A one-year default probability, a lifetime mortality probability, and the chance of at least one system failure during a decade answer different questions. If events arrive as a homogeneous Poisson process with rate (\lambda) per unit of time, the probability of at least one event over duration (t) is (1-e^{-\lambda t}). The formula is useful because its assumptions are explicit. It cannot be copied to a process with time-varying intensity or clustered arrivals without changing the model.
A survival function (S_T(t)=P(T>t)) describes the waiting time to an event. A hazard rate, where defined, is (h(t)=f_T(t)/S_T(t)). The word hazard here has a technical survival-analysis meaning: an instantaneous conditional event rate, not the general physical source of danger. This difference is important in interdisciplinary research, where the same word may have several disciplinary meanings.
Long-term contracts require attention to time value, inflation, discounting, claim settlement, and changes in the risk-generating environment. A promise with a low expected payment far in the future may have substantial sensitivity to interest rates and tail events. Dynamic asset–liability mismatch can create solvency pressure even when anticipated claims remain close to their expected values.
A temporal risk assessment should distinguish an event's onset, detection, reporting, claim development, payment, and recovery. It should also distinguish the date of historical evidence from the horizon of a prediction. Projections that assume fixed building codes, fixed assets, or fixed consumer behavior are conditional scenarios. Their relevance weakens as exposures or policies change.
Dependence, Correlation, and Diversification
Pooling reduces idiosyncratic fluctuations when individual losses are sufficiently independent. Suppose (n) losses each have variance (\sigma^2) and every distinct pair has correlation (\rho). Their sum then has variance
[ \operatorname{Var}!\left(\sum_{i=1}^{n}L_i\right)=n\sigma^2+n(n-1)\rho\sigma^2. ]
The average loss has variance (\sigma^2/n+(n-1)\rho\sigma^2/n). With independence, (\rho=0), that variance decreases as (1/n); with a common positive correlation that remains constant, the variance of the average approaches (\rho\sigma^2) as (n) becomes large. This elementary identity reveals why diversification cannot eliminate persistent common exposures.
Pearson correlation summarizes linear association. It is not a complete description of dependence, particularly for skewed or heavy-tailed variables. Two portfolios can have similar marginal distributions and ordinary correlation but markedly different probabilities of simultaneous extreme loss. Copulas link marginal distributions into a joint distribution and allow explicit investigation of dependence. The basic mathematical idea traces to Sklar's work on multivariate distributions, described in specialist histories of copulas. Schweizer (1991).
In insurance, common causes include windstorms affecting neighboring properties, a shared software dependency across insured firms, inflation simultaneously increasing many claim costs, or correlated market declines and policyholder withdrawals. Risks may become more dependent during stress than in routine observations. Hence model validation needs tail co-exceedance analysis, conditional scenarios, causal knowledge, and sensitivity to alternative dependence structures.
The practical conclusion is that effective diversification requires mechanisms that reduce common outcomes, not merely a larger number of contracts. Geographical spread can improve a portfolio while leaving exposure to global supply chains, cyber infrastructure, interest-rate shocks, and climate trends. Reinsurance can redistribute correlated exposure but may create concentration in the reinsurance market. Every diversification claim should specify the relevant risk measure and stress state.
Extreme Events and Tail Modeling
Ordinary averages describe the center of a distribution; insurance solvency often depends on its tail. A small number of severe events can account for a large proportion of aggregate losses. Extreme value theory supplies tools for maxima and threshold exceedances, with mathematical results that require regularity assumptions. Embrechts, Klüppelberg, and Mikosch (1997) provide a foundational treatment for insurance and finance. Embrechts, Klüppelberg, and Mikosch (1997).
In a block-maxima approach, one studies maxima from sufficiently comparable blocks and examines approximations by generalized extreme-value distributions. In a peaks-over-threshold approach, sufficiently high excesses (Y=X-u), conditional on (X>u), may be approximated by a generalized Pareto distribution. Its shape parameter (\xi) governs tail thickness; under the corresponding exact generalized Pareto model, a positive shape parameter produces a power-law-type tail. The mean excess is infinite when (\xi\ge1), and variance is infinite when (\xi\ge1/2), subject to the usual model parameterization. These mathematical properties are conditional on the selected model, not verified facts about all insured losses.
Threshold choice creates a bias–variance tradeoff. A threshold too low may violate the asymptotic approximation; a threshold too high leaves few observations. Fitted curves may be unstable under censoring, reporting thresholds, exposure changes, and inconsistent claims inflation. Dependence in temporal or spatial extremes further complicates inference.
Return periods require special care. An estimated annual exceedance probability (p) is sometimes expressed as an annual return period (1/p). This describes an inverse modeled rate under the relevant assumptions, not a timetable or a guarantee that an event occurs once per stated interval. In a nonstationary climate or changing portfolio, a historical return period need not describe future annual exceedance. The reference period and the source of uncertainty must be explicit.
A credible tail analysis combines historical events, physical hazard science, exposure inventories, engineering vulnerability functions, statistical uncertainty, and stress testing. Its priority is the causal credibility and decision relevance of extreme scenarios. Apparent numerical accuracy at the 99.9th percentile can exceed what the data justify.
Risk Measures and Their Mathematical Properties
A risk measure maps a loss distribution or position into a numerical assessment relevant to an objective. Common summaries include the expectation, variance, standard deviation, quantiles, expected shortfall, probability of ruin, and expected utility. Choosing a measure is a substantive part of decision making. It states which departures from desirable outcomes count and how the consequences are ranked.
For a loss (L), value-at-risk at confidence level (\alpha) is the lower quantile
[ \operatorname{VaR}_{\alpha}(L)=\inf{x:P(L\le x)\ge\alpha}. ]
VaR specifies a threshold, not the average severity beyond that threshold. If a portfolio has a 99% VaR of one million dollars under a specified annual distribution, that means the modeled loss does not exceed that threshold with probability at least 99%, using the quantile convention shown. It does not tell us how large losses can become in the remaining tail.
Expected shortfall can be defined for an integrable loss by averaging quantiles:
[ \operatorname{ES}{\alpha}(L)=\frac{1}{1-\alpha} \int{\alpha}^{1}\operatorname{VaR}_{u}(L),du. ]
For a continuous distribution under suitable conditions, this equals the expected loss conditional on exceeding the (\alpha)-quantile. For discrete distributions, careful treatment of probability mass at the threshold is needed; the quantile-integral definition avoids a frequent ambiguity. Rockafellar and Uryasev (2000) show how conditional value-at-risk can be optimized with tractable methods. Rockafellar and Uryasev (2000).
Artzner and colleagues (1999) identify four axioms of a coherent monetary risk measure: monotonicity, translation invariance or cash additivity, positive homogeneity, and subadditivity, subject to the chosen sign convention. Subadditivity expresses that the assessed risk of a combined position does not exceed the sum of the assessed risks of its components. Expected shortfall has coherence properties on appropriate integrable domains; VaR can fail subadditivity. Artzner et al. (1999).
A counterexample is elementary. Let two independent losses each equal 100 with probability 0.04 and zero otherwise. For either loss, (\operatorname{VaR}{0.95}=0) because the probability of zero is 0.96. For their sum, the probability of zero is (0.96^2=0.9216); the probability that the sum is at most 100 is 0.9984. Thus (\operatorname{VaR}{0.95}(L_1+L_2)=100), which exceeds the sum of the individual VaRs, zero. This exact discrete example illustrates the failure without relying on an empirical dataset.
Coherence is an axiomatic criterion, not an assurance that any estimate is accurate. Expected shortfall may be very hard to estimate reliably with scarce tail observations. Capital requirements also depend on exposure dynamics, asset–liability interactions, available funds, risk appetites, and regulatory choices. Banking market-risk standards have employed stressed expected shortfall, but their rules are not automatically insurance capital regulations. Basel Committee (2019).
Bayesian Inference, Estimation, and Predictive Validation
Risk models rely on data gathered under imperfect observation. Estimation therefore requires a distinction between the stochastic process, the recorded sample, and the assumed sampling mechanism. Missing claims, claims reported late, insurance limits, deductibles, inflation adjustments, and policy selection can distort a loss sample. A high-quality estimation method begins with the data-generating process and tests whether observed records represent the intended population.
In Bayesian inference, parameter uncertainty is expressed by a prior distribution (\pi(\theta)), a likelihood (p(D\mid\theta)), and a posterior
[ p(\theta\mid D)\propto p(D\mid\theta)\pi(\theta). ]
Posterior predictive loss distributions integrate over parameter uncertainty. Their results still depend on model structure and the prior, particularly with limited observations. Frequentist estimators instead often evaluate performance through repeated-sampling properties and uncertainty intervals. Both approaches require explicit assumptions; method selection should reflect the decision and the available evidence.
A prediction should be assessed against outcomes using proper scoring rules, calibration, discrimination, and stability. For a binary event with forecast (p\in[0,1]) and observed result (y\in{0,1}), the Brier score is ((p-y)^2). Averaged over cases, a lower value indicates better performance under this score, but evaluation requires a representative test sample. A calibrated model assigns events that occur about 10% of the time to a collection of cases forecast near 10%, under appropriate grouping and time stability. A model can discriminate risk levels while being badly calibrated.
Temporal validation is particularly important. Training on future observations relative to the evaluation date causes leakage. Catastrophe models also need spatial validation because nearby observations share hazard drivers. Claims models require careful treatment of development and right censoring; a dataset of settled claims can underrepresent long-tail liabilities. External validation across geographies, contract forms, and time periods reveals whether the model generalizes.
A scientific risk report should include data definitions, extraction period, exclusions, transformations, sampling uncertainty, model comparisons, calibration evidence, stress behavior, and known performance limitations. Uncertainty intervals should cover the uncertainty actually modeled and should not imply that excluded structural possibilities have been measured.
Model Risk and Robust Decision Making
Every risk model is a selective representation. It includes some mechanisms and excludes others. Model risk arises from incorrect distribution families, unstable parameters, omitted dependence, faulty data pipelines, inaccurate exposure inventories, misunderstood contracts, feedback effects, or inappropriate use of an otherwise sound tool.
Three complementary checks are necessary. First, conceptual validation asks whether the model represents the causal and contractual problem. Second, empirical validation compares its predictions with evidence, using out-of-sample observations where feasible. Third, decision validation asks whether the model's errors are material to the intended action. A forecast can have excellent average accuracy yet fail exactly where an insurer needs protection from extreme accumulation.
Sensitivity analysis varies plausible parameters; scenario analysis explores specified causal pathways; stress testing asks whether the system remains viable under severe but credible conditions. Robust optimization searches for decisions that perform adequately over a chosen uncertainty set. Adaptive strategies preserve options to revise actions as knowledge changes. These approaches are complements, not substitutes for investigation into the plausibility of the underlying scenarios.
Models should also be evaluated as components of institutions. A risk score that affects underwriting may change who buys insurance; that selection feeds back into the loss data used for later estimation. A capital rule may alter institutions' portfolios and correlations. An early-warning indicator can become less predictive once actors respond to it. The problem is endogenous change, not merely statistical error.
For high-consequence decisions, the documentation should identify assumptions, boundaries of application, versioned input datasets, independent challenge, thresholds for intervention, and procedures for model retirement. Numerical reproducibility establishes that a calculation can be repeated; it does not establish that the calculation describes the relevant world.
Insurance Pooling and Risk Transfer
Insurance is a system of contingent financial claims supported by premiums, capital, contractual definitions, and institutions. In an idealized large pool of comparable independent insured losses, aggregate experience becomes more predictable relative to scale. This stabilizing effect makes collective protection economically possible. Its strength depends on the diversity of exposures and on the tail behavior of their joint distribution.
In its simplest form, the pure premium per risk equals the expected covered loss. A sustainable charged premium also needs to account for expenses, distribution, capital costs, uncertainty, policy terms, investment income, competition, and regulatory constraints. A policy's indemnity function may differ radically from the insured's gross economic loss because of attachment points, deductibles, limits, waiting periods, and exclusions. A calculation using gross losses is not automatically a calculation of the insurer's payment.
Risk transfer changes who bears losses; its social value comes from improved allocation, loss prevention incentives, liquidity after shocks, and the ability to finance recovery. It need not eliminate the underlying physical hazard. Insurers can also influence hazard reduction through pricing, underwriting conditions, engineering surveys, and risk-control services. Such incentives can be constrained by affordability and differences in consumers' ability to invest in protection.
The concept of insurability is multidimensional. A risk may be probabilistically estimable yet economically difficult to cover because losses are highly correlated, premiums are unaffordable, adverse selection is strong, or capital demand is extreme. Conversely, public risk-sharing arrangements may expand cover where the private market cannot carry the full tail. Insurability therefore belongs jointly to probability theory, contract design, economic demand, institutional capacity, and public policy.
These distinctions define the relation between Risk Theory and neighboring Insurance Science fields. Risk Theory examines the general structure of uncertainty and its consequences; actuarial science develops specialized valuation and reserving methods; insurance economics studies market allocation and incentives; solvency regulation addresses the financial adequacy of authorized insurers. The boundaries clarify different research questions while allowing findings to be integrated.
Information Asymmetry, Moral Hazard, and Adverse Selection
Insurance markets face asymmetry when insured parties, insurers, intermediaries, and regulators hold different information. Akerlof's analysis of quality uncertainty demonstrates how asymmetrically distributed information can impair exchange, even where mutually beneficial trades would otherwise be possible. The insurance application concerns risks whose characteristics are known differently to applicants and providers. Akerlof (1970).
Adverse selection describes the tendency for insurance demand and coverage choice to vary with privately known risk. In a simple competitive market, risk types that cannot be distinguished may purchase contracts in different patterns, shifting average claims cost and potentially destabilizing a common premium. Rothschild and Stiglitz (1976) formally examined competitive insurance equilibria under imperfect information, showing the importance of separating contracts and the possibility of missing equilibria under their model assumptions. Rothschild and Stiglitz (1976).
Moral hazard concerns changes in behavior or resource use associated with insurance protection. It can take the form of reduced prevention effort, altered claims behavior, or greater demand for services when marginal out-of-pocket prices decline. Pauly's 1968 discussion of medical insurance emphasizes that consumption decisions may be price responsive after coverage. Moral hazard is an incentive problem within a specified environment; it does not establish that insured people are dishonest or that all additional care is socially wasteful. Pauly (1968).
Arrow's 1963 examination of medical care showed why uncertain demand, professional information, trust, and institutions make health services different from textbook commodities. Together, these analyses demonstrate that risk markets are shaped by what participants know and can observe, and by which behaviors contracts can monitor. Arrow (1963).
Responses include deductibles, coinsurance, experience rating, screening, policy conditions, verification, prevention incentives, and mandatory or subsidized pooling. Every response has costs. More granular pricing can reduce cross-subsidies while excluding high-risk households from affordable coverage. Stronger verification can reduce opportunism while increasing administrative burdens and privacy concerns. Formal equilibrium results describe particular assumptions; empirical policy evaluation must test their relevance.
Ruin, Liquidity, Capital, and Reinsurance
An insurer must meet liabilities when they fall due. Expected aggregate losses can be accurately forecast while a concentrated event still exceeds liquid funds or available capital. Capital adequacy is therefore a tail and timing problem as well as an expected-value problem.
A classical surplus process illustrates the issue:
[ U(t)=u+ct-\sum_{i=1}^{N(t)}X_i, ]
where (u) is initial surplus, (c) is premium inflow per unit time, and (N(t)) is the claim count by time (t). The probability of ruin concerns whether (U(t)) falls below zero over the stated horizon. In this stylized model, the assumptions about premium flow, claim arrival, severity, investments, and interventions are substantial. Real insurer solvency adds assets and liabilities whose values change together, capital eligibility rules, settlement lags, and management responses.
A large capital buffer can absorb modeled shocks but has a cost. Solvency decisions balance security, policyholder protection, availability of coverage, competition, and the opportunity cost of capital. Regulatory capital is a policy-defined requirement; economic capital is an internal model-based assessment. The two can differ in horizon, confidence standard, asset valuation, scenario set, and permissible mitigations.
Reinsurance transfers part of an insurer's loss distribution to another risk bearer. Proportional arrangements share premiums and claims in stated proportions. Excess-of-loss contracts respond above an attachment threshold, subject to a limit and other conditions. Aggregate stop-loss covers may respond after accumulated losses cross a stated threshold. Reinsurance can reduce net tail loss and capital pressure, while introducing counterparty, basis, liquidity, and accumulation risks.
The International Association of Insurance Supervisors' Insurance Core Principles describe insurer risk management and control expectations, including risk appetite, identification, assessment, monitoring, mitigation, and reporting. Their meaning depends on the adopted edition and the relevant supervisory regime; the IAIS online tool documents the December 2024 revision, which provides the international reference point rather than a substitute for national law. IAIS (2024).
Catastrophe, Climate, and Environmental Risk
Catastrophe risk combines hazardous events with the distribution of exposed people, assets, and ecological systems. Catastrophe models commonly separate event generation, hazard intensity, exposure, vulnerability, financial policy terms, and aggregation. Their output can include annual expected losses, exceedance-probability curves, event-loss distributions, and scenario-based stress results. Each component contributes uncertainty.
A coastal flood risk estimate depends on storm characteristics, sea level, topography, defenses, land use, building vulnerability, and policy coverage. Multiplying an average hazard probability by an average vulnerability is generally inadequate when vulnerability depends on hazard intensity, exposure is spatially concentrated, and losses are nonlinear. Physical models and probabilistic financial models address different links in the causal chain.
The IPCC Sixth Assessment defines climate risk in relation to the interaction of climate-related hazards, exposure, and vulnerability. Its Working Group II assessment documents compound and cascading risks and the dependence of future outcomes on warming, development, and adaptation. The 2023 synthesis report similarly identifies interacting climatic and non-climatic risks. These assessments support the conclusion that historical stationarity cannot be assumed in long-horizon climate insurance without an explicit argument. They do not furnish a universal actuarial loss distribution for a particular contract. IPCC (2022); IPCC (2023).
Climate-change attribution studies distinguish changes in physical hazard from changes in exposure, vulnerability, and insured values. A higher aggregate insured loss can result from property growth or inflation even when physical hazard intensity is unchanged; the reverse can also occur. Consistent attribution requires spatially and temporally aligned data and appropriate counterfactuals.
Mitigation and adaptation affect risk through different mechanisms. Emissions reductions influence long-term physical trajectories; defenses, building codes, land-use planning, risk awareness, and emergency response influence exposure and vulnerability. Insurance can support adaptation through risk signals and recovery financing, but it also faces affordability pressures, correlated exposures, and uncertainty about rapidly evolving hazards.
Cyber, Technological, and Emerging Risk
Cyber risk links information systems, organizational dependencies, incentives, and contractual liabilities. Events may include data compromise, service interruption, ransomware, software failure, and third-party provider outages. Frequency and severity records are shaped by underreporting, changes in technology, evolving legal obligations, coverage interpretation, and attackers' adaptation. Shared infrastructure creates common-cause risk.
Biener, Eling, and Wirfs (2015) examined 994 cyber-loss cases from an operational-risk database and identified challenges involving dependence, limited data, and information asymmetry. The dataset is a particular historical sample, not a census of global cyber losses or a forecast of today's incident distribution. Their findings help establish why conventional pooling assumptions require close examination in cyber insurance. Biener, Eling, and Wirfs (2015).
Technological adoption can alter both the frequency and severity of harm. Automation may prevent routine errors while increasing the scale of a shared software defect. Predictive systems can improve underwriting classification while creating model concentration, opaque decisions, data discrimination, and correlated failure modes. Eling, Nuessle, and Staubli's 2022 journal article synthesizes evidence on artificial intelligence across the insurance value chain and its effect on insurability. Its review context is informative, while individual implementation outcomes require separate evidence. Eling, Nuessle, and Staubli (2022).
For emerging risks, scenario elicitation, red-team testing, dependency mapping, incident taxonomy, and adaptive monitoring may initially be more defensible than precise low-frequency probabilities. Historical analogies should be evaluated for similarity in mechanisms, scale, incentives, and reporting. A rapidly evolving threat requires repeated reassessment of its own evidence base.
Systemic Risk, Contagion, and Network Effects
Systemic risk concerns adverse effects transmitted through interconnected institutions or infrastructures such that the functioning of a wider system is impaired. It may arise through direct obligations, common asset holdings, shared technology, financing constraints, supply chains, or feedback between expectations and behavior. A portfolio-level correlation matrix can miss directional contagion, threshold responses, and the reconfiguration of networks under stress.
Adrian and Brunnermeier's CoVaR work studies the risk of the financial system conditional on the distress of an institution and compares stressed conditional states with a reference state. It contributes an approach to marginal systemic-risk assessment, not a universal measurement of all forms of system failure. Adrian and Brunnermeier (2016).
Acharya, Pedersen, Philippon, and Richardson (2017) develop an analysis of systemic expected shortfall linked to undercapitalization of the financial sector. Their framework illustrates how an institution's contribution to system-wide distress can differ from its stand-alone loss distribution. The empirical application concerns financial-sector episodes and its assumptions must be examined before being transferred to insurers, utilities, or cyber networks. Acharya et al. (2017).
Insurance arrangements can absorb local shocks and distribute losses across capital providers, but cross-border reinsurance concentrations and correlated investments can transmit shocks. Risk transfer that appears effective at the original insurer level may shift rather than extinguish systemic exposure. Network analysis should therefore track counterparties, common assets, contractual triggers, collateral, liquidity, and operational dependencies.
The central modeling challenge is endogeneity. During a systemic event, participants sell assets, reduce coverage, change demand, and alter operations. Their actions change prices and loss transmission. An analysis based exclusively on historical independent observations will miss effects created by collective response. Dynamic stress scenarios and mechanisms-based models are essential complements to statistical fit.
Risk Management and Institutional Governance
Risk management is the continuous process of relating uncertainty to decisions, resources, and accountability. A credible institution identifies material risks, chooses transparent assessment methods, evaluates possible treatments, monitors consequences, and revises its decisions when conditions change. ISO 31000:2018 offers general principles and guidelines, while the IAIS principles address insurer governance and supervisory expectations. Their status, scope, and intended users should remain distinguished. ISO (2018); IAIS (2024).
The range of risk treatments includes avoidance, preventive control, loss mitigation, contractual transfer, hedging, liquidity protection, capitalization, contingency planning, and explicit acceptance. These treatments act on different parts of the risk process. Installing a fire suppression system may lower the probability or severity of property loss; purchasing insurance reallocates part of the residual financial consequence; holding liquid assets improves the ability to meet payments. A governance report should identify which mechanism each treatment changes.
Risk appetite is a decision by an institution about acceptable variation and potential loss in relation to its objectives and duties. It is usually expressed through several limits rather than one scalar. A company can simultaneously constrain the probability of ruin, earnings volatility, maximum catastrophe retention, counterparty exposure, and service outages. Constraints may conflict and require explicit prioritization.
Scenario selection, model validation, audit trails, incident learning, and independent challenge are organizational requirements because risk knowledge can be distorted by incentives. A board receives aggregated risk figures; analysts retain model assumptions; underwriters know contract details; claims teams detect emerging patterns. Failure to connect these layers can produce apparently coherent reporting that is disconnected from operations.
Risk reporting should distinguish observed losses, estimated distributions, plausible scenarios, unknown mechanisms, and normative judgments. The governance system should have documented triggers for recalibration or escalation, and a process for acknowledging when a metric ceases to represent the changing exposure.
Equity, Ethics, and Social Acceptability
Risk and insurance distribute burdens across people and generations. A loss distribution measured in currency may underrepresent irrecoverable ecological damage, loss of life, displacement, cultural destruction, or harms to individuals with few resources. The selection of a consequence metric is therefore an ethical and institutional decision as well as an analytical one.
Consider two households facing the same monetary expected loss but very different wealth and recovery capacity. Equal expected loss does not imply equal welfare loss. The social value of risk protection can be higher for the more vulnerable household, while actuarially indicated premiums may be least affordable to that household. Distributional analysis must accompany aggregate efficiency claims.
Risk-based pricing can improve information about exposure and encourage prevention. It can also produce accessibility problems if risk correlates with poverty, historical settlement patterns, or limited adaptation options. Pooling and subsidies can improve affordability while changing incentives and requiring funding. The evaluation of these arrangements depends on their objectives, fiscal sustainability, distributional effects, and legal setting.
Privacy and fairness issues arise in the collection and use of personal risk information. Predictive variables may be statistically associated with claims yet socially contested as underwriting determinants. Good governance requires legitimate data uses, transparent contract conditions, validation across relevant populations, avenues for contesting consequential decisions, and an account of who benefits from increased predictive precision.
Social acceptability cannot be derived mechanically from expected value or the willingness to pay of wealthier parties. A public risk decision must explain whose consequences are measured, whose preferences count, how uncertainty is distributed, and what institutions are responsible for remedy and review.
Worked Analytical Cases
Case: an independent versus correlated insurance pool
Assume 1,000 exposures each have a 1% probability of a covered loss of 100 and variance (10{,}000^2(0.01)(0.99)=990{,}000\). Under independence, the expected aggregate claims are100,000, and the aggregate variance is 31,464. The coefficient of variation of aggregate claims is approximately 0.3146.
Now suppose every pair shares a constant positive correlation of 0.01 in their binary loss variables, assuming a feasible joint distribution with that exchangeability structure. The aggregate variance becomes (990{,}000[1{,}000+1{,}000(999)(0.01)] = 10{,}880{,}100{,}000\), with standard deviation approximately104,308. Expected claims remain $100,000; dispersion rises because the pool shares common variation. The contrast demonstrates why exposure count alone cannot establish capital adequacy. It is an illustrative mathematical scenario, not empirical data.
Case: the limitations of quantile capital
The two independent 4%-probability losses described earlier each have 95% VaR of zero while the portfolio has 95% VaR of 4 and for the aggregate is $8. The same risk can yield different rankings under expected loss, VaR, and expected shortfall because those measures emphasize different features. A metric should be chosen with its decision purpose in view, rather than treated as a universal definition of danger.
Case: premium and utility
Suppose an individual has initial wealth of 40,000 loss. The expected loss is 400. Under a stylized logarithmic utility function, expected utility without insurance is \(0.99\ln(100{,}000)+0.01\ln(60{,}000)\). The certainty-equivalent wealth is \(100{,}000^{0.99}60{,}000^{0.01}\), approximately99,490.5, giving a maximum full-insurance premium of roughly 509.5 and the $400 expected loss is a certainty-equivalent risk premium, separate from an insurer's expense and capital loading. The figures are illustrative and assume a particular preference function and exact coverage.
Case: climate-induced model change
Imagine a coastal insurer calibrates annual flood losses on a historical period when assets, defenses, and sea level were relatively stable. A new development corridor, changing protective infrastructure, and higher mean sea level alter the mapping from storms to insured loss. Adding more historical data from the earlier exposure regime may reduce sampling error about the past while failing to resolve the future regime. A defensible assessment updates the exposure inventory, examines physical scenarios, and reports model sensitivity. Probability estimation alone cannot repair a misspecified future exposure process.
Critical Discussion and Research Frontiers
Risk theory's strengths and limitations are inseparable. Probability provides coherent calculations, but inference depends on the credibility of data and structure. Expected utility clarifies choice under specified axioms, but observed decisions involve reference points, trust, liquidity, and institutional constraints. Diversification provides powerful protection against independent fluctuations, but it cannot neutralize shared hazards or endogenous contagion. Tail metrics focus attention on catastrophic outcomes, yet the tails are frequently the least observable part of the distribution.
One persistent disagreement concerns how to represent uncertainty that resists reliable probabilities. Subjective probability traditions allow coherent beliefs for uncertain propositions; Knightian and ambiguity-sensitive traditions emphasize the limits of treating uncertain worlds as known distributions. Both approaches illuminate real decisions. The practical research task is to show how a proposed representation can be challenged and how conclusions change when plausible alternatives are admitted.
Another tension arises between solvency protection and social welfare. Greater capital and more restrictive underwriting can protect claim-paying capacity while increasing costs and leaving some exposures uninsured. Subsidized pooling may expand protection while introducing fiscal obligations and incentive effects. There is no universal technical optimum independent of the institution's objectives, financial constraints, and the people who bear its consequences.
The research frontier includes compound climate hazards, global accumulation across reinsurance and capital markets, cyber common-cause incidents, operational dependence on shared digital systems, algorithmic model concentration, changing legal and liability exposures, and the interaction of physical and transition risks. The IPCC assessments provide a scientific basis for treating compound climate risks seriously; contemporary cyber-insurability research supplies evidence of common exposures and data limitations; systemic-risk studies show how conditional distress exceeds stand-alone measurement. These results create an interdisciplinary research agenda grounded in existing work rather than a new proprietary theory.
The evidentiary standard for future progress is demanding: scenario definitions must be reproducible; data coverage and exclusions must be visible; uncertainty about distributions must be distinguished from uncertainty within distributions; validation must include adverse conditions; causal pathways must be examined; and policy consequences must be evaluated across exposed populations. Claims of predictive novelty require independent testing and comparison with existing baselines.
The original analytical synthesis of this monograph consists in connecting these established findings through one recurring question: what is known about the event, what is valued in its consequences, how is dependence introduced, which actor retains the loss, and what changes when the underlying world changes? This is an editorial integration of established theory and evidence, without asserting a novel named doctrine or unprecedented scientific priority.
Conclusion
Risk theory is the organized study of consequential uncertainty. Its scientific content arises from making events, consequences, probabilities, dependencies, decisions, and institutional responsibilities explicit. Historical contributions from probability, utility, statistical inference, insurance economics, and risk measurement have produced a powerful but conditional analytical toolkit. The mathematical formulas for compound loss, risk aversion, diversification, expected shortfall, and ruin yield rigorous conclusions when their assumptions are met. Their usefulness depends on the match between those assumptions and observed or plausible mechanisms.
The central practical result is that no single risk score can replace the joint analysis of hazard, exposure, vulnerability, valuation, time, uncertainty, dependence, and response. An insurer's expected claims, the tail of its loss distribution, its available liquidity, the stability of its reinsurance, and the affordability of coverage answer different questions. Governance succeeds when these answers are connected without allowing one to impersonate the others.
Climate change, cyber interconnection, financial contagion, and technological change intensify the need for that discipline. They place greater weight on model validation, scenario analysis, adaptive institutions, explicit uncertainty, and public accountability. The future of risk science depends on better evidence, improved causal understanding, more credible extreme-event analysis, and clearer communication of what a result warrants. Risk theory provides the foundation for that work throughout insurance, finance, engineering, public policy, and the wider study of uncertain futures.
Research Dossiers
83 approved research subjects are organized into 7 thematic families. Individual titles become links when their complete scholarly monographs are published and verified.
Foundations, History, and Concepts3
Law, Regulation, and Governance1
Risk Science, Data, and Modeling17
- Aggregate Loss and Compound Risk Models
- Bayesian Risk Assessment and Decision Updating
- Climate Risk Uncertainty and Nonstationarity
- Cohort-Level Life Insurance Exposure and Credibility Adjustment
- Compound and Cascading Hazard Risks
- Dependency-Free Life Expectancy and Long-Term Care Exposure
- Dynamic Risk and Time-Dependent Exposure
- Insurance Data, Estimation, and Model Risk
- Insurance Statistical Estimands and Prediction Targets
- Moral Hazard and Insurance Incentives
- Proper Scoring Rules for Insurance Prediction
- Risk as Hazard, Exposure, and Vulnerability
- Risk Model Sensitivity and Robustness
- Road Traffic Crash Exposure and Insurance Risk Denominators
- Simulation and Numerical Risk Evaluation
- Solvency Capital Calibration and Tail Probability Assumptions
- Stress Testing and Scenario-Based Risk Evaluation
Finance, Capital, and Markets9
- Adverse Selection in Insurance Markets
- Claims Reserve Risk Margins and Uncertainty Allowances
- Cross-Risk Aggregation and Enterprise Risk Portfolio Views
- Cyber Insurance Rate Adequacy under Data Scarcity
- Deductible, Limit, and Layer Pricing Adjustments
- Nonstationary Risk and Forward-Looking Insurance Pricing
- Pricing of New Coverages and Sparse Insurance Experience
- Reverse Stress Testing for Insurer Capital Failure
- Risk Measures and Portfolio Capital Allocation
Insurance Products, Coverage, and Claims22
- Affect, Dread, and Insurance Risk Perception
- Ambiguity Perception and Insurance Decision Confidence
- Cohort Longevity Tail Risk and Benefit Duration
- Context Dependence and Stability of Insurance Risk Preferences
- Dependence and Accumulation in Insurance Portfolios
- Insurance Contracts and Risk Transfer
- Insurance Enterprise Risk Decision Processes under Uncertainty
- Insurance Enterprise Risk Taxonomy and Materiality Assessment
- Insurance Risk Expert Judgment and Qualitative Assessment
- Life Insurance Cause-of-Death and Excess Mortality Analysis
- Mortality Data Provenance and Coding Quality in Life Insurance
- Mortality Trend Validation and Backtesting in Life Insurance
- Probability Weighting in Insurance Coverage Choices
- Reference Dependence and Loss Aversion in Insurance
- Risk Aversion and Insurance Demand
- Risk Retention, Self-Insurance, and Self-Protection
- Scientific Methods in Insurance Research
- State-Contingent Insurance and Risk Allocation
- Stochastic Claim Processes
- Structural Identification of Behavioral Insurance Preferences
- Subjective Risk Beliefs in Insurance Decisions
- The Behavioral Insurance Paradox of Small and Catastrophic Losses
Institutions, Operations, and Technology1
Specialized and Interdisciplinary Research30
- Age-Period-Cohort Models for Insured Mortality
- Aleatory and Epistemic Uncertainty
- Availability Heuristics and Rare-Loss Judgments
- Coherent and Convex Risk Measures
- Decision-Making under Ambiguity
- Dependence Structures and Copula Models
- Expected Utility and Risk Preferences
- Extreme Old-Age Mortality and Annuity Payment Uncertainty
- Insurability and Its Boundaries
- Insurability of Catastrophic and Systemic Risks
- Insured-Life Mortality Experience and Selection Effects
- Knightian Uncertainty and Ambiguity
- Loss Frequency Models
- Loss Severity and Heavy-Tailed Distributions
- Mortality Heterogeneity and Frailty in Life Portfolios
- Network Models of Risk Propagation
- Nonstationarity and Time-Varying Risk
- Objective and Subjective Interpretations of Probability
- Optimal Indemnity and Coinsurance Economics
- Probability and Risk under Limited Information
- Risk Aversion and Certainty Equivalents
- Risk Ordering and Stochastic Dominance
- Risk Pooling and Diversification
- Risk Scenarios and Consequence Models
- Robust Decision-Making under Uncertainty
- Social Risk Evaluation and Acceptability
- Systematic Retirement Drawdowns and Rule Backtesting
- Systemic Risk and Contagion
- Tail Dependence and Extreme Joint Losses
- Value at Risk and Expected Shortfall
References
- Acharya, V. V., Pedersen, L. H., Philippon, T., & Richardson, M. (2017). Measuring systemic risk. The Review of Financial Studies, 30(1), 2–47. https://doi.org/10.1093/rfs/hhw088
- Adrian, T., & Brunnermeier, M. K. (2016). CoVaR. American Economic Review, 106(7), 1705–1741. https://doi.org/10.1257/aer.20120555
- Akerlof, G. A. (1970). The market for "lemons": Quality uncertainty and the market mechanism. The Quarterly Journal of Economics, 84(3), 488–500. https://doi.org/10.2307/1879431
- Arrow, K. J. (1963). Uncertainty and the welfare economics of medical care. American Economic Review, 53(5), 941–973. https://www.jstor.org/stable/1812044
- Artzner, P., Delbaen, F., Eber, J.-M., & Heath, D. (1999). Coherent measures of risk. Mathematical Finance, 9(3), 203–228. https://doi.org/10.1111/1467-9965.00068
- Basel Committee on Banking Supervision. (2019). Minimum capital requirements for market risk. Bank for International Settlements. https://www.bis.org/publications/201901-standards-minimum-capital-requirements-market-risk
- Bernoulli, D. (1954). Exposition of a new theory on the measurement of risk (L. Sommer, Trans.). Econometrica, 22(1), 23–36. (Original work published 1738.) https://doi.org/10.2307/1909829
- Biener, C., Eling, M., & Wirfs, J. H. (2015). Insurability of cyber risk: An empirical analysis. The Geneva Papers on Risk and Insurance—Issues and Practice, 40(1), 131–158. https://doi.org/10.1057/gpp.2014.19
- Eling, M., Nuessle, D., & Staubli, J. (2022). The impact of artificial intelligence along the insurance value chain and on the insurability of risks. The Geneva Papers on Risk and Insurance—Issues and Practice, 47, 205–241. https://doi.org/10.1057/s41288-020-00201-7
- Ellsberg, D. (1961). Risk, ambiguity, and the Savage axioms. The Quarterly Journal of Economics, 75(4), 643–669. https://doi.org/10.2307/1884324
- Embrechts, P., Klüppelberg, C., & Mikosch, T. (1997). Modelling extremal events for insurance and finance. Springer. https://doi.org/10.1007/978-3-642-33483-2
- Intergovernmental Panel on Climate Change. (2022). Climate Change 2022: Impacts, Adaptation and Vulnerability (Working Group II contribution to the Sixth Assessment Report, Summary for Policymakers and Technical Summary). https://www.ipcc.ch/report/ar6/wg2/chapter/summary-for-policymakers/
- Intergovernmental Panel on Climate Change. (2023). Climate Change 2023: Synthesis Report (Summary for Policymakers). https://www.ipcc.ch/report/ar6/syr/summary-for-policymakers/
- International Association of Insurance Supervisors. (2024). Insurance Core Principles and Common Framework for the Supervision of Internationally Active Insurance Groups (adopted December 2024). https://www.iais.org/icp-online-tool/
- International Organization for Standardization. (2018). ISO 31000:2018—Risk management: Guidelines (2nd ed.; confirmed 2023). https://www.iso.org/standard/31000
- Kahneman, D., & Tversky, A. (1979). Prospect theory: An analysis of decision under risk. Econometrica, 47(2), 263–291. https://doi.org/10.2307/1914185
- 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
- Knight, F. H. (1921). Risk, uncertainty and profit. Houghton Mifflin Company. Library of Congress digitized original edition. https://www.loc.gov/item/21012860/
- Pauly, M. V. (1968). The economics of moral hazard: Comment. The American Economic Review, 58(3), 531–537. https://www.jstor.org/stable/1813785
- Pratt, J. W. (1964). Risk aversion in the small and in the large. Econometrica, 32(1/2), 122–136. https://doi.org/10.2307/1913738
- Rockafellar, R. T., & Uryasev, S. (2000). Optimization of conditional value-at-risk. The Journal of Risk, 2(3), 21–41. https://doi.org/10.21314/JOR.2000.038
- Rothschild, M., & Stiglitz, J. (1976). Equilibrium in competitive insurance markets: An essay on the economics of imperfect information. The Quarterly Journal of Economics, 90(4), 629–649. https://doi.org/10.2307/1885326
- Savage, L. J. (1954). The foundations of statistics. John Wiley & Sons. Original-edition bibliographic record: https://books.google.com/books/about/The_Foundations_of_Statistics.html?id=AGkoAAAAMAAJ
- Schweizer, B. (1991). Thirty years of copulas. In G. Dall'Aglio, S. Kotz, & G. Salinetti (Eds.), Advances in probability distributions with given marginals (pp. 13–50). Springer. https://doi.org/10.1007/978-94-011-3466-8_2
- Slovic, P. (1987). Perception of risk. Science, 236(4799), 280–285. https://doi.org/10.1126/science.3563507
- Von Neumann, J., & Morgenstern, O. (1944). Theory of games and economic behavior. Princeton University Press. Publisher's historical record and later commemorative-edition description: https://press.princeton.edu/books/paperback/9780691130613/theory-of-games-and-economic-behavior