Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Aleatory and Epistemic Uncertainty

Abstract. Aleatory and epistemic uncertainty are indispensable categories in risk science because a predictive distribution depends both on variability among possible outcomes and on the knowledge used to construct that distribution. Their scientific meaning emerges within an explicit model, observation regime, information set, and decision horizon. Aleatory uncertainty describes variability represented by the stochastic mechanism conditional on a model and its parameters; epistemic uncertainty describes unresolved knowledge about parameters, mechanisms, model structure, measurement, dependencies, and the relevance of past evidence to future conditions. The distinction is analytically useful when its conditional character is preserved. This monograph reconstructs the conceptual history of the distinction, develops formal probability and Bayesian interpretations, derives uncertainty decompositions, examines nonidentifiability and observational limitations, and traces uncertainty through insurance frequency–severity systems, dependence, aggregate loss, quantiles, expected shortfall, solvency capital, and catastrophe exposure. Original worked examples demonstrate three consequences: common uncertain parameters induce predictive dependence even when individual losses are conditionally independent; predictive count variance can differ sharply from plug-in variance; and rare-loss value at risk can change discontinuously near a probability threshold. The analysis distinguishes probabilistic integration from ambiguity or incomplete model specification, and separates uncertainty about the world from uncertainty about data that purport to describe it. Inference and governance are treated together: appropriate loss metrics, validation, information acquisition, and institutional disclosure depend on what is known, how it was learned, and how errors affect action. The central conclusion is that a scientifically meaningful risk estimate is an explicitly conditional, auditable claim whose stochastic assumptions, epistemic limits, and decision use must remain visible.

Keywords: aleatory uncertainty; epistemic uncertainty; risk science; insurance; Bayesian inference; parameter uncertainty; structural uncertainty; model discrepancy; measurement error; predictive distribution; dependence; uncertainty propagation; aggregate loss; value at risk; expected shortfall; catastrophe risk; actuarial modeling; decision analysis; model validation.

Introduction

Uncertainty is present at every stage by which insurance transforms future losses into present contractual obligations. Whether a vehicle crashes, a policyholder dies within a year, or a hurricane makes landfall contains a component of variation among possible outcomes. The estimated frequency of such events, the classification of exposed assets, the credibility of historical records, the severity model, and the transferability of observed experience to a changing portfolio contain further uncertainty about what can be known. Risk science requires an account of both components because a numerical result can be arithmetically exact while being epistemically fragile.

The older distinction between measurable risk and uncertainty appears in Frank H. Knight's Risk, Uncertainty and Profit (1921). Contemporary technical literature gives the aleatory–epistemic division a different operational role. Der Kiureghian and Ditlevsen (2009), in their analysis of aleatory and epistemic uncertainty, argue that the classification is meaningful relative to a model and the analyst's possibility of reducing uncertainty by acquiring information or improving models. The classification guides the allocation of research effort; it also changes calculations whenever epistemic variation creates shared dependence across events.

An insurer may know the conditional claim probability of a homogeneous risk class to sufficient accuracy for an immediate decision. Another insurer may quote the same numerical probability while relying on five years of incomplete observations, an unstable risk definition, or a model calibrated in another jurisdiction. Those two numerical statements have different evidential status. A distinction between stochastic variability and limited knowledge makes that difference expressible, but it must be backed by an identifiable information structure rather than labels alone.

This Research Dossier is anchored in the published Risk Theory Science monograph and is connected, under its single canonical identity, to Actuarial Science, Insurance Statistics and Predictive Modeling, Enterprise Risk Management in Insurance, and Catastrophe Risk. Its focus is the classification and propagation of uncertainty. Extended treatments of insurance data governance, ambiguity-based choice axioms, particular risk measures, and catastrophe-model engineering remain distinct scientific subjects.

Research questions and contribution

The investigation asks what mathematical assumptions distinguish randomness from unresolved knowledge; when that distinction is stable across model levels; how inference and measurement affect predicted losses; how uncertainty is transferred or amplified by nonlinear models and dependence; and which conclusions follow for capital, pricing, validation, and decisions. The method combines conceptual reconstruction, probability identities, simple exact counterexamples, comparisons of independently verified scholarly claims, and assessment of primary institutional documents.

The original contribution here is an integrated analytical synthesis and independently derived examples. The formulas rely on established probability theory. The manuscript introduces no proprietary theory, newly named model, doctrine, or priority claim. Empirical assertions are confined to the scope of the linked primary publications and standards, and hypothetical numerical illustrations are identified as such.

Types of Uncertainty

The scientific role of a conditional taxonomy

A risk analysis begins with a target quantity: a future claim amount, an insured loss count, a liability cash-flow trajectory, the probability of default, or the maximum loss across a catastrophe portfolio. The analyst then chooses a model and an information set. Within that specification, variability remaining after fixing parameters and mechanisms is conventionally called aleatory. Uncertainty about parameters, structures, data quality, and the validity of assumptions is epistemic. The same quantity can change its characterization when the analysis conditions on new information or changes scale.

Formally, let (Y) be a future outcome, (M) a model specification, (\theta) its parameters, and (D) available observations. A conditional distribution (p(y\mid\theta,M)) represents stochastic variation under the selected model and parameters. A distribution (p(\theta,M\mid D)), where a coherent posterior can be specified, represents knowledge-dependent uncertainty about those choices. The predictive distribution is

p(y∣D)=∑M∫p(y∣θ,M,D) p(θ,M∣D) dθ.p(y\mid D)=\sum_M\int p(y\mid\theta,M,D)\,p(\theta,M\mid D)\,d\theta .

The model sum applies when a countable set of candidate models and associated probabilities is defensible; the formula becomes an integral for continuous model spaces. Conditional independence of future outcomes and observations given ((\theta,M)) is an additional assumption when the first factor is simplified to (p(y\mid\theta,M)). The equation itself is a coherent Bayesian predictive construction, not a demonstration that the selected set of models contains the actual data-generating mechanism. Hoeting et al. (1999) describe the rationale and limitations of Bayesian model averaging.

An epistemic quantity can be assigned a probability under a Bayesian interpretation, represented by bounds or sets of distributions, or described as a structured scenario whose probability is unsettled. These choices express different inferential commitments. Helton, Johnson, and Oberkampf (2004) compare probability, evidence, possibility, and interval methods under uncertainty in model inputs. Their comparison supports explicit selection of a representation according to information and decision needs; it does not establish the universal superiority of one formalism.

Historical and disciplinary distinctions

Knight's 1921 economic argument concerns the limits of known probabilities and the role of judgment in economic organization. Technical aleatory–epistemic classifications usually assume a specified analytical model and then ask whether a source of uncertainty is treated as irreducible variation or reducible lack of knowledge. The two traditions intersect, but their questions are different. Knightian uncertainty and ambiguity require separate attention when the admissible probability judgments themselves remain incomplete.

Paté-Cornell (1996) examines six levels of uncertainty treatment, including the practical question of when a two-tier analysis is warranted and how expert judgments can enter epistemic assessment. A two-tier structure permits one to simulate stochastic outcomes conditional on uncertain inputs and then examine how conclusions vary across those inputs. It is especially valuable when policy decisions depend on a remote tail or on the probability that a safety constraint will be breached.

Der Kiureghian and Ditlevsen (2009) identify a further difficulty: epistemic uncertainty shared by multiple exposures can induce unconditional dependence. A variable may seem aleatory when treated as a draw from a specified population, yet epistemic when interpreted as the unknown fixed property of one realized system. A building's resistance estimated before construction, the realized resistance of that building, and its observed performance under future loads illustrate how the information and unit of analysis matter. The scientific response is to declare the level at which uncertainty has been classified.

A more exact analytical vocabulary

Several distinctions sharpen the central pair. Variability concerns differences among realized outcomes or population units. Parameter uncertainty concerns incomplete knowledge of numerical quantities inside a chosen model. Structural uncertainty concerns incomplete knowledge of model form, omitted mechanisms, or causal dependencies. Observation uncertainty concerns what an instrument, administrative record, or classification procedure actually reveals. Scenario uncertainty concerns contingencies whose future realization depends partly on social choices, interventions, or changing external conditions. Model discrepancy concerns systematic difference between a representation and the process it aims to approximate. These dimensions can overlap; they specify what evidence would narrow the uncertainty and which assumptions would need modification.

Uncertainty also differs from heterogeneity. Suppose two age classes have different mortality rates. If age is observed and the correct class-specific rate is known, observed heterogeneity can be incorporated into the conditional distribution without being classified as parameter uncertainty. If ages are missing or classes mix in unobserved proportions, uncertainty about composition becomes epistemic. If a model merges distinct classes and calls all residual variation random, it has traded an incomplete representation for a larger stochastic error term. That operation may be convenient for prediction within the calibrated population; it is hazardous when the insured population changes.

Probability, confidence, and precision

A posterior probability and a frequentist coverage statement have different conditioning. A Bayesian credible interval summarizes a posterior distribution given data and assumptions. A frequentist confidence procedure has a specified long-run coverage property under a sampling model, with its interval random before observing the data. Neither automatically quantifies structural model inadequacy. A narrow interval derived from a misspecified likelihood can convey misleading apparent precision.

The probability of an event, the confidence assigned to evidence supporting the event estimate, and the reliability of an entire forecasting procedure therefore require distinct descriptions. IPCC's guidance on consistent uncertainty language illustrates an institutional effort to distinguish evidence and agreement from calibrated likelihood terms. Its calibrated expressions serve assessment communication; they are not direct replacements for actuarial sampling distributions or regulatory capital probabilities.

Aleatory Stochastic Variation

Conditional randomness and model construction

A useful stochastic model specifies an event, a population or exposure, a time horizon, and a probability mechanism. Let (X_i) be a claim indicator for policy (i) over one year. Under conditional independence and an identical known probability (p),

Xi∣p∼Bernoulli⁡(p),N=∑i=1nXi,X_i\mid p\sim\operatorname{Bernoulli}(p),\qquad N=\sum_{i=1}^{n}X_i,

and (N\mid p\sim\operatorname{Binomial}(n,p)), with (E[N\mid p]=np) and (\operatorname{Var}(N\mid p)=np(1-p)). Here stochastic variation describes which contracts produce claims even when (p) is fixed. The formulas depend on homogeneity, the specified horizon, and conditional independence; they cannot be transported to a portfolio with clustered hazards or common shocks without modification.

For a count process (N\mid\lambda\sim\operatorname{Poisson}(\lambda)), conditional expectation and variance equal (\lambda). In insurance, a Poisson assumption may approximate rare, weakly dependent events over an exposure period. It is an analytical choice to be examined against actual count dispersion, seasonality, operational reporting, and the presence of common causes. Unexplained overdispersion can arise from population heterogeneity, clustering, omitted state variables, or uncertain intensity. Those possibilities have different implications for risk control and forecasting.

In a compound insurance-loss model, (S=\sum_{j=1}^{N}Z_j), where (N) is claim frequency and (Z_j) are severities. Assuming (N) is independent of identically distributed severities with finite first and second moments,

E[S]=E[N]E[Z],Var⁡(S)=E[N]Var⁡(Z)+Var⁡(N)(E[Z])2.E[S]=E[N]E[Z],\qquad \operatorname{Var}(S)=E[N]\operatorname{Var}(Z)+\operatorname{Var}(N)(E[Z])^2 .

Both expressions are exact consequences of conditional expectation and variance. If severe catastrophe years also cause more claims, or the severity law changes after an event, the independence assumption is false and further covariance or conditional-model terms enter. The selection of the stochastic structure is therefore already an epistemic decision.

Randomness is conditional on scale and information

At a granular level, a traffic collision or individual claim is uncertain. At an aggregate level, repeated independent exposures may exhibit partial stabilization of the average. When variance exists and (X_i) are independent with common variance (\sigma^2), the sample mean has variance (\sigma^2/n). The reduction follows from independence and finite second moments. Common shocks, unmeasured heterogeneity, and shared parameter uncertainty can prevent analogous stabilization.

Even a large insurance book can remain sensitive to a common driver: weather systems, regional inflation, litigation changes, network outages, or judicial interpretation. The number of policies is then an incomplete proxy for diversification. A portfolio can contain many contracts while having few effectively independent sources of extreme loss. Variability that diversifies at the policy level can coexist with persistent uncertainty at the portfolio level.

Frequency and severity are different stochastic objects

A count model determines how many claims occur; a severity distribution determines the cost conditional on an event. Their combination governs aggregate loss. A fitted average claim size cannot establish a tail's shape, and a stable count mean cannot establish the absence of clustering. For instance, two severity distributions can share the same mean but have radically different high quantiles and expected shortfalls. Capital conclusions require information about more than the first moment.

The distinction is especially important for insurance limits and deductibles. A payout function (g(Z)=\min{(Z-d)_+,u}), with deductible (d) and limit (u), maps underlying damage into a contractual payment. Its distribution depends on the damage tail and on exact contractual terms. The stochastic physical event and the financial claim are related through an exposure and contract model. Uncertainty about the interpretation of coverage wording or an unobserved peril exclusion belongs to the epistemic and institutional analysis even if a simulator can generate precise claim amounts.

Repeatability, nonstationarity, and hazard regimes

Stochastic regularities are learned from some historical process. Their predictive relevance requires stable features across the horizon of interest. Historical frequencies for a coastal insured portfolio may not extrapolate after migration, reconstruction, altered building codes, reinsurance changes, or a changing hazard. The IPCC's discussion of sources of climate projection uncertainty separates internal climate variability, physical response uncertainty, and future forcing scenarios. These components cannot be collapsed into a single timeless sampling error without obscuring their causes.

A process with time-varying intensity (\lambda(t)) may still possess a well-defined stochastic model, but a constant-rate fit ignores its structure. The natural analytical sequence is to ask whether the target quantity has changed, whether the underlying mechanism has changed, which covariates are predictive, and what future states remain plausible. Increased data volume from an obsolete regime may make a parameter estimate more precise while making the inference less relevant.

Parameters and Structural Knowledge

Parameter uncertainty and posterior prediction

Suppose the claim probability (p) is not known. A Bayesian analyst may represent information using a Beta distribution (p\sim\operatorname{Beta}(a,b)). Given (x) claims in (n) independent Bernoulli observations generated with a fixed common (p), the posterior is (\operatorname{Beta}(a+x,b+n-x)). For a new policy, the posterior predictive probability of a claim is the posterior mean. For a group of new policies sharing the same unknown (p), the predictive count is beta-binomial, not ordinary binomial.

Write (a,b) for posterior shape parameters and (\mu=a/(a+b)). For (m) future conditional Bernoulli exposures,

E[N∣D]=mμ,E[N\mid D]=m\mu,
Var⁡(N∣D)=mμ(1−μ)+m(m−1)Var⁡(p∣D),Var⁡(p∣D)=μ(1−μ)a+b+1.\operatorname{Var}(N\mid D)=m\mu(1-\mu)+m(m-1)\operatorname{Var}(p\mid D), \qquad \operatorname{Var}(p\mid D)=\frac{\mu(1-\mu)}{a+b+1}.

Equivalently (\operatorname{Var}(N\mid D)=m\mu(1-\mu)[1+(m-1)/(a+b+1)]). The derivation uses the law of total variance and the common latent (p): future claims are independent conditional on (p), but positively associated after integrating over shared parameter uncertainty. This is an established mixture-model consequence, closely related to the predictive dependence described by Der Kiureghian and Ditlevsen.

As an original illustration, let (a=2,b=8,m=100). Then (\mu=0.2), (\operatorname{Var}(p)=16/1100), and the conditional binomial variance at a plug-in (p=0.2) would be (100(0.2)(0.8)=16). The beta-binomial predictive variance is (16+100(99)(16/1100)=160). The predictive standard deviation is approximately (12.65) claims rather than (4) claims under the fixed-parameter plug-in. Both calculations are correct for their explicitly different assumptions. The numerical difference does not mean that one physical book contains ten times as much irreducible variation; it records uncertainty about a common rate across that future book.

The illustration also shows how a modeling decision determines whether parameter uncertainty appears as cross-policy correlation. If each policy instead has its own independent (p_i) drawn from the same Beta distribution, the unconditional indicators remain independent and the aggregate variance is (m\mu(1-\mu)). A common latent parameter and independent latent policy parameters can produce the same marginal claim probability but sharply different aggregation risk. The latent structure must be justified by how the exposures share causes or knowledge.

What additional observations can and cannot accomplish

Under a correctly specified stationary beta-binomial model, more informative independent observations update the posterior and generally reduce posterior variance. This is one reason the aleatory–epistemic distinction has practical value: it separates an irreducible conditional claim from a component that investigation may reduce. It does not imply that every uncertainty disappears with increasing sample size. Omitted variables, selection effects, changing hazard regimes, and incorrect dependence assumptions can persist or become more consequential.

Consider an insurer observing zero claims in 100 exposures, starting from a uniform Beta(1,1) prior. The posterior is Beta(1,101), with mean (1/102), rather than a claim probability of zero. For another 100 future policies sharing the same unknown (p), the exact predictive probability of at least one claim is

P(N≥1∣D)=1−B(1,201)B(1,101)=1−101201=100201≈0.4975.P(N\ge 1\mid D)=1-\frac{B(1,201)}{B(1,101)} =1-\frac{101}{201}=\frac{100}{201}\approx0.4975 .

The probability is a mathematical consequence of a deliberately stated prior, conditional independence, a fixed rate, and a future exposure size. It is not an empirical estimate of any real portfolio. The example demonstrates why a zero observed count alone cannot justify a categorical zero-risk declaration. Different scientifically justified prior knowledge and dependence structures can change the result.

Structural uncertainty and model averaging

Parameter uncertainty is defined relative to a selected model. Structural uncertainty asks whether the model contains the mechanisms needed for the target decision. Candidate frequency models may be Poisson, negative binomial, self-exciting, or hierarchical. Candidate severity models may differ in tail class, threshold, inflation dependence, or treatment of catastrophe limits. A portfolio distribution calculated under one model can be sensitive to choices whose uncertainty is hidden by precise parameter fitting.

Under a model family ({M_k}), the predictive cumulative distribution can be represented as

F(y∣D)=∑kwk(D)Fk(y∣D),∑kwk(D)=1,F(y\mid D)=\sum_k w_k(D)F_k(y\mid D),\qquad \sum_k w_k(D)=1,

when posterior model weights (w_k(D)) are defined. Hoeting et al.'s model-averaging tutorial (1999) explains how accounting for model selection uncertainty can change inference and prediction. Averaging cannot compensate for a shared omission across all candidates. If every candidate excludes a novel cyber failure mechanism, the ensemble can agree precisely and still omit an important loss pathway.

Model discrepancy can be formulated as (Y=f(X,\theta)+\delta(X)+\varepsilon), where (f) is a chosen representation, (\delta) describes systematic discrepancy, and (\varepsilon) a stochastic or observation term. Identifiability becomes difficult when (\delta) and (\theta) can compensate for one another. For instance, a high hazard rate with low damage sensitivity and a low hazard rate with high damage sensitivity may produce similar aggregate mean losses. Observing the mean alone will not identify the separate mechanisms.

Nonidentifiability and observational equivalence

Inference requires more than a sufficiently sophisticated optimizer. Suppose a compound-Poisson loss model has rate (\lambda) and mean severity (\mu_Z). The observed aggregate expected loss is (\lambda\mu_Z). If only that product is observed, pairs ((\lambda,\mu_Z)) on the same product curve are observationally equivalent for the mean. Their count variability, deductible effects, and high-loss tails can nevertheless differ. Separating frequency and severity requires information or assumptions beyond the aggregate mean.

A second ambiguity arises in correlation: a measured cross-policy covariance may reflect persistent common exposure, transient catastrophe states, common estimation uncertainty, or observation-system synchronization. A covariance matrix by itself need not identify a causal source. Analytical conclusions about diversification or reinsurance structure must therefore be conditional on a justified interpretation, not merely on a fitted numerical dependency coefficient.

Probability ranges and alternative uncertainty representations

When a single posterior is underdetermined, a bounded model family may represent current knowledge more faithfully. Let (\Theta) contain a defended set of parameters or models, with each (\theta) generating (F_\theta). One may report a range ([\inf_{\theta\in\Theta} r(F_\theta),\sup_{\theta\in\Theta}r(F_\theta)]) for a risk functional (r), where the extrema exist or are interpreted appropriately. The range is a sensitivity statement conditional on the admissible set; it is not itself a frequentist confidence interval. Wider admissible sets yield broader conservative possibilities, sometimes including implausible extremes if set construction is careless.

The alternative uncertainty representations investigated by Helton, Johnson, and Oberkampf provide a reason to separate method choice from substantive evidence. A probability density introduces comparative weights over possibilities. An interval may express only limits. Expert-defined plausible scenarios can describe pathways with no defensible frequency. Risk decisions should reflect the semantic difference instead of converting every uncertainty representation into an apparently comparable percentile.

Observation and Measurement

From physical events to administrative records

Insurance data are produced by physical events, contractual definitions, reporting behavior, organizational processes, and information systems. The observed record is therefore a mediated representation of the risk target. A catastrophe damage event may generate multiple claims, no claims, delayed claims, or a disputed claim. A fraud classification may depend on investigation intensity. A claim amount may be incurred, reported, paid, reserved, gross of reinsurance, net of deductible, or adjusted for inflation. Each quantity has a different observation rule.

An illustrative observation model is (Y_{\mathrm{obs}}=h(Y_{\mathrm{true}},C,T)+\varepsilon), where (C) denotes contract and classification conditions, (T) timing, and (\varepsilon) residual measurement error. Additive noise is often a simplification. Missing claims, reporting delays, limits, threshold censoring, and selection have mechanisms that cannot in general be reconstructed by assuming an independent zero-mean error.

Distinguish data-generation uncertainty from true heterogeneity. A damaged property assigned to a wrong geographic cell can create geocoding error; an unknown building material can create exposure uncertainty; a building's actual vulnerability under a random event is part of the modeled hazard–damage process. A catastrophe estimate can inherit all three. Better geocoding reduces one source of epistemic uncertainty without changing the underlying windstorm.

Measurement uncertainty and metrological standards

The Joint Committee for Guides in Metrology's Guide to the Expression of Uncertainty in Measurement (JCGM 100:2008) establishes rules for evaluating and expressing measurement uncertainty, and its Monte Carlo propagation supplement (JCGM 101:2008) describes propagation of distributions through a measurement model. These standards address measurement modeling with explicitly characterized inputs and outputs. Their methods illustrate a discipline of identifying measurands, assumptions, uncertainty components, and propagation, but they do not by themselves verify an insurance catastrophe model or resolve contractual coverage questions.

Suppose a property valuation (V) is estimated with random assessment error and a possible systematic regional valuation bias (b). Repeated appraisals may reduce independent assessment noise, while leaving the shared bias largely untouched. Averaging ten valuations that use the same flawed replacement-cost index can yield a narrow sampling interval around a systematically distorted value. Distinguishing repeated-measurement variability from common method bias is central to accurate reporting.

Missingness, truncation, and selection

An insurer often learns from observations conditional on entry into a database. Only reported claims appear in a claims table; only insured exposures appear in a portfolio; only historically affected properties may have detailed loss-adjustment records. If inclusion depends on event severity, location, wealth, or policy conditions, the observed sample need not represent the target population. A consistent parameter estimator for the selected sample may be inconsistent for the future insurance book.

For a severity (Z) with recording threshold (u), observed payments may correspond to (Z\mid Z>u), or to a payment transformation of (Z), depending on accounting conventions. Treating observed threshold exceedances as an untruncated random sample distorts likelihoods and aggregate predictions. Recovering the underlying distribution requires the truncation mechanism and exposure denominator to be explicit. Where these are unavailable, scenario ranges or partial identification may be more defensible than a single calibrated tail parameter.

A change in reporting software can alter observed frequency while the underlying physical event intensity remains constant. Conversely, improved fraud detection may increase detected fraud incidence without implying that fraud occurrence increased. Scientifically useful time-series comparison therefore includes a record of definitions, coverage, system migrations, and revisions.

Validation: inference about future performance

Model validation is a claim about the relationship between a model, the data used to assess it, and the intended decision. Reproducing in-sample counts is weak evidence for future extreme-loss calibration. In strictly proper scoring rules, Gneiting and Raftery (2007) explain how scoring probabilistic forecasts can encourage distributions that genuinely reflect predictive judgment. Gneiting, Balabdaoui, and Raftery's (2007) calibration and sharpness framework separates statistical consistency of forecasts and observations from concentration of predicted distributions.

For continuous outcomes, a predictive interval advertised at nominal level 90% should be evaluated against an appropriate out-of-sample population and horizon. Coverage alone is incomplete: a trivially wide interval can cover nearly everything. Sharpness, conditional calibration across risk groups, proper scoring rules, decision-specific tail diagnostics, and comparison with simple baselines reveal additional properties. In high-quantile insurance applications, a limited historical sample may have few exceedances; this limits what backtesting alone can establish.

The ASOP No. 56 modeling standard establishes professional considerations for actuarial model purpose, assumptions, inputs, output review, and communication within its United States scope. It should be interpreted according to its applicability and adoption rules rather than projected as binding law across every jurisdiction. Its relevance here is methodological: the limitations and purpose of a model matter as much as its numerical calculation.

A data-quality counterexample

Consider two books with the same observed average annual claim frequency. Book A has complete claims and exposure records for a stable, precisely defined portfolio. Book B has a large but changing deductible, delayed reporting, and an exposure denominator assembled from unmatched policy systems. Equal sample means do not give equal support to a prediction of future paid losses. A larger Book B database may contain more rows and less relevant information.

The appropriate object for evidence assessment is the entire observation process, not the row count. Accuracy of events, definition of exposures, completeness, time alignment, external validity, and decision-specific error costs determine how strongly past data can justify future risk claims. Statistical uncertainty intervals constructed from the recorded values require separate treatment of these foundational limits.

Uncertainty Propagation

The law of total variance as a starting point

Let (Y) have finite variance and (\Theta) summarize uncertain parameters or a latent state. The law of total variance gives

Var⁡(Y)=E[Var⁡(Y∣Θ)]+Var⁡(E[Y∣Θ]).\operatorname{Var}(Y)=E[\operatorname{Var}(Y\mid\Theta)]+\operatorname{Var}(E[Y\mid\Theta]).

The first term is conditional stochastic variation averaged over (\Theta); the second is variation in the conditional mean as (\Theta) changes. It is convenient to call the first an aleatory component and the second an epistemic component when (\Theta) represents epistemically uncertain parameters. The naming remains conditional: if (\Theta) represents an objectively fluctuating environmental state, the second term is variability across states instead. Algebra supplies an exact decomposition; scientific interpretation supplies its meaning.

For future insurance losses with model parameter (\theta), the fully predictive calculation draws (\theta) from its justified distribution or set, then outcomes conditional on that draw. Replacing (\theta) with its estimated mean changes the variance, dependence, and sometimes the shape of the tail. The bias in a tail measure is not generally equal to the extra variance term because quantiles are nonlinear functionals of the entire predictive distribution.

Shared latent rates and portfolio dependence

Suppose (N_1,\ldots,N_k) are conditionally independent Poisson variables with a common latent annual rate (\Lambda): (N_i\mid\Lambda\sim\operatorname{Poisson}(\Lambda)). The tower property yields

E[Ni]=E[Λ],Var⁡(Ni)=E[Λ]+Var⁡(Λ),E[N_i]=E[\Lambda],\quad \operatorname{Var}(N_i)=E[\Lambda]+\operatorname{Var}(\Lambda),

and for distinct (i,j),

Cov⁡(Ni,Nj)=Var⁡(Λ).\operatorname{Cov}(N_i,N_j)=\operatorname{Var}(\Lambda).

The covariance follows because the conditional covariance is zero while (E[N_i\mid\Lambda]=E[N_j\mid\Lambda]=\Lambda). If the (N_i) instead have independently drawn latent rates, their cross-covariance can vanish. Hence the same one-portfolio marginal distribution does not uniquely determine systemwide diversification. One must specify which unknown factors are shared and how they act on different insured units.

This result is relevant to risk pooling, reinsurance, and portfolio diversification, but its interpretation depends on the meaning of (\Lambda). A common stochastic catastrophe state, an uncertain global rate parameter, and a mixture of both can generate similar predictive covariance. Data may not uniquely distinguish them, and their possible responses to new information or interventions may differ.

Nonlinear propagation, Monte Carlo, and numerical precision

For an output (Y=f(X_1,\ldots,X_d)), a local first-order approximation around a point (x_0) gives (\operatorname{Var}(Y)\approx \nabla f(x_0)^\top\Sigma_X\nabla f(x_0)) when the approximation is valid. Strong curvature, discontinuities, threshold effects, heavy-tailed inputs, and dependencies can make such an approximation inaccurate. The JCGM's distribution propagation by Monte Carlo provides a general measurement-model example of evaluating outputs by propagating input distributions, subject to its stated scope.

In insurance, a simulation can draw claim counts, severities, economic inflation, catastrophe intensities, and uncertain parameters in nested or joint fashion. The simulation design must preserve common factors within a simulated year. Independent resampling of a shared latent rate for each individual contract can artificially remove dependence. Conversely, reusing a common parameter draw across scenarios can be appropriate when a parameter describes the same unknown environment. Simulation is a computational realization of assumptions; increasing the number of Monte Carlo paths reduces numerical error under suitable conditions but does not remove structural or input uncertainty.

A disciplined simulation report identifies at least the target functional, stochastic laws, parameter uncertainty representation, correlation and common-factor design, number of simulations, numerical convergence evidence, and model alternatives. For tail estimators, uncertainty from rare-event sampling must be separated from epistemic input uncertainty. Nested simulation may have considerable computational expense; sensitivity analysis can identify which inputs warrant more investigation.

Extreme nonlinearities and fragile thresholds

A simple threshold calculation illustrates the limits of variance-only reporting. Let (L) equal a fixed severe loss (M>0) with probability (p), and zero otherwise. Define (\operatorname{VaR}_q(L)=\inf{\ell:F_L(\ell)\ge q}). Then

VaR⁡q(L)={0,p≤1−q,M,p>1−q.\operatorname{VaR}_q(L)= \begin{cases} 0,& p\le 1-q,\\ M,& p>1-q. \end{cases}

At (q=0.995), an event probability of (0.004) produces (\operatorname{VaR}{0.995}=0), while (p=0.006) produces (\operatorname{VaR}{0.995}=M). The two models differ by only two tenths of one percentage point in event probability. At (p=0.004), expected shortfall defined by integrated quantiles, ((1-q)^{-1}\int_q^1\operatorname{VaR}_u,du), is (0.8M). These are exact results of the hypothetical two-point model. They show a decision threshold at which limited knowledge of (p) is more important than extra numerical digits in the computed VaR.

This is not a general theorem that VaR is always unsuitable or that expected shortfall eliminates model uncertainty. Both measures remain functions of a modeled tail. For a nonlinear or discontinuous functional, sensitivity across defensible alternative inputs can be more informative than a point estimate and its local derivative. A capital report should expose assumptions that can move a tail threshold.

Climate, catastrophe, and scenario propagation

A catastrophe model may be abstracted as a chain from event generation through local hazard intensity, exposure, vulnerability, insurance financial terms, and aggregate financial loss. Each component has stochastic and epistemic dimensions, and errors may be dependent. A small spatial error in locating a high-value concentration can change estimated accumulation. A vulnerability curve calibrated for one construction class may systematically misrepresent another. A future sea-level or precipitation pathway may depend on changing external conditions rather than sampling fluctuation around a fixed historical law.

The IPCC's regional projection uncertainty analysis identifies uncertainty in future forcing, climate response, internal variability, and the regional modeling chain. Insurance losses add socioeconomic exposure, behavioral adaptation, contract terms, and financial inflation. The resulting loss distribution cannot be inferred from a hazard projection alone. The time horizon determines how much scenario uncertainty matters, and information about an imminent event has different value from information about a distant development pathway.

Double counting and false decomposition

A recurrent modeling error is to introduce parameter uncertainty into predictive simulations and then add a separate margin for the same uncertainty as though it were independent. If an epistemic effect is already in the predictive distribution, an additional parameter-risk load must correspond to a distinct objective, a distinct risk aversion, or a genuinely omitted uncertainty; otherwise the treatment may count the same source twice. The converse error is to report a conditional variance and forget model uncertainty entirely.

Accounting standards and regulatory frameworks define their own risk adjustments and capital objectives. A generic variance decomposition cannot decide an IFRS 17 nonfinancial risk adjustment, a market-consistent valuation, or a Solvency II capital requirement without their governing definitions. Scientific clarity begins by separating probability modeling from the valuation or regulation applied to its outputs.

Risk Measurement Consequences

Expected loss, pricing, and reserves

Expected loss is a first-moment estimate, while uncertainty around that estimate affects pricing, reserve adequacy, liquidity, and capital. Under a correctly specified Bayesian model, (E[L\mid D]) integrates uncertainty in parameters and future realizations. A plug-in estimate (E[L\mid\hat\theta]) may differ from the integrated expectation if the expectation function is nonlinear in (\theta). If (E[L\mid\theta]) happens to be linear, the means may agree while predictive variance and tails diverge substantially.

Pricing introduces additional quantities: expenses, acquisition costs, profit objectives, risk load, reinsurance costs, and competition. Uncertainty alone does not determine a unique premium. A risk margin may be derived from a specified utility, cost-of-capital method, coherent risk measure, or other justified valuation objective. The analyst must state which uncertainty components enter the loss prediction and which enter the decision-maker's compensation for bearing that prediction.

For reserves, observational development patterns, reporting delays, inflation, and case estimates can generate both process and parameter uncertainty. A precise central reserve estimate may conceal a broad predictive range for ultimate claims. The range can be reduced by improved data for some lines; it may grow when latent social inflation or judicial changes become more plausible. Governance requires tracking assumption revisions and reasons, not only updating the final estimate.

Quantiles, tail loss, and solvency

The European Union's Solvency II Article 101 specifies a one-year 99.5% value-at-risk calibration of basic own funds for the Solvency Capital Requirement within the applicable legal framework. The historical consolidated text linked here is identified by its version date; implementation and amendments must be checked for a real jurisdictional determination. It supplies a concrete institutional illustration of how a high-quantile risk measure can become a governing target. The result is affected by predictive distribution assumptions, diversification, and model uncertainty; the regulation's numerical confidence level does not validate any particular insurer's model.

The mathematics of the previous chapter shows why a low-probability event near a quantile cutoff may create a large change in reported VaR. Heavy-tail choices, catastrophe event rates, dependency structures, and temporal exposure adjustments can have comparable effects. Expected shortfall integrates quantiles beyond a threshold, supplying information about average tail severity under a specified distribution. It too depends on what the model includes and how well tails are supported by observations.

Risk measurement should therefore distinguish at least three uncertainties: the distribution of realized losses under a chosen model, the uncertainty in the distributional estimate, and the adequacy of the model family itself. A single plotted confidence band seldom resolves all three. Scenario and sensitivity analysis can expose dimensions omitted from a fitted statistical standard error.

IFRS 17 and the insurance liability

The IFRS Foundation's IFRS 17 Insurance Contracts describes measurement of insurance contracts using current estimates of future cash flows and the risk-adjusted fulfillment cash flows framework. Its nonfinancial risk adjustment concerns compensation required for bearing insurance risk, not a universal numerical surcharge derivable from the posterior variance of a risk parameter. Distinguishing estimation from the entity's required compensation helps avoid a false equation between epistemic spread and an accounting margin.

Different accounting and supervisory purposes can legitimately use different summaries of the same modeled loss distribution. An accounting liability, a solvency capital measure, a reinsurance retention limit, and an operational stress test answer different decision questions. Their comparisons require a common exposure description, transparent valuation conventions, and explicit treatment of model uncertainty. One number cannot automatically serve all objectives.

Model governance and reporting responsibility

The International Association of Insurance Supervisors' Insurance Core Principles and ComFrame (December 2024) describe enterprise risk management for solvency purposes, including ongoing identification, measurement, reporting, and management of risks. They provide supervisory guidance rather than evidence that a particular insurer's data or loss model are correct. Professional and institutional controls add a second question to the statistical one: who understands the limitations of the estimate, and who is authorized to act on it?

The ASOP No. 56 modeling requirements reinforce the importance of model purpose, inputs, assumptions, review, and disclosure within their professional scope. ISO's Risk management guidelines provide a wider organizational framework for integrating risk management into decisions. These sources support a general finding: the usefulness of an uncertainty calculation depends on decision context, accountable review, and communication of consequential assumptions. They do not prescribe a single international quantitative decomposition.

A defensible risk report identifies the insured object, relevant time horizon, definitions of covered loss, empirical basis, conditional model, parameter uncertainty, uncertainty about structure, measurement limitations, scenario assumptions, main sensitivity results, chosen risk functional, and the decisions it is intended to inform. Important unresolved dependencies or extrapolations should be visible in a prominent place rather than confined to internal model documentation.

Value of information and decision relevance

Epistemic uncertainty matters because some of it can be reduced by observation, study, experiment, or an improved mechanism model. The value of additional information depends on the decisions it can change. For an action (a), uncertain state (\theta), and loss (C(a,\theta)), the expected value of perfect information for a loss-minimizing decision is

EVPI⁡=min⁡aE[C(a,Θ)]−E[min⁡aC(a,Θ)]≥0.\operatorname{EVPI}=\min_a E[C(a,\Theta)]-E[\min_a C(a,\Theta)]\ge0.

The inequality follows because choosing an action after observing the state cannot worsen the optimal expected cost compared with choosing once before observing it, assuming the state information and action are available without cost. Real information is imperfect and usually costly, so its net value may be lower.

Consider an original two-state example. An exposure has loss probability (0.01) in state A and (0.20) in state B, with the two states equally weighted for decision analysis. A loss costs 100 units. Full protection costs 5 units and eliminates this modeled loss. Without further information, expected unprotected cost is (0.5(1)+0.5(20)=10.5), so protection at cost 5 is optimal. With perfect free state information, the optimal action costs 1 in state A and 5 in state B, for expected cost 3. The EVPI is (5-3=2) units. This is a conditional illustration, not an estimate of information-market value. Partial testing is worthwhile only if its expected decision benefit exceeds its cost and if the model of its informational reliability is credible.

The example clarifies why not every reducible uncertainty requires further research. If a decision is optimal across all plausible model states, additional precision may have little immediate decision value. Conversely, moderate uncertainty near a regulatory threshold, a catastrophe retention limit, or an underwriting exclusion can warrant intensive examination. Sensitivity to model assumptions should guide investigation priorities.

Robustness and the risk of misplaced confidence

A model is decision-robust when the recommended action performs acceptably across scientifically defensible alternatives, using an explicitly stated acceptability criterion. That assessment can be achieved with sensitivity analysis, multiple calibrated models, stress cases, or robust optimization. It does not require declaring a new decision theory or replacing probabilities wholesale. A probability-weighted decision and a worst-case decision need not agree because they encode different judgments about the credibility and cost of adverse possibilities.

Model agreement is insufficient when every model shares the same data defect, exposure definition, or omitted loss mechanism. Numerical convergence is insufficient when simulation paths implement the wrong dependence. Historical forecast calibration is insufficient when conditions have shifted. The defensible institutional response is a traceable chain of conditional claims, independent challenge for material models, and explicit escalation when deviations would alter pricing, reserving, or solvency decisions.

International and social implications

An uncertainty estimate also has distributional consequences. Insured and uninsured households, small firms, governments, and reinsurers may face the same hazard with different information, financial buffers, claims rights, and recovery capacity. Reporting uncertainty as one portfolio-wide coefficient can obscure the concentration of epistemic gaps in undermeasured communities or rapidly changing exposure classes. A risk model that performs well in insured urban data may extrapolate poorly to sparse rural exposures or jurisdictions with different building codes and claims processes.

Scientific transparency therefore includes the geography, period, population, and contractual scope of every major inference. Information asymmetry among insurer, policyholder, regulator, and model vendor affects who bears the consequences of a model error. Uncertainty communication should identify whose decisions are changed by the estimate and whose losses remain outside its data-generating population.

Integrated comparative findings

The investigation yields six connected conclusions. First, aleatory uncertainty is a property of a conditional representation of possible outcomes, whereas epistemic uncertainty concerns the supported knowledge behind that representation. Second, a latent factor's interpretation and sharing structure determine whether its variation enters marginal variance, dependence, or both. Third, data volume cannot replace identification, reliable measurement, and relevance to the future population. Fourth, nonlinear functionals such as extreme quantiles can amplify apparently small epistemic differences. Fifth, probabilistic calibration assesses observed predictive behavior under its test design and cannot alone establish the completeness of future hazard scenarios. Sixth, research effort is most valuable where additional evidence can change material decisions rather than merely narrow a published interval.

These are analytical conclusions drawn from established probability, scholarship, and verified institutional sources. Their force lies in the stated assumptions and transparent counterexamples. They establish a coherent way to evaluate risk assertions without claiming a new theoretical system.

Conclusion

Aleatory and epistemic uncertainty form a productive scientific distinction when each is attached to a precisely described model, information set, target, and decision. Random variation continues within a model whose parameters and mechanisms are fixed. Incomplete knowledge enters through those parameters, the model's structure, the observation process, the dependence between exposures, and the relationship between historical data and future conditions. These elements interact in prediction and insurance valuation, and their interactions cannot be read from a single summary statistic.

The formal findings establish why this matters. Integrating a common unknown claim probability produces beta-binomial dispersion and predictive dependence that a fixed-probability calculation conceals. Integrating a shared latent event rate creates covariance among conditionally independent counts. Threshold-based quantile risk can switch sharply when a rare-event probability crosses the confidence-level boundary. The same body of data can support different conclusions when model form, measurement validity, and decision horizon differ. Each demonstration is conditional on explicit assumptions, and each is independently reproducible from the equations given.

The scientific and editorial consequence is that risk estimates should state what varies, what is unknown, what evidence reduces uncertainty, and what remains outside the chosen model. Strong inference requires suitable observations, tested assumptions, a relevant horizon, calibrated predictive procedures, and serious attention to structural alternatives. Risk governance adds an institutional requirement: disclose material uncertainty in forms that correspond to actual underwriting, reserve, reinsurance, capital, and public decisions.

The resulting account preserves both mathematical rigor and the limits of available evidence. It strengthens the interpretation of insurer risk measurements and cross-disciplinary uncertainty analyses, while leaving model-specific estimates, jurisdictional implementation, and future empirical research to their appropriate independently documented studies.

References