Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Modeled Loss Triggers and Calculation Agent Governance

Abstract. A modeled-loss trigger converts observed characteristics of a qualifying catastrophe into a contractual loss amount by applying a designated catastrophe model to a prescribed reference portfolio. The resulting estimate may determine whether insurance-linked securities (ILS) principal is impaired and how much collateral becomes payable to a protection buyer. Its operative financial variable is a constructed loss, contingent on the contract's selected exposure data, event definition, model, computational procedures, and adjustment provisions. These features make modeled-loss risk transfer simultaneously a problem in catastrophe science, financial contracting, measurement, and institutional governance. This monograph examines the allocation of authority over reference data, model versions, hazard observations, loss computation, attachment, exhaustion, annual resets, calculation agents, verification, disputes, and reproducibility. Its evidentiary core combines a publicly filed 2005 reinsurance agreement specifying escrowed models and calculation-agent procedures; institutional descriptions of catastrophe-bond triggers; the World Bank's 2019 Philippines sovereign modeled-loss bond; an earlier Philippines parametric-insurance evaluation; catastrophe-model actuarial standards; and peer-reviewed work on basis risk and information asymmetry. A formal payout derivation and three explicitly hypothetical sensitivities explain how contract design and measurement perturbations transmit into investor principal risk and sponsor recovery. The synthesis establishes that a trigger is legally and economically defined by the entire calculation protocol rather than by a headline peril or model name. Reference-portfolio changes, model-version changes, event-data revisions, timing cutoffs, and conflicts of interest can cause materially distinct determinations for identical underlying disasters. Contractual finality, independent review, and reproducibility must therefore be designed together. The conclusions are analytical interpretations of established structures and evidence, without claiming a new theory or an empirical causal result.

Keywords: Catastrophe bonds; insurance-linked securities; modeled-loss trigger; calculation agent; catastrophe modeling; event footprint; reference portfolio; escrow model; basis risk; attachment point; exhaustion point; annual reset; model validation; contractual finality; data provenance; independent verification.

Introduction

Catastrophe-linked instruments turn insurance contingency into a specified distribution of capital-market cash flows. A sponsor purchases protection against a defined risk; an investor accepts the possibility of losing part or all of principal in return for remuneration. In many transactions a bankruptcy-remote special-purpose structure holds the collateral, while in World Bank Capital at Risk Notes the institutional issuer assumes a different intermediation role. The World Bank Capital at Risk Notes program describes the latter distinction. The NAIC overview of insurance-linked securities likewise explains securitized catastrophe protection and its place alongside other ILS.

The trigger is the link between the hazardous event and the allocation of capital. Indemnity structures reference the sponsor's own claims; industry-loss structures reference an industry index; parametric structures rely on defined physical measurements; modeled-loss structures apply a loss model to actual event parameters and a reference exposure. The OECD's catastrophe-bond trigger comparison distinguishes these approaches, including hybrids. Lalonde and Poliquin's 2012 catastrophe-bond issuance primer defines a notional portfolio that need not reproduce the sponsor's own book. That distinction supplies the central investigative question: whose loss is measured, by what model, through which process, and with what authority?

A modeled-loss trigger can use reliable external observations and pre-agreed assumptions, permitting determinations less dependent on the sponsor's claims adjustment than an indemnity trigger. Nevertheless, the loss result is sensitive to hazard-field reconstruction, reference data, vulnerability assumptions, model code, numerical settings, contract conditions, and operator choices. An independent calculation agent is therefore a fundamental institutional participant in the promised allocation of risk. The Philippines program evaluation describes the initiator, calculation agent, and escrow agent as separately identifiable post-event roles; its recommendation that a calculation performed by a transaction counterparty receive independent verification provides comparative institutional evidence.

The present analysis studies the public record as it stands on November 15, 2026. Open marketing material is distinguished from executed contractual disclosure; public descriptions of individual catastrophe bonds are not used as substitutes for unavailable offering documents; supervisory principles are interpreted within their institutional and jurisdictional scope. Numerical illustrations below are independently derived hypotheticals, not reported market observations. No access to confidential catastrophe-model code, private bond indentures, nonpublic event determinations, or proprietary reference portfolios is asserted.

Three research questions govern the inquiry. First, which data, rights, and contractual definitions constitute a modeled-loss determination? Second, how do portfolio and model selections allocate basis risk and change investor and sponsor payoffs? Third, which operational and institutional controls make the resulting determination reproducible, contestable, and credible to all parties? The six chapters proceed from contractual evidence through computation and institutional governance to an integrated synthesis.

Definitions, Source Record, and Scientific Boundary

A contractually constructed measure of catastrophe loss

Let e denote a covered event, H(e) the event observations or reconstructed footprint, P the prescribed reference portfolio, V the contractually applicable model version and associated settings, and M the contractual loss-calculation procedure. The modeled loss can be written descriptively as

L_m(e) = M[H(e), P, V, C],

where C includes financial terms, policy conditions, covered geography, treatment of secondary perils, data-cutoff rules, aggregation conventions, currency adjustments, and permitted determinations of the calculation agent. This equation is an analytical representation of an existing modeling workflow, not a newly proposed model. The result L_m is an agreed loss index expressed in currency or another contractual unit; it is not necessarily an estimate of the sponsor's final claims.

This difference matters throughout the contract. The reference portfolio can be a representative or invented exposure schedule. The sponsor's ultimate insured losses L_s depend on the actual portfolio, coverage, legal liabilities, claims handling, inflation, reinsurance recoveries, and post-disaster development. The signed difference L_s - L_m is a convenient analytical indicator of loss mismatch; it is not automatically the contract's definition of basis risk, which properly refers to discrepancies between protection payments and the sponsor's underlying economic loss or hedging objective.

The 2014 SEC-filed fund disclosure distinguishes a catastrophe-model firm's hypothetical-company or model-portfolio losses from industry loss and physical parametric triggers. The OECD (2024) also explicitly recognizes modeled-loss model risk. These classifications are consistent with a precise criterion: the decisive settlement variable is an event-conditioned model estimate of a specified reference loss, rather than the observed sponsor indemnity itself.

The evidence hierarchy

Contractual hierarchy is central to research reliability. The strongest publicly accessible example found for this analysis is the 2005 agreement between PXRE Reinsurance Ltd. and Atlantic & Western Re Limited, filed with the U.S. Securities and Exchange Commission. Its modeled-loss and reset provisions define event parameters, a notional portfolio, the escrow model, a calculation agent, notice dates, the conversion of modeled loss into event loss amounts, and loss-payment notices. The agreement names AIR Worldwide Corporation as calculation agent through the associated calculation-agent agreement. The SEC archive is evidence of the text publicly filed; contractual interpretation here remains confined to the disclosed instrument.

The 2005 agreement is especially instructive because it operationalizes concepts often left abstract in secondary commentary. Following an event notice, its Article VI directs the calculation agent to obtain designated event parameters and withdraw the model and notional portfolio from escrow. A duplicate may substitute for unavailable escrow materials if the agent certifies identity. Article VII permits a defined model-based reset that updates the reference portfolio and determines revised attachment and exhaustion points; Article VIII then converts a modeled result into amounts payable; Article IX establishes loss-payment notice mechanics. This arrangement makes data custody, version identity, time, and calculational authority elements of the economic promise.

The 2019 Philippines catastrophe-bond case study gives a public sovereign counterpart. It identifies two tranches totaling USD 225 million, a modeled-loss trigger, and a stepped payment schedule with 0%, 35%, 70%, and 100% outcomes subject to specified levels. These are issuer-reported transaction characteristics; they do not disclose every model input or substitute for an investor's complete offering terms. The World Bank's 2020 public account independently places the issuance in November 2019 and separates the USD 75 million earthquake protection from USD 150 million tropical-cyclone protection.

A second comparison comes from California workers' compensation catastrophe bonds. The Golden State Re II 2014 directory entry attributes its modeled-loss calculation to a notional workers' compensation portfolio, earthquake severity, geography, building type, time of day, and day of week; the source is an independent industry directory with stated data limitations. A 2011 RMS transaction announcement likewise describes ground-motion observations and time-of-day effects in California earthquake casualty modeled-loss protection. These public accounts demonstrate an important general point without requiring unverified private specifications: the same earthquake footprint can interact with the distribution of people at work, not only the replacement value of damaged buildings.

Scientific scope and exclusions

The investigated object is the modeled-loss determination of contingent payment in ILS, especially catastrophe bonds and directly comparable insurance-linked risk-transfer contracts. Closely related mechanisms enter only to identify differences in observability, incentives, settlement delay, and basis risk. The subject is therefore distinct from conventional excess-of-loss reinsurance claims settlement, catastrophe-bond credit or liquidity pricing as such, and general actuarial model governance detached from contractual settlement rights.

The documentary record does not support assigning one fixed calculation-agent governance template to every ILS transaction. A privately negotiated contract can allocate appointment, oversight, discretion, dispute, and correction powers differently from a publicly filed U.S. agreement or an intergovernmental financial instrument. Classification must be contract-specific and jurisdiction-aware. The international literature and standards inform judgment; their existence does not establish automatic incorporation into a given deal.

Reference Portfolio and Model Version Selection

Defining the exposure that bears the modeled event

A reference portfolio is a structured specification of values and risks against which a modeled event is evaluated. Depending on the contract, entries may include location coordinates, occupancy and construction types, insured values, coverage deductibles, limits, geographic weights, policy conditions, business interruption assumptions, or exposed personnel. Reference exposures can derive from the sponsor's actual database, a synthetic representative book, or an industry exposure database. The Verisk ILS practice description specifically distinguishes client-provided exposure data and proprietary industry exposure databases when performing transaction risk analysis.

The reference portfolio governs two separate tasks. Before issuance it supports estimation of a probability distribution of modeled contingent losses, informing risk spread, attachment level, exhaustion, expected principal impairment, and sponsor hedge fit. After a qualifying event it becomes the loss-measurement denominator and subject matter, holding particular exposure assumptions constant or resetting them under agreed conditions. Confusing these phases obscures a basic design problem: a model and portfolio may price an instrument accurately under the information available at inception while becoming less representative of the sponsor's evolving insured book over the term.

Suppose a hypothetical portfolio contains three zones with replacement-value exposures of USD 600 million, USD 900 million, and USD 400 million. For a specified historical-style event footprint, the illustrative conditional damage ratios are 5%, 12%, and 8%. Without deductibles or secondary losses, modeled zone losses are USD 30 million, USD 108 million, and USD 32 million, totaling USD 170 million. If a hazard-data correction or change in vulnerability mapping revises only the second ratio to 15%, the same exposure schedule yields USD 197 million. The USD 27 million difference has no direct claim-adjustment component; it arises entirely from a modeled-input or assumption change. The figures and calculation are hypothetical, and therefore establish an algebraic sensitivity rather than the empirical magnitude of real catastrophe-model variation.

In a modeled-loss cover attaching at USD 180 million, such a revision crosses the attachment threshold. The relevant question is not merely whether the later run is scientifically preferable; it is which hazard record, vulnerability setting, and version the contract authorizes the calculation agent to use. A discretionary scientific update made without a corresponding contractual rule can alter the distribution of capital after investors have accepted risk. A frozen model that cannot accommodate a demonstrable input error can also frustrate the intended measurement. The source and the timing of the authority to correct data are consequently inseparable from the chosen model.

Locking, depositing, and identifying the model

Model identity is richer than a product label. Reproduction may depend on software release, peril module, hazard catalog, vulnerability functions, exposure-preprocessing routines, financial module, random seeds when simulation is used, numerical precision, supplemental event files, geocoding rules, rounding, and run settings. The 2005 SEC-filed escrow-model provisions provide direct historical evidence that contractual parties have treated model preservation as integral to the agreement. A duplicate's certified identity when escrow materials are unavailable is an unusually clear acknowledgement that custody failure creates settlement risk.

The Actuarial Standards Board's ASOP No. 38 sets U.S. professional guidance for actuaries selecting, using, reviewing, or evaluating catastrophe models. Its 2021 revision, effective December 1, 2021, emphasizes suitability, input review, output validation, limitations, communications, and documentation. It binds covered actuarial professional work in its defined scope; it is not by itself the governing contract of every catastrophe bond, nor a globally enacted insurance regulation. Nevertheless, it gives a professionally grounded reason to document model version and material adjustments.

A robust contract should distinguish five types of change as an analytical governance recommendation: a correction of demonstrably erroneous source data; an authorized periodic exposure reset; a vendor model upgrade; a change of financial or coverage terms; and a replacement of the designated calculation agent. Each can have different notice, approval, consent, or investor-protection consequences. The evidentiary foundation is the revealed reset and escrow structure, the model standards, and the economic impact of sensitivities; the list is a practical reading of existing contractual problems rather than a universal statutory checklist.

Scheduled portfolio resets

A portfolio that was representative at inception may cease to track changing geographic concentrations, insured values, inflation, construction, occupancy, or business volumes. An exposure reset can improve hedge matching but also change the bond's conditional risk. A reset is economically consequential even when nominal principal and coupon spread remain unchanged. A sponsor with authority to select data after issuance may prefer revised parameters that increase expected protection; investors need to understand how the contract constrains that authority.

The filed 2005 PXRE arrangement shows one explicit resolution. Article VII provides a defined right to deliver reset portfolio data meeting data criteria. The calculation agent computes a revised loss distribution and specified reset attachment and exhaustion measures using the escrowed model, with an identified notice and reporting sequence. This is a source-specific fact, not evidence that all modeled-loss contracts reset annually or adopt that precise timing. The example demonstrates that change control can be specified in the same instrument that allocates principal risk.

For a reproducible reset, governance should retain the previous and new portfolios, define the effective risk period, document why the reset was permissible, and preserve evidence of the transformation from exposure input to attachment and exhaustion measures. A disagreement over which dataset applied to a catastrophe near a reset boundary cannot be settled by comparing only headline output numbers. It requires the event date, coverage inception and expiration, calculation chronology, and effective-version record.

Representative risk and sponsor hedge fit

A notional portfolio creates a trade-off. Its abstraction can reduce disclosure of sensitive policyholder exposure information and constrain interference by the sponsor; at the same time it introduces basis risk if the sponsor's actual losses respond differently to the same catastrophe. The OECD's 2024 assessment discusses modeled-loss model risk and the broader trigger trade-offs. Lee and Yu's analysis of moral hazard and basis risk provides a theoretical financial-economics treatment of how these conditions can affect cat-bond valuation.

The comparison should be conditional, rather than a claim that one trigger type always dominates. Sponsor losses and index losses can differ because of exposure composition, vulnerability, exclusions, operational interruption, reporting methods, inflation, or judicial developments. An indemnity trigger follows contractually eligible actual loss but may entail long claims-development periods and additional claims-handling incentives. A parametric trigger offers an observable input rule but may have substantial financial mismatch. Modeled loss inserts calibrated modeled consequences between physical event observations and settlement, making its hedge quality dependent on both event science and exposure representativeness.

An economically useful pre-issuance study therefore compares sponsor losses and reference-portfolio index losses under the same simulated event set, using a consistent risk period, loss basis, and policy terms. Expected payment alone is insufficient: conditional shortfalls during the sponsor's most severe loss scenarios often matter more to liquidity and capital objectives. Correlations calculated from a common model can be overstated when model misspecification affects both series. Re-running with a second modeling approach and stress scenarios is a defensible sensitivity test, consistent with ASOP No. 38's output-validation discussion.

Event Footprint and Modeled-Loss Computation

From reported catastrophe to model-ready event

The event footprint is the physical description required to estimate losses: a spatial field of wind speed, ground motion, rainfall, flood depth, or other covered intensity, together with temporal and peril-specific attributes. Such fields can combine instrument observations, agency bulletins, interpolation, numerical hazard models, and engineering assumptions. Earthquake magnitude and epicenter alone do not specify shaking at every insured property; cyclone track, central pressure, and wind measurements alone do not uniquely determine localized damage. Reconstructing a loss therefore requires a prescribed chain of transformations.

In the 2005 SEC-filed agreement, earthquake event parameters are specified by reference to an identified U.S. Geological Survey data provider and include date and time, moment magnitude, epicenter, focal depth, rupture azimuth, and dip angle. The agreement also defines an event-parameters date and event notice period. Those terms matter because physical measurements may be refined after the initial alert; the contract must decide whether the calculation uses preliminary, updated, or fixed-date values. A reported revision can alter payout even if the underlying physical event has not changed.

The chain from observation to payable loss can be represented in stages: identify a covered event and period; obtain authorized event characteristics; construct the prescribed spatial and temporal footprint; allocate hazard intensity to reference exposure; apply the designated vulnerability and financial models; aggregate eligible loss; and apply the contract's attachment, exhaustion, step, and payment conventions. Each stage has distinct potential errors. Combining all stages into one unqualified number labeled "model output" conceals the location of the error and the party authorized to correct it.

Payout algebra: a transparent proportional layer

Consider a hypothetical, single-event proportional modeled-loss layer with modeled loss L, attachment A, exhaustion E, and maximum principal exposed N. Assume E is strictly greater than A and that there are no deductibles, reinstatements, annual aggregation adjustments, separate loss ratios, stepped triggers, or other contract-specific conditions. The principal-loss fraction is

q(L) = min(1, max(0, (L - A)/(E - A))).

The investor principal reduction is N q(L), and the protection payment is N q(L), assuming the financial contract passes that reduction entirely to the sponsor without additional adjustments. Outstanding investor principal is N[1 - q(L)]. This simple layer equation is an illustrative interpretation of a proportional attachment-to-exhaustion mechanism. Actual contracts may prescribe different loss-payment factors, ratios, and time conventions, as does the 2005 PXRE instrument.

Set N = USD 100 million, A = USD 250 million, and E = USD 500 million. Modeled losses of USD 230 million, USD 300 million, USD 425 million, and USD 500 million produce principal reductions of USD 0, USD 20 million, USD 70 million, and USD 100 million respectively. These amounts follow solely from the declared hypothetical formula. A model revision that changes L from USD 300 million to USD 330 million increases investor principal loss from USD 20 million to USD 32 million—a USD 12 million consequence of a USD 30 million modeled-loss change. Within the layer, the marginal contract payment is N/(E - A) = 0.4 dollars per dollar of modeled loss. Near attachment and exhaustion that sensitivity changes abruptly at the kink.

The piecewise structure reveals how model governance becomes financially nonlinear. A small modeled-loss change below attachment may have no effect; the same change can move an event across attachment and start a payout; above exhaustion further increases produce no additional principal loss. Sensitivity reporting should examine these boundaries rather than relying on a single expected modeled loss. Threshold proximity also magnifies incentives to contest seemingly minor data decisions.

For a hypothetical sponsor loss L_s = USD 360 million in this example, a modeled index L = USD 300 million produces a USD 20 million payment, whereas L = USD 425 million produces USD 70 million; the actual sponsor loss is unchanged. Both differ from what a hypothetical indemnity trigger could provide under its own terms. Because the policy details of any indemnity comparator are absent, this numerical demonstration establishes basis-risk transmission only; it makes no claim about the fair or efficient payment for any real sponsor.

Stepped structures and sovereign evidence

The Philippines 2019 structure illustrates why proportional-layer algebra must not be assumed universal. The World Bank's transaction case study specifies step outcomes of 0%, 35%, 70%, and 100% according to modeled-loss thresholds, rather than continuous interpolation. At a hypothetical USD 100 million notional amount, the step factors would yield USD 0, USD 35 million, USD 70 million, and USD 100 million, but the actual 2019 issuance consisted of USD 75 million and USD 150 million classes. The hypothetical USD 100 million computation must not be read as a reported payout or term sheet for either class.

A stepped payout creates regions of zero local sensitivity separated by threshold jumps. It may be easier to communicate certain pre-agreed discrete outcomes and may respond differently to noisy loss estimates. Yet an observation revision near a step threshold can shift the payment by a large amount. Threshold precision, the permissible rounding rule, and the final data cutoff thus acquire considerable importance. A stepped contract can be administratively simple to explain while relying on a technically elaborate modeled loss to decide which step applies.

The Philippines had earlier experimented with modeled-loss-based parametric insurance. The World Bank's 2020 lessons-learned evaluation describes a distinct 2017–2018 pilot with a partial-payout setting of 40% under its particular design and reports that modeled-loss calculations proved difficult for stakeholders to understand. This prior program should not be conflated with the later 2019 capital-markets bond or its 35% and 70% steps. Its value for this study is institutional: transparency requires more than fast payments or familiar terminology; stakeholders must be able to trace the event, model, rule, and payout.

Event aggregation, timing, and financial transformations

A qualifying event is also a legal unit. Earthquake clusters, multiple storm landfalls, aftershocks, consecutive rainfall episodes, and hours clauses can determine whether observations are combined or treated separately. In the 2005 instrument, Article VIII includes detailed treatment of related California earthquakes within an event period and applies a defined maximum-loss treatment to certain clusters. That rule prevents an analyst from summing all modeled loss results as if every event necessarily generates a separate full attachment and payment.

Other financial transforms can also matter. The filed agreement sets contract-specific European windstorm currency conversion conventions and ceilings or floors for certain conversions. These clauses demonstrate that an event's nominal modeled damage and the eventual note principal reduction may depend on more than the physical model. Financial module settings, currency dates, stated payout ratios, aggregate limits, and payment calendars are part of the operative function.

A research study that checks catastrophe-model output against a final note loss amount should therefore reconstruct intermediate amounts: raw hazard-driven damage; net modeled loss under contract conditions; event loss above attachment; applicable payout ratio; currency transformation; previous principal reductions; and timing-dependent notices. The SEC-filed loss-payment notice form supplies a documentary example of reporting several such amounts separately. This approach allows the scientific and financial components to be independently challenged.

Expected loss and tail quantities

For a transaction's risk analysis, a simulation produces a distribution of possible modeled principal reductions under the event catalog and annual risk assumptions. Let Q denote the fraction of principal lost in a risk year; the modeled annual expected principal loss is E[Q], and modeled annual expected dollar loss is N E[Q]. Attachment probability is Pr(Q > 0), while exhaustion probability is Pr(Q = 1). These are different statistics. A transaction may have low attachment probability and high conditional severity, or frequent partial losses and rare total losses, even at the same expected loss.

The Verisk risk-analysis description explains the use of simulated and historical modeled losses for trigger design and basis-risk work. The AIR discussion of global modeled losses makes an important measurement distinction: average annual losses can be aggregated over appropriately consistent regions, whereas regional exceedance-probability points are not generally additive. The distinction carries over to trigger studies involving multiple geographic portfolios.

Expected loss is model-conditional, not a directly observed long-run frequency for each bespoke structure. Hazard catalogs, historical data, exposure changes, vulnerability curves, correlation and nonstationarity all matter. The 2026 preprint by Ery and Koch on parametric-modeled windstorm loss illustrates current research on transparent hazard–vulnerability–exposure design, but is identified here as a preprint, not as independently peer-reviewed validation of a settled market method.

Calculation Agent Oversight and Independence

Institutional tasks and agency relationships

A calculation agent executes the defined decision procedure on behalf of, and under the constraints of, the transaction documents. The role can encompass obtaining physical event parameters, loading preserved models, validating submitted exposure data, running catastrophe calculations, performing contractual payment arithmetic, certifying assumptions, and distributing notices to designated parties. Its powers are those actually conferred by the instrument. It should not be presumed to have general discretion to improve a model, redefine a covered event, or negotiate an economically fair result after the event.

The SEC-filed 2005 contract is concrete. The designated agent follows the escrowed data process and calculation-agent agreement, receives event and reset notices, and communicates modeled loss, event loss, payment, and principal reduction. Its reset timetable and notice obligations show that authority, reporting, and chronology can be contractually specified in measurable terms. These are enforceable terms to the extent provided by the actual relevant contracts and applicable law; the study does not opine on their ultimate enforceability in a particular dispute.

The initiator decides when to request the calculation under the applicable contract, the independent data provider supplies physical observations, the custodian or escrow agent controls preserved model materials, and the calculation agent performs the computations. One corporate entity can combine roles if permitted, but concentration creates distinct conflicts. The World Bank Philippine pilot evaluation discusses precisely this arrangement and the logic of independent calculation or independent verification.

Independence, incentives, and reliance on experts

Independence is a functional property, not merely the adjective assigned to an external vendor. Material considerations include who appoints and pays the calculation agent, which party can instruct it, whether its model was also used in deal structuring, whether it can change methodology after inception, who can request recalculation, what information it may withhold as proprietary, and how the agent may be replaced if unavailable or conflicted.

The same specialist may advise on risk modeling at issuance and calculate losses after an event. The 2011 RMS Golden State Re account describes RMS as both a risk modeler and calculation agent for that deal. This demonstrates role combination, not misconduct. Such combination can improve technical consistency, but it makes independent reproducibility, clear constraints on discretion, and conflicts management especially relevant.

In a contract with uneven informational access, the sponsor may know more about its changing portfolio than investors, while the model provider may know more about proprietary code than either party. Potential conflicts are not automatically eliminated by putting an unaffiliated company between them. An agent with a fixed fee but renewal dependence on the sponsor can face different incentives from one appointed jointly by transaction parties; those are analytical possibilities requiring deal-specific evidence before being attributed to any real actor.

Lee and Yu (2002) examined the pricing significance of moral hazard and basis risk. Götze and Gürtler's empirical study of ILS moral hazard focuses on sponsor incentives in indemnity catastrophe-bond structures. Their results should not be transferred mechanically to modeled-loss agents. The relevant insight is that trigger selection reallocates control and observability; evaluating one agency problem does not remove every other agency problem.

Professional oversight and jurisdictional qualifications

The International Association of Insurance Supervisors' Insurance Core Principles and ComFrame, adopted in December 2024, address risk management, internal controls, and outsourcing governance. These are international supervisory standards, implemented differently through national insurance regimes; they are not automatically binding on every catastrophe-bond agent or issuer. Their most useful relevance is where an insurer has outsourced a material risk-management or control function and remains accountable under applicable regulation.

ASOP No. 38 directs an actuary within its professional scope to evaluate models and reliance on experts and document material adjustments and limitations. It does not transform an actuarial modeler into a contract adjudicator. The ASOP requirements for output validation support periodic reconciliation and independently documented professional judgment while leaving specific investor rights to the transaction documents.

A credible governance arrangement may separate three functions: scientific quality assurance of the hazard and loss model; contractual determination of the amount payable; and independent review of whether the contractually prescribed process was followed. Although these functions can share information, they answer different questions. A technically improved model result is not necessarily the contractually correct result; a properly executed contractual algorithm may still generate a poor hedge for the sponsor. Both facts should be visible to the affected parties.

Oversight controls through the determination cycle

Before issuance, counterparties can specify qualifications, prohibited conflicts, permitted subcontracting, version custody, input responsibility, validation requirements, and termination or succession. At each exposure reset, the agent can document input receipt, adherence to data criteria, authorized effective date, calculation outputs, and approval notices. After an event, an auditable record can identify the source and timestamp of observations, exact event identifier, frozen configuration, calculations performed, exception decisions, amounts, and delivery dates.

The World Bank pilot's post-event workflow clarifies that initiation, calculation, and custody should be specified before a loss. The 2005 escrow agreement illustrates an early contractual solution to model-preservation risk. Together, these sources justify the analytical recommendation to assign explicit custody and review responsibilities to every material digital artifact in a modern calculation chain.

Escalation provisions should state which disagreements are technical errors, which are definitional disputes, who can request corrections, the deadline for doing so, whether uncontested amounts can be paid, and what happens if a model vendor ceases operations. The 2005 agreement supplies evidence that duplicates and successors can be addressed by contract. It does not establish that a particular independent expert review, court, or arbitration forum is universally required. Those rights must be read from the relevant instrument and governing law.

Contract Dispute and Reproducibility

Five classes of potential dispute

Disagreements can arise even when every party agrees that a severe catastrophe occurred. First are event-identity disputes: whether the observations belong to the defined peril, geographic area, event period, and accumulation unit. Second are data disputes: which observations, provider, revision date, interpolation, and data-quality exceptions control. Third are exposure disputes: whether the prescribed reference portfolio was accurately preserved, updated, or validly reset. Fourth are modeling disputes: whether the right version, configuration, and financial assumptions were used. Fifth are payment disputes: whether calculated modeled loss was properly translated into the contract's attachment, step, aggregate, currency, and payment provisions.

The 2005 filed contractual definitions and notices allow a reader to see several classes separately: the event-parameters date, escrowed model, reset data criteria, event loss amounts, currency conventions, and formal loss-payment notice. A party contesting the final figure should identify the specific input, rule, or intermediate calculation alleged to be incorrect. A generalized objection that an event caused greater or smaller actual damage is insufficient where the contract selected a notional modeled-loss mechanism.

Reproducibility as an evidentiary property

Reproducibility means that a qualified person possessing the authorized inputs, version, methods, contract terms, and computational environment can regenerate the determination within stated numerical tolerances. It is broader than obtaining the same rounded payout from different software versions. If two runs coincide only after undisclosed manual adjustments, their coincidence does not establish an auditable calculation.

The minimum technically meaningful record consists of the legally controlling text and amendments; reference-portfolio snapshot; model release and configuration; hazard data and documented cutoff; preprocessing steps; all material financial terms; run log and intermediate outputs; final arithmetic; exception and signoff record; and notices to affected parties. Where proprietary constraints prevent full public disclosure, evidence can be escrowed or made available under contractual confidentiality controls. The point is preserved inspectability by authorized reviewers, not an assumption that every commercial model must be publicly open source.

Cryptographic hashes of input datasets, software packages, and output reports can establish identity of archived bytes, provided the materials were appropriately captured and their provenance is independently documented. A SHA-256 value alone cannot prove that inputs were correct or chosen in accordance with the contract. Signed event data and retained versions may allow the agent's claimed process to be compared with the actual executed process. A change log should distinguish algorithm upgrades from error corrections and investor-approved amendments.

The earlier Philippines pilot supplies a warning about interpretability: the World Bank evaluation's lessons on comprehension report difficulties understanding modeled-loss-based triggers. Governance therefore requires two levels of explanation: a concise public or investor-readable description of what causes payment and a detailed technical record capable of independent recalculation. Neither level substitutes for the other.

The legal boundary of finality

Many financial contracts specify how a calculation agent's determination becomes final and which errors or misconduct permit challenge. The present source set does not establish that all modeled-loss deals confer an identical standard of finality, a single dispute resolution forum, or an overarching legal duty of independence. The parties' express terms, legal characterization of the risk-transfer arrangement, choice of law, governing securities documentation, and supervisory conditions determine their rights.

The difference between a contractually wrong result and a scientifically inaccurate model deserves emphasis. The calculation agent can correctly implement a frozen model that understates a real catastrophe; the sponsor's resultant under-hedging is a basis-risk outcome, not necessarily a breach. Conversely, the agent can use a newer scientifically preferred footprint when the contract requires fixed-date data; that may be a contractual error despite its plausible physical accuracy. Courts, arbitrators, independent experts, and contractual review panels address different questions according to their jurisdiction and mandate.

A valid research conclusion must therefore preserve conditional language when describing the resolution of actual disputes. Without an individual bond's complete operative documents and applicable law, this work can specify governance issues and testable questions but cannot pronounce contractual liability. This limitation is particularly important when public deal summaries omit nonpublic definitions, agency agreements, and side letters.

Model failure, vendor cessation, and emergency continuity

A catastrophe can disrupt the same communications infrastructure needed to observe and model it. Missing station data, delayed agency revisions, conflicting geospatial files, inaccessible compute facilities, and unavailable vendors can interrupt determination. The escrow and certified-duplicate provisions in the SEC-filed 2005 agreement show that some continuity risks were anticipated explicitly. A contemporary arrangement can additionally define alternative data hierarchies, contingency compute environments, disaster recovery, and succession rights without giving one party unilateral power to substitute economically different rules.

Independent reproducibility requires separation of technical capability and contractual authority. A second model vendor may be technically capable of estimating loss but lacks the right to determine payout unless the contract empowers it. A replacement agent may hold legal authority but require access to licensed models and archived datasets. Both legal entitlement and practical access must survive the failure of the initial provider.

The research implication is operationally concrete: design the evidentiary chain at issuance. Once the physical event has happened and counterparties face opposing financial incentives, reconstructing a missing reference dataset or an undocumented model configuration may be costly or impossible. Retrospective procedural negotiation can become a conflict over capital rather than a neutral repair of scientific measurement.

Integrated Comparative Synthesis and Research Implications

Comparison across observable instruments

The 2005 PXRE–Atlantic & Western agreement, 2011–2014 Golden State transactions, and the 2019 Philippines catastrophe bond provide complementary cases. The first documents fine-grained legal machinery for escrow, reset, calculation, and notice. The second illustrates modeled exposure to earthquake-related workforce injury, including time-dependent vulnerability rather than building property loss alone. The third illustrates sovereign fiscal protection in which modeled-loss thresholds feed discrete payment steps. Their institutional contexts differ, but the comparison supports a common descriptive fact: observed disaster data alone do not determine principal loss without the preselected model, portfolio, and financial rule.

Evidence case Publicly evidenced trigger feature Governance question illustrated Evidentiary boundary
PXRE–Atlantic & Western, 2005 Escrow model, reference portfolios, data-provider parameters, reset clauses, calculation-agent notices Identity, custody, timing, reset authority, payment trace Filed agreement; not a template imposed on other deals
Golden State Re, 2011 Earthquake ground motion and time-sensitive modeled worker exposure Hazard reconstruction and casualty-vulnerability choice Modeling-agent public account
Golden State Re II, 2014 Notional workers' compensation loss with geographic and temporal factors Representative portfolio and third-party calculation Industry deal-directory summary
Philippines sovereign cat bond, 2019 Modeled-loss basis, USD 75m / USD 150m classes and stepped payouts Sovereign risk transfer and threshold sensitivity World Bank case study; confidential model specifics unavailable
Philippines parametric insurance pilot, 2017–2018 Modeled-loss-based parametric trigger and specified calculation roles Comprehension and agent independence Distinct earlier insurance program, not the 2019 bond

The cases do not warrant market-wide causal estimates or a universal ranking of governance methods. They establish the real existence of several institutional choices. The strongest contractual evidence concerns a 2005 instrument and therefore cannot be treated as proof of current standard terms. The later sovereign and workers' compensation descriptions broaden the comparison but provide less complete investor-level documentation.

The economics of control rights

A bondholder's exposure is affected by two distributions: the physical distribution of catastrophe events and the distribution of contractually determined results produced by model and governance choices conditional on those events. At issuance, a risk model generally treats the second as fixed by specification. If important inputs or agent decisions are ambiguous, additional uncertainty enters the investor payoff. The distinction can be expressed in elementary statistical terms through conditional variance:

Var(Q) = E[Var(Q | H)] + Var(E[Q | H]),

where Q is the principal-loss fraction and H represents observed physical hazard information. The first term captures uncertainty remaining after the selected hazard characterization; it can include exposure or model uncertainty when those are random under a study's assumptions. The second captures variability across hazard states. This is the standard law of total variance applied for interpretation, not an original theory and not an estimable empirical decomposition without data.

Contractual fixes to the reference portfolio, model and error-handling rules can reduce avoidable uncertainty about which Q will be determined from a given authorized H. They do not eliminate underlying catastrophe-model misspecification or guarantee perfect sponsor hedge matching. Conversely, a completely fixed but poorly selected proxy may produce consistent calculations and poor protection. The governance objective is therefore joint: contractual reproducibility and ex ante risk-transfer suitability.

Materiality should be measured near payment boundaries

The hypothetical calculations show why a model change expressed as a small proportional change in industrywide loss may still be material to a particular bond. Materiality depends on its impact on the contract's payment function, not merely its effect on modeled gross damage. At a step threshold, an arbitrarily small revision under exact mathematical assumptions can move the payout between discrete levels; at a proportional-layer boundary, it can change whether a loss is payable. In practice, finite measurement precision and contract rounding rules control how near-boundary revisions are handled.

A defensible sensitivity appendix for any deal should report, at minimum, the principal effect of plausible event-parameter revisions, alternative vulnerability functions, reference exposure changes, and timing or currency rules. Such work can prioritize cases where uncertainty matters most to investor rights and sponsor liquidity. The ASOP No. 38 validation guidance provides an established professional basis for sensitivity and reasonableness checks within its professional scope.

Information asymmetry and the limits of outside verification

External modeling adds specialized expertise but can make the determination opaque when models and reference portfolios are proprietary. An investor can know the principal-loss function and still be unable to reproduce the modeled variable. Independent experts may test process compliance without access to every source-code element if they have controlled access to archived software and inputs, but their assurance scope must be made explicit. A purely public summary can support market comprehension without demonstrating full verification.

Under the IAIS governance principles, insurer accountability for material outsourced functions remains relevant where implemented by applicable supervisory law. A bond issuer's contractual calculation agent, an insurer's outsourced actuarial function, and a securities trustee have different legal roles; institutional analogy does not merge them. The research conclusion is to identify who owes what duty to whom in the actual transaction, rather than treating the generic term "independent" as sufficient disclosure.

Limits of quantitative and comparative inference

The illustrations establish arithmetic consequences conditional on chosen inputs. They do not estimate event probabilities, calibrate a catastrophe vendor model, reconstruct claims, or calculate an actual 2005 or 2019 bond payout. No market-wide database was independently audited. The SEC contract is a historical filing; the World Bank case study is an issuer description; Artemis transaction pages are industry secondary records; Verisk and AIR materials describe vendor practice; the OECD is institutional synthesis; and the cited peer-reviewed papers address economic models or empirical questions with their own samples. Treating all these sources as interchangeable would produce false confidence.

Current research can make the question of transparency more precise. The 2026 Ery and Koch preprint examines a transparent parametric-modeled loss design for German windstorm. Its status and scope indicate an active methodological direction, not evidence that commercial catastrophe bonds generally use its proposed technique. Future empirical work should compare frozen contract documentation, subsequent permitted model resets, actual event data revisions, and realized sponsor losses. Such data would permit an estimate of the distribution of settlement delays and model-driven basis risk rather than a merely theoretical account.

The research agenda implied by the evidence

Four research tasks follow. A comparative contract corpus can code whether the model is escrowed, whether inputs are independently sourced, how resets operate, how often model upgrades are permitted, and who possesses challenge rights. A transaction-level event study can compare initial and revised hazard inputs with actual issued calculation notices. A cross-model exercise can examine how multiple sanctioned configurations generate different loss amounts under identical reference exposures. A governance study can assess whether independence protections and dispute procedures correlate with observed correction, settlement time, and basis risk, after controlling for peril, trigger type, and contract vintage.

These are proposed research directions, not claims of findings already established. Public documentation is uneven; confidentiality may prevent full causal analysis. The strongest attainable contribution of the present dossier is to articulate testable units of comparison, trace the documentary and mathematical sources of modeled-loss determinations, and make clear where general evidence ends and contract-specific legal authority begins.

Conclusion

Modeled-loss triggers allocate catastrophe risk through a prescribed computational determination. Their contractual object is the loss that an agreed model assigns to an agreed exposure portfolio under authorized observations and financial conventions. This structure can transfer defined catastrophe exposure without awaiting complete adjustment of the sponsor's actual insurance claims, while exposing both investor and sponsor to the effects of model choice, exposure representativeness, cutoff dates, and contract interpretation.

The primary record demonstrates that reference-data custody and calculation-agent procedures are substantive elements of risk transfer. A 2005 SEC-filed contract identifies escrowed model materials, event data providers, reset rights, and loss-payment notices. Later modeled-loss structures extend the practice to time-dependent workers' compensation exposures and to sovereign protection using stepped payout levels. These cases illustrate institutional diversity and confirm the need to analyze the full determination chain in each instrument.

The mathematical analysis finds that changes in modeled loss can have disproportionate payment consequences near attachment, exhaustion, and step thresholds. This is an ordinary implication of contingent-contract mathematics and supplies a concrete test for model-change materiality. A reproducible technical calculation, a valid contractual determination, and an effective hedge of sponsor losses are related properties with separate evidentiary tests.

Calculation-agent governance is consequently a matter of authority, controlled data, documented computation, independence, correction rights, continuity, and reviewable notices. Professional and supervisory principles help evaluate these arrangements within their proper spheres; the particular contract and applicable law establish the operative rights. A sound analytical assessment traces the route from event observation to final principal reduction, records each authorized intermediate transformation, and compares the resulting protection with the exposure actually intended to be hedged.

This research contributes an original documentary and mathematical synthesis of modeled-loss trigger governance. It claims no novel theoretical system, no independently calibrated catastrophe probability distribution, no confidential contract reconstruction, and no external peer review. The finished scientific findings remain available for the project's controlled research-package and later separately authorized publication process.

References