Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Unsupervised Risk Segmentation and Clustering in Insurance

Abstract. Insurance portfolios display overlapping forms of heterogeneity: insured objects differ in observable characteristics, behavior, contract terms, geography, exposure duration, and loss-generating processes. Unsupervised segmentation creates candidate groupings from observed attributes without using an individual claim label as the direct clustering target. Its scientific value depends on the object of segmentation, the representation of mixed insurance records, the sampling mechanism, the algorithmic definition of similarity, and independent evidence of relevance to a specified insurance risk. This monograph develops those conditions from statistical first principles and evaluates partitioning, hierarchical, density-based, graph-based, and finite-mixture approaches against the distinctive structure of insurance data. Its core analytical finding is that a reproducible partition of policy records, a latent mixture component, an exposure group, an insurance risk class, and a permissible decision group are different objects requiring separate validation. Portfolio clustering can summarize exposure composition, reveal previously overlooked patterns, guide exploratory analysis, and generate testable hypotheses; any claim about future frequency, severity, tail risk, or lawful differential treatment requires additional outcome-based, temporal, and governance evidence. The monograph synthesizes primary methods papers, recent actuarial applications, and jurisdictionally bounded supervisory materials. Two fully declared calculations demonstrate exposure-weighted loss-rate contrasts and the danger of selecting groups using the same loss information later offered as independent validation. No new named theory, independently collected insurer dataset, or population-level causal effect is claimed.

Keywords: Insurance risk segmentation; unsupervised learning; cluster analysis; latent classes; finite mixtures; mixed-type dissimilarity; k-means; medoids; density-based clustering; exposure normalization; insurance telematics; portfolio heterogeneity; cluster stability; actuarial validation; algorithmic fairness; model governance.

Introduction

The modern insurance portfolio is a record of relationships among contracts, insured persons or assets, periods at risk, hazards, claims, payments, and decisions. Its statistical structure is shaped by both external loss processes and the insurer's own selection, pricing, reporting, and administrative arrangements. Clustering examines how such records resemble or differ from one another. It thereby makes a claim about structure in the representation being analyzed. Whether that structure is financially meaningful is an additional, empirically answerable question.

A useful starting point is Hennig's (2015) analysis of the meaning of clusters: cluster validity depends on the purpose and characteristics attributed to a group. Applied to insurance, this implies that the analyst must specify whether a cluster represents similarity of exposure, similarity of behavior, similarity of claims-development histories, spatial concentration, contractual products, or similarity of estimated loss distributions. These objects can coincide in some portfolios and diverge sharply in others. The analytical aim determines the definition of the observational unit and the relevant metric long before an algorithm is fitted.

Two substantive insurance contributions clarify the field. Jamotton, Hainaut, and Hames (2024) examine clustering methods for heterogeneous insurance attributes and compare ways of representing motor insurance portfolios containing numerical and categorical fields. Hannon and O'Hagan (2025) investigate mixture-based telematics risk assessment, combining Gaussian mixture grouping with downstream cluster-specific claim regressions. Their studies demonstrate distinct forms of clustering-based analysis, not a general claim that a cluster label is itself a validated premium factor. Zhu and Wüthrich (2021) use image-based driving-style clustering with pretrained feature extraction; that pipeline illustrates a further distinction between the unsupervised final grouping and the provenance of features used to form it.

The research questions are therefore explicit. First, under which representations and sampling conditions do data clusters provide interpretable descriptions of portfolio heterogeneity? Second, which objective functions and assumptions make clustering procedures suitable for insurance's mixed, longitudinal, dependent, and selected observations? Third, what independent evidence is necessary to claim that a discovered grouping matters for future insured losses, prudent operations, or authorized classification?

The method is a critical analytical synthesis of original statistical works and primary institutional instruments, supplemented by mathematical deductions and disclosed hypothetical illustrations. Methods papers establish technical properties; actuarial application papers support claims about those particular empirical settings; supervisory materials establish only their documented institutional and territorial scope. The review distinguishes an original analytical inference from an experimentally observed result. A scientific research dossier does not substitute a portfolio-specific actuarial study or a regulator's legal determination.

Unsupervised Learning and Insurance Risk

The observational unit defines the estimand

Let an observed portfolio consist of units indexed by i. The unit may be a policy, a contract-year, a person-year, a vehicle-month, a building, a claim, or a territorial exposure cell. Assign to each unit an attribute vector X_i, exposure denominator E_i where relevant, observation window T_i, and an optional future loss outcome Y_i. Unsupervised segmentation applies a rule to X rather than minimizing a loss function directly on Y. Its output may be a hard assignment Z_i in {1,...,K}, a probability vector over K components, a hierarchy, or a set of density-connected groups with outliers.

The scientific estimand is not automatically an insurance claim probability. For example, clusters obtained from vehicle age and engine displacement characterize those recorded vehicle attributes. Clusters obtained from speed–acceleration summaries characterize patterns of observed usage. Clusters of open-claim administrative trajectories describe claim-handling histories and may already contain downstream information unavailable at underwriting. The same word, risk, must be attached to an exact target—frequency per unit exposure, severity conditional on a claim, aggregate loss per policy-year, catastrophic concentration, insolvency loss, or another defined quantity.

The formal difference is visible in the relationships

Zi=gθ(Xi),λk=E[Ni∣Ei=1,Zi=k],Z_i = g_{\theta}(X_i), \qquad \lambda_k = \mathbb{E}[N_i\mid E_i=1,Z_i=k],

where the second expression requires a specific exposure normalization and an outcome-generating assumption. A clustering procedure estimates or constructs g; it does not determine lambda_k without claims observation, a model of claims frequency, and evidence that the conditional expectation is stable in the intended population. When exposure duration varies, a more appropriate illustrative frequency specification is E[N_i | Z_i=k,E_i]=E_i lambda_k, subject to the chosen stochastic model.

The distinction between a risk indicator and a risk estimand is essential. An unsupervised group may be useful because it identifies a concentration of construction types even where its historical average claim count is indistinguishable from neighboring groups. Conversely, a group with high historical claim counts may be driven by unusually long exposure or incomplete claim settlement rather than any persistent latent hazard.

Insurance records arise through institutional selection

Observed insurance data are sampled after quoting, marketing, underwriting, acceptance, renewal, cancellation, and policyholder decisions. The data-generating process is therefore conditional on coverage being offered and purchased. A hypothetical high-hazard class priced out of the insurer's portfolio is absent from the dataset; a newly introduced telematics product may contain drivers who opt in for reasons correlated with unobserved behavior. Cluster centers estimated among selected policyholders describe that selected population.

An analyst seeking a broader insurance-market statement must ask what the recording mechanism excludes and whether inclusion probabilities vary with X or the unobserved outcome. This is a statistical identification question, not a cosmetic adjustment. If S_i indicates participation, then the available feature distribution is P(X | S=1). It generally differs from P(X) for the intended market unless suitable conditional independence, transportability, or weighting assumptions are justified. Clustering P(X | S=1) cannot identify missing segments outside the recorded support.

Measurement is further shaped by contractual and operational rules. Deductibles, limits, benefits, waiting periods, subrogation, repair networks, fraud controls, inflation, and case reserving can alter Y even when underlying physical hazards are unchanged. Mixing contract cohorts can produce groups organized around coverage design. A loss-linked cluster therefore requires attention to the distinction between exposure risk, insured contractual loss, recorded claim, and settled payment.

The importance of purpose-dependent data fitness is reflected in the U.S. Actuarial Standards Board's Data Quality practice standard, which concerns selection, review, and use of data in actuarial services within its professional scope. A scientifically sound analysis preserves source provenance, data dictionaries, as-of dates, transformations, and known reporting incompleteness. A standardized dataset whose substantive origin cannot be reconstructed is inadequately specified for an insurance inference.

Granularity, dependence, and repeated observations

Policies are not necessarily independent observations. One insured may have several policies, one fleet may contribute hundreds of vehicles, and one catastrophe may generate claims across numerous contracts. A segmentation algorithm that treats every duplicated row as an independent customer can overweight large accounts and turn data architecture into artificial statistical density.

Different levels of analysis should be modeled explicitly. At policy level, the analyst might cluster coverage configurations; at vehicle level, physical and use attributes; at household level, interdependent assets; at spatial cell level, hazard accumulation; and at claim level, settlement pathways. The resulting group labels belong to different unit types and may not be meaningfully compared. Hierarchical sampling can require group-aware resampling at account, household, broker, event, or geographic levels.

A portfolio assembled over several years also contains temporal regimes. Weather patterns, supply-chain prices, road networks, medical practices, underwriting criteria, and reporting systems can change. A cluster defined by calendar-era indicators may be highly stable within the historical extract yet irrelevant to future prospects. Period indicators may still be scientifically informative, but the analysis should state whether it investigates secular evolution or an invariant exposure typology.

Three separable meanings of risk group

The phrase risk group has at least three scientifically distinct uses. A descriptive exposure group is a set of records that are similar under declared attributes and a declared distance or likelihood. A predictive risk group has demonstrated a relationship with a specified future loss distribution under an independent test. A decision group is a class an institution is authorized and willing to use in underwriting, pricing, reserving, claims operations, or prevention.

Movement from one meaning to the next requires evidence. Descriptive geometry does not confer actuarial credibility. Predictive association does not automatically establish causation, fairness, transparency, or legal admissibility. This separation is an analytical clarification for interpreting established methods; it is not advanced as a new theoretical doctrine. It prevents a common inferential leap in which a software-generated cluster number acquires the authority of a validated insurance category.

The most defensible early role of clustering is hypothesis generation and structured portfolio understanding. Its stronger roles arise where independently observed outcomes, explicit uncertainty measures, and sound governance support them.

Distance Metrics and Scaling

Similarity is a substantive modeling decision

Distance governs which portfolio records appear near each other. Consider numerical attributes measured in kilometers, vehicle age in years, building replacement value in currency, and ratio variables. Direct Euclidean distance

d2(xi,xj)=∑r=1p(xir−xjr)2d_2(x_i,x_j)=\sqrt{\sum_{r=1}^{p}(x_{ir}-x_{jr})^2}

is sensitive to measurement units. Multiplying one column by one thousand changes the geometry, even though the underlying insured object is identical. Standardizing numerical fields can correct arbitrary scale differences, but it can also assign excessive influence to a rare, noisy variable. Neither unscaled Euclidean distance nor automatic z-scoring has an insurance-specific priority independent of analytic purpose.

Suppose a motor portfolio contains annual mileage, nighttime driving share, vehicle age, and a binary usage-purpose flag. If the research question concerns driving behavior, vehicle age may be a control or supplementary descriptor rather than a dominant coordinate. A scientifically explicit similarity rule can place greater weight on mileage-adjusted nighttime share and on driving-pattern summaries. Such weights should be fixed from the substantive research aim or evaluated under sensitivity analysis; tuning them against the same claims later described as an untouched external validation set would compromise independence.

A weighted Euclidean measure may be written as

dw(xi,xj)2=∑r=1pwr(xir−xjrsr)2,wr≥0,d_w(x_i,x_j)^2=\sum_{r=1}^{p} w_r\left(\frac{x_{ir}-x_{jr}}{s_r}\right)^2,\qquad w_r\geq 0,

where s_r is a prespecified scale and weights express research priorities. An associated Mahalanobis distance uses an inverse covariance matrix to account for correlated directions, but covariance estimation becomes unstable when the number of variables approaches or exceeds the effective sample size; heavy-tailed variables and mixed categorical codes introduce further difficulties. Regularization and robust estimation are substantive decisions rather than defaults.

Mixed numerical and categorical features

Insurance information is often intrinsically mixed: age and mileage are continuous, property type and vehicle use are categorical, coverage options are binary or multivalued, and geographic jurisdictions are structured rather than ordered. The number assigned to a categorical label is an encoding artifact. Treating occupation category 4 as twice occupation category 2 creates a fictitious metric.

Gower's (1971) general similarity coefficient for mixed attributes provides a classical approach in which attribute-level contributions are combined with appropriate ranges and availability indicators. A simplified dissimilarity is

dG(i,j)=∑r=1pwr δijr dijr∑r=1pwr δijr,d_G(i,j)=\frac{\sum_{r=1}^{p}w_r\,\delta_{ijr}\,d_{ijr}}{\sum_{r=1}^{p}w_r\,\delta_{ijr}},

where d_ijr is a bounded contribution, delta_ijr marks comparability, and the denominator must be positive. Numerical differences may be scaled by an observed or meaningful range; nominal categorical differences may be zero for a match and one otherwise. The actual Gower family requires correct variable-type handling, including asymmetrical binary characteristics where appropriate.

The denominator's dependence on observed fields creates a subtle issue: two pairs may have distances computed from different subsets of variables. A driver record missing usage purpose might appear similar to another record because the disputed field did not contribute at all. This is not the same as demonstrating similarity on the full intended attribute space. Missingness may be analytically informative, and the missingness pattern itself may reflect business processes. The analyst should investigate which pairs share sufficient comparable coordinates.

High-cardinality categories require particular caution. Postal codes, vehicle models, intermediary identifiers, diagnostic codes, and repair shop identifiers can produce many sparse modalities. A direct exact-match penalty may cause groups to reflect arbitrary administrative identifiers. Collapsing categories, using carefully justified domain hierarchies, or learning representations under a documented process can be preferable. The same holds for broad location proxies that may covertly capture socioeconomic structure.

Dependence-aware categorical dissimilarity

Jamotton, Hainaut, and Hames (2024) present a Burt-distance approach to insurance clustering, using dependence among categorical modalities for mixed-feature segmentation, and benchmark alternatives in a motor portfolio. Their contribution is useful because insurance variables are often associated: vehicle age and type, driver experience and vehicle power, or property construction and geography are not independent categories. Counting their mismatches equally can exaggerate multiple measurements of one underlying exposure characteristic.

The empirical performance reported in that study is conditional on its portfolio, preprocessing, algorithms, and assessment criteria. Its results do not establish that Burt-based dissimilarity universally dominates Gower or other distances. Inference about optimal geometry requires new comparisons on the intended dataset, with consistent targets, runtime constraints, and independently defined evaluation criteria.

The categorical association structure can be unstable under selection, sparse cells, and changing product mixes. A rare combination may appear highly distinctive in one year and ordinary after an underwriting rule changes. If a learned dissimilarity uses a global contingency table, temporal transportability requires evaluation. Furthermore, a metric that captures rich dependence among X attributes may have weak association with future Y losses; the portfolio's most distinctive descriptive axis is not necessarily the most predictive one.

Exposure-normalized behavioral and financial variables

A claims count of five over ten insured-years describes a different rate from five over one insured-year. The feature engineer must decide whether to cluster counts, rates, event time distributions, or the underlying exposure-generating characteristics. Rates computed from tiny denominators may have high variance. Replacing counts with raw rates can make short-duration policies appear extreme merely through sampling noise.

For behavioral telematics, hard acceleration events per trip, per driving hour, or per kilometer answer different questions. A driver who travels many kilometers can have more recorded harsh-braking events without exhibiting a higher event rate. Device model, data-collection frequency, sensor outages, and matching of trips to policies can produce artificial proximity. Aggregation to a monthly or policy-term feature vector must be documented, including the information cutoff relative to underwriting decisions.

For claim analytics, gross amount, net amount, paid amount, incurred amount, and inflation-adjusted loss occupy distinct financial states. A mixed dataset that combines claim payment at different development maturities may identify administrative progress clusters rather than underlying loss severity. In reserving, this may be intentional; in underwriting, it is often a type of future-information leakage.

Dimensionality and sensitivity

As the number of weakly informative dimensions rises, some distance measures concentrate, making nearest neighbors increasingly difficult to distinguish. Even without appealing to a universal dimensionality theorem, an insurance portfolio can be tested directly: inspect distributions of nearest-neighbor and between-cluster distances, compare results after dropping weak variables, and assess whether a few features dominate the assignment.

Principal component analysis, multiple correspondence analysis, autoencoder embeddings, and similarity graphs can make large mixed data tractable, but each changes the definition of an observable group. A component maximizing feature variance is not optimized for a particular claim outcome. If feature extraction uses pretraining or supervised information, the entire pipeline should be described accurately. Zhu and Wüthrich's (2021) speed–acceleration heatmap study is instructive: its final k-means grouping followed feature extraction based on pretrained image recognition, so the full representation process has a history beyond the unsupervised objective.

Sensitivity analysis should vary scales, feature weights, missingness treatment, categorical representation, outlier handling, unit aggregation, and time horizon. An apparently robust K is inadequate when cluster membership changes radically under plausible alternative representations. The target is a justified meaning of similarity rather than the visual attractiveness of a two-dimensional projection.

Clustering and Mixture Models

Partitioning around centers

The classical k-means objective partitions n points into K groups to minimize within-cluster squared Euclidean dispersion:

min⁡μ1,…,μK, z1,…,zn∑i=1n∥xi−μzi∥22.\min_{\mu_1,\ldots,\mu_K,\ z_1,\ldots,z_n}\sum_{i=1}^{n}\|x_i-\mu_{z_i}\|_2^2.

For insurance exposures, a justified weighted objective could replace each summand by v_i times the squared distance, where v_i captures the chosen sampling or exposure emphasis. This weighting alters the scientific unit: emphasizing policy-years is different from giving each customer one vote. Weight selection should therefore be attached to the estimand. Algorithms based on iterative assignments and center updates typically reach a local optimum, and random initializations can yield differing partitions. Arthur and Vassilvitskii's (2007) careful k-means seeding method improves initialization behavior but does not make the fitted segmentation a unique economic truth.

K-means minimizes a geometric criterion; it does not directly optimize predictive insurance pricing. Its centroids are means of continuous coordinates, making one-hot categorical interpretation awkward. A medoid-based method instead uses observed representative records and can operate on suitable general dissimilarities. This can help when a representative policy profile is easier to audit than a synthetic centroid containing fractional categories. However, the medoid is still an algorithmic summary under the selected distance.

The practical trade-off involves stability, computational scaling, explanatory usefulness, and resistance to outliers. A cluster may be tight merely because one high-variance monetary variable was truncated, and a rare catastrophic-exposure subpopulation may be swallowed by a large central group. Evaluating economically important minorities demands more than minimization of average dispersion.

Hierarchies and nested insurance structures

Hierarchical clustering organizes observations into nested merges or splits. Single linkage can recover chains of near-neighbor points but may connect distinct exposure types through bridge observations. Complete linkage emphasizes maximum within-group distance and can favor compact structures. Average linkage averages pairwise dissimilarities, while Ward-type approaches depend on squared-Euclidean variance logic; a generic Gower dissimilarity cannot be assumed to preserve Ward's original objective.

Hierarchies are attractive when insurers want a coarse portfolio summary with progressively finer subgroups. Yet a dendrogram can visually imply meaningful nesting even when scale changes produce alternative branches. Selecting a cut height should be a documented research decision and, for risk claims, connected to external information. Very small terminal nodes often lack enough claims exposure to estimate outcome rates credibly.

Nested labels must not be confused with the actual dependence structure of insured objects. A fleet's vehicles may share a contract and risk-management process, regardless of whether their individual feature vectors end up in the same cluster. Hierarchical statistical sampling and hierarchical clustering address different mathematical objects.

Density-connected and graph-based methods

Density-based algorithms identify regions of relatively high observation density separated by lower-density areas. Ester and colleagues' (1996) original DBSCAN clustering procedure introduces density-connected groups and noise labeling suitable for nonconvex geometries. This can be useful when telematics behavior forms elongated patterns or property locations form irregular spatial concentrations.

For insurance data, density is inseparable from the sampling scheme. A densely populated postcode cluster can reflect the insurer's distribution network, not heightened physical hazard. A sparse cluster can represent an important low-volume catastrophe exposure. Global density thresholds can also make groups with different natural concentrations difficult to recover consistently. Applying DBSCAN to a badly scaled mixed representation merely produces density relations in that defective geometry.

Spectral clustering starts with a similarity graph and uses graph eigenstructure to represent connectivity before partitioning. It can recover structures missed by center-based methods, but the graph definition, neighborhood scale, and computational cost become central. Jamotton and colleagues' insurance spectral clustering comparison includes graph-oriented experiments; interpretation depends on the chosen graph and its preprocessing. A spectral partition describes the chosen similarity network and must still be linked to an insurance question.

Finite mixture models and latent assignments

A statistical mixture models the observed feature density as

f(x)=∑k=1Kπkfk(x∣θk),πk≥0,∑kπk=1.f(x)=\sum_{k=1}^{K}\pi_k f_k(x\mid\theta_k), \qquad \pi_k\geq0,\quad \sum_k\pi_k=1.

Under a fitted mixture, an observation can have posterior membership probabilities

τik=πkfk(xi∣θk)∑ℓ=1Kπℓfℓ(xi∣θℓ).\tau_{ik}=\frac{\pi_k f_k(x_i\mid\theta_k)}{\sum_{\ell=1}^{K}\pi_\ell f_\ell(x_i\mid\theta_\ell)}.

These probabilities express component uncertainty conditional on the model, feature vector, and fitted parameters. They are not probabilities of filing an insurance claim. Fraley and Raftery's (2002) framework for model-based clustering formalizes how model choice and density assumptions enter clustering. Dempster, Laird, and Rubin's (1977) expectation–maximization estimation algorithm supplies the general iterative framework used in many latent-variable likelihood problems.

For a Gaussian mixture, covariance restrictions affect whether components represent spherical, elongated, or otherwise oriented shapes. Mixture identification may be sensitive to starting values, small component weights, and local likelihood maxima. With mixed insurance fields, naive Gaussian treatment of nominal categories is inappropriate. Mixtures with multinomial, ordinal, count, and continuous component distributions can be more faithful, but conditional independence assumptions and model complexity become additional scientific commitments.

An important identification limitation follows: a probability density may be approximated by several components without there being the same number of physically real risk populations. Conversely, a single continuous gradient of exposure can be represented by many Gaussian components. The component count is a feature of the specified density model and selection method. Posterior uncertainty should be retained when group assignment is ambiguous, particularly in downstream analyses that might otherwise turn an uncertain component into an absolute individual classification.

Schwarz's (1978) dimension-penalized model selection criterion motivates the widely used BIC expression, BIC = -2 log L + q log n, where q is the parameter count. In mixture settings, standard asymptotic regularity can be delicate at boundary weights and coincident components; a smaller BIC does not certify stable risk classes or accurate future claim prediction. Different plausible covariance structures may select different K.

How many clusters?

K is a modeling parameter requiring a scientific rationale. Elbow plots are descriptive and often ambiguous. Rousseeuw's (1987) silhouette measure of cluster separation compares within-group cohesion with the nearest competing group. Tibshirani, Walther, and Hastie (2001) develop the gap statistic against a reference distribution, while mixture-based methods may select models by information criteria. Arbelaitz and colleagues (2013) show in their comparison of internal cluster validity indices that commonly used indices can behave differently under noise and overlap.

A large silhouette can reward groups that are geometrically separated on irrelevant administrative variables. A gap statistic depends on the reference distribution selected for comparison. An information criterion depends on the fitted likelihood family. Each method answers its own statistical question. If the objective is to characterize future insured loss, independent claims validation should accompany internal fit metrics.

The model comparison should report a plausible range of K, minimum support for each group, sensitivity to preprocessing, stability under seeds and resampling, and any substantive gain from an additional split. A fine-grained classification with dozens of small clusters can be visually impressive while impossible to support with credible claims estimates or practical explanation. The best statistical partition for exploratory data description may therefore differ from the parsimonious grouping useful for a subsequent decision analysis.

Soft membership, overlapping exposures, and nonexclusive hazards

Hard assignments allocate each policy to exactly one cluster. Insurance risk sources often overlap: a coastal property may be exposed simultaneously to flood, wind, wildfire smoke, and supply-chain interruption; a driver may combine long commutes with sporadic high-risk night trips. A single categorical label can obscure such multidimensionality.

Soft mixture assignments are one representation of overlap, but posterior component probabilities do not guarantee decomposition by real hazard. Factor models, multiple memberships, or separate dimension-specific clustering can sometimes serve the substantive purpose better. The analyst should avoid presenting component probabilities as calibrated event probabilities or combining them arithmetically with insurer prices without a valid outcome model.

The underlying lesson from the methods literature is a controlled form of pluralism: clustering algorithms encode different candidate structures. Method choice is part of the scientific claim, and no method can independently supply the missing evidence about claims, causality, or admissible insurance practice.

Latent Exposure Subgroups

Motor insurance: vehicles, drivers, and usage patterns

Motor portfolios offer an unusually rich test bed for unsupervised segmentation because policies describe vehicle characteristics, driver histories, geographical regions, coverages, and increasingly telematics behavior. An exploratory question may concern the structure of usage rather than claims: are there recurring patterns of night driving, mileage, smoothness, trip length, or acceleration distribution? The observation window and sensor measurement process must be fixed before clustering.

Jamotton, Hainaut, and Hames (2024) examine motor portfolio mixed-feature segmentation and propose a Burt-based method that accounts for dependencies among categorical modalities. Their paper's motor data and assessment results provide concrete evidence that heterogeneous insurance fields can be organized using alternatives to straightforward Euclidean k-means. They do not establish that every cluster is a legally usable rating factor in every jurisdiction or that the method is uniformly optimal.

Zhu and Wüthrich (2021) investigate clusters of speed–acceleration heatmaps represented as images. The study illuminates the importance of behavior summaries and feature-extraction architecture. The output concerns distinctions in observed driving styles. Any translation into prospective claim effects requires suitable claims, exposure denominators, and observation timing.

Hannon and O'Hagan (2025) take a further step in cluster-specific motor claim modeling: they use Gaussian mixtures and subsequent binary regressions to evaluate how telematics and traditional information interact. The separation of an unsupervised grouping stage from an outcome regression stage is especially instructive. Their results arise from a synthesized telematics dataset emulated from real insurance data and from the study's particular modeling strategy. Reuse in another insurer or jurisdiction requires independent transportability and conduct assessment.

The insurer should distinguish behavioral grouping from a judgment of driver character. Differences in shift-work schedules, road environments, device quality, and distance traveled can affect behavior measurements. A person with an elevated nighttime-driving share may have no correspondingly elevated claim frequency after accounting for exposure and road conditions. An apparently accurate risk predictor can also operate as a proxy for protected or otherwise legally constrained attributes. Descriptive clustering is a reason to investigate such channels, not a shortcut around them.

Property and catastrophe portfolios

For property insurance, clustering can summarize construction material, building age, elevation, occupancy, retrofit history, surrounding vegetation, replacement cost, and location attributes. The scientific purpose could be asset typology, concentration mapping, missing-data diagnosis, or disaster preparedness. A spatial cluster of insured buildings is neither necessarily a zone of homogeneous peril intensity nor a sufficiently detailed accumulation unit.

Catastrophe losses exhibit event-level dependence. A storm may affect many geographically adjacent risks at once, while two widely separated assets can be linked through infrastructure, reinsurance arrangements, or supplier relationships. Therefore, cluster validation for property concentration should include relevant hazard scenarios and event aggregation rather than comparing only independent annual claim counts.

A useful property grouping may reveal that similar construction styles are concentrated within one drainage basin. To infer flood risk from that grouping, the analyst must connect it to flood elevation, hydrology, defenses, insurance terms, and observed or modeled event losses. If a cluster contains both insured and uninsured buildings, institutional selection must be addressed before drawing conclusions about the broader built environment.

For catastrophe portfolio management, the unit may be a geospatial cell with insured value and building exposure rather than an individual policy. An area-level grouping cannot automatically assign the same expected loss to each property in the area. This ecological inference problem makes scale and spatial aggregation central to scientific interpretation.

Health, life, and disability insurance

Health and life portfolios contain age, duration, coverage, socioeconomic, medical, and administrative information with potentially strong selection and regulatory sensitivity. The scientific purpose might involve claim-service patterns, benefit utilization, policy persistency, morbidity trajectories, or underwriting process differences. Each task requires its own unit, time window, allowed variables, and endpoint.

Clustering past utilization may discover different service pathways while being strongly affected by provider access, coding regimes, care availability, or benefit design. A high-utilization group does not necessarily represent a high underlying disease-risk group. In life insurance, cohorts separated by coverage vintage may have different observable profiles because underwriting practice and mortality exposure changed over time.

In these lines, segmentation is particularly likely to touch sensitive information. In the European Union, Annex III(5)(c) of Regulation (EU) 2024/1689 identifies specified AI systems for risk assessment and pricing in relation to natural persons in life and health insurance, subject to the Regulation's Article 6 classification provisions and applicable commencement rules. This classification does not mean that every descriptive clustering exercise is automatically a high-risk deployed AI system, and it cannot be generalized to all insurance lines or jurisdictions.

Scientific investigation must distinguish descriptive research, supervised decisions, operational use, and legally regulated high-risk systems. The data protection, insurance conduct, and antidiscrimination questions depend on particular applications and legal contexts. When the project is exploratory, reducing identifiability and reporting aggregate findings can be appropriate design choices, but privacy cannot be inferred merely from replacing names with numerical cluster IDs.

Claims handling and operational process segmentation

Claims records offer rich event sequences: notification, assessment, repair, litigation, reserve revision, partial payment, and final settlement. Unsupervised process grouping can identify operational pathways and bottlenecks. A cluster of long-settling claims may be relevant to staffing and workflow redesign even when the ultimate insured peril is shared with faster-settling claims.

Timing defines the permissible inference. If the task is to allocate handlers at first notice of loss, the clustering features must be available at first notice. Final settlement time, ultimate incurred amount, and outcome of litigation cannot enter the first-notice clustering representation without creating future-information leakage. A historical retrospective taxonomy can legitimately use them, provided the manuscript states that it is retrospective.

Because claims handling itself changes the observed outcome, administrative clusters should not be casually interpreted as natural hazard strata. A slow process may reflect internal queue management or policy complexity. A high-reserve group may arise from company procedures rather than insured hazard. The proper validation endpoint may be elapsed handling time, reopened claim rate, customer service outcome, or error incidence, with attention to human intervention.

Portfolio economics and meaningful small groups

Insurance capital and risk-management decisions often depend on the tails of loss distributions. A rare peril combination can be economically important even when it represents a negligible fraction of policies and has low contribution to a global clustering objective. A segmentation focused on minimizing average within-cluster dispersion may ignore it. Discovery of unusual observations should be managed alongside clustering, without equating statistical outlier status with fraudulent behavior or elevated expected loss.

Clustering can also uncover redundancy: several nominal coverage-product categories may form one coherent exposure profile. This may improve reporting and portfolio understanding. However, smoothing or merging categories for convenience can obscure contractually significant differences. Product conditions and covered events belong in the explanatory context of every proposed group.

A group's economic significance therefore has at least three dimensions: portfolio weight, loss-relevant exposure, and practical decision consequence. A small, high-severity cluster might deserve separate investigation; a large and extremely stable cluster might have little explanatory value for loss variability. The choice depends on the research question rather than an automatic ranking by cluster size or average claim cost.

Worked illustration: exposure weighting and independent outcomes

Consider two hypothetical clusters formed exclusively from pre-period vehicle attributes. Suppose the next independent calendar period supplies the following fully observed claim counts and exposures:

Group Policy-year exposure Claims Claims per policy-year Approximate Poisson rate standard error
A 600 24 0.040 0.0082
B 200 20 0.100 0.0224
Combined 800 44 0.055 0.0083

The combined claim frequency is 44 / 800 = 0.055 per policy-year. The unweighted average of the two cluster rates is (0.040 + 0.100) / 2 = 0.070 and is inappropriate for the portfolio aggregate because exposure weights differ. Under independent Poisson counts with fixed exposures, an elementary estimated standard error for a cluster rate is sqrt(C_k)/E_k; thus B's rate is less precise than A's despite its larger point estimate. Dependence, overdispersion, claim-reporting delay, and large-loss heterogeneity would require a richer uncertainty model.

For an illustrative Poisson comparison with one common rate, the pooled rate is 0.055 and expected counts are 33 for A and 11 for B. The standard Poisson deviance against saturated group-specific rates is

D=2∑k∈{A,B}[Cklog⁡(CkEkλˉ)−(Ck−Ekλˉ)]≈8.628.D=2\sum_{k\in\{A,B\}}\left[C_k\log\left(\frac{C_k}{E_k\bar\lambda}\right)-(C_k-E_k\bar\lambda)\right]\approx8.628.

This calculation is arithmetic on an invented, declared dataset; no real claim-frequency result or generalizable statistical discovery is asserted. The numerical deviance measures a difference in the example under its Poisson setup; it is not a cluster-validity index, a causal effect, or proof of practical pricing eligibility. No threshold conclusion is drawn without a specified test protocol and assumptions.

The critical design condition is that the 24 and 20 outcome counts did not enter cluster fitting, distance selection, K tuning, or pre-period subgroup labeling. Had the analyst searched among many segmentations and selected the one maximizing exactly this claim contrast, the same contrast would no longer be independent confirmation. It would be an optimized training statistic, likely overstating future differentiation. The distinction between discovery and out-of-sample validation is more valuable than the numerical size of the contrast.

Validation of Risk Groups

Internal validity addresses geometric structure

Validation should be divided into internal geometry, computational reproducibility, sampling stability, and externally observed insurance outcomes. These are complementary tests, not interchangeable quality scores. An internal index answers whether the observed grouping behaves as intended under the chosen metric, while an external analysis asks whether the grouping corresponds to a separate property of interest.

For a point i with mean dissimilarity a(i) to its own cluster and smallest mean dissimilarity b(i) to another cluster, the conventional silhouette is

s(i)=b(i)−a(i)max⁡{a(i),b(i)},s(i)=\frac{b(i)-a(i)}{\max\{a(i),b(i)\}},

with the usual method-specific treatment of singleton clusters. Average silhouette width can assist partition comparison, as established by Rousseeuw (1987), but only within a coherent metric and analysis population. Comparing silhouettes across incompatible feature definitions may be misleading.

The gap statistic procedure of Tibshirani and colleagues compares observed within-cluster dispersion with a reference distribution. Its reference support must be defined carefully for constrained insurance features; independently sampling categorical indicators or physically impossible attribute combinations can create an unrealistic null. Likewise, the comparative study of cluster validation under noise and overlap demonstrates why no single internal index should be treated as the scientific verdict.

A dimensionality-reduced plot is an interpretive aid. A method may produce well-separated colors in an embedding because the projection was optimized for local neighborhood separation. The visual impression does not establish equally separated original-space risks or different future claim distributions. Report the original-space criteria and preprocessing alongside any figure.

Stability as an empirical object

A reproducible cluster must survive relevant variation in computational initialization and in the data-generating sample. Repeating a deterministic algorithm on exactly the same file tests almost nothing about robustness. More informative checks vary random seeds, bootstrap samples, collection periods, subsamples of insured entities, and plausible perturbations to measurement and preprocessing.

Hubert and Arabie's (1985) chance-corrected comparison of partitions provides a standard approach to comparing hard partitions, after matching records across runs. Hennig's (2007) cluster-wise Jaccard stability assessment focuses on whether individual clusters recur under perturbation, avoiding sole reliance on one overall stability number. Von Luxburg's (2010) review of clustering stability limitations clarifies that stable solutions can occur for reasons other than recovery of uniquely real groups.

For insurance panels, resampling policy rows independently may violate event, vehicle, household, or fleet dependence. Cluster stability can appear artificially high when copied or correlated records are repeatedly sampled. The resampling unit should follow the target population and data structure. Event-level clusters may require event-block resampling; fleet-level findings may require fleet blocks; longitudinal telematics may require customer or driver blocks.

Stability should be reported at several levels: K agreement; frequency with which individual groups are recovered; assignment probability for ambiguous records; changes in medoids or centroids; interpretation of the group profile; and sensitivity of any downstream loss contrast. A stable profile that changes its label number is a labeling issue, not substantive drift. A cluster that persists geometrically but changes completely in actuarial loss experience is descriptively stable and predictively unstable.

Outcome validation requires an independent protocol

For a claimed insurance-risk application, the relevant validation target might be next-period claim frequency, claim severity, pure premium, time to event, catastrophe contribution, or operational performance. The choice should precede the analysis. A cluster with a meaningful difference in past claim frequency may fail on next-period severity, and a group with high average cost may merely contain larger sums insured.

A defensible protocol fixes the observation unit, as-of date, training sample, preprocessing, candidate methods, hyperparameter-selection rule, group assignment procedure, and validation period. It then assesses a genuine holdout. Reusing observations for exploratory feature selection can contaminate an apparently prospective test even if the final clustering was technically unsupervised. Grouping can be fitted without Y while still being selected after examining Y across dozens of candidate partitions.

Out-of-sample analyses should compare a baseline actuarial model and a baseline without clusters, using the same policy population, exposure terms, and outcome definition. A simple evaluation asks whether a cluster label adds stable predictive value conditional on conventional covariates. That comparison can be carried out with Poisson, negative-binomial, binary, severity, or survival formulations as appropriate to the loss-generating process.

A cluster might appear predictive only because it summarizes an already modeled feature; then its additional information may be negligible even if groups have dramatically different raw rates. The scientific question is the incremental contribution of a robust descriptive grouping. Model complexity, extra validation effort, instability, and administrative costs must be considered along with nominal fit improvement.

Claims maturity, credibility, and uncertainty

Insurance outcomes arrive with delays. A recent policy cohort may have incomplete reported claims or immature incurred losses. Comparing such a cohort directly with a fully developed older cohort can create a false temporal trend in group performance. The analyst should either use a mature endpoint, explicitly model development, or state the limitation.

Small group claims rates have substantial sampling uncertainty. For count outcomes, uncertainty depends on both the exposure denominator and the event variance. For loss severity, heavy-tailed events can dominate averages; ordinary Gaussian confidence intervals may be inappropriate. Hierarchical or partially pooled models can help with sparse groups, but their assumptions and priors or shrinkage penalties must be disclosed.

For a clear external test, record group sizes, effective exposure, event counts, distributional summaries, uncertainty intervals, and the degree of dependence among observations. Report absolute and relative differences, not just statistical significance. A slight yet statistically precise difference in a massive portfolio might lack practical importance; a large difference from ten claims might be too imprecise for pricing or risk governance.

Validation should also assess claim definition. A paid claim under a deductible is not identical to an insured accident, a reported claim under one notification policy, or an ultimate loss under another coverage. If the group is supposed to characterize physical hazard, a contract-contingent claims endpoint must be understood as an indirect measure.

Drift, transfer, and transportability

A segmentation fitted before a change in product rules, underwriting appetite, mobility patterns, climate exposure, or regulatory practice may become obsolete. Monitoring should distinguish input drift P(X), cluster-assignment drift P(Z), outcome drift P(Y | Z), and changes in the relationship between the representation and the target. These can move independently.

Cross-insurer and cross-jurisdiction transfer is especially demanding. Two companies may use different coding systems, deductible structures, policy terms, distribution channels, or claim settlement practices. A cluster name such as urban high-mileage commercial can mask different empirical populations. Transportability requires comparable definitions and independent checks on outcome relationships.

The date of initial clustering is a model artifact; the date of a subsequent portfolio monitoring report is a separate one. Version control should preserve data windows, code hashes, transformation choices, and assignment rules. Re-fitting a grouping without a lineage record makes historical comparisons ambiguous, because a group with the same human-readable label may represent a different subset of contracts.

Interpretable validity audit

A useful scientific dossier specifies a sequential set of questions for each proposed group:

  1. What is the unit, population, feature cutoff, and exposure denominator?
  2. Which columns determine proximity, how are they scaled, and what type of missingness is present?
  3. Does the group recur under realistic re-estimation and dependence-aware resampling?
  4. Does it differ on a prespecified, independently observed outcome of relevance?
  5. Is that difference robust after adjustment for obvious confounders and contract differences?
  6. Is the distinction interpretable, stable over time, and practical for the proposed action?
  7. Can the action be justified under the governing professional, legal, and conduct requirements?

The audit is a transparent application of established scientific reasoning. Passing only the geometric or stability conditions authorizes a description of portfolio structure. Stronger actuarial claims require additional evidence; downstream decisions require a further institutional review. No cluster number is entitled to bypass these steps.

Ethical and Practical Interpretation

Research use, decision use, and regulatory scope

The consequences of unsupervised clustering differ by use. A researcher who clusters anonymized aggregated exposure profiles to understand portfolio composition is conducting a descriptive analysis. An insurer who assigns individual customers to clusters and uses that label to set coverage eligibility or premiums is making or supporting a consequential operational decision. The latter requires examination of the entire decision chain, even when the grouping algorithm was fitted without claim labels.

The European Union's Artificial Intelligence Act classification provisions specify AI-system obligations and categories of high-risk use; Annex III identifies certain natural-person life and health insurance risk assessment and pricing applications. Classification depends on intended use and Article 6. The legal effect of an algorithmic segmentation cannot be inferred from its being called unsupervised. Its deployment context, profiling function, and downstream decisions matter.

EIOPA's 2025 Opinion on AI governance clarifies existing insurance-sector supervisory expectations for relevant AI systems, including fairness, record keeping, explainability, and human oversight, using a proportional approach. EIOPA expressly distinguishes the Opinion's scope from AI Act high-risk and prohibited systems. The Opinion is addressed to European national supervisors and does not impose globally uniform rules on insurers.

The IAIS 2025 AI supervision application paper discusses risk-based supervision, governance and accountability, robustness, transparency, and fairness and redress in relation to existing Insurance Core Principles. It is international supervisory guidance, not a single law operative unchanged in every jurisdiction. A global encyclopedia should preserve this document-type distinction.

In the United States, the NAIC's model bulletin on insurer AI systems was adopted by the NAIC in December 2023 and describes governance and documentation expectations. A model bulletin becomes applicable through relevant state action and existing law, with jurisdiction-specific wording and enforceability. The U.S. professional context also includes ASOP No. 56 on modeling and ASOP No. 23 for actuaries performing services within those standards' scopes. Treating these instruments as globally interchangeable would corrupt comparative analysis.

Fairness and proxies

Clustering is capable of organizing personal and asset information along correlated social dimensions. Postal code, occupation, travel time, device adoption, medical utilization, and payment patterns can convey information about demographics, income, or access to services. Removing a protected field from X does not guarantee that correlated fields cease to act as proxies. Clusters may also be based on incomplete records that disproportionately omit certain populations.

Fairness assessment requires defining both the relevant group interests and the insurance decision context. A descriptive segmentation can be examined for disparate composition and unstable representation; a predictive segmentation requires validation of outcomes; a decision application requires assessment of treatment, justification, applicable law, contestability, and adverse consequences. These obligations cannot be compressed into a single statistic such as silhouette width.

The analyst must also examine feedback loops. A customer classified into a higher-priced segment may change coverage, exposure, or willingness to share telematics. Future data then reflect the decision rule that used the clusters, and the segment can appear self-confirming. Withdrawal from coverage and market selection distort the observed outcomes. Where lawful differential pricing is considered, governance should investigate these behavioral consequences in addition to static predictive fit.

A cluster can serve inclusive purposes as well: identifying coverage gaps, poor access to mitigation resources, or administrative barriers can guide service improvement. Such use demands careful definition of target population and avoidance of unwarranted individual inference. The scientific value of identifying a service gap is independent of whether the same segmentation would be legitimate as a price multiplier.

Explainability and governance

Cluster explainability starts with an accurate description of the features, preprocessing, and objective function. It should describe representative profiles and distributions rather than attach psychological or moral labels to people. A label such as high-risk driver must be justified by independently defined risk evidence; smooth, late-night, regional, or short-trip descriptors are closer to the actual observed behavior when those were the clustering inputs.

Governance should document the accountable owner of the model, allowed applications, sensitive variables, training and monitoring windows, version identifiers, key assumptions, user permissions, quality assurance, and procedures for challenge or correction. A human reviewer needs the ability to understand why a policy was grouped and what evidence supports any subsequent decision. For soft assignments, the uncertainty interval or posterior component distribution may be relevant to the review.

The Model Bulletin and the EIOPA and IAIS publications support risk-based governance expectations in their respective institutional settings. Their specific legal force and supervisory implementation require local review. This monograph uses them to analyze the structure of responsible deployment, not to provide individualized legal advice.

Professional modeling practice reinforces attention to intended use, validation, and disclosure. The U.S. ASB's actuarial guidance on data fitness and actuarial model governance standard are concrete examples of institutional controls with specified scope. Applying an analogous documentary discipline in research is prudent, even where a particular professional standard does not legally govern a given researcher.

Reproducible empirical protocol

A complete independent study should preserve the following materials: population and exclusion rules; data lineage and snapshot timestamps; sampling weights and exposure construction; raw-data quality observations; feature construction specifications; missingness and categorical handling; scaling constants; algorithms and software versions; random seeds; initialization counts; hyperparameter search; group profiles and distributions; resampling and agreement measures; independently held-out claims results; subgroup fairness review; and governance signoff for any proposed use.

A reproducible split must respect the operational as-of date. For motor telematics, behavioral features used to predict a later claim can come from a pre-claim observation interval, while features recorded after the outcome cannot. For property exposures, hazard and insured-value snapshots must be contemporaneous with the test interval. For claim-process segmentation, the relevant available-state record depends on the stage at which a decision would occur.

A useful reporting table separates three rows of evidence: geometry fit, sampling stability, and external risk performance. Geometry metrics may include dispersion and silhouette. Stability metrics may include cluster-wise bootstrap Jaccard values or chance-adjusted partition agreement. Risk-performance metrics may include independently estimated frequencies, loss costs, uncertainty intervals, or holdout predictive scores. These rows make the inferential boundaries visible to subsequent reviewers.

Computational efficiency matters in large portfolios. Dense pairwise distance matrices can require order n-squared memory, and some graph-based procedures have costly eigencomputations. Mini-batch or approximate procedures can reduce runtime, but the approximation must be checked against accuracy and stability on a representative subset. Efficiency is a real research consideration, yet it cannot substitute for suitability of the underlying metric.

International research agenda

Further scientific studies should compare segmentation across life, health, motor, property, agricultural, and specialty insurance without presuming the same latent structures. Multi-insurer evaluation can separate universal technical regularities from company-specific data and contract selection. Studies of catastrophe exposure should incorporate spatial and event dependence. Telematics studies should test device and enrollment selection, pre-period stability, and incremental prediction after conventional covariates. Claims studies should distinguish retrospective workflow typologies from prospectively available features.

The most demanding open methodological problem is inference after adaptive discovery: an analyst may fit many unsupervised configurations, inspect outcomes, and finally present the most favorable segmentation. The separation between unsupervised fitting and supervised validation is therefore a property of the entire research protocol, not merely of the final algorithm's objective function. Prespecified outcome selection, locked holdouts, nested evaluation, and transparent reporting of candidate searches are essential for credible claims of actuarial relevance.

Another open question concerns meaning under changing conditions. A group defined by a current social, technological, or climate exposure regime may be a transient feature of the portfolio. Evaluations should examine the stability of descriptions and of outcome associations separately. Apparent stability of K or of geometric centroids is insufficient when the underlying insurance claim mechanism or institutional selection changes.

The appropriate conclusion is empirical discipline. Unsupervised methods enable discovery of candidate structures, and subsequent statistical and institutional work determines what those structures mean. International insurance research benefits most when the layers of description, prediction, and decision are documented with equal precision.

Conclusion

Unsupervised risk segmentation is an established family of methods for discovering structure in insurance data. Its objects include descriptive partitions, density-connected sets, hierarchical groupings, and latent mixture components. The choice among them depends on the observational unit, feature representation, similarity criterion, sample selection, and research purpose. Insurance data introduce specific complications through mixed categorical and numerical fields, variable exposure denominators, contractual differences, correlated insured risks, time-dependent claim observation, and institutional selection.

The fundamental scientific result of this synthesis is an inferential separation: similarity among recorded exposures does not by itself establish similarity in future insured losses, while a demonstrated loss association does not by itself authorize a classification decision. Geometric validity, reproducibility, independent claims validation, and accountable institutional use are cumulative evidence domains. This conclusion follows from the statistical properties of clustering methods, the insurance-specific empirical literature, and the differentiated scopes of professional and supervisory instruments; it is presented as analytical synthesis, not as a newly named theory.

The published actuarial studies of heterogeneous motor portfolio clustering and telematics demonstrate research feasibility and illustrate distinct model architectures. They do not supply a universal optimal algorithm for all insurance lines. A valid application therefore begins by declaring its target, ends with independent evaluation under the intended population, and preserves the reasoning connecting raw records to a consequential interpretation.

The hypothetical exposure-weighted calculation shows how easily segment comparisons can be distorted by denominator choices, and how outcome separation can be misrepresented when the same claims are used to select and validate a partition. Proper scientific practice records both the opportunities for portfolio insight and the boundaries of what the evidence establishes. External peer review remains a separate process, and no public publication or population-level empirical discovery is asserted for this research edition.

References

  • Gower, J. C. (1971). A General Coefficient of Similarity and Some of Its Properties. Biometrics, 27(4), 857–871. Source and bibliographic record.

  • Rousseeuw, P. J. (1987). Silhouettes: A Graphical Aid to the Interpretation and Validation of Cluster Analysis. Journal of Computational and Applied Mathematics, 20, 53–65. Source and bibliographic record.

  • Tibshirani, R., Walther, G., & Hastie, T. (2001). Estimating the Number of Clusters in a Data Set via the Gap Statistic. Journal of the Royal Statistical Society: Series B, 63(2), 411–423. Source and bibliographic record.

  • Fraley, C., & Raftery, A. E. (2002). Model-Based Clustering, Discriminant Analysis, and Density Estimation. Journal of the American Statistical Association, 97(458), 611–631. Source and bibliographic record.

  • Dempster, A. P., Laird, N. M., & Rubin, D. B. (1977). Maximum Likelihood from Incomplete Data via the EM Algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39(1), 1–22. Source and bibliographic record.

  • Hennig, C. (2015). What Are the True Clusters? Pattern Recognition Letters, 64, 53–62. Source and bibliographic record.

  • Hennig, C. (2007). Cluster-Wise Assessment of Cluster Stability. Computational Statistics & Data Analysis, 52(1), 258–271. Source and bibliographic record.

  • von Luxburg, U. (2010). Clustering Stability: An Overview. Foundations and Trends in Machine Learning, 2(3), 235–274. Source and bibliographic record.

  • Hubert, L., & Arabie, P. (1985). Comparing Partitions. Journal of Classification, 2, 193–218. Source and bibliographic record.

  • Jamotton, C., Hainaut, D., & Hames, T. (2024). Insurance Analytics with Clustering Techniques. Risks, 12(9), Article 141. Source and bibliographic record.

  • Zhu, R., & Wüthrich, M. V. (2021). Clustering Driving Styles via Image Processing. Annals of Actuarial Science, 15(2), 276–290. Source and bibliographic record.

  • Hannon, J., & O’Hagan, A. (2025). Statistical Models for Improved Insurance Risk Assessment Using Telematics. British Actuarial Journal, 30, e16, 1–39. Source and bibliographic record.

  • Ester, M., Kriegel, H.-P., Sander, J., & Xu, X. (1996). A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD-96) (pp. 226–231). AAAI Press. Source and bibliographic record.

  • Arthur, D., & Vassilvitskii, S. (2007). k-means++: The Advantages of Careful Seeding. In Proceedings of the Eighteenth Annual ACM–SIAM Symposium on Discrete Algorithms (pp. 1027–1035). SIAM. Source and bibliographic record.

  • Arbelaitz, O., Gurrutxaga, I., Muguerza, J., Pérez, J. M., & Perona, I. (2013). An Extensive Comparative Study of Cluster Validity Indices. Pattern Recognition, 46(1), 243–256. Source and bibliographic record.

  • Schwarz, G. (1978). Estimating the Dimension of a Model. The Annals of Statistics, 6(2), 461–464. Source and bibliographic record.

  • European Insurance and Occupational Pensions Authority (EIOPA). (2025, August 6). Opinion on Artificial Intelligence Governance and Risk Management (EIOPA-BoS-25-360). Source and bibliographic record.

  • International Association of Insurance Supervisors (IAIS). (2025, July 2). Application Paper on the Supervision of Artificial Intelligence. Source and bibliographic record.

  • European Parliament and Council of the European Union. (2024). Regulation (EU) 2024/1689 Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act), consolidated text as of July 27, 2026, especially Article 6 and Annex III(5)(c). Source and bibliographic record.

  • National Association of Insurance Commissioners (NAIC). (2023, December 4). Model Bulletin: Use of Artificial Intelligence Systems by Insurers. Source and bibliographic record.

  • Actuarial Standards Board. (2016; effective April 30, 2017). Actuarial Standard of Practice No. 23: Data Quality. Source and bibliographic record.

  • Actuarial Standards Board. (2019; effective October 1, 2020). Actuarial Standard of Practice No. 56: Modeling. Source and bibliographic record.