Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Parametric Hazard Exceedance Frequency Estimation

Abstract. Parametric hazard exceedance frequency estimation connects measurable natural hazards, statistical event catalogs, and predetermined insurance payments. This monograph distinguishes event rates, annual probabilities, and contractual settlement frequencies; develops their mathematical relationships under explicit count and tail assumptions; examines missingness, censoring, truncation, detection and declustering; and evaluates annual-maxima, peaks-over-threshold, nonstationary and observation-aware approaches. Comparative evidence from precipitation, floods, cyclones, earthquakes and drought illustrates the importance of preserving the physical and contractual definition of a trigger. Primary documentation from regional disaster risk pools shows why modeled impact thresholds require an additional transformation of hazard data. Original analytical synthesis derives the expected cost of binary, stepped and bounded linear payout schedules from the relevant exceedance curves and quantifies the effects of alternative synthetic event definitions, incomplete observation and tail assumptions. The study establishes an auditable methodological basis for pricing and validation, without claiming a new statistical theory, empirical hazard calibration or external peer review.

Keywords: Parametric insurance; hazard exceedance; event frequency; annual exceedance probability; return period; extreme value theory; generalized Pareto distribution; annual maxima; censoring; truncation; declustering; event catalog; observation completeness; nonstationarity; expected payout; catastrophe risk.

Introduction and Research Method

A parametric insurance contract turns a physical or modeled hazard into a contingent payment by specifying an observation, eligibility condition, threshold, and payment rule before the risk period. The actuarial problem begins with estimating how often the legally relevant quantity will pass the specified threshold. The result determines the expected incidence of payments, informs attachment and exhaustion selection, and enters pricing, liquidity, and capital calculations. It is therefore necessary to distinguish the hazard recurrence process from the observation process and from the settlement process, even when all three are represented by a single historical index.

The present investigation develops the statistical foundations of that estimation problem within the parent Science Parametric and Index-Based Insurance. Its object is the frequency curve for an operationally defined trigger, including event catalog construction, incomplete observation, threshold estimation, and transformation from probability to financial consequences. The subject connects hydrology, meteorology, seismic risk, extreme value statistics, survival analysis, actuarial mathematics, and institutional disaster financing. It retains a distinct boundary from comprehensive treatments of basis risk, climate calibration, payout optimization, and insurance law.

The research method combines original mathematical derivations, critical comparison of established statistical methods, examination of primary technical and institutional documentation, and explicitly hypothetical arithmetic scenarios. The technical sources address precipitation frequency, flood records, threshold excess distributions, cyclone best tracks, and seismic catalogs. Scholarly abstracts are used only for statements they explicitly establish. Institutional documents establish the described design, not independent proof of effectiveness. No proprietary insurer contract, actual payout history, raw hazard observations, or fitted catastrophe model is supplied in this study. The numerical exercises accordingly demonstrate the structure of estimation under specified assumptions, rather than claim empirical rates for a location or product.

The central research question is how a parameter selected for payment can be assigned an estimable event frequency, with uncertainty that is both physically interpretable and financially relevant. Three subsidiary questions concern the denominator of frequency in an incomplete catalog, the statistical consequences of connected exceedance observations, and the conversion of intensity-frequency information into a payment distribution. The argument proceeds from the specification of the estimand to the conditions for counting events, modeling their intensity, evaluating data, and carrying the resulting uncertainty into financial estimation. Its independent contribution is an integrated analytical comparison and reproducible illustrations using established mathematics. No priority claim, new named statistical theory, or proprietary model is advanced.

Scientific Foundations and Recurrence Measurement

The physical variable and contractual estimand

Let X be a scalar hazard intensity produced by an observation rule for an eligible event. The specification includes units, measurement technology, spatial support, averaging interval, time zone, duration, aggregation, event identifier, observation vintage, and the threshold comparison. Examples include maximum sustained wind across a prescribed zone, maximum twenty-four-hour precipitation at a named gauge, peak ground acceleration at a specified site, or the negative of seasonal water adequacy where low values are adverse. Define the adverse direction consistently so that larger X indicates a more severe contractual outcome. With a discrete or rounded index, X > a and X ≥ a are different events whenever probability mass lies at a.

A hazard threshold a has a rate λ(a) defined by the expected number of distinct eligible exceedance events per unit time, commonly one year. A yearly exceedance probability p(a) is the probability that at least one such event occurs during the year. An annual maximum M has exceedance probability P(M > a), with M defined on the same event population and observation support. A third object, the expected number of positive payments, can differ from the expected number of eligible hazard events when the contract applies a waiting period, aggregate limit, occurrence cap, reinstatement, or other settlement rule. A frequency estimate that fails to name the estimand cannot be reliably interpreted by an actuary or policyholder.

The NOAA precipitation frequency glossary explicitly distinguishes an annual maximum series from a partial duration series and defines recurrence between exceedance years separately from recurrence between events. The definitions transfer directly to parametric design: a contract that pays only once per year naturally requires the yearly occurrence probability, whereas an unlimited per-event schedule requires an event count and corresponding intensity distribution. Continuous-time stationarity describes invariance of the probabilistic law under time translation; it does not guarantee independent events, and independence alone does not guarantee constant intensity.

Poisson conversion and the two meanings of return period

For a homogeneous Poisson count N over time t with exceedance intensity λ, the count probabilities are P(N = k) = exp(−λt)(λt)^k/k!. Consequently the zero-event probability is exp(−λt), and the chance of at least one event is 1 − exp(−λt). The annual exceedance probability is p = 1 − exp(−λ). The mean time between Poisson events is 1/λ years. The reciprocal 1/p instead describes the expected number of independent identically distributed calendar-year trials until a year with an exceedance, including the successful year. These quantities only approximately coincide in the rare-event limit.

At λ = 0.2 per year, the mean event interval is five years, p is 18.13%, and 1/p is approximately 5.52 years. At λ = 1 per year, the corresponding annual probability is 63.21%, not 100%, and the reciprocal annual probability is about 1.58 years. If an annual exceedance probability is exactly 20%, the Poisson intensity inferred under the model is −ln(0.8) = 0.22314 per year. In an independent-year process with p = 0.01, the probability of at least one exceedance over thirty years is 1 − 0.99^30, approximately 26.03%. This is a probability conditional on the model and fixed annual probability, not a date prediction.

The distinction is also emphasized in the NOAA precipitation frequency FAQs (n.d.). The frequent-event divergence has an immediate pricing implication. A fixed payout B under an unlimited per-event Poisson cover has annual expected payment Bλ. A fixed payout B under an annual one-payment cap has expected payment Bp. At λ = 0.5 and B = $100,000, these are $50,000 and approximately $39,347. The difference is due purely to contract structure, even if both products share the same physical threshold.

A return level is a magnitude determined by inverting a decreasing exceedance curve. A level associated with a mean event interval T solves λ(a) = 1/T. A level defined by an annual exceedance probability 1/T solves p(a) = 1/T. Ambiguous use of the phrase “T-year event” obscures this choice. The appropriate disclosure states the units, whether the number refers to event rate or annual probability, event definition, time horizon, baseline dates, and uncertainty.

Beyond the Poisson assumption

The mapping λ to p is not distribution-free. For any nonnegative integer count N, E[N] ≥ P(N ≥ 1). Equality holds when at most one counted event can occur. If N takes value 4 with probability 0.1 and zero otherwise, the mean is 0.4 and the occurrence probability is 0.1. A Poisson count with the same mean gives 0.3297. Grouping four correlated observations into a single contractual episode could therefore change the expected payment in a different way from changing the hazard intensity itself.

When yearly intensities vary randomly, a mixed Poisson model gives P(N ≥ 1) = 1 − E[exp(−Λ)]. By convexity of exp(−x), this is no larger than 1 − exp(−E[Λ]). The direction applies when the conditional count is Poisson; it is not asserted for every dependent process. Count dispersion, recurrence time distribution, and within-year grouping must be investigated rather than inferred from the observed rate alone. Earthquake renewal, storm clustering, and seasonal rainfall can require different process formulations.

A rate curve must be nonincreasing in its threshold when eligibility and event population remain fixed. If λ(70) exceeds λ(60) for identical units and observation rules, the estimated curve violates set inclusion. Separate models at each attachment may exhibit this error through sampling noise. Joint fitting, monotone estimation, or constrained smoothing can preserve the required relationship, but a smoothing remedy does not cure a catalog whose event definitions change with threshold.

Event Catalogs, Censoring, and Observation Architecture

Defining exposure time

A catalog is an organized representation of event evidence. A statistical sample comprises more than a file of event rows: it includes the observation windows during which events could have been detected, rules determining inclusion, measurements and uncertainty, and non-event intervals. The denominator T is effective observation time for the target event definition, not automatically the elapsed time between the first and last recorded event.

For a completely observed homogeneous Poisson process, the likelihood for N = n in time T is proportional to exp(−λT)λ^n, giving the maximum likelihood estimate n/T. If a twenty-year calendar record contains only fifteen fully observable years and nine independent events in those years, the complete-window estimate is 9/15 = 0.6 per year. Using 9/20 = 0.45 treats unobserved years as no-event years and introduces downward bias when the missing windows are truly uninformative. The fifteen-year calculation itself assumes the excluded intervals are independent of the process or that the remaining windows represent the target intensity. Storm-caused outages violate that assumption.

For observation probability q(t) independent of event intensity conditional on time and known covariates, thinning a Poisson process yields observed intensity q(t)λ(t). If λ is constant and q is known, the effective exposure is the integral of q(t) over time; the estimate is n divided by that integral. For variable q and variable λ, the product is what the observed count identifies. If q is unknown, the same count can be generated by a low hazard rate with good detection or a high rate with poor detection. Independent information about observation quality is necessary to distinguish the two.

An event catalog can be incomplete because detection instrumentation improves, a governing agency changes, station coverage expands, earlier records are digitized selectively, or only severe events are transcribed. A homogeneous-looking list may contain a heterogeneous denominator. A record of the coverage, detection threshold, and quality flags, including non-event periods, is thus part of the scientific evidence for a priced trigger.

Missing, censored, truncated, and rounded measurements

An exact observation identifies an index value x. An interval-censored observation identifies x within (l, r]; a left-censored record states x ≤ r, while a right-censored record states x > l. Missingness indicates the absence of sufficient observation, which may be independent of hazard severity or depend on it. Truncation concerns which values appear in the data: if only events with x > d can enter a catalog and d is known, a mark likelihood conditional on inclusion divides the original density by P(X > d).

If F is a continuous mark CDF with density f, the mark likelihood contributions are f(x) for an exact observation, F(r) − F(l) for an interval, F(r) for left censoring, and 1 − F(l) for right censoring. For interval rounding, using narrow intervals around reported values retains information about the probability mass implied by measurement resolution. These formulas assume the interval or censoring mechanism is ignorable given modeled covariates; informative censoring requires a joint measurement model. In a catalog of historical events identified only above a perception threshold, a year known to contain no event above that threshold conveys different information from a year whose history is entirely unknown.

The official USGS summary of Bulletin 17C (England et al., 2018) describes its representation of interval and censored flood records, Expected Moments Algorithm, low-outlier treatment, and confidence intervals. The Expected Moments Algorithm study guide of the U.S. Army Corps of Engineers also identifies perception thresholds as relevant to fitting. These sources support the importance of using graded observation evidence. The Bulletin 17C moment-based approach to annual flood peaks is a specific hydrological methodology; the generic likelihood expressions here are illustrative statistical derivations, not a claim to implement that procedure.

Joint occurrence and severity likelihood

For an observable marked Poisson process with intensity ν(t, x), a region A of event time and marks subject to detection has total expected observed count μ_A = integral over A of ν(t, x) dt dx. A likelihood for exact observed points is proportional to exp(−μ_A) times the product of ν at each point. A known detection function q(t, x) changes the observed intensity to q(t, x)ν(t, x). Integration over the appropriate mark interval replaces an exact mark density for a censored event. The compensator must reflect the same observed population as the product terms.

Suppose the underlying intensity of events exceeding 40 units is λ_40 and their excess marks have survival S. If a historical subperiod detects only marks above 60, its observable event intensity is λ_40 S(20), rather than λ_40. Combining recent complete data above 40 with earlier data above 60 requires representing each period's exposure and observable mark support. Treating all historical years as complete above 40 depresses estimated moderate-event frequency. Conversely, dividing severe counts only by a recent exposure period discards usable old information for rare events.

A single common threshold can simplify a pooled analysis, at the expense of suppressing moderate data from better-observed periods. A period-specific observation likelihood preserves more information but depends on reliable documentation of perception thresholds, station availability, and changes in measurement technology. When documentation is limited, fit alternative plausible completeness windows and disclose the induced range instead of presenting a false single precise estimate.

Distinct events and serial dependence

Hourly wind observations are not hourly independent storms. A moving rainfall window can exceed attachment for several consecutive timestamps during one rainstorm. A seismic sequence can have multiple aftershocks that either count separately or are grouped under an occurrence definition. A discrete count of raw threshold crossings is therefore an intermediate observation. The contractual rule may identify a storm, a maximum over a fixed window, a geographically contiguous episode, or an explicitly defined period of sustained exceedance.

Ferro and Segers (2003) identify sensitivity to how clusters are separated and present an approach based on times between extremes and the extremal index. Their result motivates diagnostic attention to declustering rather than a universal fixed separation period. Statistical clustering and contractual grouping have different purposes: the former seeks a tractable recurrence representation; the latter determines liability. If a policy pays on each contractually defined event but a statistical method groups several events for independence, the grouped severity must encode the within-cluster set of payments or be accompanied by a process for within-cluster counts.

A run rule groups exceedances separated by fewer than a selected number of calm observations. Its duration should be compared with physical storm duration, observation sampling frequency, and contract wording. A diagnostic exercise changes the run length, counts clusters, and estimates their peaks. If the count changes sharply while the peak distribution is comparatively stable, the estimated attachment frequency is particularly dependent on event identity. Reporting only the fitted tail parameter conceals this source of uncertainty.

The extremal index θ, under suitable asymptotics for stationary sequences, lies between zero and one in common clustering interpretations and modifies the independent-observation extreme limit. It can approximate a cluster rate relative to exceedance intensity at high thresholds in appropriate settings; it is not automatically a contractual aggregation factor at finite thresholds. Application requires checking dependence conditions and the relation between statistical clusters and actual payable occurrences.

Catalog provenance and event ledger

A reproducible ledger records stable event identifiers, start and end times, relevant geography, original agency identifier, unit and averaging interval, raw and derived values, quality flags, source version, detection regime, transformation, grouping decision, and exclusion reason. An observation calendar separately records periods when the index could be calculated. For revised observations, it retains the value known at contractual determination and the later revised value so that operational replay and retrospective scientific reconstruction remain distinct.

Deduplication should identify the physical occurrence before eliminating rows: several best-track points or multiple agency estimates are measurements of one cyclone, while two hazardous episodes within the same system may remain distinct under the policy. “No match” across catalogs is not definitive proof of a missing event because agencies may use different classification and inclusion rules. Metadata and reconciliation should precede comparisons of event counts. The estimator is valid for the represented observation system and must be retested after changing that system.

Statistical Estimation of Exceedance Curves

Empirical counts and annual maxima

For an observed complete interval and fixed threshold a, the event-rate estimator is N(a)/T. For n independent complete years and annual indicator I_i(a) equal to one when the threshold is exceeded, the empirical annual probability is the sum of indicators divided by n. These estimators answer different questions. Their sample sizes are elapsed observation time and number of independent years respectively. Under a Poisson model, the estimated variance of N/T is approximately N/T²; the plug-in standard error of p from independent Bernoulli years is √(p(1−p)/n). Dependence, incomplete years, and estimated observation probabilities change these formulas.

For zero events in T years, the maximum likelihood estimate zero is uninformative about the attainable tail. A one-sided 95% Poisson upper confidence bound solves exp(−λT) = 0.05, giving λ = −ln(0.05)/T. For twenty completely observed years with zero events, the upper bound is about 0.1498 per year, or p = 13.91% under the modeled Poisson conversion. The result follows by direct inversion; it illustrates that zero observed payments in a short record does not establish zero risk.

For n years containing at least one counted event, a binomial model can give intervals for the annual exceedance probability. For independent Poisson counts, exact two-sided confidence limits result from inverting lower and upper Poisson tails; the corresponding χ² quantile representation uses 2n and 2(n+1) degrees of freedom for n positive, scaled by 2T. When n is zero the lower endpoint is zero. These intervals are conditional on the specified count law and complete observation; they do not quantify climate change, measurement disagreement, threshold selection, or event-definition risk.

The annual maximum estimator selects the largest admissible observation in each complete year. It is suitable for annual occurrence probability and can mitigate within-year serial dependence. However, it discards information about additional storms and cannot identify expected per-event counts without supplementary assumptions. For a seasonal cover, use the maximum within the covered season and define the number of comparable observed seasons; the calendar-year maximum may include irrelevant winter or off-season events. Year-boundary and cross-boundary events need a consistent assignment.

Generalized extreme value and threshold excess models

Extreme value statistics provides limiting approximations under appropriate domain-of-attraction and dependence conditions. A generalized extreme value (GEV) distribution for block maxima with location μ, scale σ > 0, and shape ξ is

G(x) = exp{−[1 + ξ(x − μ)/σ]^(−1/ξ)}

on the support 1 + ξ(x − μ)/σ > 0 for nonzero ξ. For ξ = 0 the continuous limit is G(x) = exp{−exp(−(x − μ)/σ)}. A return level with annual exceedance probability p is the quantile G⁻¹(1 − p). The assumption that annual maxima are modeled by this family must be assessed against the sampled block lengths, distribution, and physical setting. A statistically fitted upper endpoint should not be interpreted as an established physical maximum without independent physical justification.

Davison and Smith (1990) developed an analysis that treats excess magnitudes together with exceedance occurrence, including models for temporal structure. A peaks-over-threshold (POT) representation exploits observations above a sufficiently high statistical threshold u. Conditional on X > u, let Y = X − u. Its generalized Pareto survival approximation is

S_u(y) = (1 + ξy/β)^(−1/ξ), for y ≥ 0 and 1 + ξy/β > 0,

with scale β > 0, and S_u(y) = exp(−y/β) at ξ = 0. For negative ξ, the support ends at y = −β/ξ and the survival beyond the endpoint is zero. For positive ξ, the tail decays polynomially. The NIST generalized Pareto documentation (2023) supplies the GPD form and discusses mean residual life as a diagnostic; the conditional mean excess exists when ξ < 1. The diagnostic's approximate linearity is useful but never sufficient to establish validity of extrapolation.

Let λ_u be the rate of distinct admissible events exceeding u. If the conditional tail is GPD and the same event population is used, then for a ≥ u:

λ(a) = λ_u S_u(a − u).

This identity is the rate version of conditional probability. It holds for first-moment intensities without requiring Poisson arrivals; the conversion of λ to annual probability requires a count law. For a Poisson process and an annual probability p, the target intensity r = −ln(1−p). If r ≤ λ_u, a = u + (β/ξ)[(λ_u/r)^ξ − 1] for nonzero ξ, and a = u + β ln(λ_u/r) for ξ = 0. The formula applies within the fitted tail and cannot estimate a threshold below u. For a rate-based return period T, substitute r = 1/T. For ξ < 0, as r approaches zero the implied level approaches the finite endpoint from below.

A statistical threshold u is a modeling decision. A contractual attachment a is a settlement decision. They can coincide but need not. If the contract attaches below the region where a GPD approximation is credible, its frequency should be estimated with the body of the distribution or an appropriate broader model. Fitting a high tail and extrapolating backward to attachment is a different and generally unsupported step. The insurance calculation should specify which parts of the payout curve are interpolated within data, extrapolated into tails, or unsupported.

Likelihood, threshold choice, and diagnostics

For independent excesses y_i under a GPD with nonzero ξ, log likelihood is −m ln β − (1 + 1/ξ) times the sum of ln(1 + ξy_i/β), restricted to positive support factors. The exponential limit gives −m ln β − sum(y_i)/β. The maximizer in that exponential model has β equal to the sample mean of excesses. Numerical optimization with shape parameter near an endpoint requires explicit support checking; unconstrained steps that place an observation outside the support invalidate the evaluated likelihood. Moment estimators, probability-weighted moments, and Bayesian methods can offer alternatives depending on sample size, parameter restrictions, and prior evidence.

The rate component is estimated from counts and exposure while the conditional severity component is estimated from observed marks. With complete marked Poisson observations and independent identically distributed mark law, the likelihood factorizes into count and mark parts. Censoring dependent on severity, heterogeneity, or nonstationary selection breaks that simple separation. A joint uncertainty assessment must preserve dependence between estimates induced by data selection or shared covariates.

Threshold choice balances approximation bias against sampling variance. A low threshold supplies more observations but can include values for which a limiting excess model fits poorly. A high threshold may produce unstable shape and rate estimates. Inspect mean residual life, parameter stability after accounting for the changing GPD scale, quantile plots, exceedance-rate stability, and sensitivity of the actual payout expectations. Under an exact threshold-stable GPD tail, the scale at a higher threshold v is β_v = β_u + ξ(v−u), as follows from conditional survival ratios. Compare ξ and adjusted scale across candidate thresholds; do not compare raw scale parameters as though they should be invariant.

Assess fit in the range that matters financially. A visually acceptable probability plot over moderate observations may conceal systematic errors near contractual exhaustion. A small change in ξ can shift an extrapolated return level much more than it shifts an interpolated moderate threshold. Quantify the spread of estimates across credible thresholds, models, observed time windows, and clustering choices. Profile likelihood or a bootstrap that resamples suitable blocks can convey parameter uncertainty; neither method cures model misspecification.

Regional pooling can borrow strength when multiple sites share sufficiently comparable processes after normalization or covariate adjustment. A regional frequency estimate should disclose assumptions about the homogeneous group and spatial dependence. Treating many gauges exposed to the same storm as independent years mechanically inflates apparent effective information. Hierarchical models can partially pool local rate and tail parameters while retaining differences; transfer to an unsampled site requires validation that physical drivers, exposure support, and measurement procedures correspond.

Nonstationarity and prediction horizons

Milly and colleagues' Milly et al. (2008) argues that stationarity should cease to be a default premise in long-horizon water-resource risk assessment under climate change. The actuarial implication is to identify the predictive law for a specified contract period rather than treat an arbitrary historical average as invariant. A short one-year product and a long-duration public infrastructure program require different horizons. Current physical conditions, observation regime, adaptation, and the identity of the insured index may also differ.

For a nonhomogeneous Poisson process with deterministic intensity λ_a(t), the expected count over interval (s, e] is Λ = integral from s to e of λ_a(t)dt, and the probability of any exceedance is 1 − exp(−Λ). A log link λ_a(t) = exp(α + b z(t)) can express a changing covariate z; the exposure offset in a binned count regression ensures counts are compared on the same time basis. A season with significant hazard activity but no coverage contributes to physical climatology, not contract exposure.

For an illustration, suppose the next annual rate could be 0.10 under a baseline or 0.15 under a stress scenario. The probabilities of at least one Poisson event are 9.52% and 13.93%. For a fixed uncapped per-event payout of $200,000, expected annual payout changes from $20,000 to $30,000. This is conditional scenario arithmetic, not an inferred climate trend. Under a time-varying or regime-dependent law, a single reciprocal rate should be replaced or supplemented by a stated period-specific probability curve and explicit scenarios.

A covariate trend may be confounded with a detection break, station relocation, improved remote sensing, or change in the trigger algorithm. Correlating the observed count with calendar time alone does not establish a change in hazard. Refit using a stable observation regime, independent physical covariates, and sensitivity to plausible catalog changes. With sparse extremes, a simple stationary baseline can remain a useful benchmark even when an assumption of perpetual stationarity is untenable; complex alternatives earn credibility through physical plausibility and predictive evaluation.

Comparative Hazard Measurement and Institutional Evidence

Precipitation and river flooding

A rainfall trigger must state depth or intensity, accumulation duration, clock convention, spatial support, observation product, and eligibility window. A threshold of 100 mm in a rolling twenty-four-hour accumulation differs from the same threshold at a calendar-day gauge. The NOAA HDSC definitions distinguish constrained observations from unconstrained duration maxima. As the duration grows, a maximum accumulation based on nonnegative precipitation cannot decrease under nested, comparable windows, although mean intensity may decrease. This physical restriction offers a consistency diagnostic for depth-duration-frequency estimates.

A point precipitation return level cannot be equated with an areal accumulation return level. Rainfall at different pixels is dependent, and the maximum over a region increases opportunities for an extreme. A contract triggering when any station crosses a threshold prices the union of station events. A contract using area-average precipitation prices a different functional of the hazard field. Both may be labeled a regional rainfall policy but have different event frequencies. Changing the set of stations after inception changes the measurement object if membership enters the trigger.

Flood discharge is related to precipitation through catchment conditions, routing, rainfall spatial pattern, antecedent soil moisture, storage, and water management. A threshold on river stage depends additionally on the local stage-discharge relationship and its stability. Consequently an Atlas-style precipitation-frequency value is not a substitute for the annual frequency of a river-stage or discharge trigger. USGS flood frequency guidance provides a primary example of a discipline-specific approach to incomplete historical peak data, while the required contract index remains determined by the actual measurement rule.

Where seismic or weather hazards produce significant historic events without instrumental peaks, threshold evidence can be valuable for rare-event inference. A paleoflood indication of water above a certain elevation can contribute an interval bound when converted through a defensible hydraulic reconstruction. Such a reconstruction adds uncertainty to the observation; it does not create an exact instrument reading. A contractual trigger based on a modern digital gauge may require scientific translation of old evidence to the gauge-equivalent variable, including changes in channel geometry and rating relationships.

Tropical cyclone and coastal hazards

The NOAA IBTrACS documentation brings together agency best-track information and identifies version-specific data products. It is a valuable starting point for event identity and historical position and intensity. Best-track archives describe storms; a contract might instead use a site-specific wind estimate, a polygon crossing, maximum gust, pressure deficit, or modeled impact. Track interpolation, radius assumptions, wind-field transformation, and sampling interval affect whether a threshold is crossed at a particular insured site.

An analyst should reconcile agency intensity definitions and measurement conventions before counting exceedances. Maximum wind along a storm track is not the same as maximum wind at a contract location. A storm that is intense offshore may never generate a local trigger. A lower-intensity cyclone passing close to the insured site may trigger a local threshold. If the observation interval is coarse relative to rapid intensification or close passage, threshold-crossing status can be uncertain even when a storm is unambiguously documented.

For storm surge, the relevant variable might be residual surge, total water level relative to a specified datum, or an inundation depth. Astronomical tide and sea-level baseline can change total water-level exceedance without identical changes in meteorological surge. A port contract with operational closure eligibility differs further from a pure water-level trigger. A multihazard policy that pays on wind or surge needs occurrence matching because wind and surge within one cyclone are correlated. Separate marginal hazard curves alone cannot price payout incidence for the union.

An event set produced from simulated cyclone tracks may be much longer than the observational archive. Its statistical sampling error conditional on its generation model can be small, while uncertainty about track distribution, frequency, intensification, and vulnerability remains substantial. Simulation years are modeled opportunity units and not independent historical evidence. A documented separation between observational validation and model sampling precision is essential.

Earthquakes and local shaking

The USGS ComCat documentation describes source parameters and related products supplied by seismic networks and acknowledges incomplete loading of some historic regional material. An earthquake catalog has geographic, magnitude, time, and network coverage boundaries. A source magnitude threshold estimates the occurrence frequency of earthquakes of a given source size; a ground-motion threshold estimates shaking at a location. The latter depends on distance, depth, rupture characteristics, ground-motion variability, and site conditions.

A simplified probabilistic seismic hazard curve at site s can be written λ_s(a) = sum over source categories j of integrals of P(Y_s > a | m, r, j) with respect to the annual occurrence-rate measure of source magnitudes m and distances r. This integrates conditional ground-motion exceedance across source occurrence scenarios. The USGS earthquake hazard explanation (USGS, n.d.) describes annual-rate ground-motion curves and the Poisson law for zero occurrences; the integral here is a generic analytical representation. A regional count of earthquakes of magnitude at least six cannot replace this site-specific calculation.

Some contracts define one event through an agency event identifier and permit separate aftershock payments; others aggregate a sequence during a stated hours clause. Statistical declustering intended for independent mainshock hazard need not coincide with those definitions. If a catalog excludes aftershocks to fit a background Poisson model but contracts pay aftershocks, the expected claims frequency will be understated unless the omitted process is separately represented. The opposite problem arises if multiple shocks are aggregated contractually but priced as independent full-limit payments.

Catalog completeness thresholds may differ across regions and periods. Apparent increases in small earthquakes can follow improved instrumental coverage. For a high shaking trigger, unrecorded smaller events may still be irrelevant in a given exposure domain, but this conclusion depends on the site distance and ground-motion model. A completeness analysis should address the trigger's physical sensitivity, not just a single national magnitude cutoff.

Seasonal drought and model-generated response costs

Seasonal drought contracts may use rainfall deficit, standardized precipitation, vegetation indicators, soil moisture, or crop water adequacy. An adverse lower-tail threshold is transformed to an upper-tail problem by defining X as the negative of the original indicator, or by directly modeling a lower-tail probability with the correct inequality. A season is frequently the natural unit of observation because crops and livelihoods respond to cumulative timing and water stress. Daily rainfall exceedances cannot determine the frequency of season-level drought losses without the associated temporal transformation.

The Africa RiskView methodology of the African Risk Capacity Group describes its use of rainfall estimates, crop water adequacy expressed through WRSI, vulnerability information, and modeled drought response costs. The ARC risk transfer parameters states that payout thresholds are selected with risk transfer parameters including attachment, limit, and ceding percentage. These institutional descriptions exemplify a modeled impact variable as the payment index. For inference, historical weather must be transformed consistently with the model configuration whose risk is being priced.

Historical reconstruction raises a difficult choice: applying current vulnerability and response assumptions to earlier weather estimates the hazard-driven risk of the current modeled exposure; using historical exposure at each date estimates a historical sequence of modeled response costs. Both exercises can be scientifically useful, but they answer different questions and may generate different exceedance curves. A change in estimated response costs could reflect modeled vulnerability or policy costs even if seasonal hazard remains constant.

In Caribbean and Central American disaster risk financing, the revised 2025 CCRIF brief describes CCRIF policy elements: payments attach to pre-agreed modeled event loss and depend on the selected layer terms. The public document uses return-period language for selection of attachment and exhaustion. It does not, by itself, supply the complete proprietary event set, member policy wording, or numerical frequency curves. Reconstructing a specific policy's payout distribution from the public description alone is therefore beyond the evidence. The example establishes the institutional relevance of translating a physical event catalog through a loss calculation before applying the threshold.

The comparative conclusion is that international hazard disciplines share mathematical ideas about extreme frequency but require distinct physical definitions and data architectures. Regional risk pools demonstrate the additional transformation between weather or ground motion and a payout-relevant model quantity. A portable methodology preserves the exact trigger, detection regime, grouping and statistical estimand; it does not transplant a flood distribution into drought insurance or treat a cyclone maximum as a direct proxy for local wind liability.

Risk Transfer, Payment Frequency, and Actuarial Consequences

Binary, stepped, and continuous payout functions

A binary per-event cover has payment g(x) = B if x > a and zero otherwise. If eligible event occurrence has finite annual intensity measure ν(dx), the expected total payment of an uncapped per-event cover is integral g(x)ν(dx). Therefore the expected annual payout is Bλ(a). This expression is an application of linearity of expectation for an integrable marked point process and does not require Poisson independence. Count dependence matters for the distribution and risk capital, while first moments depend on expected occurrence and marks.

For a bounded linear layer with attachment a, exhaustion b > a, and limit B, let g(x) = B min(max((x−a)/(b−a), 0), 1). Using the identity min(max(x−a,0), b−a) = integral from a to b of the indicator that x exceeds z dz, expected annual uncapped payout is

E[S] = B/(b−a) integral from a to b of λ(z) dz.

This equation integrates the event frequency curve across the covered layer. It is dimensionally consistent: λ is events per year; dz is hazard units; B/(b−a) is currency per hazard unit; the output is currency per year. Because the hazard rate decreases with threshold, Bλ(b) ≤ E[S] ≤ Bλ(a), with boundary refinements for discrete equality rules as necessary. The expected payout for a gradient layer is thus determined by the entire curve between attachment and exhaustion, not attachment frequency alone.

For stepped payout amounts B_1 < B_2 < … < B_k at increasing thresholds a_1 < … < a_k, write g(x) as the sum of increments ΔB_i multiplied by the relevant threshold indicators. The expected annual uncapped payment is the sum of ΔB_i λ(a_i). This provides a transparent diagnostic: separately estimating each tier should preserve an increasing severity schedule and nonincreasing exceedance rate. The contribution of a layer to expected cost follows its incremental payment, not the entire face amount at every threshold.

If the contract pays based on the annual maximum M instead of each event, replace the rate curve with p(z) = P(M > z) in the corresponding expectation. For the bounded linear schedule paid once using M, E[g(M)] = B/(b−a) times the integral of p(z) over the layer. For an annual aggregate cap applied to a sum of event payouts, neither this maximum formula nor the uncapped integral gives the exact answer. The schedule must be simulated or the aggregate distribution computed subject to the cap, reinstatements, and event rules.

A synthetic event-frequency and payment exercise

Assume a hypothetical complete thirty-year catalog contains twelve eligible independent cluster events above u = 40 index units. Their excesses, in units, are 2, 3, 4, 5, 6, 7, 8, 9, 10, 12, 16, and 38. The sum is 120, so the exponential tail maximum likelihood estimate is β = 10; λ_u = 12/30 = 0.4 events per year. Under an assumed homogeneous marked Poisson model with exponential excesses, the fitted rate at 50 units is 0.4 exp(−1) ≈ 0.14715, and at 60 units is 0.4 exp(−2) ≈ 0.05413. Corresponding annual probabilities are approximately 13.68% and 5.27%. These are estimates from a designed synthetic example; the assumed exponential form is not validated by twelve observations.

Consider a bounded linear layer from 50 to 60 with limit $1,000,000. Substituting the exponential frequency curve into the integral gives

E[S] = $1,000,000 × 0.4 × (exp(−1) − exp(−2)) ≈ $93,018 per year.

The result is between the exhaustion-rate bound of $54,134 and attachment-rate bound of $147,152. A binary $1,000,000 per-event product attaching at 50 would instead cost about $147,152 in expected payouts before loads. An annual one-payment binary cover at the same threshold has expected payment about $136,837. The contract choices are responsible for these differences under an identical hazard recurrence assumption.

For the empirical count strictly above 50, excesses greater than 10 are 12, 16, and 38; the count is three and the rate is 0.1. Excesses greater than 20 contain the single value 38, rate 0.03333. The fitted rate curve smooths and extrapolates relative to empirical counts. The difference is a reminder that the exponential estimate depends on the twelve observed marks and the assumed distribution. It should not be interpreted as more reliable solely because it is smooth.

For sensitivity, preserve λ_u = 0.4 and β = 10 but allow shape ξ = −0.2 or +0.2 without claiming that these alternative shapes are fitted. At 60 units the GPD rate is 0.4(0.6)^5 = 0.031104 for negative shape, and 0.4(1.4)^−5 ≈ 0.07437 for positive shape, compared with 0.05413 under the exponential tail. The fixed binary $1 million expected yearly payout at 60 changes accordingly from about $31,104 to $74,374. This conditional stress exercise identifies the sensitivity of exhaustion risk to tail shape; it is not a confidence interval.

The rate estimator's plug-in standard error at u is √12/30 ≈ 0.11547 events per year, approximately 29% of its estimate. Rate uncertainty is already consequential before considering shape, scale, observation completeness, or clustering. A twelve-mark catalog cannot justify a precise quotation for far rarer triggers. A reliable pricing exercise would present a candidate distribution, its statistical uncertainty, alternative tail models, and independent physical plausibility assessments.

Frequency, loss relevance, and basis risk

Let L denote actual economic loss or an explicitly defined protection need, and X the index. A threshold exceedance probability estimates P(X > a), while the probability that protection is needed is P(L > ℓ) for a chosen material threshold ℓ. The quantities P(X ≤ a, L > ℓ) and P(X > a, L ≤ ℓ) describe different trigger-protection mismatches. These joint probabilities cannot be inferred from the marginal hazard exceedance curve alone.

A product can be precisely calibrated to a published precipitation frequency but poorly related to local damage. Likewise a physically relevant index can have an uncertain rare-event probability because records are sparse. Statistical uncertainty about X is distinct from dependence uncertainty between X and L. For model-based triggers, the loss estimate computed by the contractual model is itself part of X; the divergence between that modeled loss and actual losses remains relevant to coverage performance. ARC and CCRIF illustrate why this distinction must survive when a payout index is expressed in currency rather than physical units.

An informative analysis estimates the trigger frequency, expected payout, expected shortfall of protection relative to need, and the distribution of payout timing under the operational observation rule as separate outputs. More precise estimation of the first output does not automatically improve the others. The appropriate joint evaluation includes matched hazard, payment and loss observations or an independently justified joint simulation. This monograph does not estimate empirical basis-risk rates or claim observed welfare impacts.

Aggregate distribution and uncertainty for capital

Suppose N is Poisson with annual mean λ, and severities U_i = g(X_i) are independent identically distributed, independent of N. For S = sum of U_i over N events, E[S] = λE[U] and Var(S) = λE[U²], obtained by conditioning on N and using the law of total variance. One can choose λ for a population above a base threshold even if some marks produce zero payouts. These compound Poisson formulas are exact for the stipulated model. For clustered or dependent arrivals, additional covariance terms change aggregate dispersion; for common regional events, cross-policy covariance can be substantial.

The expected value is not a complete premium. An actuarial amount may include acquisition and administrative costs, risk and capital charges, expenses for data and settlement, profit or contingency provisions, reinsurance effects, and taxes as applicable. A frequency estimate is consequently a component of a broader financial calculation rather than a market quotation. Expected payout error generally propagates directly into the expected-cost component; capital-cost error may be more sensitive to severe tail assumptions and common-event accumulation.

For multiple policies linked to one event set, expected aggregate payment remains the sum of expected payments in the absence of portfolio-level caps, regardless of dependence. The probability that any policy triggers, variance of aggregate cost, and exhaustion of joint reinsurance depend on joint outcomes. Summing independent single-site probabilities is inappropriate for a regional trigger or portfolio capital model. If two triggers have annual probabilities p_A and p_B, the probability of at least one is p_A + p_B − P(A∩B). The intersection cannot be obtained from the marginals without dependence information; it lies between max(0,p_A+p_B−1) and min(p_A,p_B). These bounds follow from set inclusion and are useful as a minimal consistency check.

Validation, Error Analysis, and Scientific Synthesis

A hierarchy of uncertainty

The prediction target depends on several coupled sources of uncertainty. Sampling uncertainty concerns finite numbers of events. Parameter uncertainty concerns estimates of rate, scale and shape under a selected model. Structural uncertainty concerns the choice among count processes, tails, observation mechanisms and dependence structures. Measurement uncertainty concerns the accuracy and comparability of the index. Scenario uncertainty concerns future physical conditions, exposure and data regimes. Contract uncertainty concerns ambiguous settlement windows or calculations and is best addressed by clear specification rather than absorbed indiscriminately into a statistical error bar.

A confidence interval that conditions on exact observation and a homogeneous Poisson process quantifies only a portion of total uncertainty. A model-averaged predictive distribution can include alternative models with stated weights; those weights require justification. Stress scenarios should remain individually interpretable. If a parameter cannot be identified because of unknown detection rates, presenting a narrow parametric interval can create an illusion of accuracy; bounds or sensitivity sets more honestly describe the available information.

Measurement error near attachment deserves special attention. Let true physical intensity be X and observed contractual index W = X + E. The relevant payment probability is P(W > a), obtained by integrating the conditional exceedance P(E > a−x | X=x) over the law of X. Even unbiased measurement error can change this probability because the event indicator is nonlinear. Calibration to X while settling on W therefore prices a different object. Rounding to whole units generates point masses in W; comparison signs and interpolation policies can materially affect the set of payable outcomes.

The observation process can also fail precisely during severe events. A gauge destroyed during a flood creates severity-related missingness. Deleting its missing interval assumes a relationship between observation and severity that may be false. A fallback model may preserve operational settlement but introduces a second index observation law. Historical replay should implement the same fallback hierarchy and record which estimates arise from primary versus substitute sources. If the fallback did not exist in older periods, a retrospective reconstruction may be informative but cannot recreate all operational decisions without assumptions.

Out-of-sample tests and scoring

A validation design first defines time blocks or independent event groups rather than randomly splitting correlated observations. Fit on an earlier or nonoverlapping part of the record; predict event counts, annual exceedances, threshold marks, and if possible payouts for withheld observations. Repeated rolling-origin splits can reveal deterioration with changing data or physical conditions. Spatial validation should leave out regions or event groups when neighboring observations are strongly related. Random observation-level cross-validation may leak the same storm into training and test data.

For held-out years with predicted probabilities p_i and binary outcomes y_i, the mean squared probability error, or Brier score, is the average of (p_i−y_i)². If the true Bernoulli probability is q, the expected squared error is (p−q)² + q(1−q); it is minimized at p = q. This derivation illustrates proper scoring. Gneiting and Raftery (2007) discuss the general principle that proper scores incentivize honest probability forecasts. For counts, assess a suitable predictive likelihood or proper count-distribution score, rate calibration, dispersion, and temporal residual dependence. For continuous severity, compare predictive distributions and exceedance coverage at decision-relevant thresholds.

Rare-event scores can be insensitive to a limited test set; a product predicting a low probability each year may see no triggers in a short held-out window. Compare models using calibration across meaningful threshold ranges and longer observational or independent regional records when justified. A validation result must disclose the number of test periods, number of events, effective dependence, and the range in which validation was possible. “No observed failure” in a small sample is weak evidence of high-tail calibration.

Count diagnostics can compare observed variance to conditional mean over comparable blocks. Overdispersion may signal heterogeneity, clustering, omitted seasonality, or changing observation. Underdispersion may result from aggregation or constrained opportunities. A residual diagnostic is a pointer to investigation, not proof of a causal mechanism. For an inferred point process, one can examine interevent times against the fitted model and transform them by the modeled integrated intensity when appropriate; systematic departures indicate a need for alternative timing structure or exposure modeling.

A reproducible settlement-frequency sensitivity design

The following synthetic comparison specifies a reproducible calculation with no external data. Take twenty calendar years, of which fifteen have complete coverage. Among them, nine contractual storm episodes occur; raw observation crossings number twenty-seven. Assume the five unobserved years have unknown counts and that the complete periods are representative under a stationary Poisson episode law. Raw crossings divided by fifteen yield 1.8 observations per year; distinct episodes divided by fifteen yield 0.6 events per year. Dividing the nine episodes by twenty yields 0.45 events per calendar year, an estimator inappropriate under the stated missing-period assumptions.

Under the modeled episode rate 0.6, the annual occurrence probability is 1−exp(−0.6) ≈ 45.12%. Using 0.45 gives 36.24%. For a $100,000 annual one-payment policy, the resulting expected annual payouts are approximately $45,119 and $36,237 respectively. Treating every raw crossing as an independently payable event produces $180,000 expected annual uncapped payout instead of $60,000 for episodes. The scenario changes precisely one classification or exposure choice at a time.

If the true average detection probability during the fifteen nominally covered years is only 0.8 and is known, with independent Poisson thinning, the corrected rate is 9/(15×0.8) = 0.75. The corresponding annual probability is about 52.76%. The comparison demonstrates confounding between observation completeness and the inferred hazard. The eight-tenths value is expressly assumed. It cannot be estimated from the nine recorded episodes alone, and applying an arbitrary detection correction would add hidden judgment to pricing.

For a sensitivity study with a real catalog, freeze the raw inputs and observation calendar, define at least one physically justified baseline event rule and alternative grouping, fit rates and marks under the same contract, compare empirical and fitted rates, vary the supported tail shape and baseline, propagate each into frequency and price, and report the resulting range. Publish or securely archive input versions and deterministic transformations, subject to data rights. This procedure is a synthesis of established statistical controls, not a proposed new theory.

Governance and traceability of the estimated curve

A scientific frequency estimate must be traceable from the priced threshold back to an observation definition and exposure denominator. The minimum review record identifies the contract trigger and adverse direction, eligible time and geography, source agency and dataset version, completeness evidence, event grouping, estimator and assumptions, sample size, uncertainty, diagnostics, and pricing transformation. It should also identify the distinction between scenario simulations and observations so that future reviewers do not count synthetic events as corroborating historical observations.

An insurer or model operator should freeze the contract-relevant data source and algorithm version for settlement, with an explicit mechanism for correction and fallback if applicable. Scientific recalibration may occur between underwriting periods, yet the new prediction should be tested under the new observation and exposure definition. Version changes in the index are potentially changes in the random variable itself. Historical trigger rates produced by replacing the settlement algorithm throughout an archive are counterfactual reconstructed rates; they are not necessarily the historical observed payment frequency.

The legal implications of eligibility, fallback, and finality depend on governing law and actual policy wording. Institutional summaries explain mechanisms, but they cannot resolve disputes over particular contracts. The research therefore uses a common mathematical representation of payments and leaves jurisdiction-specific enforceability to a separately supported analysis. The practical value of the frequency estimate consists in making the contract risk quantitative and reviewable rather than supplying a universal legal classification.

Integrated analytical interpretation

The mathematical and comparative evidence supports four connected conclusions. First, observed extreme measurements must be transformed into the unit of payable occurrence before frequency is estimated. A long catalog with poorly specified observation windows can be less informative than a shorter stable record. Second, the event rate, annual exceedance probability, and contract payment frequency are separate quantities even if one model connects them. Third, a financial layer requires a scientifically defensible frequency curve throughout its active range, rather than a single return-period label. Fourth, physically meaningful observation data do not by themselves establish adequacy of loss protection; the joint behavior of index and need remains a separate empirical question.

These conclusions are especially important in internationally distributed insurance programs. A common payout rule across regions can conceal differences in weather observation, historical completeness, physical hazard, or spatial scale. Conversely, a common science of extreme-value inference can be adapted to multiple hazard types when the estimand, event identity and observation regime remain explicit. The transferable content is the logical relation among measurements, occurrence counts, predictive uncertainty, and settlement rules.

A future empirical program would independently collect versioned hazard records and observation calendars, reconstruct contract-relevant indexes for identified locations and coverage periods, and compare predicted frequency curves with withheld data. It would evaluate detection thresholds, record changes and event grouping across hazards, compare candidate tails and dependence models, and match simulated payouts to actual policy determinations only where lawful access exists. The present work establishes the mathematical and evidentiary requirements for such research while avoiding unsupported location-specific frequencies or proprietary contract claims.

Conclusion

Parametric hazard exceedance frequency estimation is the statistical reconstruction of the recurrence of a specified adverse index crossing under a defined event and observation architecture. It combines physical hazard measurement with event catalog construction, treatment of incomplete and censored observations, recurrence and tail estimation, and propagation into financial payment rules.

An annual probability is the appropriate object for the chance of a yearly trigger, while the event rate is the appropriate first-moment object for unlimited per-occurrence payouts. Their Poisson relationship is exact under the stipulated count process and ceases to be a universal conversion when dependence or heterogeneous occurrence changes the distribution of counts. Different definitions of exposure, eligible event identity, measurement duration, geographical support, and threshold inequality must therefore be treated as changes in the estimand.

The analytical examples establish how missing observation periods, cluster grouping, tail shape, and payout caps can generate materially different expected payments under otherwise similar hazard information. The integration of a frequency curve across a bounded linear payout layer demonstrates why attachment rarity alone is insufficient for pricing. A reliable frequency calculation retains source provenance and uncertainty across the full range of contractual severity, including the distinction between actual observations and model-generated events.

The study's comparative evidence demonstrates the varying observational requirements of precipitation, flooding, tropical cyclones, earthquakes and seasonal drought. Primary descriptions of ARC and CCRIF demonstrate an additional modeled-cost or loss transformation in certain sovereign arrangements. Their documentation does not establish policy-specific recurrence rates, and the study's illustrations are synthetic. The resulting synthesis defines a sound basis for research, validation and transparent actuarial reasoning without asserting new empirical calibration or a new statistical theory.

References

  • African Risk Capacity Group. n.d. Africa RiskView. Methodology: Rainfall, Drought, and Tropical Cyclone. Accessed November 18, 2026. Source document.

  • African Risk Capacity Group. n.d. How ARC Works. Section How a Country Gets a Payout. Accessed November 18, 2026. Source document.

  • CCRIF SPC. 2025. Brief on CCRIF SPC. July 2024, revised April 2025. Sections How a CCRIF Policy Is Triggered and Elements of a CCRIF Parametric Insurance Policy. Source document.

  • Davison, A. C., and R. L. Smith. 1990. Models for Exceedances Over High Thresholds. Journal of the Royal Statistical Society: Series B (Methodological) 52 (3): 393–425. DOI: 10.1111/j.2517-6161.1990.tb01796.x. Publisher record.

  • England, John F., Jr., Timothy A. Cohn, Beth A. Faber, Jery R. Stedinger, Wilbert O. Thomas, Jr., Andrea G. Veilleux, Julie E. Kiang, and Robert R. Mason, Jr. 2018. Guidelines for Determining Flood Flow Frequency—Bulletin 17C. U.S. Geological Survey Techniques and Methods 4-B5. DOI: 10.3133/tm4B5. Source document.

  • Ferro, Christopher A. T., and Johan Segers. 2003. Inference for Clusters of Extreme Values. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 65 (2): 545–556. DOI: 10.1111/1467-9868.00401. Publisher record.

  • Gneiting, Tilmann, and Adrian E. Raftery. 2007. Strictly Proper Scoring Rules, Prediction, and Estimation. Journal of the American Statistical Association 102 (477): 359–378. DOI: 10.1198/016214506000001437. Publisher record.

  • Milly, P. C. D., J. Betancourt, M. Falkenmark, R. M. Hirsch, Z. W. Kundzewicz, D. P. Lettenmaier, and R. J. Stouffer. 2008. Stationarity Is Dead: Whither Water Management? Science 319 (5863): 573–574. DOI: 10.1126/science.1151915. University of Washington publication record. Publisher record.

  • National Institute of Standards and Technology. 2023. Conditional Mean Exceedance (CME). Dataplot Reference Manual. Created June 5, 2001; updated December 11, 2023. Accessed November 18, 2026. Source document.

  • National Oceanic and Atmospheric Administration, National Centers for Environmental Information. n.d. International Best Track Archive for Climate Stewardship (IBTrACS). Version 4r01 documentation and data access. Accessed November 18, 2026. Source document.

  • National Oceanic and Atmospheric Administration, National Weather Service, Hydrometeorological Design Studies Center. n.d. HDSC FAQs. Section 3, Precipitation Frequency Analysis Methods and Definitions. Accessed November 18, 2026. Source document.

  • National Oceanic and Atmospheric Administration, National Weather Service, Hydrometeorological Design Studies Center. n.d. HDSC Glossary. NOAA Atlas 14 precipitation-frequency terminology. Accessed November 18, 2026. Source document.

  • U.S. Army Corps of Engineers, Hydrologic Engineering Center. n.d. Expected Moments Algorithm Study Guide: Overview. HEC-SSP Tutorials and Guides. Accessed November 18, 2026. Source document.

  • U.S. Geological Survey, Earthquake Hazards Program. n.d. ANSS Comprehensive Earthquake Catalog (ComCat) Documentation. Accessed November 18, 2026. Source document.

  • U.S. Geological Survey, Earthquake Hazards Program. n.d. Earthquake Hazards 101—The Basics. Accessed November 18, 2026. Source document.