RESEARCH DOSSIER
Insurer Emerging Risk Identification and Horizon Scanning
Abstract. Emerging risk identification is the continuous institutional practice through which an insurer recognizes possible material changes in the sources, transmission, concentration, and consequences of risk before those changes are sufficiently established to enter routine actuarial measurement. Horizon scanning supplies observations and alternative futures; enterprise risk management determines how evidence becomes interpretation, prioritization, decision authority, and learning. This monograph examines that transition as a research problem in insurance governance, integrating international supervisory texts, risk-management standards, corporate foresight scholarship, expert judgment research, actual emerging-risk publications, and contemporary insurance-sector monitoring. Its analysis differentiates an observation from a hypothesis, a scenario from a forecast, evidence strength from financial materiality, and supervisory guidance from jurisdiction-specific law. It develops a practical comparison of signal acquisition, epistemic assessment, insurer-specific exposure mapping, proportionate escalation, action selection, and retrospective evaluation using established techniques rather than instituting a new theoretical system. Three illustrative applications consider generative artificial intelligence and cyber concentration, extreme heat and health liabilities, and uncertain long-tail casualty exposure. Decision thresholds are derived under openly stated hypothetical assumptions, while real supervisory observations retain their measured scope. The synthesis finds that effective horizon scanning is a governed conversion of uncertain information into timely and revisable insurance decisions, with attention to correlated exposures, false negatives, source provenance, conflicting judgments, and the cost of premature intervention. It also specifies falsifiable performance measures and limitations for empirical study. Scientific claims are attributed to the verified original sources; proposed procedures and numerical examples are identified as analytical interpretations.
Keywords: Emerging risk; horizon scanning; insurance; insurer enterprise risk management; weak signals; strategic foresight; expert elicitation; materiality screening; ORSA; risk taxonomy; risk interdependence; risk appetite; scenario analysis; data provenance; uncertainty; early warning; escalation; governance; cyber accumulation; climate risk; systemic risk.
Introduction
An insurer makes long-lived promises under distributions of loss that change through technology, legislation, climate, behavior, geopolitics, demographic structure, and the organization of markets. An event that appears remote when a contract is priced can become consequential while the insurer continues to collect premiums, invest assets, provide coverage, settle claims, and renew business. New underwriting conditions also transmit through supply chains, litigation, capital markets, counterparties, and the firm's own operational infrastructure. A scientific account of emerging risk therefore begins with exposure and causal pathways, not with a popular ranking of alarming events.
The International Association of Insurance Supervisors (IAIS) situates risk identification and interdependencies within Insurance Core Principle 16, including enterprise-wide risk types, their characteristics, and aggregated financial impact. ISO's general risk-management guidance connects risk identification with assessment, treatment, communication, and review. The specialized emerging-risk technical specification complements that broader approach. These texts help locate horizon scanning within insurer management; they do not themselves provide empirical validation of every proposed scan frequency, scoring rule, or information source.
The historical insight that organizational surprise may arise despite apparently sophisticated planning dates at least to Ansoff's (1975) response to weak signals. Later futures scholarship questioned how a weak signal is even recognized: Hiltunen (2008) highlighted the ambiguity of future signs. The insurer-specific challenge is more demanding than recognizing novelty. An observation must be translated into the exposure of a particular book, legal entity, asset class, policy wording, territory, risk layer, or operational dependency, and then acted upon under identifiable authority.
The inquiry addresses three linked questions. First, what precisely is identified when an insurer discovers an emerging risk, and how does this object differ from a known loss factor or an attractive narrative? Second, how can heterogeneous expert judgments, statistical anomalies, credible documents, and scenarios inform materiality without false precision? Third, how do boards, risk committees, actuarial and underwriting functions, and legal-entity management transform uncertain indications into proportionate and reviewable decisions?
Methodologically, this is a documentary and analytical monograph. Evidence consists of original institutional standards and official regulatory materials; primary industry disclosures describing their own scanning practices; scholarly publisher records of relevant studies; and sectoral supervisory observations. Source-access distinctions are recorded in the accompanying evidence file: where a journal article, technical standard, or book is restricted, the claim is limited to the publicly verified abstract, official metadata, or publisher description. It does not use confidential carrier data, perform a representative survey of insurers, estimate new causal effects, or assert a newly verified authorial theory. All decision examples are explicitly hypothetical. The study's international coverage demands a clear distinction between worldwide IAIS principles, voluntary standards, EU obligations and supervisory guidelines, and U.S. state-based implementation.
Definitions, History, and Documentary Foundation
Scientific object and temporal scope
An emerging insurer risk is a potentially material adverse deviation in an insurer's outcomes arising from a newly identified source of risk, a changing condition surrounding a familiar source, a new pathway of propagation, or a new concentration of exposures, where the relevant consequences or parameters remain incompletely understood. The International Risk Governance Council (IRGC) explicitly treats both novel conditions and familiar risks in new contexts as emerging. This description establishes the subject of investigation; it leaves the existence, severity, and timing of any particular risk to substantive evidence.
Horizon scanning is the deliberate, repeatable observation of relevant developments beyond currently modeled loss experience. A scan identifies candidate signals across scientific publications, supervisory announcements, court decisions, technical incidents, competitor and reinsurer disclosures, policyholder behavior, macroeconomic data, demographic studies, and internal claims or operational anomalies. The resulting portfolio of hypotheses can be organized over a short operational horizon (weeks to months), medium business-planning horizon (one to three years), and long contract or structural horizon (five to several decades). These are illustrative working bands, not universal regulatory limits. Life and long-tail liability business may require considerably longer evaluation than an annually renewable product.
A signal is an observed datum, statement, pattern, or event with attributable source and observation date. A risk hypothesis is a proposition that links that observation to a causal channel and a potentially material insurer consequence. A scenario is a coherent conditional account of how interacting assumptions could generate that consequence. A forecast assigns a probability, distribution, or predictive ranking under a stated forecasting method. An early warning is a monitored indicator whose change is connected to a pre-agreed decision. These distinctions prevent a headline from automatically becoming a quantified exposure.
Horizon scanning should also recognize beneficial opportunities: emerging protective technology can reduce loss severity or enable coverage. Nevertheless, a favorable opportunity can conceal an emerging concentration or model risk. In the present dossier, the core outcome remains the identification and governance of adverse insurer exposure, including the trade-offs of innovation.
Intellectual development
Ansoff's weak-signal work provides an early argument that management should build graduated responses to incomplete strategic information rather than wait for definitive proof. Hiltunen's semiotic analysis indicates why one observer may perceive novelty while another sees noise. Structured foresight then adds explicit alternatives; Schoemaker (1995) described scenario planning for strategic thinking as a means of confronting uncertainty and overconfidence. The 2024 UK Government Office for Science Futures Toolkit summarizes methods that can be adapted to institutional foresight. These are methodological sources, not demonstrations that any one insurer's scanning program reduces insolvency risk.
The IRGC's (2015) emerging-risk governance guidelines join scanning to adaptable response options and identifiable intervention points. Its later systemic-risk governance guidelines place system boundaries and transitions at the forefront of inquiry. The 2018 document's own research focus concerns physical and socio-ecological systems; applying its method to financial and reinsurance contagion requires explicit analytical extension rather than misrepresenting it as an insurance solvency rule.
Normative location within insurance supervision
IAIS ICP 16 is internationally influential supervisory material. Its treatment of ERM, risk appetite, risk relationships and ORSA forms a defensible institutional foundation for emerging-risk oversight, but its legal effect depends on domestic implementation and supervisory powers. In the European Union, Solvency II Article 45 requires undertakings to consider short- and long-term risks they face or may face, overall solvency needs, and deviations from solvency capital model assumptions. The article establishes a forward-looking legal demand without prescribing a single scanning algorithm. The EIOPA ORSA guidelines further situate assessment in a company's own risk profile, governance, and supervisory practice.
In the United States, the National Association of Insurance Commissioners (NAIC) describes ORSA and prospective capital assessment as an insurer's continuing evaluation of material and reasonably foreseeable exposures. NAIC Model #505 is a model for state enactment: details are determined by applicable state law and regulatory practice. Accordingly, a cross-jurisdictional analysis can compare functional expectations about emerging-risk identification while maintaining differences in reporting thresholds, confidentiality, legal form, and group supervision. COSO's (2017) ERM and strategy guidance contributes a management perspective on organizational performance and risk governance, without replacing binding insurance law.
The essential relationship is now clear. Scanning feeds ERM and ORSA; it is not equivalent to an ORSA report, solvency capital model, stress-testing program, or underwriting referral. ORSA evaluates material risks and capital sufficiency over a firm's business planning context. Scanning may detect an embryonic or distant possibility whose immediate response is gathering information, adjusting contractual monitoring, or exploring scenarios. A proportionate record should therefore connect each signal to the stage of knowledge, plausible exposure, and authorized decision, including a documented judgment that no immediate action is justified.
Evidence classes and strength
The inquiry uses an evidence hierarchy sensitive to purpose. A legal requirement is established by current legislation and competent official interpretation. A technical method's published scope is established by its official standard or publisher text. An empirical incidence or loss effect requires population, measurement dates, units, assumptions, and appropriate analysis. A company report can substantiate what that company says it scans, but not the worldwide prevalence of the underlying risk. An expert survey documents beliefs of its respondents rather than objective event frequencies.
This last distinction is important in interpreting the World Economic Forum's 2026 Global Risks survey methodology. The survey collected responses from over 1,300 experts across sectors about multiple future horizons. Its value to an insurer lies in generating candidate questions and comparing perspectives, provided the user does not turn a salience ranking into an actuarial probability or apply global severity scores directly to a proprietary portfolio.
Emerging Risk Signal Taxonomy
From broad themes to traceable observations
A robust taxonomy preserves three different pieces of information: what changed, how the change could propagate, and which insurer exposures might be affected. Consider six broad observation domains: physical and biological conditions; technology and digital infrastructure; law and social institutions; macro-financial and geopolitical conditions; behavioral and demographic changes; and insurance-market architecture, including reinsurance capacity. These domains are sufficiently broad to avoid premature confinement to traditional risk silos. They are analytical categories for organizing sources, not a newly named conceptual system.
Within each observation, a secondary classification should record the change mechanism: novel hazard, parameter drift, increased concentration, changed dependency, changed legal interpretation, accelerated diffusion, behavioral adaptation, or a conjunction of existing risks. A third classification should map affected operations: underwriting, claims, assets, liabilities, investments, distribution, technology, governance, reserves, liquidity, reinsurance, and policyholder protection. A single signal can occupy more than one class; forcing exclusivity would disguise connected exposure.
The Swiss Re Institute's SONAR scanning practice offers a documented industry example. It collects internal contributions from underwriters, client managers, and other specialists, and considers impact, emergence horizon, and lines of business. This is an operational illustration of distributed scanning. Swiss Re expressly warns that selected themes do not comprise the entire emerging-risk landscape. Its SONAR 2025 findings include heat, fungal threats, plastics-related litigation, deepfake-enabled fraud, and changing technologies alongside structural concerns. The scientifically defensible inference is that a multi-domain scanning process can identify diverse candidate pathways; the report does not identify which specific insurers will lose how much.
A minimum viable signal record
For every candidate, the following fields are analytically necessary: signal identifier; exact source or document URL; retrieval and observation dates; source type and independence; what was actually observed; hypothesized causal pathway; affected jurisdictions and legal entities; direct and indirect insurance lines; insured and investment exposures; estimated earliest materialization; evidence-quality statement; uncertainty type; expert dissent; proposed next observation; responsible owner; and decision log. These fields facilitate reconstructability and later error analysis. Source material that cannot be retrieved or is commercially restricted should retain a bibliographic pointer and a clear access limit.
A record should preserve observations separately from the analyst's interpretation. For instance, a scientific report about heat-related mortality supports a physical and biological risk indicator. A causal claim that the insurer will incur additional life claims requires information about covered populations, morbidity or mortality sensitivities, contractual terms, pricing assumptions, geographic distribution, time horizon, and adaptive behavior. Collapsing these stages would generate false confidence from an authoritative source that never examined the insurer.
Detecting the unexpected and avoiding selection bias
Scanning only familiar categories produces category lock-in. A search that starts exclusively from a traditional insurer risk register will tend to retrieve evidence of already recognized risks. A useful practice is therefore to preserve an exploratory stream independent of current capital-model categories, while later mapping promising candidates back to accountable risk owners. Conversely, unbounded scanning generates a flood of low-value material. Neither maximal information collection nor maximal ranking precision is an end in itself.
Sampling risks arise through media attention, source duplication, and selection of English-language or high-income-jurisdiction materials. A burst of syndicated reporting may create many apparent signals from one original source; the correct effective source count is closer to one. Conversely, scientific reports from underrepresented regions may expose vulnerabilities missing from international industry rankings. A provenance graph or simple duplicate-citation field can distinguish original documents from commentary and repeated summaries. The proposed metadata requirement is a practical audit device, not a claim of universal technological necessity.
Dependence and accumulation
Insurers' risk taxonomies often assign risk to single lines or functions, even when a technological or environmental shock crosses them. A cyber incident at a widely used service provider can affect claims operations, delegated underwriting, investment execution, policy administration, and simultaneously insured cyber losses. Climate-related heat can affect mortality, health claims, worker compensation, agriculture, property demand, and asset values. The IPCC's assessment of compound and cascading risks provides a source-grounded reason to examine interaction rather than independent hazards.
The analytic objective is to identify common drivers and pathways and to represent them in a graph of exposures where data allow. The graph needs a distinct node for each legal entity, counterparty, product, contractual limit, and material dependency. It must also distinguish correlation due to a common underlying shock from duplicate bookkeeping of the same loss. Network language is useful only when it clarifies financial transmission and can be checked against actual data.
Signal lifecycle
A practical lifecycle begins with observation and deduplication, advances through hypothesis development and exposure mapping, and proceeds to triage, escalation, interventions and monitoring. It must also allow consolidation, withdrawal and reopening. A signal that later proves mistaken should remain in the historical record with an explicit disposition, because removal would bias retrospective evaluation in favor of supposed foresight accuracy. Likewise, early accurate warnings should record when an action was actually taken, allowing a distinction between detection and organizational response. This historical discipline makes a scanning program scientifically auditable.
Information architecture and source lineage
Scanning quality is constrained by the architecture through which information is ingested, interpreted, retained, and reused. A single observation may reach an insurer as an academic preprint, an insurance intermediary's alert, a regulatory discussion paper, a trade-press headline, or a claims adjuster's field report. These are distinct records of an underlying development and should carry distinct provenance. The first practical question is therefore which document contains the original observation and which documents merely repeat it. A system that counts all repetitions as independent corroboration will systematically overstate evidential strength.
The minimum data lineage begins at the original document, preserving the document title, publisher, edition, publication time, retrieval time, jurisdiction, and a stable identifier or URL. It then records the insurer analyst's extracted factual proposition and the precise section of the source that supports it. The next stage holds a risk hypothesis, expressed as a proposed cause, transmission pathway, portfolio location, time horizon, and potential outcome. Finally, the governance record documents materiality judgments, actions, and review decisions. Separating these fields makes it possible to revise a hypothesis without silently rewriting the initial observation. It also allows another analyst to audit a decision using the information available at that time, rather than later evidence.
Source timestamps deserve particular attention. A study's publication date may differ from the period when observations were collected; an insurer's claims file may contain events incurred earlier and reported later; a regulatory warning may summarize information received before its own issuance. These clocks answer different questions. For timeliness studies, the earliest verifiable signal date, the insurer's ingestion date, the internal-review date, and the decision date should be recorded separately. Otherwise an apparently rapid response may simply reflect late identification of an earlier warning.
For a source that materially affects a decision, version control is especially valuable. Scientific papers can be corrected, laws amended, and institutional dashboards updated. A citation to a current landing page may remain useful to the reader while the archival record preserves the precise version consulted. The insurer should retain enough metadata to distinguish the version on which a historic decision relied from a subsequently changed publication. Confidential data can be represented through restricted access and cryptographic or document identifiers, with access assigned according to legal and security obligations. Public traceability and protection of sensitive insurer information can coexist when the system preserves exact references without exposing protected underlying records.
A structured risk register should therefore represent the relationship between a single signal and multiple hypotheses, and between multiple signals and a single hypothesis. For instance, a scientific finding, a change in claims coding, and a new legal ruling may each support different elements of a possible liability pathway. The claims observation might also have an unrelated explanation. Retaining this many-to-many structure helps prevent automatic acceptance of a causal chain and preserves competing interpretations.
Machine learning can support deduplication and retrieval, but its output must be evaluated against a manually verified set. A clustering algorithm may merge genuinely different legal developments because they share terminology, or separate identical reports because an organization has reformatted them. The proper metric is not simply the percentage of items assigned to clusters, but the error in source independence, causal classification, and information reaching the right risk owner. This is particularly relevant to AI-enabled scanning systems discussed by the NIST AI risk framework. The proposal here is a practical validation requirement grounded in data-quality logic, rather than an assertion that any specific software architecture has been scientifically proved optimal.
Spatial and temporal sampling
International insurers face uneven observability. A risk may emerge first in regions where digital records are sparse, legal databases are costly, or insurance penetration is limited. Academic and supervisory reports from highly documented markets may dominate a global search, while less visible exposures remain in the portfolio. The firm should document geographic, linguistic, and product coverage of its source stream, and state openly when conclusions are based on only a subset of jurisdictions.
Time-series anomalies also require attention to the denominator. An increase in the number of cyber claims might accompany an increase in policies in force, expanded coverage, changed deductibles, claims inflation, revised incident reporting, or more effective detection. Without normalizing the exposed population and coding conventions, a scanning indicator could be directionally misleading. A well-maintained denominator and an explicit alternative-explanation field may contribute more to a risk decision than a complex novelty score. The research implication is to evaluate source coverage and measurement change as part of risk identification itself.
Expert Elicitation and Confidence
The role and limits of expert judgment
Emerging risks often precede relevant claims history. Experts can identify plausible physical mechanisms, contractual interpretations, operational dependencies, and future adjustments before a stable statistical sample exists. Experts also bring professional incentives, selective experience, correlated information sources, group conformity, and uneven ability to assess rare events. Their evidence should be elicited as structured, attributable judgment rather than converted mechanically into objective frequency.
Cooke's (1991) study of expert uncertainty addresses the representation, assessment, and aggregation of subjective probability. Rowe and Wright's (1999) review of Delphi forecasting identifies supportive comparisons for Delphi groups but does not establish superiority over every other structured group procedure. This distinction matters when insurers select a method: the value of a Delphi exercise depends on question definition, panel diversity, evidence, iteration, feedback, and how the outcome changes decisions.
A structured panel should first define the event precisely. “Cyber risk will increase” is not an elicitable event with a measurable probability; “a shared cloud dependency causes an outage exceeding 24 hours across at least three named material processes during the next 12 months” is closer to one, subject to reliable event recording and a defined population. Different experts may be asked to assess conditional pathways rather than supply a single percentage. In many embryonic settings, an ordinal confidence category and a well-constructed scenario range are more defensible than an invented distribution.
Separating three forms of uncertainty
Aleatory variability concerns stochastic outcomes under an adequately defined process; epistemic uncertainty concerns imperfect knowledge about process, parameters, or causal pathways; ambiguity concerns which conceptual account or data-generating structure applies. The distinctions are analytical conveniences with overlapping real-world cases. A frequency estimate supported by repeated, comparable observations addresses variability. A novel litigation theory with uncertain judicial treatment is more naturally evaluated through legal scenarios. A dependency whose architecture is not disclosed may first require fact-finding. The response method should follow the uncertainty rather than an across-the-board demand for numerical probabilities.
Assessments should record separate confidence in source authenticity, causal plausibility, exposure relevance, loss quantification, and time horizon. A scientific observation can be highly reliable while the financial implications for a particular insurer remain uncertain. Experts who disagree about legal interpretation should have their alternative assumptions carried forward. An averaged score without the underlying disagreement creates unjustified consensus.
Panel design, feedback, and calibration
Expert selection should mix internal underwriting, actuarial, claims, legal, investments, IT, operations, reinsurance, and compliance specialists with independent external scientific or policy expertise as appropriate. Disclose dependence when panel members rely on the same proprietary model or external report. Provide anonymous first-round estimates where feasible, document reasons rather than only numbers, offer structured feedback, and preserve final dissent. A risk committee should not impose numerical convergence as an end in itself.
Where repeatable outcomes exist, performance can be measured through a proper scoring rule, such as the Brier score for binary forecasts. For events with realized outcomes y in {0,1} and forecast p in [0,1], the score (p − y)^2 is lower for better forecasts. This is a well-established rule applied here as an illustrative audit metric, not as a new scientific model. It is applicable only to clearly settled events with identical resolution criteria. Rare events, changing insured populations and selective question resolution can make small retrospective samples unstable. For early-stage risks with no measurable binary resolution, qualitative forecasting records and reasoned postmortems remain appropriate.
The organization can also evaluate the information quality of a scan by measuring duplicate rates, proportion of claims supported by primary documentation, time from source publication to review, and the share of material warnings that reach the responsible decision-maker. These process measures are not identical to long-run loss reduction. A rapid scanner that fails to inform underwriting still has limited strategic value.
Adversarial interpretation
Every material signal benefits from a serious alternative explanation. A reported frequency increase may reflect greater detection rather than more hazards. A novel litigation theme may draw many headlines but few insured defendants. An apparent cyber trend may derive from one vendor's changing telemetry. A newly published scientific hypothesis may rest on small samples or uncertain transfer across regions. Asking experts what evidence would falsify their interpretation increases the explanatory value of the process.
Corporate foresight research is relevant here: Rohrbeck and Gemünden's (2011) study of corporate foresight roles, based on 19 case studies and 107 interviews, identifies a role that challenges established innovation projects. The same institutional principle can be adapted to insurance risk review: assign an independent challenger with explicit standing to test the assumed causal channel, data selection, and proposed response. The case-study evidence supports conceptual usefulness, not a measured universal decrease in losses.
Materiality Screening
Materiality belongs to a specific insurer
A risk is financially material through its plausible effect on the insurer's obligations, capital resources, liquidity, earnings, franchise, policyholders, or operational ability to perform. Materiality is conditional on product, geography, portfolio duration, deductibles, limits, exclusions, contract language, reserves, asset allocation, counterparties, reinsurance, and risk appetite. A global severity ranking may help discover a topic but cannot determine whether a regional specialist, life carrier, reinsurer, or mutual insurer is exposed.
Screening should proceed in stages. The first stage verifies source provenance and causal plausibility. The second identifies exposure channels and their direction. The third describes severity, timing, aggregation, and control options. The fourth compares available evidence with the decision that actually must be taken. Insufficient evidence may justify deeper analysis, contractual monitoring or an exposure freeze under defined risk appetite; it need not force a binary conclusion about the risk's existence.
IEC's risk-assessment techniques guidance recognizes technique selection under uncertainty. It is therefore reasonable to use qualitative matrices for early appraisal, scenarios for structural ambiguity, stress tests for plausible shocks, frequency–severity analysis where statistical samples exist, and probabilistic portfolios when data support calibrated distributions. Each technique has a domain of validity, which should be recorded alongside its output.
Conditional loss and accumulation
An insurer's conditional loss under scenario s can be written for analytical clarity as L(s) = Σ_i E_i(s) × V_i(s) × C_i(s), where E_i denotes an exposure measure, V_i a loss response conditional on the scenario, and C_i a net contractual or recovery factor. The representation is illustrative, not a generally estimable universal formula. Contractual limits, nonlinear indemnity schedules, reinsurance towers, interaction effects, and counterparty failures may require explicit claims cash-flow simulation instead of such a product expression.
The amount potentially transferred to reinsurers should not simply be subtracted at face value. Its timing, exhaustion, dispute risk, reinstatement cost and counterparty exposure affect the insurer's retained loss. Group exposure must also respect fungibility and restrictions on intercompany asset transfers. An emerging-risk register becomes material to capital governance when it can show how newly recognized pathways affect existing assumptions, dependencies or stress scenarios.
A transparent hypothetical decision example
Suppose a provisional protective action costs 2 units immediately and would avert 15 units of loss if a precisely defined adverse condition materializes. Assume for this isolated example that the action is perfectly effective, has no other benefits or costs, and that the decision-maker's subjective probability is p. Its estimated net benefit is p × 15 − 2, and the break-even p is 2/15, approximately 13.3%. If p = 0.20, the illustrative net benefit is 1 unit; if p = 0.05, it is −1.25 units. These values are invented instructional quantities, not empirical insurer rates or policy recommendations.
More realistically, the loss-avoidance estimate may itself have a range and intervention might be reversible. The decision could then depend on the expected value of additional information, irreversibility, the consequences of waiting, and the possibility of partial protection. A scan can have immediate value even when no action is implemented because it changes which data an insurer collects before a more costly commitment. The correct threshold depends on the cost of false negatives, the cost of false positives, the action's actual efficacy, and who bears the consequences.
Prioritization without manufactured precision
An ordinal scale can categorize severity, proximity, trend, reversibility, information quality, and controllability. Combining these by arbitrary multiplication can create a false sense of cardinal accuracy. A score of twelve produced from three ordinal ratings is not automatically twice as serious as a score of six. A defensible triage method uses such scores only to order review, discloses thresholds, tests rank sensitivity to different weights, and reserves board decisions for a documented analysis of tail outcomes and material dependencies.
A useful comparison separates three decision questions. Is the risk scientifically plausible enough to monitor? Is it sufficiently material to this insurer to escalate? Is the evidence strong enough to authorize a particular intervention? Different thresholds answer different questions. A low-confidence, potentially catastrophic exposure may warrant inexpensive monitoring or resilience tests. A high-confidence, small exposure may stay with routine line management. Explicit decision rules reduce both overreaction to novel headlines and inaction when quantitative models lag material change.
Capital models, scenarios, and model uncertainty
An existing solvency capital requirement can omit an emerging pathway, represent it only partially, or contain implicit correlation assumptions that fail under new conditions. EU ORSA requirements specifically recognize possible deviations between a firm's risk profile and capital-model assumptions. Identifying such deviations is often more immediate than estimating a single new capital charge. A disciplined sequence documents the suspect assumption, quantifies the exposure where possible, runs sensitivity and reverse stress scenarios, records governance interpretation, and decides whether risk appetite or capital planning requires revision.
A horizon scan is not itself validation of a predictive model. Claims experience may be too short, reporting processes may change, or exposure denominators may be unobserved. Any scenario analysis should distinguish observed parameters, externally sourced estimates, expert assumptions, and stress choices. Without that separation, a carefully formatted loss distribution may be nothing more than undocumented opinion.
Sensitivity, dependence, and stress construction
Where the exposure can be measured, materiality screening should reveal how much the conclusion depends on contestable assumptions. A baseline portfolio projection may use historical event rates, severity distributions, contractual limits, expenses, discount rates, and reinsurance recoveries. An emerging condition can challenge one or several of these simultaneously. Sensitivity analysis should change each assumption within an explainable range, then consider joint changes where physical or financial pathways support dependence. A scenario whose assumptions cannot be traced to observations or explicitly labeled expert judgments is unsuitable for a definitive capital conclusion.
Simple aggregation by adding standalone tail losses is generally too conservative when losses are mutually exclusive and can be seriously misleading when it overlooks common-cause shocks or exhaustion of shared reinsurance cover. The appropriate treatment depends on joint exposure and contractual structure. Where robust estimation is infeasible, a conditional scenario table can identify shared drivers and potential bottlenecks without claiming a reliable joint probability distribution. A risk concentration warning may be useful even when a calibrated capital number cannot yet be produced.
Consider a purely hypothetical insurer with three exposure channels linked to one digital platform: delayed claims settlement, interruption of policy issuance, and insured business interruption. If an outage causes all three, treating each as an independent event obscures the common trigger. Conversely, a reported financial loss may appear in both a claims estimate and a broader operational disruption estimate; adding both mechanically double-counts it. A causal account should distinguish the original triggering event, gross losses by channel, costs common to several channels, offsets, policy and reinsurance terms, and total retained loss. This exercise is part of exposure accounting, not an original statistical theory.
Sensitivity testing must also accommodate changes in the insurer's behavior. Rising premiums can change demand and selection; new exclusions may reduce future exposure but affect policyholder outcomes or introduce litigation; a change in reinsurance price can alter retention strategy; and loss-prevention activity may change hazard realization. Static assumptions can therefore misstate long-horizon results. The analyst should indicate when an outcome is conditional on holding strategy fixed and when a scenario explicitly incorporates managerial or customer adaptation.
Value of further information
Early decisions often concern what to learn next. Suppose management considers an inexpensive data study that may reveal whether a hypothesized dependency is concentrated among a small group of vendors. The expected value of perfect information provides an upper bound on what management should be willing to spend to resolve uncertainty in a simplified expected-value decision. With several feasible interventions a, uncertain states s, and current information I, this upper bound can be written as E[max_a U(a,s) | I] − max_a E[U(a,s) | I]. The expression is nonnegative because deciding with knowledge of the state cannot yield a worse optimal decision than choosing an action without that knowledge under the same utility assumptions. This mathematical relationship is an established decision-theoretic benchmark used illustratively here.
Actual investigations provide imperfect information and may arrive too late. Their value is lower than the perfect-information bound after accounting for acquisition cost, conditional signal accuracy, implementation delay, and the risk that an action becomes unavailable. Consequently, a scan should sometimes accelerate a reversible protective measure while deeper research proceeds; at other times a short delay for a reliable study is economically preferable. The distinction depends on the shape of the loss and action-cost distributions, not on an arbitrary emerging-risk score.
When probabilities are too poorly supported for a meaningful expected-value computation, the analysis can compare decisions over a family of plausible scenarios. A decision that works acceptably across several adverse futures may be more robust than one that is optimal only under a favored forecast. Such comparative scenario analysis accords with the logic of Schoemaker's strategic scenarios and the IRGC's approach to flexible emerging-risk responses. Robustness should still be defined in terms of explicit objectives, scenario coverage and costs; it is not a substitute for evidence.
Avoiding retrospective reconstruction
A mature risk team may be tempted to judge identification by asking whether a topic eventually appeared in a quarterly register. That metric rewards registration and says little about warning timeliness or decision quality. The relevant record is the history of observations, analytic judgments, and actions at their original dates. To study predictive usefulness, an insurer could prospectively designate measurable event definitions and monitoring horizons, archive each signal and anticipated outcome, and allow independent evaluators to resolve outcomes from specified data. This would identify forecasts that proved accurate, forecasts that failed, and hypotheses that were never resolved. Missing outcomes should not be silently removed from accuracy statistics.
In studying organizational performance, the greatest difficulty is causal attribution. A carrier may start a scanning process after a major loss, change its underwriting appetite simultaneously, and then experience fewer claims because exposures have been reduced. Comparing before and after performance without accounting for portfolio change falsely attributes the whole improvement to scanning. More informative designs would compare matched portfolios, staged implementation where appropriate, and ex ante measures of action latency, whilst recording interventions and institutional changes. These study designs remain proposed empirical tests; no proprietary carrier performance data have been collected for this monograph.
Enterprise Escalation and Action
Decision rights and institutional accountability
Identification becomes a governance function when every material signal can reach a responsible decision-maker with authority to request evidence, constrain exposure, or alter strategy. A practical allocation begins with decentralized discovery by product and operating units; a central emerging-risk or ERM function performs deduplication, cross-line aggregation, and consistency review; specialist committees examine actuarial, claims, legal, cyber, investment, and reinsurance questions; the chief risk officer or equivalent coordinates materiality and appetite; senior management selects interventions; and the board oversees material exposures and risk appetite. Actual allocations vary with jurisdiction, insurer size, governance form and group structure.
The escalation record should identify the threshold crossed, decision deadline, evidence relied upon, uncertainties, dissent, alternatives rejected, action owner, implementation schedule, and trigger for reassessment. Decisions can include monitor, commission research, request data, run scenario analysis, revise underwriting guidelines, adapt policy wording prospectively under applicable law, test operational continuity, diversify a dependency, adjust pricing or reserves where justified, obtain reinsurance, engage supervisors, or alter capital plans. These are candidate responses, each subject to its own legal, actuarial, commercial and customer-conduct constraints.
The IAIS global market review for 2025 illustrates the diversity of contemporary supervisory attention: private credit, geoeconomic fragmentation, and AI adoption require different source systems, models and enterprise decisions. One generic hazard score cannot replace the specialized evidence demanded by all three.
Cadence, challenge, and right to reopen
Scanning frequency should match information dynamics. A cyber service dependency may justify rapid alerts; legal precedent may require monitoring by jurisdiction and case milestone; long-horizon climate or demographic change may be reviewed through periodic portfolio and scenario updates. Fixed committee calendars should coexist with event-driven escalation. A board receiving a quarterly emerging-risk slide should be able to inspect the underlying source and learn whether a new risk is increasing, declining, unresolved, or merely better observed.
A closed signal record should specify why it was closed and under what conditions it will be reopened. Closure may follow disconfirmation, absorption into a conventional risk inventory, completed mitigation, portfolio exit, or loss of materiality. This prevents the emerging-risk function from becoming a permanent warehouse of unresolved labels. Reopened signals can reveal the value of early interpretations, but only if records of prior decisions remain intact.
Operational and model-risk interface
The adoption of advanced analytics and artificial intelligence for scanning can increase coverage of scientific, regulatory and market sources, facilitate clustering, and surface cross-line topics. It can also produce incorrect summaries, invented citations, repeated false alerts and automation bias. NIST's AI risk-management framework offers voluntary general guidance on AI risks; EIOPA's AI governance opinion clarifies relevant insurance supervisory expectations and its stated scope. Neither source warrants autonomous acceptance of machine-produced risk findings without checking the original document.
A practical human-controlled workflow requires exact source retrieval, clear separation of quoted observation from inferred impact, duplicate detection, sensitive-data controls, accountable approval, and retention of corrected records. Machine-discovered novelty should increase analytic attention, not automatically trigger a materiality or capital conclusion. Technology expands sensing capacity only when verification resources and governance scale with it.
Performance and learning
Four distinct outcomes are worth measuring: detection quality, interpretive quality, decision transmission, and intervention effectiveness. Candidate metrics include source-to-review latency; proportion of traceable original sources; duplicate and unsupported-signal rates; time from materiality threshold to risk-committee consideration; agreement and dissent records; retrospective event-resolution scores for testable forecasts; and completed actions against recorded decisions. Outcome measures may examine loss and service impacts, but attribution requires credible counterfactual comparison. An insurer that suffers fewer losses may have a better portfolio or favorable exposures rather than a superior horizon-scanning mechanism.
An empirical design could prospectively timestamp signals and board decisions across multiple carriers, normalize for book composition, collect independently coded risk outcomes, and compare decision timing under comparable exposures. Pre-registered event definitions and blinded hindsight audits would reduce selective retrospective success stories. Access to confidential governance decisions and sufficiently long observation periods remains a central feasibility limit. Until such evidence exists, the manuscript's contribution is an explicit source-grounded account of mechanisms, assessment problems, and testable operational criteria.
Governance controls and policyholder consequences
Emerging-risk governance affects external stakeholders as well as shareholders and solvency ratios. A decision to withdraw cover, narrow underwriting eligibility, raise deductibles, or revise claims processes can reduce insurer exposure while altering the availability and affordability of protection. The purpose of scanning is to identify those trade-offs early enough for lawful, evidence-based policy choices. Decision documentation should therefore include consequences for policyholder protection and market conduct, alongside gross and net loss estimates.
For example, an insurer observing heat-related health risks may investigate price adequacy, covered-population vulnerabilities and preventive services. Different adjustments may affect communities unevenly, and the evidence supporting risk differentiation may be uncertain. Supervisory and legal obligations concerning fair treatment, nondiscrimination and transparency must be assessed in their applicable jurisdiction. Similarly, an AI-based fraud-detection control may lower claims leakage while exposing customers to erroneous denials or unreasonable delays. EIOPA's AI governance opinion makes data governance, fairness, explainability and human oversight relevant to insurer risk oversight. The specific legal effect and scope of the opinion remain distinct from binding EU rules.
Board reporting can expose these trade-offs without converting every uncertainty into a large numeric dashboard. A concise decision paper can state the verified signal, source date, plausible pathway, affected customer and company exposures, alternative explanations, pending information, worst relevant scenarios, intervention options, and proposed review date. It should make clear which statements have empirical support and which remain expert assumptions. This format permits informed challenge and produces a durable record that later can be tested against outcomes.
Research agenda for institutional effectiveness
A scientifically informative multicompany study would require several coordinated datasets. The first is a prospective timestamped record of candidate signals and sources. The second is an exposure dataset normalized for business mix and changes in coverage. The third is a governance dataset recording which signals were reviewed, escalated, acted upon or closed. The fourth is an outcome dataset capturing realized events, ultimate losses where mature, operational disruption, and policyholder outcomes. The datasets require legal protections because confidential risk registers and underwriting decisions may reveal commercially sensitive exposures.
The principal empirical outcomes should be chosen before inspecting results. Detection precision, decision timeliness, independently verified exposure identification, and preventable loss could each change in different directions. A comprehensive scanner may correctly flag many low-probability threats and therefore appear to generate false alarms by a simplistic event-count measure. An intervention may avert a loss, making the absence of a realized event a poor indication of the warning's quality. Rigorous evaluation needs explicit causal assumptions and, where possible, matched or quasi-experimental comparisons.
Research should also examine organization-specific failure points: failure to ingest an available primary source, misinterpretation of an accurate source, omission of a legal entity from the exposure map, delayed escalation, lack of decision authority, unsuccessful implementation, and failure to revisit an outdated decision. These failure points demand different corrective actions. A single retrospective risk score conceals their location. Analysis of such institutional mechanisms can ultimately establish whether improvements in scanning contribute to effective protection and financial resilience, provided the findings are supported by credible observed evidence.
Comparative Synthesis and Research Conclusions
Three insurer-relevant illustrations
The first illustration concerns artificial intelligence, shared digital infrastructure and cyber concentration. A signal may originate as a documented increase in dependency on a common third-party service, a new class of model-driven claim decisions, or observed novel fraud techniques. Its immediate risk hypothesis may involve operational interruption, customer harm, underwriting errors, reputational loss or cyber claims accumulation. EIOPA's October 2026 risk dashboard judges digitalization and cyber risk high at sector level with a worsening outlook. This supplies contemporary supervisory context, not a probability for the individual carrier. A valid insurer investigation would map vendor dependencies, critical functions, coverages, legal entities, claim limits and incident response. Alternative explanations for rising incident counts, including changing monitoring intensity, must be tested. Possible interventions are concentration limits, recovery tests, independent vendor assurance and portfolio stress analysis. Such choices are justified only after exposure-specific analysis.
The second illustration concerns extreme heat and long-duration life and health liabilities. Scientific evidence about heat-related health effects and mortality is relevant to product assumptions, geographic concentration, occupational composition, public-health adaptation and changing exposure. EIOPA's September 2026 climate and health risk report identifies extreme heat as a leading climate-related liability concern for its studied sectors, while noting current limiting factors. A carrier cannot automatically transfer the European sector's qualitative outlook to a global book. It should identify insured cohort, age profile, healthcare coverage, benefit design, regional exposure, mortality and morbidity data, policy duration and selection effects. Long-run scenario analysis can examine several adaptation pathways. The immediate governance decision may concern data acquisition and assumption monitoring rather than a numerical reserve adjustment unsupported by evidence.
The third illustration concerns long-tail casualty claims from a newly scrutinized substance or behavior. The signal could be new toxicological research, a change in legal admissibility, emerging litigation, or social expectations about corporate responsibility. The pathway from laboratory effect to insured claim includes exposure history, identification of defendants, duty and causation standards, limitation periods, policy wording, allocation across years, defense costs and coverage disputes. Insurers must distinguish scientific hazard identification from proof of covered liability. Swiss Re's 2025 discussion of potential plastics or ultra-processed-food liability themes is useful for generating competing hypotheses, not determining ultimate legal outcomes. Expert panels should involve claims and legal specialists from relevant jurisdictions, with explicit disagreements retained. Scenario analysis may encompass delayed emergence, clustered filings, changing legal interpretation and reinsurance aggregation.
In all three cases, the central challenge is the same: source credibility is necessary yet insufficient for insurer-specific loss quantification. Each case requires an exposure map, a legally and operationally coherent causal pathway, and a decision proportional to the expected value and reversibility of action.
Comparing institutional information systems
Global sources differ by design. The Swiss Re SONAR material reflects an internal specialist signal-generation process that selects potentially consequential themes. EIOPA's dashboard draws on supervisory and market data to describe an existing sector risk profile, past changes and outlooks. The WEF survey reflects respondent judgments about broad global risk. IAIS publications emphasize supervisory priorities and industry-wide stability. Each informs different stages of scanning. Treating them as interchangeable estimates creates an epistemic error: source purpose, population, temporal unit and outcome variable differ.
This leads to a practical synthesis. The insurer should maintain an acquisition layer covering broad signals, a verification layer establishing provenance, an interpretation layer mapping pathways, an exposure layer testing relevance, and a governance layer recording decisions. These “layers” describe an implementation sequence drawn from existing ERM and foresight practices; they are not proposed as a named new theory or universal regulatory architecture. Their analytical value lies in making the weakest evidence link visible before management acts.
Jurisdiction and organizational differences
An internationally active insurance group must distinguish global cross-entity signals from local legal obligations, pricing powers and capital transfers. Solvency II ORSA places explicit emphasis on short- and long-term risk identification; NAIC ORSA standards function through state adoption; IAIS ICP 16 offers supervisory principles. A small domestic mutual may draw on fewer specialists but can still maintain sourced signals, explicit escalation criteria and a documented ORSA interface. A global reinsurer may need stronger treatment of retrocession, accumulation, common vendors and cross-border litigation. Proportionality changes implementation cost and institutional roles while preserving source verification and accountability.
Comparability across carriers also depends on common denominators. “Number of emerging risks” is not interpretable without a detection rule. “Risks mitigated” is not interpretable without an action definition. A firm with a long candidate list might have greater observation capacity, not a riskier balance sheet. Public benchmarking should therefore disclose the process boundary, categories, time horizons, exposure scale and exclusions before inferring superiority.
Scientific limitations and testable claims
This monograph offers original analytical synthesis in the form of explicit distinctions and source-based integration. It does not assert empirical superiority of a particular scanning method, calculate a new industry loss distribution, or claim a newly validated model. The strongest documentary findings are normative or descriptive: supervisory principles require enterprise attention to material and prospective risks, risk-management standards contemplate managing emerging uncertainty, foresight literature offers structured techniques, and industry and supervisory sources show varied operational ways to identify and discuss emerging issues. Their causal effect on underwriting quality, loss ratios or solvency remains a separate empirical question.
Several hypotheses can be tested without elevating them to a new doctrine. A documented exposure-mapping stage may reduce unsupported risk escalation compared with headline-only review. Independent expert challenge may improve recorded source-to-claim accuracy. Faster escalation may improve response timeliness when intervention rights exist, although its effect on ultimate losses depends on event severity and effective actions. Cross-line dependency mapping may reveal concentrations absent from single-line reviews. Each proposition requires appropriate comparison groups and defensible outcome definitions; the present sources establish plausibility, not proof.
Retrospective research is especially vulnerable to survivorship and hindsight bias. Visible crises attract stories about missed warnings, while unmaterialized warnings rarely attract equivalent investigation. A strong empirical program should archive all candidate signals prospectively, including those later rejected. It should evaluate error costs, not only apparent prediction successes, and distinguish detection, warning delivery, and action. Only then can studies estimate whether horizon scanning actually improves insurers' outcomes beyond ordinary risk-management controls.
Conclusion
Insurer emerging risk identification and horizon scanning is a governed process of converting uncertain, heterogeneous information into documented propositions about potential material exposure and decisions that can be revised as evidence develops. Its scientific object combines risk origin, transmission pathway, insurer-specific materiality, decision authority and observed outcome over time. The process is most useful when its evidence can be reconstructed and its judgments can be challenged.
The documentary record supports five principal conclusions. First, novelty belongs to the changing conditions of exposure and knowledge, not merely to new words in a risk register. Second, supervisory and managerial sources establish the relevance of forward-looking ERM, while their requirements and recommendations differ by jurisdiction and authority. Third, weak signals and expert judgments must be separated from calibrated probabilities and portfolio loss estimates. Fourth, a scan becomes operationally significant through causal exposure mapping, aggregate and dependent risk assessment, and timely decisions under specified risk appetite. Fifth, an insurer's scanning program can be judged scientifically only when historical records preserve false alarms, missed events, decisions, and plausible counterfactual comparisons.
The resulting research agenda seeks better evidence on which sources identify actionable risks, how differences in expert judgment affect decisions, when precautionary action is economically justified, and how cross-line concentrations are found before losses accumulate. These are research questions and analytical conclusions derived from established literature and supervisory practice, not assertions of a new proprietary theoretical system. The practical standard for success is a traceable chain between observed change, verified meaning, relevant exposure, authorized intervention, and subsequent learning.
References
International Association of Insurance Supervisors (2024–2026). Insurance Core Principles and ComFrame: ICP 16, Enterprise Risk Management for Solvency Purposes. https://www.iais.org/icp-online-tool/
International Organization for Standardization (2018). ISO 31000:2018, Risk Management—Guidelines. https://www.iso.org/standard/31000
International Organization for Standardization (2023). ISO/TS 31050:2023, Risk Management—Guidelines for Managing an Emerging Risk to Enhance Resilience. https://www.iso.org/standard/54224.html
International Electrotechnical Commission (2019). IEC 31010:2019, Risk Management—Risk Assessment Techniques. https://webstore.iec.ch/en/publication/59809
International Risk Governance Council (2015). IRGC Guidelines for Emerging Risk Governance. https://irgc.org/risk-governance/emerging-risk/a-protocol-for-dealing-with-emerging-risks/
International Risk Governance Council (2018). Guidelines for the Governance of Systemic Risks. https://irgc.org/risk-governance/systemic-risks/guidelines-governance-systemic-risks-context-transitions/
European Parliament and Council (2009). Directive 2009/138/EC on the Taking-Up and Pursuit of the Business of Insurance and Reinsurance (Solvency II), Article 45. https://eur-lex.europa.eu/legal-content/EN/ALL/?uri=CELEX%3A32009L0138
European Insurance and Occupational Pensions Authority (2015). Guidelines on Own Risk and Solvency Assessment (ORSA). https://www.eiopa.europa.eu/publications/guidelines-own-risk-solvency-assessment-orsa_en
National Association of Insurance Commissioners (2026). Own Risk and Solvency Assessment: Insurance Topics and Model Act #505 Overview. https://content.naic.org/insurance-topics/own-risk-and-solvency-assessment
Committee of Sponsoring Organizations of the Treadway Commission (2017). Enterprise Risk Management—Integrating with Strategy and Performance. https://www.coso.org/guidance-erm
Swiss Re Institute (2025). SONAR 2025: New Emerging Risk Insights. https://www.swissre.com/institute/research/sonar/sonar2025.html
Swiss Re Institute (2025). SONAR—Systematic Observation of Notions Associated with Risk. https://www.swissre.com/institute/research/sonar.html
European Insurance and Occupational Pensions Authority (2026). Insurance Risk Dashboard: October 2026 (EIOPA-BoS-26-495). https://www.eiopa.europa.eu/assets/insurance-risk-dashboard/EIOPA-BoS-26-495%202026%20October%20Insurance%20Risk%20Dashboard.html
International Association of Insurance Supervisors (2025). Global Insurance Market Report 2025: Key Supervisory Priorities. https://www.iais.org/2025/12/iais-global-insurance-market-report-2025-highlights-growth-of-investments-in-private-credit-geoeconomic-fragmentation-and-ai-adoption-as-key-supervisory-priorities/
World Economic Forum (2026). Global Risks Report 2026: Overview of Methodology. https://www.weforum.org/publications/global-risks-report-2026/in-full/global-risks-report-2026-overview-of-methodology/
UK Government Office for Science (2024). The Futures Toolkit for Policymakers and Analysts. https://www.gov.uk/government/publications/futures-toolkit-for-policy-makers-and-analysts
Ansoff, H. Igor (1975). Managing Strategic Surprise by Response to Weak Signals. https://cmr.berkeley.edu/1975/02/18-2-managing-strategic-surprise-by-response-to-weak-signals/
Hiltunen, Elina (2008). The Future Sign and Its Three Dimensions. https://www.sciencedirect.com/science/article/pii/S0016328707001085
Rowe, Gene, and George Wright (1999). The Delphi Technique as a Forecasting Tool: Issues and Analysis. https://www.sciencedirect.com/science/article/pii/S0169207099000187
Cooke, Roger M. (1991). Experts in Uncertainty: Opinion and Subjective Probability in Science. https://academic.oup.com/book/49909
Schoemaker, Paul J. H. (1995). Scenario Planning: A Tool for Strategic Thinking. https://shop.sloanreview.mit.edu/store/scenario-planning-a-tool-for-strategic-thinking
Rohrbeck, René, and Hans Georg Gemünden (2011). Corporate Foresight: Its Three Roles in Enhancing the Innovation Capacity of a Firm. https://www.sciencedirect.com/science/article/pii/S004016251000140X
Intergovernmental Panel on Climate Change (2023). Climate Change 2023: AR6 Synthesis Report, Summary for Policymakers. https://www.ipcc.ch/report/ar6/syr/summary-for-policymakers/
European Insurance and Occupational Pensions Authority (2026). Report on Impacts of Climate Change on Life and Health and Potential Emerging Risks. https://www.eiopa.europa.eu/publications/eiopa-report-impacts-climate-change-life-and-health-and-potential-emerging-risks-insurance-and-iorp_en
European Insurance and Occupational Pensions Authority (2025). Opinion on AI Governance and Risk Management. https://www.eiopa.europa.eu/eiopa-publishes-opinion-ai-governance-and-risk-management-2025-08-06_en?prefLang=lt
National Institute of Standards and Technology (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). https://www.nist.gov/itl/ai-risk-management-framework
International Risk Governance Council (2010–2015). Emerging Risks: Definitions, Sources, Drivers, and Governance. https://irgc.org/risk-governance/emerging-risk/