RESEARCH DOSSIER
Objective and Subjective Interpretations of Probability
Abstract. Probability assigns a numerical structure to uncertainty, but its interpretation determines what a numerical assertion means, what evidence warrants it, and which decisions may follow. This research dossier investigates the classical, logical, frequentist, propensity, personalist, and decision-theoretic interpretations of probability, with particular attention to insurance, actuarial inference, catastrophe modeling, risk classification, and solvency. The inquiry distinguishes mathematical probability measures from their substantive interpretation, physical chances from observed frequencies, rational credences from unexamined impressions, and conditional forecasts from unqualified claims about the future. Original readings of Ramsey, de Finetti, Savage, Knight, Popper, and the probability-calibration literature are integrated with actuarial standards and a selected European solvency provision. Formal demonstrations address frequency convergence, heterogeneous risk portfolios, exchangeable Bernoulli observations, posterior predictive estimates, model mixtures, the Brier score, and the effect of dependence on portfolio loss uncertainty. The central analytical finding is that a defensible insurance probability statement requires a defined event, exposure and time horizon; an identifiable target population or physical process; a warranted transfer from evidence to the contemplated future; and disclosure of the interpretation and assumptions governing its use. These are operational questions for documentary scrutiny rather than a new probability theory. Frequency evidence and personal probabilistic judgments often interact productively, while coherence and model calibration test different properties. Financial and regulatory consequences depend on distributions, coverage and dependence as well as event probabilities. This monograph provides a comparative interpretation, documented examples, explicit methodological limits, and source-traceable conclusions. It does not present a new doctrine, a universal actuarial procedure, or a probability estimate for any actual insurer.
Keywords: Probability interpretation; objective chance; subjective probability; personal probability; frequentism; relative frequency; propensity; logical probability; coherence; exchangeability; de Finetti representation; Bayesian inference; calibration; proper scoring rules; uncertainty; epistemic uncertainty; aleatory uncertainty; insurance loss probability; actuarial credibility; catastrophe risk; model risk; Solvency II; risk assessment; probability communication.
Introduction: Research Object, Method, and Boundaries
A probability such as “the annual probability of a covered loss is 2%” appears to convey a self-contained fact. In practice, its meaning depends on the event counted, the denominator of exposure, the period, the mechanism that generates loss, the information available, and the reason a probability is being asserted. A 2% observed historical proportion is a statistic; a 2% estimate for an upcoming contract period is a predictive claim; a 2% physical chance assumes features of an underlying process; and a 2% personal judgment expresses an informed evaluator's uncertainty. These statements can have the same numerical form and different evidential warrants. The Stanford account of probability interpretations provides the general philosophical taxonomy; the present study examines its practical consequences for insurable and systemic risks.
The object of the monograph is the interpretation and justification of probabilistic statements that enter insurance and risk decisions. Formal probability theory supplies the relations among events and measures. Insurance practice must also determine whether the model describes policyholders, claims, hazards, individual exposure, annual aggregates, or the balance-sheet impact of a portfolio. Confusion between these objects leads to misleading comparisons even when every calculation is arithmetically correct.
The investigation proceeds by textual and conceptual analysis of identifiable foundational works, comparison of mathematical requirements, reconstruction of transparent hypothetical examples, and contextual assessment of professional and supervisory material. It distinguishes the original arguments of historical authors from contemporary reconstructions, publisher abstracts from fully inspectable texts, and binding European law from nonstatutory U.S. actuarial standards. Direct source links accompany the claims for which they are relevant; detailed identities and access limitations appear in References and the evidence ledger.
Five questions organize the analysis. What makes a probability assignment meaningful under each interpretation? How can repeated observations justify a prospective statement? What constrains a personal probability when unique or sparsely observed events are at issue? Which statistical diagnostics reveal the strengths and weaknesses of probability forecasts? Finally, which interpretive distinctions are necessary when those forecasts affect insurance pricing, capital, or claims-paying capacity? Mathematical examples are specified as hypothetical and carry no inference about the realized experience of a named insurance market.
Why Interpretation Matters for Insurance
Insurance operates by converting uncertain future states into present prices, contractual entitlements, reserves, and capital allocations. Quantification is useful because an insurer must aggregate contingent consequences across exposures and time. The conceptual problem precedes estimation: an actuary cannot evaluate an estimator until the target estimand is defined. The target might be the mean number of claims per policy-year, the probability of at least one claim, an individual's probability of loss conditional on characteristics, the annual occurrence rate of a hazard, a tail quantile of aggregate loss, or an underwriter's subjective prediction after reviewing sparse evidence.
The distinctions matter because familiar transformations require assumptions. If claim counts follow a homogeneous Poisson process with annual intensity λ, then the probability of at least one claim is 1 − exp(−λ). The annual count expectation λ and the event probability are closely related at low rates but are not identical. If the data instead arise from a mixture of Poisson intensities, applying the transformation to an average rate does not generally give the portfolio's average event probability. Similarly, a 99.5th-percentile loss measure is a statement about a modeled distribution, not a stand-alone physical law of solvency.
The risk triplet analysis of Kaplan and Garrick (1981) describes scenarios, likelihoods, and consequences as connected aspects of risk analysis. Its practical importance for this dossier is that a likelihood has meaning only in relation to a specified scenario, and a price or decision depends on consequences as well as that likelihood. Treating event probability as an exhaustive description of risk therefore discards essential information about severity, aggregation, uncertainty, and controllability.
An insurance probability statement commonly rests on several sources of judgment: realized portfolio experience, physical information about hazards, economic and behavioral assumptions, model design, and expert assessment of what will change. These components can be mutually informative. The editorial task is to identify their respective roles and resist transferring certainty from a transparent numerical answer to premises that were only partially observed.
Historical Probability Concepts
The historical development of probability brought together problems of equally possible cases, repeated trials, rational belief, mathematical measure, and decision. These strands intersected without collapsing into a single definition. The classical approach associated with games of chance employs symmetry among alternatives. Frequentist interpretations attach probability to collectives or long-run frequencies. Logical accounts concern evidential support. Personalist accounts associate probability with coherent betting commitments, preferences, or conditional judgments. Propensity approaches seek a physical basis for a chance attached to an experimental or causal arrangement.
John Maynard Keynes's A Treatise on Probability (1921) develops the relation between evidence and rational degrees of support. The mathematical comparability of probabilities is not assumed in every case. Ramsey's Truth and Probability, written in 1926 and published posthumously in 1931, interrogates both frequency and Keynesian logical approaches before investigating the logic of partial belief. His willingness to distinguish probability in logic from probability in physical science remains important: a disagreement about probability sometimes concerns different objects of discourse, not a disputed arithmetic identity.
Frank H. Knight's Risk, Uncertainty and Profit (1921) distinguishes classes of uncertainty that can be measured by grouping or recognized regularity from situations whose relevant uniqueness makes measurement difficult. This distinction is historically important for insurance, where the practical issue is often whether observed groups are sufficiently relevant and stable for prospective pooling. It should not be converted into a claim that nonrepeatable events can never receive disciplined probabilistic judgment.
Richard von Mises's Probability, Statistics and Truth articulates the frequency-based collective tradition. Andrey Kolmogorov's Foundations of the Theory of Probability provides the measure-theoretic foundation for modern probability mathematics. Kolmogorov's axioms, however, leave open what the measure represents in a particular domain. An internally valid measure can describe physical randomization, a survey respondent's degrees of belief, or an abstract integration problem. Substantive interpretation remains necessary.
Bruno de Finetti's 1937 work on foresight integrates personal probability with the mathematical consequences of exchangeability. Leonard Savage's The Foundations of Statistics (1954) pursues decision-theoretic representation of personal probability and utility under explicit postulates. Karl Popper's propensity interpretation (1959) offers a different answer: probability can characterize the physical tendencies of an experimental situation. These approaches supply complementary questions for insurance: what is repeated, what is believed, what is physically structured, and what can be rationally chosen?
Formal Probability and the Interpretation Problem
A probability space consists of a sample space Ω, an event collection F that is a sigma-algebra, and a measure P satisfying nonnegativity, normalization P(Ω) = 1, and countable additivity for disjoint events. From these properties follow bounds 0 ≤ P(A) ≤ 1, P(A complement) = 1 − P(A), and additivity over mutually exclusive events. Conditional probability is P(A|B) = P(A ∩ B)/P(B) when P(B) > 0. These mathematical relations are stable across several interpretations. They provide a rigorous language for expressing coherence, frequency models, and physical chances, while the interpretive question concerns what justifies P.
Countable additivity matters for limiting operations and conventional stochastic modeling. Some forms of personalist probability, including parts of de Finetti's framework, begin with finite additivity rather than accepting every measure-theoretic commitment as a rationality requirement. The existence of a mathematical convention therefore does not resolve the philosophical question of which assumptions are compulsory for every conceivable judgment. In ordinary actuarial models, countably additive probability is the prevailing working calculus; the distinction explains why formal compatibility alone cannot select a substantive interpretation.
Events also require a fixed description. Let A denote “at least one insured storm-loss claim during calendar year t under wording W in territory R.” Modifying the deductible, territory, insured population, event threshold, reporting delay, or year changes the event. If two analysts report different probabilities for the same verbal label but different definitions, their numerical disagreement may be entirely explained by design. Conversely, identical probabilities do not ensure identical risks when financial consequences differ.
Conditional probability offers an exact example of interpretation-independent mathematics and interpretation-dependent inputs. For a hazard H and claim C, P(C) = P(C|H)P(H) + P(C|not H)P(not H), provided H and its complement partition the possibilities. The identity follows from additivity; the components demand judgments about hazard frequency, coverage, vulnerability, secondary causes, and measurement. If one component is a fitted frequency and another comes from engineering judgment, the resulting number is neither purely empirical nor arbitrary. It is a conditional model whose status must be documented.
Classical Symmetry and Logical Probability
Classical probability assigns equal probability to alternatives judged equally possible. A fair six-sided die has P(even) = 3/6 given symmetry, mutually exclusive faces, and the construction or modeling assumption of a fair process. In insurance, symmetry is rarely available as a direct empirical premise. Policyholders, buildings, storms, illnesses, driving habits, and contract terms differ. Using equal weights because analysts have not yet distinguished alternatives introduces an epistemic assumption; equality of knowledge does not demonstrate physical equality of chance.
The well-known ambiguity of the principle of indifference can be made concrete without relying on a puzzle about geometric random chords. Suppose an analyst defines two exhaustive events for a new hazard: any material loss and no material loss. Assigning 1/2 to each by indifference would imply a 50% loss probability. Repartition material loss into minor and major loss produces three verbal categories. Assigning 1/3 to each now implies a 2/3 probability of some loss. The discrepancy appears because changing the descriptive partition changed the allocation rule while leaving the underlying insurance event conceptually unchanged. An evidentially defensible estimate therefore needs more than the mere count of labels.
Logical or evidential probability asks how strongly specified evidence supports a proposition. Keynes's historical argument matters because evidence may support comparative rankings without determining precise numerical degrees. A claims analyst might reasonably conclude that one portfolio is more exposed than another while lacking grounds to attach exact annual probabilities. Epistemic precision should reflect information and methodology, not just spreadsheet formatting.
There is no general conflict between using symmetry and requiring empirical review. Symmetry may be a justified feature of an experimental mechanism or a deliberately neutral model under stated restrictions. Once differentiated data become available, an apparently symmetric prior or reference class can be revised. The burden of proof belongs to the asserted symmetry and its relevance to the insured event.
Frequency and Repetition
The frequentist family interprets probability through regularity in repeated trials, particularly limiting relative frequencies. Given indicators X₁,…,Xₙ for an event, the empirical frequency is fₙ = (X₁ + ··· + Xₙ)/n. If trials are independent and identically distributed Bernoulli variables with event probability p, then E(fₙ) = p and Var(fₙ) = p(1 − p)/n. Chebyshev's inequality yields P(|fₙ − p| ≥ ε) ≤ p(1 − p)/(nε²). The mathematical convergence follows from a model, and the observed sequence furnishes information about p when its conditions are credible. The theorem does not prove from observations alone that the physical or economic process will remain unchanged.
Two important ambiguities require attention. A finite observed frequency is not necessarily the limiting probability, and the existence of a stable long-run limit is not established by seeing a short stable run. A population may also be nonstationary: underwriting changes, legal reforms, climate patterns, medical technologies, and reporting incentives can shift the distribution. A mathematically convergent estimator of an obsolete parameter may be a poor estimator of future claims. Thus, the frequency interpretation must address the definition of the repeatable collective and the continuity between its observed and future members.
Suppose 12 claims occur across 6,000 exposure-years. The observed claim-count rate is 12/6,000 = 0.002 per exposure-year. If counts are modeled by independent homogeneous Poisson processes with mean 0.002 per exposure-year, the probability of at least one claim is 1 − exp(−0.002) ≈ 0.0019980013. The rate and occurrence probability are nearly equal because λ is small, yet their units and event meanings differ. Replacing the Poisson assumption with clustered claims, heterogeneous exposures, or multiple-claim dependence alters the transformation.
A repeated event need not be identical in everyday language to count statistically. Insurance uses structured comparability, class adjustment, covariates, and exposures to construct populations suitable for inference. The philosophical problem persists in operational form: how much similarity is enough to make one loss informative about another? Sampling design, underwriting classes, claim definitions, and temporal drift are part of the answer. None can be supplied by a frequency ratio alone.
Sampling Uncertainty, Stability, and Predictive Transfer
Frequency-based statements require an account of sampling error and predictive transfer. If a sample of 100 independent identically distributed exposures contains 2 claims, the observed proportion is 0.02. Its binomial standard error estimated by the plug-in formula is sqrt(0.02 × 0.98/100) ≈ 0.014. This rough calculation shows that treating 0.020 as an exact immutable physical parameter would overstate the support of the data. For very small counts, a plain symmetric normal interval is generally inadequate; exact or more appropriate binomial procedures, Bayesian predictive analysis, or hierarchical pooling may be preferable according to the question and assumptions.
Statistical uncertainty has several locations. Counting variation describes fluctuations conditional on a sampling model. Parameter uncertainty arises because a parameter must be estimated. Structural uncertainty arises because the selected model family may fail. Population uncertainty concerns whether the measured exposures represent those the insurer will actually cover. Reporting uncertainty can emerge from incomplete development, late notifications, fraud screening, or data recording changes. These mechanisms can coexist and should not be concealed in one confidence label.
A practical insurance forecast is a mapping from past observations to future conditions. Let D be historical data and M a model. A predictive distribution can be written P(Y_future|D,M). A more complete exercise considers the possibility that model assumptions are uncertain, represented by alternative models or scenarios with justified weights. Mathematical notation makes explicit a conditioning step; it cannot independently validate the transfer from D to future conditions. Historical homogeneity, consistent exposure units, and stability of causal mechanisms require their own review.
An important institutional distinction follows. A model may achieve excellent retrospective fit by reproducing known clusters while being unable to generalize across new territories or weather regimes. Conversely, a physically informed model may outperform simple historical frequency when severe events are rare, but it depends on the credibility of hazard, vulnerability, and financial components. The choice is contextual and evidential. Neither high apparent sophistication nor long observed history confers unconditional authority.
Physical Chance and Propensity
A propensity interpretation associates probability with a physical setup or disposition toward outcomes. Popper's 1959 propensity article belongs to a broader attempt to account for probabilities of individual outcomes within experimental circumstances. The appeal for insurance is evident: a particular building may face physical hazard processes even before any claim has occurred. Engineers can model wind load, flood water depth, material fragility, or other mechanisms to estimate event and damage chances.
Nevertheless, the translation from physical conditions to covered claims is mediated by human and institutional variables. The same flood depth can produce different claim outcomes under different deductibles, insured values, exclusions, mitigation measures, notification practices, and settlement rules. A physical probability of a hazard cannot automatically serve as the probability of an insurance payment. The conditional chain must be represented: hazard occurrence, exposure, vulnerability, contract response, and claims development.
Physical accounts also face questions of specification. Is a propensity a property of the storm system, the geographic site, the insured asset, or the entire coupled configuration? A physical model can be correctly calibrated to a defined observable while misapplied to another. The correct level of analysis follows from the event under consideration, not from a general commitment to objectivity.
The adjective “objective” can reasonably describe a physical mechanism, an agreed reproducible procedure, or a public statistical record; these meanings are related but distinct. Public observability makes a dataset inspectable; it does not make every probability derived from it a unique feature of the world. Even estimates of physical chances depend on measurement conventions, observation models, uncertainty about parameters, and decisions about what environmental changes to include. Rigor consists in identifying these assumptions and examining their consequences.
Single-Case Events and Causal Conditions
Insurance routinely covers particular events: this shipment may be lost, this patient may require treatment, this property may burn, and this regional catastrophe may occur during a specified year. A strict repeated-event interpretation must relate a single event to a reference class. The class can be operationally useful, but there is no unique class given by the identity of the object. A building is simultaneously part of a regional group, a construction-type class, a flood-zone class, and an underwriting segment. Each membership can suggest a different estimate.
A propensity approach may focus on the physical circumstances of the individual case, and a personalist approach can combine reference-class data, local evidence, and competing models into a conditional credence. The difference is not whether the single case has consequences; the insured event will either occur or not occur. The difference lies in what a probability assertion claims before the event is resolved. Probability 0.1 does not mean the realized event will happen one-tenth of the way; it describes a chance or uncertainty assignment whose operational merit emerges across coherent relations, evidence, or comparable forecasts.
A useful counterexample concerns an unprecedented statutory change. A liability insurer might face an uncertain court interpretation with no genuinely repeated prior cases under the new rule. Assigning an exact objective frequency from an unrelated jurisdiction would hide a transfer assumption. A structured personal assessment can still inform reserves and scenario preparation, provided the evidential status and sensitivity are displayed. A range of plausible outcomes and their consequences may be more informative than an unsupported unique decimal.
Causation should be represented carefully. An insured loss may depend jointly on a hazard, protective conditions, human conduct, and coverage. Probability conditional on an intervention can differ from probability conditional on merely observing a correlated feature. A model trained on correlations between building attributes and past claims does not establish the effect of changing those attributes. Causal interpretation matters when a risk probability is used to justify prevention, underwriting incentives, or public mitigation measures.
Personal Probability, Preference, and Coherence
Personal probability represents a degree of confidence about an event held by an evaluator whose judgments satisfy appropriate consistency requirements. Ramsey's logic of partial belief connects rational preferences, beliefs, and action. Savage's 1954 decision-theoretic treatment investigates how personal probabilities and utility can be represented under substantive preference postulates. De Finetti's 1937 subjective foundation links probability to fair or coherent betting commitments, and clarifies the importance of exchangeability for learning from apparently similar cases.
The coherence argument is usually illustrated by the impossibility of accepting a combination of stakes that produces a guaranteed loss under mutually compatible prices. If a decision-maker offers to buy or sell a unit payoff on A at price p, an incoherent collection of prices for A, its complement, and disjoint decompositions creates opportunities for a sure-loss combination under suitable betting assumptions. Probability axioms describe the resulting consistency conditions. The betting interpretation is a conceptual device; real insurance transactions have transaction costs, limited liability, liquidity constraints, market power, and differing risk aversion. Market premiums therefore do not directly reveal unique personal probabilities.
Coherence is a requirement on the internal relations among judgments, not an empirical guarantee that a judgment is well informed. A forecaster may give a coherent probability of 0.8 to a rare claim on the basis of false data. The assignment can obey probability laws and perform very badly under external verification. Likewise, a model may be empirically well calibrated in aggregate while its stated conditional probabilities omit important subgroup information. These diagnoses address different aspects of warrant.
Personalist probability is especially relevant to decisions concerning sparse catastrophic events, emergent technologies, unique legal outcomes, or rapid regime changes. In these settings experts inevitably decide which data and analogies matter. Their judgments can be documented, elicited, challenged, and updated, but no general theorem turns those activities into an empirically measured physical frequency. Good practice makes the judgment legible to other analysts and distinguishes informed uncertainty from impressionistic confidence.
Exchangeability and Predictive Induction
Exchangeability provides a precise bridge between the appearance of repeated observations and a personal predictive model. A sequence of binary variables is exchangeable when its joint probability remains unchanged under finite permutations. For example, the probability assigned to the sequence claim–no claim–claim is the same as that for no claim–claim–claim, assuming the same sequence length and count. Exchangeability does not require unconditional independence: observations can be positively dependent through uncertainty about a shared latent claim probability.
De Finetti's original exchangeability investigation underlies the representation in which an infinitely exchangeable Bernoulli sequence has joint distribution expressible as a mixture of independent Bernoulli models. In modern notation, for outcomes x₁,…,xₙ and a mixing distribution F on [0,1], P(x₁,…,xₙ) = ∫ θ^s(1 − θ)^(n−s)dF(θ), where s is the number of ones. The representation explains why a latent parameter can serve as a mathematical device within a personalist predictive account. It does not show that every exchangeable portfolio has an observed physical constant θ, nor that arbitrarily selected risks are exchangeable.
Insurance populations often violate naive exchangeability. Exposures differ in occupation, location, age, property value, hazard mitigation, and coverage. Conditional exchangeability within justified risk groups may be useful; arbitrary pooling can induce misleading shrinkage. Hierarchical models can accommodate variation among groups, but the distribution chosen for that variation remains an assumption to evaluate.
As a simple illustrative model, put a Beta(2,8) distribution on a Bernoulli probability θ for a specified risk class. Its prior predictive loss probability is 2/(2+8) = 0.2. With 3 losses in 10 conditionally independent, equally defined exposures, the posterior distribution is Beta(5,15) and the next-trial predictive probability is 5/20 = 0.25. The raw observed proportion is 0.3. The change from 0.2 to 0.25 is an exact consequence of the stated prior, likelihood, and sample; it is not an objective actuarial result independent of those conditions. Another defensible prior can give a different posterior, especially with only 10 observations.
This example shows how experience and prior judgment combine in a transparent calculation. It also illustrates a limitation: a stable model may borrow strength across scarce data, but borrowing strength from an inapplicable reference class creates systematic error. The appropriate evidential question is what validates class comparability and the prior concentration, not whether the posterior fraction looks mathematically respectable.
Bayesian Updating and Prior Sensitivity
Bayes' rule states P(H|D) = P(D|H)P(H)/P(D) for events with positive conditioning probability. In statistical inference, a prior over parameters or hypotheses is updated by a likelihood model to produce a posterior. Frequentist procedures may also use Bayes' identity as a mathematical relation; the philosophical distinction concerns whether uncertainty about hypotheses is represented by probability distributions and how those distributions are warranted.
The Beta–Bernoulli example can be generalized. Under a Beta(a,b) prior and s losses in n conditionally independent exposures, the posterior is Beta(a+s,b+n−s) and the next-observation predictive probability is (a+s)/(a+b+n). The formula reveals how the prior equivalent sample size a+b shapes the prediction. With a weak prior the observed data receive more influence; with a highly concentrated prior the forecast remains close to its original value. Reporting only the posterior mean conceals this prior-to-data weighting and can give an exaggerated impression of data determination.
An actuarial analyst can compare several candidate priors, likelihood structures, and reference classes; the sensitivity of the decision then becomes part of the scientific result. If pricing or reinsurance purchase depends strongly on the prior at plausible values, a single point estimate is an inadequate summary. The relevant output can include distributions of aggregate loss, scenario comparisons, and material thresholds at which the preferred action changes. These are standard analytical tools rather than an original doctrine.
Bayesian updating also requires an accurate information date. A claim received late, a revised catastrophe catalog, and an additional hazard observation all change the conditioning information. Reusing a posterior without recording the evidence included risks double counting, unacknowledged omission, or retrospective contamination. The Bayesian formalism is exact conditional on its inputs; institutional controls make those inputs reproducible.
Probability, Decision, and Expected Loss
A decision requires payoffs and constraints in addition to probabilities. The expected indemnity for a contract with nonnegative payment X is E(X), calculated under a specified loss model. A premium can incorporate expected claims, expenses, risk margins, capital costs, profit objectives, taxes, and commercial constraints. None of these additions is itself an event probability. The relationship between a forecast and a financial action must therefore specify whether the probability refers to a trigger, claim occurrence, aggregate loss, or contractual benefit.
Suppose two hypothetical insured events have identical annual occurrence probability 0.02. One produces a fixed payment of 1 million. Their annual expected payments are respectively 20,000 under the stated models. If the second payment is affected by a catastrophe that simultaneously triggers many contracts, its portfolio tail implications can be much greater still. Numerical equality of occurrence probabilities supplies no equivalence of financial risk.
Expected-value comparisons are often insufficient for decisions involving ruin, tail exposure, limited liquidity, or risk aversion. Knight's discussion of uncertainty and enterprise provides a foundational reason to examine both distributional properties and the limits of enumeration. Modern regulatory and actuarial practice likewise uses distributions, assumptions, capital metrics, and constraints. However, one must distinguish the meaning of a probability from the normative objective of a decision. Two coherent decision-makers can agree on the probability distribution and rationally choose different contracts because their assets, obligations, utility functions, or risk tolerances differ.
In a state-contingent decision, define a finite set of outcomes j with probabilities p_j and financial consequences L_j. Expected loss is Σp_j L_j. A tail quantile VaR_α is an α-quantile of the loss distribution under an explicitly stated quantile convention. A tail mean or expected shortfall is a different statistic, requiring care when the distribution has discrete mass at the threshold. A model that correctly estimates expected loss can still understate a high quantile if joint dependence or tail severity is misspecified. Interpretation therefore requires attention to the functional of the distribution being communicated.
The authors of the quantitative definition of risk stress the combination of scenarios, likelihoods, and consequences. Their framing supports a practical discipline: identify the scenario first; state whether its likelihood comes from historical frequencies, physical mechanisms, structured belief, or an ensemble of models; then show the consequences and the uncertainty surrounding both. A decision table without this linkage invites a misleading perception of completeness.
Insurance Probability Statements and Exposure Design
The first test of an insurance probability is semantic. Every probability assertion should identify the event, exposure denominator, insured population, contract conditions, observation or forecast period, and information set. A statement such as “there is a 1% chance of a loss” is incomplete when it fails to specify whether loss means any physical damage, reported claim, paid claim, total annual cost above a threshold, or insurer insolvency. These are not interchangeable outcomes.
For claim frequency, the analyst must distinguish exposures, policies, covered persons, policy-years, insured values, and hazard opportunities. A ratio of claims to policies is not automatically a per-year probability where periods differ, where repeat claims are permitted, or where entrants are selected over time. Inconsistent denominators can create an illusion of changing risk even when underlying conditions are stable. A change in reporting rules can change recorded frequency without changing physical events.
Contractual interpretation has similar effects. Raising a deductible can lower the probability of an insurer payment while leaving the physical probability of damage unchanged. Narrowing covered perils can reduce paid claims without affecting the underlying hazard. Increased reporting activity can raise observed claims while physical risk is constant. A meaningful trend analysis compares consistent event definitions or explicitly models definitional change.
The ASOP No. 23 data-quality guidance emphasizes the selection, review, use, and appropriate disclosure of data in actuarial services. Its relevance is methodological: documentary integrity is part of the warrant for probability estimates. The standard applies within its U.S. professional context and does not purport to define probability philosophically or replace international jurisdiction-specific requirements. It nevertheless provides a concrete example of how the evidential premises of actuarial probabilities become an explicit professional responsibility.
A sufficiently complete probability statement can be written in a form such as: “Under model M, given information through date d and the present portfolio definition, the probability of at least one covered claim during the next twelve months for risk class C is p.” The statement permits examination of M, d, C, event wording, and the forecast horizon. It does not guarantee that the estimate is correct; it specifies what would have to be investigated to judge it.
Heterogeneity, Dependence, and Portfolio Risk
Pooling does not make insured exposures identical. Let X_i denote the payment on policy i, with mean μ_i and variance σ_i². Aggregate payment S = ΣX_i has E(S) = Σμ_i, regardless of independence, while Var(S) = ΣVar(X_i) + 2Σ_{i<j}Cov(X_i,X_j). Thus, expected aggregate payment and its uncertainty respond differently to dependence. A model can estimate the marginal event probabilities accurately and still severely misrepresent the distribution of aggregate losses if it neglects a common catastrophe trigger.
Consider two identical binary loss indicators with probability p = 0.1 and payment $100 per loss. Under independence, the expected number of payments is 0.2 and the variance of the number of payments is 2p(1−p) = 0.18; the probability of two payments is 0.01. Under perfect dependence, both policies pay together with probability 0.1; the expected number is still 0.2, but its variance is 4p(1−p) = 0.36, and the probability of two payments is 0.1. Dependence changes the chance of large simultaneous loss by an order of magnitude while leaving expected payment unchanged. The numerical contrast is exact for the stated hypothetical constructions.
Heterogeneous claim frequencies also matter when translating rates into probabilities. Suppose half the exposures have Poisson annual intensity zero and half have intensity 0.004. The average intensity is 0.002. The average probability of at least one claim is (1/2)(1−e^0) + (1/2)(1−e^(−0.004)) ≈ 0.0019960053. A homogeneous model using the average intensity gives 1−e^(−0.002) ≈ 0.0019980013. Although the difference is small at these rates, its sign reflects the concavity of 1−exp(−λ). Larger dispersion can make the practical gap material. This example tests the assumption of a homogeneous rate, not the philosophical supremacy of any interpretation.
Exposure heterogeneity creates an additional inferential problem: the class definition is itself a research decision. Risk classification must be useful, relevant to the insured outcome, and examined within applicable legal and ethical requirements. A subgroup's observed frequency might be unstable because its sample is small; predictive borrowing can improve variance while introducing bias when groups differ. The analyst must compare the cost of instability against the cost of inappropriate pooling.
A frequentist model can represent heterogeneous populations through regression, random effects, mixture distributions, or stratified estimators. Personal probabilistic judgment may influence variable choice, prior distribution, model selection, and treatment of data gaps. Physical hazard modeling can supply covariates or component probabilities. These methods may coexist within one actuarial model. The objectivity of a result must be assessed through evidence, assumptions, and reproducibility rather than by attaching a single philosophical label to a hybrid method.
Catastrophe Probabilities and Model Ensembles
Catastrophic events strain straightforward frequency inference because extreme outcomes are sparsely observed, spatially dependent, and sensitive to the definition of a hazard or insured loss. Historical catalog length is not equivalent to a number of independent homogeneous trials. Coastal development, building standards, climate trends, inflation, policy wording, and claims practices can change observed losses even if aspects of the physical hazard remain stable. Conversely, physical processes themselves may be changing. Analysts must identify which components of an event probability are empirically estimated, physically simulated, or judgmentally specified.
Catastrophe models commonly separate a stochastic event catalog, hazard intensity, vulnerability, exposure, and financial terms. The annual probability that a specified portfolio loss exceeds threshold x belongs to the model's full loss-generation chain; it cannot be inferred from a hazard return period alone. An “annual 1-in-100 event” often denotes a modeled annual exceedance probability near 1%, yet the phrasing should not imply that an event occurs at a regular 100-year interval. Under an independent annual Bernoulli model with p = 0.01, the probability of at least one event over 30 years is 1 − (0.99)^30 ≈ 0.2603, not 30% and not certainty after 100 years. Dependence or changing annual probabilities alters the calculation.
The adopted ASOP No. 38 catastrophe-modeling standard identifies professional considerations for actuaries selecting, using, reviewing, or evaluating catastrophe models. It also clarifies the relationship to general modeling guidance. Such a standard can improve how uncertainty and reliance on model providers are documented; it does not establish that a proprietary model's event probabilities are physically verified by virtue of professional use. Inspection of model scope, input sensitivity, geographic relevance, and limitations remains material.
Suppose model A assigns a 1% annual probability to a defined loss threshold and model B assigns 3%. Assigning weights 0.6 and 0.4 yields a mixture predictive probability 0.6(0.01) + 0.4(0.03) = 0.018. That arithmetic is valid if weights are defensible probabilities for the competing models or a declared scenario-weighting convention. It does not demonstrate that 1.8% is an objective physical frequency. An analyst could equally present the two results separately, particularly when model reliability cannot be quantified. Model ensembles express disagreement; they do not eliminate it by averaging.
One consequence is interpretive discipline around tails. If decision-makers are concerned with capital adequacy under systemic exposure, a single annual catastrophe occurrence probability is only an input. Loss amplification, reinsurance recoveries, reinstatement provisions, multiple-event interaction, liquidity, and changing exposure can dominate the eventual financial outcome. A probability estimate becomes useful when its placement in that wider model is explicit.
Calibration, Resolution, and Proper Scoring
A forecast can be checked against outcomes through repeated prospective use. Glenn Brier's 1950 verification article provides the historical basis for squared-error scoring of probability forecasts. For a binary outcome y ∈ {0,1} and forecast p ∈ [0,1], the binary Brier loss is (p − y)². If the true conditional occurrence probability is q, then E[(p−Y)²] = (p−q)² + q(1−q). The expectation is minimized at p = q. The calculation demonstrates propriety under the specified Bernoulli data-generating condition.
For example, a forecast p = 0.2 incurs loss 0.64 when the event occurs and loss 0.04 when it does not. If its true conditional event probability is q = 0.3, expected Brier loss is 0.3(0.64) + 0.7(0.04) = 0.22. Forecasting p = 0.3 instead would have expected loss 0.21. The difference 0.01 equals (p−q)². This is a hypothetical verification calculation, not an assessment of an observed insurer or model.
Adrian P. Dawid's well-calibrated Bayesian analysis (1982) examines the relationship between coherent personal predictions and empirical calibration. Calibration asks, roughly, whether events assigned a probability near p occur at corresponding frequencies in a suitable series. The qualifier “suitable” is indispensable. A portfolio with rare high-impact losses may be impossible to verify tightly over a short period. Grouping predictions into broad bins can conceal systematic conditional errors, and changing exposure populations can invalidate retrospective comparisons.
The Gneiting and Raftery scoring-rule study (2007) clarifies proper and strictly proper probabilistic scores. Proper scoring supplies an incentive-compatible criterion under stated assumptions; it does not by itself select the right event or guarantee stability under a changing world. Calibration, discrimination or resolution, and sharpness each illuminate a different property of probabilistic forecasts. A forecast that always reports the overall base rate may be well calibrated in aggregate while providing little ability to distinguish relatively high- and low-risk exposures. A discriminating forecast may require recalibration before it supports prices or regulatory decisions.
Verification also faces a selection problem. If an insurer observes only accepted policies, it cannot directly infer the outcomes that would have occurred for applicants it rejected. Claims data are additionally censored by coverage, retention, notice, and settlement practices. A prospective scoring study must define which predictions were made before outcomes, which observations are included, and what happens to missing or delayed responses. Without that design, an apparently good score can reflect selective measurement.
Behavioral Judgments and Institutional Probability
Human probability judgments are influenced by representation, remembered examples, and initial reference values. Tversky and Kahneman's 1974 heuristics study identifies representativeness, availability, and anchoring mechanisms in judgments under uncertainty. Their empirical and conceptual findings matter to insurance because rare but memorable disasters can dominate perceived risk, while familiar recurrent losses may be insufficiently salient. The studies do not show that every expert judgment is biased in the same way or that personalist probability necessarily equals a psychological intuition.
Subjective probability as a formal account of coherent degrees of belief must be distinguished from an unexamined verbal feeling. Elicitation procedures can ask experts to specify events, disclose evidence, compare internally related probabilities, consider counterexamples, and revise after diagnostic feedback. Expert panels may improve coverage of uncertainty but can introduce shared assumptions and correlated errors. Averaging their numbers without inspecting independence, incentives, and expertise is a procedure, not a justification.
Institutional probabilities emerge when individual judgments enter underwriting rules, actuarial committees, vendor models, management overlays, and regulatory reporting. Decision-makers may describe a forecast as “model-based” even though training data, scenario choices, feature selection, and overrides contain human decisions. These contributions are legitimate subjects of review. Methodological transparency makes such forecasts open to correction, replication, and informed disagreement.
Insurance communication raises a further distinction between probability and concern. A person may assign a low annual probability to a devastating event and still purchase coverage because the potential loss is financially intolerable. Another may decline coverage at the same probability because the premium, assets, exclusions, or alternative protections differ. The stated probability does not determine the economic decision without its consequence structure. Confusing risk perception with event frequency can misdiagnose rational behavior as statistical error.
Actuarial Standards and Professional Documentation
Actuarial practice translates uncertainty statements into estimates that other institutions may rely upon. The Actuarial Standards Board's adopted ASOP No. 25 Credibility Procedures discusses the use of relevant subject experience and other information in estimating expected values. Credibility methods make explicit a recurrent interpretive issue: how much evidential weight should be placed on a small local experience sample relative to a broader reference class? Some procedures rely on assumptions about distributions, some on empirical relationships, and some on expert judgment. The appropriateness of a procedure depends on its purpose and underlying conditions.
The adopted ASOP No. 56 Modeling addresses responsibilities associated with selecting, developing, using, reviewing, and evaluating actuarial models. A model can contain stochastic and deterministic components, and output variability may reflect observed randomness, parameter estimation, data limitations, or structural choices. Professional review should examine the intended use, output interpretation, reliance on other work, sensitivity, and material limitations. The existence of a modeling standard is evidence of recognized professional discipline, not independent proof of an insurer-specific estimate.
The data-quality standard, ASOP No. 23, connects probability estimation to selection, review, and use of data. Its requirements should be considered within the professional and legal scope it actually governs. Insurance markets in other jurisdictions use their own actuarial, accounting, supervisory, and legal rules; this monograph neither claims universal enforceability of U.S. standards nor gives jurisdiction-specific compliance advice. The standards are used here as documentary examples of how statistical assumptions acquire institutional accountability.
Four kinds of disclosure are especially informative for a research dossier. First, state what random variable and exposure system is being modeled. Second, identify the origin and limitations of data and the model's role in turning it into future predictions. Third, document judgment, alternative hypotheses, and sensitivity where they materially affect results. Fourth, connect the probability estimate to the financial decision and distinguish mean, frequency, quantile, and tail outcomes. These are synthesized analytical reporting concerns, not a proposed new actuarial code.
Regulatory Probability: Solvency Capital Calibration
Capital regulation provides a clear illustration of a probability embedded in an institutional framework. The Solvency II Directive, Article 101, describes the Solvency Capital Requirement as corresponding to Value-at-Risk of basic own funds at a 99.5% confidence level over one year, within the Directive's specified risk coverage and assumptions. The legal calibration should be read in its applicable consolidated form and in conjunction with implementation rules and supervisory practice. Its 99.5% value is a prescribed regulatory model-calibration standard; it is not a direct empirical observation that every regulated insurer has a 0.5% annual failure probability.
The difference is substantive. A model must specify how changes in own funds arise from underwriting, market, counterparty, operational, and other covered risks. Estimation and dependence assumptions, valuation rules, management actions, and risk mitigation affect the calculated quantile. Two insurers with different portfolios can both satisfy a common calibration rule and face different practical vulnerabilities. Even a perfectly computed quantile of a misspecified model is conditional on the model rather than a guarantee about the world.
A quantile also does not summarize losses beyond it. For a hypothetical annual loss L with a 99.5th-percentile threshold of $100 million, the probability mass at and above the threshold, the distribution of losses beyond it, and the resulting financial consequences require further specification. A tail can contain outcomes vastly larger than the threshold. Reinsurance contracts and collateral arrangements may transform the insurer's net distribution. Consequently, a statement about required capital should identify the measured balance-sheet variable and risk horizon.
This regulatory example underlines why the adjective “objective” has to be used precisely. A rule can be publicly specified and mechanically applied, producing a reproducible calculation. The underlying modeled distribution may still involve expert judgment, limited observations, and uncertainty about extreme events. Legal standardization creates comparability of procedure under its scope; it does not remove interpretive uncertainty about all estimated probabilities.
Worked Comparative Examples and Counterexamples
The first example concerns an observed loss proportion. Ten exposures with three losses produce frequency 0.3. The statistic alone does not state a physical constant or a next-year forecast. Under independent identical Bernoulli trials with a Beta(2,8) prior, the posterior predictive probability becomes 0.25. Under a simple frequency plug-in forecast it remains 0.3, with substantial sampling uncertainty. If the exposures are heterogeneous or selected, neither transfer is automatically valid. The competing numbers document distinct assumptions rather than a contradiction in arithmetic.
The second example concerns probability of at least one occurrence. An annual homogeneous Poisson intensity of 0.002 implies approximately 0.001998 probability of one or more events. Averaging two populations with intensities zero and 0.004 gives the same mean intensity but a lower average occurrence probability, approximately 0.001996. The change derives from heterogeneity and a nonlinear transformation. Reporting an intensity as a Bernoulli probability would erase the event definition and a modeling assumption.
The third example concerns dependence. Two binary policies each have 10% claim probability and 20 both when independent and when perfectly dependent. The probability of a $200 aggregate payment is 1% under independence and 10% under perfect dependence. A pricing calculation based solely on expected payment could miss a material capital problem, even though every marginal probability is correctly specified.
The fourth example concerns temporal recurrence. An annual probability of 1% does not imply an event every century. Under independent yearly occurrences, the chance of at least one occurrence over 30 years is about 26%. If annual hazards rise through the period, the appropriate no-occurrence probability becomes the product of the relevant (1 − p_t) terms only under conditional independence or a suitably specified process. A single observed return period should therefore never be presented as a schedule.
The fifth example concerns forecast verification. A p = 0.2 prediction on a Bernoulli event whose true probability is q = 0.3 has expected Brier loss 0.22, compared with 0.21 for a prediction of 0.3. The difference is due to predictive inaccuracy as measured by a proper scoring rule. One realized event could make either individual score appear good or bad; a collection of prospectively recorded comparable predictions is needed for empirical evaluation.
The sixth example concerns decision sensitivity. Suppose a 10,000 to $30,000 before other pricing components. The range is financially significant even without declaring one of the numbers the correct physical chance. A decision can be robust across that range or highly sensitive to it; this is an empirical and institutional question to investigate.
The seventh example concerns event wording. Suppose a hazard has a 5% annual probability under a physical model, and only 40% of hazard events produce an insured claim under a stated coverage-and-vulnerability model, while no claims arise without the hazard. Then P(claim) = 0.05 × 0.40 = 0.02. The 5% hazard probability and the 2% claim probability are both valid for their different events. Neither value may be substituted for the other in capital, premium, or consumer communications without changing the subject.
Finally, consider a threshold catastrophe model assigning probabilities of 1% and 3% under alternatives A and B. A 60/40 weighted mixture produces 1.8%, but this is meaningful as a probabilistic model mixture only when the weights are justified. When the weights are merely a sensitivity convention, reporting both models and their decision consequences is more transparent. The numerical result demonstrates arithmetic consistency; evidence for the weights remains a separate substantive matter.
Epistemic Limits and Critical Synthesis
The preceding results support an integrated interpretation of probability use in insurance. Every method must confront a bridge from model or experience to the particular insured future. Classical symmetry requires justified exchangeability among alternatives. Frequency-based prediction requires relevant repeated exposure and a credible transfer across time and population. Physical propensity requires a specified mechanism and a link between hazard and insured consequence. Personal probability requires internal coherence, identifiable information, and accountability for assumptions and updating. Logical approaches ask whether evidence supports the asserted precision. None of these requirements is met merely by writing a number between zero and one.
Several conclusions follow from the comparative research. First, claims about the “objective probability of an insurance loss” often contain mixed elements: empirically observed rates, physical hazard conditions, selection and classification, contractual interpretation, and modeling judgment. Calling the output objective without specifying these elements may mistake reproducibility for ontological independence. Second, personalist probability is fully capable of incorporating external frequencies and scientific mechanisms; the label subjective indicates the location and rational constraints of the probability judgment, not an exemption from evidence.
Third, coherence and calibration test different things. A coherent set of probabilities may be badly calibrated to eventual outcomes, while an aggregate-calibrated model may obscure subgroup errors or exploit a selective dataset. Both testing approaches are useful and neither provides comprehensive proof of prospective accuracy. Dawid's calibration argument and the proper scoring analysis make this distinction especially important in forecast evaluation.
Fourth, probability of occurrence, expected claim frequency, expected payment, and loss quantiles are separate outputs. Dependence can leave expected total payment unchanged while multiplying the chance of simultaneous large claims. A probability statement therefore becomes financially consequential through an explicit mapping to severities, aggregation, insurance terms, and capital metrics. Finally, large numerical precision can coexist with weak epistemic support. Appropriate presentation may involve sensitivity analyses, model ranges, and clear descriptions of conditioning assumptions rather than a single unqualified estimate.
These findings synthesize established probabilistic, actuarial, and risk-science arguments. They do not constitute the declaration of a new formal interpretation of probability, a new insurer risk model, or an independently discovered theorem. The original authorial contribution here is the cross-domain reconciliation of conceptual interpretations with clearly traced insurance statements, counterexamples, and tests of institutional meaning.
Research Limitations and Verification Boundaries
This study is a comprehensive analytical monograph within a defined documentary selection, rather than a registered systematic review or a meta-analysis. It reviews original bibliographic records, accessible historic text, scholarly publisher descriptions and abstracts, an encyclopedic philosophical synthesis, selected adopted U.S. actuarial standards, and an official European legislative text. Some classic books are represented by publisher or catalog records rather than a page-by-page verification of every edition. The original Ramsey essay was available in a hosted scholarly transcription; the underlying print edition and all appended papers were not independently collated. Subscription-only scholarly articles were treated at the level of verifiable bibliographic metadata and available abstracts unless otherwise stated in the evidence ledger.
No proprietary catastrophe-model engine, policyholder microdata, claims database, examiner file, or insurer balance sheet was inspected. The mathematical examples are fully hypothetical, and no outside empirical estimate is derived from them. The analysis does not assert a particular legal duty beyond the described source scope, prescribe pricing for an actual product, or substitute for an actuarial opinion. The Solvency II provision is used to illustrate interpretive issues in a regulatory confidence level; the complete sequence of delegated regulations, national applications, and supervisory judgments lies outside this dossier.
Probability interpretations have continuing philosophical controversies, including the scope of countable additivity, reference classes, evidence dependence, ontology of chance, and rational constraints on priors. The monograph identifies their practical implications and evaluates representative positions without claiming to settle them in general. Important future studies could test model calibration in named insurance datasets, distinguish hazard and claims distributions across jurisdictions, or investigate expert elicitation and tail-weight sensitivity in documented insurance decisions. These research avenues are proposals for empirical work, not unanswered prerequisites to the present conceptual findings.
The source-to-claim and bibliography controls in the companion EVIDENCE.json identify exactly what has been checked. A hyperlink to a publisher landing page proves bibliographic identity and a visible abstract only to the extent disclosed in that source record; it is not treated as evidence of a full-text examination. Formal derivations and toy numerical calculations are independently reproducible from the stated assumptions. Internal scholarly and editorial approval concerns this completed manuscript and its declared scope, not an external peer review or public release.
Conclusion
Probability interpretation determines the meaning of an insurance probability statement and the grounds on which that statement can guide a decision. The mathematical laws of probability provide necessary consistency, while empirical frequency, physical conditions, rational judgment, and evidential support give different accounts of what a measure represents. The practical task is to identify their roles in each asserted risk probability and to justify the transfer from observations and models to future insured events.
The comparative investigation establishes that observed frequency is informative under relevant and stable exposure definitions; propensity modeling requires a defensible physical-to-contractual chain; personal probability requires coherent and accountable judgments; and logical support requires that the degree of stated precision match available evidence. Exchangeability can connect repeated observations to personal prediction when its assumptions are substantively warranted. Bayesian updating displays how prior judgments and observations combine, but it does not make prior or model choices vanish. Proper scoring and calibration offer empirical tests of forecasts, each with limitations in rare-event and changing-population settings.
Insurance adds financial and institutional demands. A probability of claim occurrence differs from a claim-count intensity, an expected payment, and a tail quantile. Dependence and severity can transform portfolio risk without altering marginal event chances. Professional actuarial guidance links probability work to data quality, credibility, and model documentation. The Solvency II example shows how a formally specified confidence level can standardize a regulatory calculation while the underlying model remains conditional on assumptions.
The ultimate analytical conclusion is that a warranted probability assertion has a defined event and horizon, an identifiable evidential basis, a transparent inferential or physical model, and an interpretation appropriate to the decision it informs. Scientific accountability resides in that complete proposition. The use of a single decimal does not dispense with any of its components.
References
Actuarial Standards Board. (2013). Actuarial Standard of Practice No. 25: Credibility Procedures (adopted December 2013; effective May 1, 2014). https://www.actuarialstandardsboard.org/asops/credibility-procedures/
Actuarial Standards Board. (2016). Actuarial Standard of Practice No. 23: Data Quality (revised December 2016). https://www.actuarialstandardsboard.org/asops/data-quality/
Actuarial Standards Board. (2019). Actuarial Standard of Practice No. 56: Modeling (adopted standard). https://www.actuarialstandardsboard.org/asops/modeling/
Actuarial Standards Board. (2021). Actuarial Standard of Practice No. 38: Catastrophe Modeling (for All Practice Areas) (revised July 2021; effective December 1, 2021). https://www.actuarialstandardsboard.org/asops/actuarial-standard-of-practice-no-38-revised-edition/
Brier, G. W. (1950). Verification of forecasts expressed in terms of probability. Monthly Weather Review, 78(1), 1–3. https://journals.ametsoc.org/view/journals/mwre/78/1/1520-0493_1950_078_0001_vofeit_2_0_co_2.xml
Dawid, A. P. (1982). The well-calibrated Bayesian. Journal of the American Statistical Association, 77(379), 605–610. https://doi.org/10.1080/01621459.1982.10477856
de Finetti, B. (1937). La prévision: Ses lois logiques, ses sources subjectives. Annales de l'Institut Henri Poincaré, 7(1), 1–68. https://www.numdam.org/item/AIHP_1937__7_1_1_0/
European Parliament and Council. (2009; consolidated January 17, 2025). Directive 2009/138/EC on the taking-up and pursuit of the business of Insurance and Reinsurance (Solvency II), Article 101. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A02009L0138-20250117
Gneiting, T., & Raftery, A. E. (2007). Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association, 102(477), 359–378. https://doi.org/10.1198/016214506000001437
Hájek, A. (2023). Interpretations of probability. Stanford Encyclopedia of Philosophy (substantive revision November 16, 2023). https://plato.stanford.edu/entries/probability-interpret/
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
Keynes, J. M. (1921). A Treatise on Probability. Macmillan (original publication; cataloged reprint accessible online). https://books.google.com/books/about/A_treatise_on_probability.html?id=6DUJAAAAIAAJ
Knight, F. H. (1921). Risk, Uncertainty and Profit. Houghton Mifflin. https://www.loc.gov/item/21012860/
Kolmogorov, A. N. (1933; English translation, 1956). Foundations of the Theory of Probability (2nd English ed.; N. Morrison, Trans.). Chelsea Publishing Company. https://books.google.com/books/about/Foundations_of_the_Theory_of_Probability.html?id=tnNEAAAAIAAJ
Popper, K. R. (1959). The propensity interpretation of probability. The British Journal for the Philosophy of Science, 10(37), 25–42. https://doi.org/10.1093/bjps/X.37.25
Ramsey, F. P. (1926; published 1931). Truth and probability. In R. B. Braithwaite (Ed.), The Foundations of Mathematics and Other Logical Essays (pp. 156–198). Kegan Paul, Trench, Trubner; Harcourt, Brace. Accessible scholarly electronic transcription. https://www.fitelson.org/probability/ramsey.pdf
Savage, L. J. (1954). The Foundations of Statistics. Wiley. https://books.google.com/books/about/The_Foundations_of_Statistics.html?id=AGkoAAAAMAAJ
Tversky, A., & Kahneman, D. (1974). Judgment under uncertainty: Heuristics and biases. Science, 185(4157), 1124–1131. https://doi.org/10.1126/science.185.4157.1124
von Mises, R. (1957). Probability, Statistics and Truth (2nd English edition, reprint). George Allen & Unwin. https://books.google.com/books/about/Probability_Statistics_and_Truth.html?id=GMCVQBzlqI8C