RESEARCH DOSSIER
High-Cardinality Feature Engineering in Insurance
Abstract. Insurance portfolios contain categorical attributes with hundreds, thousands, or millions of distinct levels: occupation codes, repair facilities, intermediary identifiers, postal micro-areas, vehicle variants, insured enterprises, clinical provider classifications, commercial activities, and combinations of risk descriptors. These variables may capture meaningful heterogeneity, yet their apparent predictive power is particularly vulnerable to sparse observations, differential exposure, delayed outcome reporting, target leakage, temporal drift, identity duplication, and legal constraints on classification. This monograph develops an integrated statistical and actuarial treatment of high-cardinality feature engineering. It defines the estimation target and the information available at prediction time, compares one-hot, hashing, grouping, smoothed target statistics, hierarchical partial pooling, ordered methods, and learned representations, derives shrinkage and leakage results, and specifies group-aware and chronological out-of-sample validation. Original mathematical illustrations examine exposure-weighted category credibility, self-inclusion bias, unexpected levels, label maturation, and deployment-time stability. The comparative discussion connects peer-reviewed actuarial work, statistical learning research, professional model and data-quality guidance, and jurisdiction-specific insurance data governance. Its principal conclusion is that the relevant unit of evaluation is an entire time-indexed feature-construction and prediction procedure, including category creation, exposure allocation, outcome availability, smoothing, downstream learning, and operational fallback. Category count alone neither establishes predictive value nor determines an encoding strategy. The contribution is an original analytic synthesis and reproducible worked reasoning, without claims to a new theory, original dataset, universal superiority of an algorithm, or external peer review.
Keywords: High-cardinality categorical variables; insurance predictive modeling; actuarial feature engineering; sparse categories; rare levels; exposure; credibility; target encoding; empirical Bayes; hierarchical pooling; leakage; cross-fitting; ordered boosting; categorical embeddings; feature hashing; unseen categories; group validation; temporal validation; insurance pricing; underwriting; claims data; fairness; model governance; distribution shift.
Introduction: Research Questions, Scope, and Method
Insurance prediction is a problem of learning from experience that is incomplete, time-dependent, unevenly distributed, and generated through contractual and administrative processes. A category recorded on a policy may name a causal environment, approximate an unobserved risk, reflect a reporting convention, or merely identify an individual entity. These possibilities have different implications for prediction. A numerical code assigned to a category has no intrinsic magnitude, and a level occurring once cannot be assumed to have a stable underlying effect. The central research question is how categorical information can become a reproducible feature of insurance risk without allowing an individual's outcome to masquerade as advance knowledge.
The immediate applied point of departure is the study by Avanzi, Taylor, Wang, and Wong (2024), whose actuarial high-cardinality analysis explicitly treats occupation-like variables and integrates random effects with neural models. The methodological lineage also includes Micci-Barreca's (2001) high-cardinality preprocessing scheme, the broad benchmark by Pargent et al. (2022) on regularized target encoding, and Prokhorenkova et al.'s (2018) ordered categorical statistics. These works address overlapping technical problems while resting on distinct empirical settings and learner assumptions. Their results supply mechanisms and comparative evidence, not a single algorithmic ranking for every insurance portfolio.
The insurance-specific methodological baseline includes the Casualty Actuarial Society's comprehensive pricing monograph by Chalk, Deacon, Guillen, and Martinelli (2026), which explicitly addresses high-cardinality rating factors, and Goldburd et al.'s insurance GLM monograph, revised in 2025. The present research examines feature construction rather than the full problem of choosing rating relativities, determining reserves, or deciding the legal permissibility of individual tariff variables. Those decisions require separate actuarial, commercial, and jurisdictional analysis after predictions have been assessed.
Methodologically, this monograph combines formal statistical reasoning, critical comparison of original research, professional standards analysis, and explicitly hypothetical arithmetic. Each derivation identifies its assumptions. The worked numbers are constructed examples, not estimates from insurance companies. No customer-level dataset was acquired, no proprietary model was run, and no invented out-of-sample performance statistics are attributed to prior authors. The evidence ledger records directly inspected publisher articles, methodological documentation, official institutional statements, or legally authoritative texts, with source-specific verification limits. The international scope encompasses insurance operations across jurisdictions while distinguishing the status of US actuarial standards, US model regulatory instruments, European supervisory opinions, and EU legislation.
Insurance High-Cardinality Predictors
The observation unit and available information
Let observation i refer to a policy-period, claim, covered item, contract, or other explicitly defined unit. Let A_i(t) denote a categorical risk attribute as it was known at the decision time t, and let Y_i(t,h) be the outcome over horizon h. A useful target could be expected claim count per unit of exposure, expected aggregate paid loss over a future period, the probability of a claim, or a conditional severity distribution. These are distinct estimands. The transformation g(A_i) must be defined from an information set available before the decision; a later claim-reserving classification cannot be treated as if it were an original underwriting attribute.
Insurance data are often relational. One household produces several policies, one policy produces multiple renewals, one catastrophe generates multiple claims, and one repair provider services many claimants. Treating these rows as independent when engineering category statistics produces artificially small uncertainty estimates and makes train/test splits optimistic. Category identifier, entity identifier, legal contract identifier, and grouping key therefore require separate fields. A postcode can refer to a micro-area, a stable geographical unit, or a string that changes meaning when postal boundaries change. A provider ID can represent an economic organization, its branch, or a local billing record. A category is not scientifically defined until its semantics and observation history are known.
Define K as the number of levels observed in development and n_k as the number of rows in level k. The raw cardinality K does not describe precision because two datasets with equal K can have drastically different occupancy distributions. Report the entire frequency distribution, the share of singletons, entropy of level frequency, and the proportion of exposures belonging to categories absent from an earlier training period. The insurer should additionally report how often categories were re-keyed, merged, or deleted. A thousand stable occupation groups and a thousand transient supplier identifiers imply different transport risks, even when a data dictionary labels both as categorical variables.
Cardinality, identity, and exposure
Count-based occupancy is only a first diagnostic. For frequency models the relevant denominator is typically time at risk, insured units, trip duration, or another contractually appropriate exposure. A category with 600 rows each having one day of coverage may contain less information about annual claim frequency than a category with 30 full-year policies. For severity models, claim count and claim selection alter the information base, and category averages of observed severities say nothing directly about the unreported nonclaiming population. Claims administration can also move a level from an external vendor taxonomy to an internal revised taxonomy without any change in the underlying hazard.
A category's observed outcome is consequently the product of risk, contract terms, exposure, observation, and the insurer's own decisions. Selection into insurance, deductibles, limits, fraud review, claims settlement, and reserving all influence a recorded target. The difference between exposure risk and recorded experience matters especially when engineered features are transferred from claims operations into prospective underwriting. Professional instructions on data assessment in ASOP No. 23 provide a US actuarial reference for documenting data appropriateness, review, limitations, and reliance on others; their legal and professional reach is tied to their stated scope.
Prediction targets and insurance applications
Consider commercial property insurance. An occupation category may contain thousands of detailed descriptions, yet the intended estimand could be expected fire claim cost for the next policy year given selected coverage and exposure. The model must distinguish occupation effects from building characteristics, location, security controls, and insured values. A categorical encoding estimated from aggregate loss may combine frequency and severity effects, while a downstream frequency GLM and a severity GLM require different target definitions. For a Poisson frequency model, the category score should represent a rate or its log, calculated with compatible exposure offsets. For positive claim severities, an insured loss definition and appropriate weights must precede smoothing. For a pure premium model, a compound distribution or direct conditional mean may be relevant, but zero mass and skew must remain visible.
The Elements of Statistical Learning by Hastie, Tibshirani, and Friedman (2009) establishes a broad framework for bias, variance, regularization, and comparison across learning families. Insurance adds a structural complication: loss observations can be revised after the original policy expires. A feature built from currently settled ultimate claims may unknowingly reflect several later years of information. A legitimate historical predictor must use the claims snapshot that would actually have existed at the chosen prediction cut-off, or explicitly explain why retrospective estimates are suitable for a different research objective.
The appropriate technical question is therefore conditional: for a specified outcome, decision time, portfolio, and future deployment distribution, which transformation and estimator yield stable incremental predictive information at acceptable operational and legal cost? This conditional framing permits coherent comparisons without assuming that a variable's high cardinality is itself a defect.
Sparse and Rare Categories
Sampling uncertainty and effective experience
Suppose a group k has N_k claims over exposure E_k and a Poisson frequency model N_k | lambda_k, E_k ~ Poisson(E_k lambda_k). The direct level rate is r_k = N_k / E_k. Under the model, Var(r_k | lambda_k) = lambda_k / E_k. Thus precision depends on exposure, and for rare events it may remain low even when hundreds of policies share a category. For severity, finite second-moment assumptions may be fragile if large claims dominate; a stable point estimate can coexist with considerable tail uncertainty. Experience credibility needs to be assessed for the outcome actually used.
Let the level effect be drawn conceptually from a population distribution. A Gamma prior lambda_k ~ Gamma(alpha,beta), with beta as rate, yields a posterior mean (N_k + alpha)/(E_k + beta). Set mu_0=alpha/beta. Then the posterior mean equals w_k r_k + (1-w_k) mu_0 with w_k=E_k/(E_k+beta). This elementary conjugate calculation exhibits exposure-based partial pooling. The expression is a statistical illustration under a Poisson-Gamma specification, not a claim that every risk factor follows a Gamma mixing law or that posterior means alone guarantee valid tariffs.
Related shrinkage logic appears in Efron and Morris's (1973) empirical Bayes estimation analysis. The connection is conceptual: estimates with little independent information should ordinarily be moderated by a relevant reference distribution. A global portfolio mean, however, may be a poor prior for a specialized industry or country. Hierarchical pooling can instead draw a rare occupation toward its parent economic activity, an insurer-specific group, or a segment mean. The degree of pooling must be estimated under training-only information and assessed in truly held-out experience.
Shrinkage, partial pooling, and uncertainty
For a binary target with n_k observations and x_k events, a Beta(a,b) prior produces posterior mean (x_k+a)/(n_k+a+b). Writing alpha=a+b and mu_0=a/(a+b) yields (x_k+alpha mu_0)/(n_k+alpha). This representation clarifies why an unsmoothed rate of 100% for a singleton event is not credible: with alpha=20 and global event probability 0.05, the encoded rate for one event is (1+1)/(1+20)=0.095238, whereas a zero-event singleton maps to 1/21=0.047619. The sharp difference between raw 0 and 1 is compressed into two estimates near the portfolio baseline. These values are mathematically exact for the specified numbers and prior, and no external claims about optimal alpha follow.
One should not substitute row counts for exposure in every actuarial task. For annualized claim rates, frequency numerator and exposure denominator must remain coherent; for rare high-severity events, the mean can be misleading and extreme quantiles may require distinct modeling. A category with low expected loss but high tail dependence could be assigned an innocuous one-dimensional target code, losing the information that matters to capital or coverage design. Multiple transparent target summaries can sometimes be useful, but each adds tuning degrees of freedom and leakage opportunities.
Pooling is a deliberate exchange between variance and structural bias. If a category genuinely differs from its parent, aggressive shrinkage underestimates that distinction. If the category is effectively noise, insufficient shrinkage produces false discrimination between risk classes. Reliable selection treats the pooling level as a hyperparameter chosen within development data, with sensitivity analysis for rare, common, and emerging levels, rather than selecting it to flatter a final holdout result.
Sparse-variable selection and feature competition
A one-hot matrix for K levels has K indicator columns before constraints or interactions are added. Sparse storage may make a large K computationally manageable, but high dimension still challenges interpretability, variance estimation, and coefficient stability. The lasso and related penalties offer coefficient control; Tibshirani's (1996) regression shrinkage and selection formally motivates one such strategy. Penalizing category coefficients is distinct from selecting a statistically stable encoding. Rare classes may be all zero in some folds, and transformations fitted before cross-validation contaminate the evaluation even when the downstream model is regularized.
Tree methods encounter another problem: a variable with many possible splits can appear attractive because of selection opportunities. Strobl et al. (2007) documented random-forest importance bias for predictors with differing measurement scales and numbers of categories in their investigated settings. The implication for insurance is to avoid treating impurity-based importance as a universal measure of genuine actuarial contribution. Conditional permutation, grouped ablation, and retraining under the same prospective validation protocol can provide additional evidence, each with its own dependence assumptions.
The separation between information content and selection opportunity is central. A huge category dictionary contains many labels and many possible fitted effects, but those degrees of freedom become useful only when the policy population supplies replicable predictive evidence. The cost of estimating a distinction should enter model choice alongside its claimed out-of-sample benefit.
Target Encoding and Leakage
The self-inclusion problem
Let Y_i be an observed binary outcome and C_i its category. The naive target mean for level k is T_k = sum_{j:C_j=k} Y_j / n_k. When this value is assigned to an observation i included in the same sum, it contains Y_i / n_k. If n_k=1, T_{C_i}=Y_i exactly. Even if C is statistically independent of a future Y, the training transformation can manufacture a perfect in-sample relationship for unique identifiers. A downstream learner may therefore assign a large coefficient or importance to what is actually a row index in disguise.
This contamination is analytically distinct from ordinary overfitting by the downstream learner. The feature itself is calculated using the held response. Adding stronger regularization at the last stage does not restore a correct train-time information set. Micci-Barreca (2001) supplied the historical smoothing motivation; Pargent et al. (2022) showed in a multi-dataset benchmark that regularized encodings performed strongly across their tested learners and tasks. Their empirical ranking should be read within that benchmark design and never promoted into an insurance-wide law.
A leave-one-out mean reduces direct self-inclusion for an observation if other rows of its category are available. It does not automatically solve entity dependence, temporal look-ahead, or correlated outcomes. If the same insured property has multiple renewals, removing the present row may leave future claims from its other records inside the encoder. If the level has a single record, a leave-one-out code must fall back to a valid prior rather than divide by zero. Moreover, leave-one-out encodings can still carry unstable class information through near-duplicate rows.
Out-of-fold and ordered encoders
Out-of-fold target encoding constructs the feature for fold f from targets and exposure belonging exclusively to allowable fitting folds, then predicts held-out fold rows. For every observation i in validation fold f, let D_{-f} represent the encoder's training data. A smoothed rate can be T_i = (N_{C_i,-f}+alpha mu_{-f})/(E_{C_i,-f}+alpha), where N and E are defined consistently and the prior comes only from D_{-f}. The full encoder for later production use is refitted on all permitted historical development data, never on final evaluation outcomes. This feature map is part of the fitted predictive pipeline and should carry a version identifier.
The TargetEncoder implementation documentation explicitly distinguishes fit_transform with internal cross-fitting from fit followed by transform, and its cross-fitting demonstration uses informative and noninformative categories to show why the distinction matters. These documented implementation behaviors make a useful independent reproducibility reference. They do not certify that the default internal folds respect insurance-specific entity, calendar, or claims-development constraints. A research design must deliberately impose those constraints.
Prokhorenkova et al. (2018) introduced a permutation-based approach for categorical target statistics and ordered boosting to counter prediction shift arising from target information reuse. Their CatBoost research paper addresses a specific algorithmic mechanism. A random permutation is nevertheless not equivalent to historical information availability. If an insurer predicts policies issued in January using claims reported the following December, a random ordering among development rows cannot by itself reconstruct the January information set. Chronological feature audits remain necessary even in models with built-in categorical safeguards.
Insurance-specific target maturity and entity leakage
A reporting triangle illustrates the problem. Policy A begins on January 1. Its loss occurs in September and is reported in October. At the January quote date, a model may legitimately use previous calendar-year occupational claims that were already available, but it cannot use the future September event of Policy A. For historical backtesting, the data warehouse must reproduce the old reporting cut-off, including the absence of claims not yet reported. Engineering from a modern final loss table while splitting by policy inception date still leaks future outcomes because old policy records have been enriched retrospectively.
Additional contamination arises when insurers calculate encodings from a data mart that includes claims after cancellations, settlement estimates revised by external legal processes, fraud conclusions made after payment, or treatment assignments influenced by the insurer's own prior prediction. A downstream model can inadvertently rediscover its predecessor's decisions instead of independent risk. Claims features should carry event date, reporting date, accounting snapshot date, and feature extraction date; these dates govern different admissibility questions.
Group leakage requires equal attention. If the scientific target is performance on new insured entities, categories and split groups should prevent an entity's history from appearing simultaneously in development and validation when that would be unavailable to the intended new-business prediction. If the target is renewal for known entities, prior experience may be legitimate, but future experience from the same entity remains inadmissible. GroupKFold documentation describes nonoverlapping group splits and supplies a technical pattern for entity separation; it is not a substitute for clearly stating which population is being evaluated.
A clean random holdout is thus necessary but often insufficient. The stronger test asks whether an analyst could have computed each category statistic on the actual historical prediction date using only the allowed input sources. This question turns data provenance into a component of estimator validity.
Grouping and Smoothing
Non-target encodings and structured grouping
Unsupervised category transformations can avoid response reuse at the encoding step. One-hot encoding creates indicators; frequency encoding maps a category to training-set prevalence or exposure share; count encoding uses occurrence counts; hashing maps categories into fixed-dimensional buckets. None of these methods makes future information available by itself, but their computation still needs a valid training period, and their output can reveal proxies for identity, scale, or market structure.
Hashed features solve a storage and fixed-dimension problem, particularly when levels arrive continually. Weinberger et al.'s (2009) feature hashing analysis establishes theoretical and empirical grounds for efficient large feature spaces in its studied setting. Hashing introduces collisions: two distinct categories can share a bucket, while related categories may remain separated. The number of buckets and interaction strategy therefore require independent selection. A bucket dimension selected after looking at final insurance holdout results is itself test-set overfitting.
Grouping by substantive taxonomies offers a different tradeoff. An occupation can be represented by a hierarchical industrial code; a vehicle model by make, series, and technical characteristics; a repair provider by independently available specialization and territory. This may yield interpretable transfer to new levels, provided parent taxonomies are stable and available at the relevant date. Grouping labels by observed outcome without cross-fitting simply moves the leakage problem up one level. Risk-relevant predeclared category groupings should therefore be documented separately from statistically optimized target-informed groups.
Insurance-weighted smoothing and numeric demonstrations
Consider a hypothetical portfolio with global annual claim frequency mu_0=0.08. For category A, exposure is 5 policy-years and claims count is 2. For category B, exposure is 500 policy-years and claims count is 42. Raw rates are 0.40 and 0.084. Set a smoothing exposure beta=50 policy-years. A Poisson-Gamma posterior-mean-style encoder becomes (N_k+beta mu_0)/(E_k+beta). Category A maps to (2+4)/(5+50)=6/55=0.1090909; category B maps to (42+4)/(500+50)=46/550=0.0836364. A rare category's extreme observed rate is moderated strongly, while B's observed rate largely survives. Both operations are a direct consequence of the declared assumptions.
This example illustrates why equal row-count Laplace smoothing may be inappropriate when exposure per row differs. If a commercial policy lasts three days and a second policy lasts twelve months, row counts do not represent the same sampling opportunity for frequency. Conversely, the number of claims may provide the effective sample size for certain severity objectives. The correct denominator follows the estimand rather than the convenience of a database table. The CAS's high-cardinality pricing treatment provides an actuarial context for credibility, penalization, and modeling tradeoffs; the numerical examples here are independently constructed.
Now imagine that category A is absent from the next policy year and a new category C enters the portfolio. A strictly prospective fitted encoder should map unseen C to a predeclared global or suitable hierarchical baseline, rather than calculate a score from outcomes that will occur in the future. Forcing an unseen code into the most similar observed category requires an externally specified similarity rule and an audit of errors. A model that has excellent retrospective accuracy but frequently defaults new categories may provide less incremental deployment value than a slightly weaker, more transportable representation.
Hierarchical, mixed-effects, and representation approaches
Hierarchical generalized linear mixed models introduce an estimated group effect, often with partial pooling across categories. Their structure can preserve a direct risk interpretation under a suitable link while accounting for sparse level experience. Avanzi et al. (2024) propose GLMMNet for actuarial categories, combining mixed effects with a neural component and using variational inference to fit their formulation. The reported research is directly relevant to insurance and supplies an important counterpoint to treating target encoding as the only serious high-cardinality option. Its modeling assumptions, computational requirements, and evidence setting must accompany any transfer to a different line of insurance.
Dense embeddings learn a low-dimensional representation of categories through a supervised prediction objective. They can share statistical strength through common downstream patterns and interaction structure. They also rely on sufficiently representative data, can embed sensitive proxies, and may represent rare levels poorly. An embedding dimension is another source of model capacity; selecting it requires nested validation. A learned vector is not a stable actuarial explanation simply because it is compact. Interpretable random effects, hierarchical probabilities, and an external independent taxonomy can be combined with learned representations when their roles are made explicit.
Tree-based learning offers different interfaces. The XGBoost system paper by Chen and Guestrin (2016) establishes a scalable framework for gradient boosting, whereas Friedman (2001) formalized the broader gradient-boosting function approximation. Implementations differ in native categorical handling, splitting rules, and missing-level behavior. Algorithm labels alone do not answer whether a trained model used a legally permitted factor, whether its input encoding leaked a future outcome, or whether a minor category had enough historical insurance exposure.
An actuarial choice among these families should compare the entire preparation-and-learning chain using consistent loss definitions, information cut-offs, model capacity controls, and prospective test periods. Hybrid pipelines are legitimate where justified, but their additional complexity must be justified by reliable gain and maintainable auditability.
Out-of-Sample Feature Construction
A temporally valid experimental design
A defensible evaluation begins by naming the deployment question. For example: use information frozen at the close of 2022 to predict claims arising from policies incepting in 2023, allowing a specified reporting-development lag for evaluation. The development period may contain earlier annual cohorts, and a subsequent period may be reserved for tuning. Every feature transform, dictionary normalization, grouping, missing-value rule, category prior, encoder, and model coefficient must be learned inside the appropriate training boundary. The final test period must not influence hyperparameter search or taxonomy edits.
For repeated policies or grouped insured entities, choose the grouping structure in light of the intended application. A new-business performance claim may require entire customer or insured-asset groups to be absent from development. A renewal prediction may legitimately use earlier outcomes belonging to the same customer, but only those known by the renewal quote. A geographic generalization claim may require held-out territories or explicitly forecasted territorial shifts. The independence assumptions must be matched to the estimand. A single random row split answers only a narrow question about similarly sampled records.
Rolling-origin evaluation trains on successive admissible historical windows and evaluates on later periods. If window j ends at time t_j and validation begins at t_j+delta, the minimum gap delta should account for feature computation and target observation delays; it is not a universal fixed number. For claims with long tails or delayed reporting, right truncation makes a recently observed cohort incomparable with a mature one unless the label development problem is modeled. Distribution shift results such as Kuznetsov and Mohri's (2017) nonstationary learning bounds emphasize the significance of training-target discrepancy under their assumptions; those formal results do not substitute for empirical monitoring of insurance hazards and book composition.
Nested selection, unseen levels, and transport
A high-cardinality feature can be selected or rejected only after a fair benchmark. Compare a simple baseline with no candidate category, a coarser domain taxonomy, one-hot or hashing where appropriate, cross-fitted smoothed target encoding, and an independently specified hierarchical or native categorical learner. Match underlying training histories, exposure offsets, loss functions, and validation periods. Select smoothing strength, rare-category threshold, number of hash buckets, interaction depth, and embedding dimension within inner development folds. Report outer-fold performance without using the outer holdout to choose these settings.
Important evaluation quantities include mean claim frequency deviance, Poisson or Tweedie deviance when model assumptions fit, held-out log likelihood or proper scores for probability forecasts, pure-premium error and calibration, and stability of results across segments and calendar periods. For classification, a strong ranking score can coexist with poor probability calibration. For pricing, a small improvement in a relative discrimination metric does not establish an acceptable technical premium, and a good aggregate score may conceal material errors in thinly populated categories.
A complete stress test should report performance for levels frequent in training, rare levels, entirely unseen levels, recently created levels, and identifiers merged or remapped between periods. State the share of future exposures in each category. An encoder that works only on repeated levels may still be valuable in a stable renewal portfolio but is unsuitable as a general solution for rapidly changing external vendors. The fallback behavior for missing and unseen levels is part of the actual predictive model, not a deployment detail to decide after sign-off.
Luteijn et al.'s (2026) British Actuarial Journal health-and-care data science framework discusses target encoding, cross-validation, and categorical embeddings in applied actuarial projects. It supports the relevance of the procedures to actuarial workflows, while its healthcare setting and general methodological scope do not confer a measured performance advantage on an unrelated insurance dataset.
Reproducibility and controls
A reproducible experiment should retain: input source names and snapshots; outcome and exposure definitions; time-zone and cut-off rules; historical dictionary versions; entity grouping rules; missingness treatment; category-frequency distribution; transformation source code and library versions; cross-fitting partitions; fitted priors and hyperparameters; random seeds; target maturity criteria; model artifacts; final evaluation reports; and the legal review of candidate variables. The record should allow an independent analyst to reconstruct what a policyholder's feature vector contained when a real decision was made.
Three controls are decisive. First, a leakage test rebuilds all encodings after permuting targets within permitted training groups; spurious categories should lose meaningful signal on independent holdout data. Second, an as-of replay checks sampled predictions against historically archived inputs and loss-development snapshots. Third, a no-candidate ablation compares the complete selected pipeline with an otherwise equivalent pipeline that excludes the high-cardinality feature. These are specified diagnostics, not claims of empirically established test results in this monograph.
The United States' ASOP No. 56 on modeling, effective for work performed on or after October 1, 2020, provides an actuarial framework for model purpose, inputs, validation, governance, and disclosure within its stated professional scope. Its requirements are relevant to feature representations because the model includes preprocessing, assumptions, and data dependencies as well as the final estimator. Independent review should be able to inspect the complete transformation chain.
Interpretability and Validation
Predictive, actuarial, and operational evaluation
Interpretability has several levels. A human-readable variable name explains what a category represents. A stable encoding documents its numerical construction. A transparent model describes how that number contributes to a prediction. An actuarial explanation connects the prediction to a meaningful risk and contract, while a governance explanation establishes who authorized the variable, what data were used, and how a customer can contest errors. These levels overlap but are not interchangeable.
A one-dimensional smoothed target statistic may be more inspectable than a large embedding, yet its value can change when categories are redefined or the outcome matures. A hierarchical random effect may carry a coherent interpretation, but it still depends on assumptions about exchangeability, group structures, and covariates. Relative risk factors calculated after fitting may be unstable when strongly correlated features compete for the same observed variation. The investigator must distinguish feature-level association from causal risk mechanisms and avoid using predictive association as automatic justification for classification.
Feature importance requires an estimand. Conditional permutation asks how much the fixed fitted model suffers when an input is disturbed under a particular perturbation distribution. Retraining ablation asks how much predictive value remains without that variable after competitors adapt. Neither quantity identifies a causal effect; each describes a distinct operation. Because high-cardinality categorical variables correlate with geography, firm size, occupation, and distribution channels, both experiments should consider realistic joint distributions. A variable with many levels must not be credited for a proxy it merely duplicates.
An insurer should also assess computational costs: feature table size, dictionary synchronization, training-inference consistency, latency, fallback rates, source-data license terms, privacy exposure, and error correction. Hashing minimizes explicit dictionary maintenance but complicates error attribution. Hierarchical encoding reduces some variance but requires a governed taxonomy. Ordered statistics reduce specific leakage mechanisms but can be costly and require careful treatment of calendar time. The operational objective is sustained reliability, not a leaderboard result from one historical sample.
Fairness, privacy, and institutional scope
High-cardinality categories often operate as fine-grained proxies for protected, sensitive, or socially consequential traits. Micro-geographic indicators, occupation labels, employer groups, and provider identifiers can reconstruct differences that were not expressly collected as protected fields. Correlation with an outcome does not establish admissibility under insurance law, and statistical shrinkage does not automatically remove disparate impacts. Fairness assessment must identify the relevant insurance line, decision, jurisdiction, contract, and statutory classification rule before approving a feature for use.
The National Association of Insurance Commissioners adopted its AI systems model bulletin in December 2023, establishing a model framework for state insurance supervision and documentation expectations. A model bulletin becomes operative through jurisdiction-specific adoption or implementation; the NAIC's action alone is not a binding federal insurance rule. The European Insurance and Occupational Pensions Authority's 2025 AI governance opinion addresses national supervisors and interprets existing sectoral governance expectations with a risk-based approach. An opinion and a national law have different legal force.
For personal data within its scope, the EU General Data Protection Regulation provides requirements including purpose limitation, data minimization, accuracy, and safeguards relating to certain automated decisions. The EU Artificial Intelligence Act lists specified life and health insurance risk-assessment and pricing AI uses in Annex III; the applicability and timing of obligations require examination of the current consolidated legal text and the precise intended use. A predictive feature created for property pricing, a system for health underwriting, and an administrative claims classification should not be presumed identical legal use cases.
These documents lead to a practical scientific distinction. A category may be technically helpful but based on unauthorized personal data; it may be legally permitted but statistically unreliable; or it may be both useful and permissible yet impossible to explain or correct after source-dictionary changes. Approval requires separate evidentiary records for statistical reliability, data permissions, decision governance, and consumer treatment. The present monograph does not supply country-by-country legal advice or a universal lawful-rating-factor list.
Integrated analytical findings and research agenda
Five findings follow from the combined reasoning. First, time is part of the mathematical definition of an engineered insurance feature: identical formulas using different information cut-offs produce different estimators. Second, experience must be measured on the estimand's proper denominator; category row count cannot substitute automatically for exposure or claim count. Third, validity belongs to the pipeline, including fitting and refitting the encoding, rather than a stored transformed column. Fourth, predictive improvement must be conditional on deployment distributions, unseen categories, and maturity of outcomes. Fifth, explanatory legitimacy requires knowledge of how a category was defined, collected, authorized, and maintained.
These are analytic conclusions synthesizing existing statistical and actuarial knowledge. They do not constitute a newly named theory, a proprietary modeling doctrine, or a demonstrated breakthrough. In particular, no evidence supports a universal claim that target encoding dominates all hierarchical models, all categorical embedding methods, or all native categorical learners in every insurance line. Pargent et al.'s benchmark and Avanzi et al.'s actuarial modeling paper show the importance of comparing alternatives under their own designs; neither eliminates the need for new-portfolio evaluation.
A rigorous reproducible research program would use multiple real insurance portfolios with permissioned data, harmonized targets and exposure, clearly separated known-entity and new-entity use cases, historical as-of snapshots, and genuinely prospective geographic and temporal holdouts. It would preregister the evaluation design; compare methods under identical outcome maturation and compute budgets; report uncertainty intervals and category-specific fallback coverage; audit proxy behavior by lawful and relevant groups; and preserve results showing little or negative predictive value. External replication across different lines and countries is necessary before generalizations beyond the portfolios under study.
Formal estimands, information sets, and category stability
To describe the full learning problem, index policy units by i and evaluation dates by t. Let (\mathcal{F}{i,t}) denote the information available to the insurer for the intended decision at (t). A permissible engineering operation is a measurable function (Z{i,t}=g_{\hat{\theta}t}(C{i,t},X_{i,t})), where (C) is the recorded category, (X) comprises other observable data, and (\hat{\theta}t) is learned exclusively from allowable historical experience. A fitted encoder that depends on a future component of Y{i,t+h} fails the information-set condition even if the insurer stores the result in an ordinary feature table. The condition holds for supervised encodings, unsupervised frequency summaries, learned embeddings, and dictionary transformations that are maintained after deployment.
The predictive quantity must also be identified. For claim frequency, an expected count conditional on observed exposure e and features X can be expressed as (\mathbb{E}[N_i\mid X_i,C_i,e_i]=e_i\lambda(X_i,C_i)). A transformation of categories intended to approximate lambda may be built from claims per exposure, but it should not automatically be reused for the occurrence probability (P(N_i\geq 1\mid X_i,C_i,e_i)), which under an assumed Poisson process is (1-\exp(-e_i\lambda)). The difference becomes important when exposures vary. A code intended to predict positive claim severity is estimated on a selected claiming population, while a code intended for pure premium reflects both claim occurrence and cost. These distinctions explain why a generic average of a column called "claim amount" can be a poorly specified target encoding.
For a portfolio with several policies per person, policy-level observations may share a latent risk U_i. Even if the observed claims are conditionally independent given U_i, a random row split that places the same person in both periods may reward the encoder for identifying this latent risk through past and future records. If the deployed application is future renewal, past records may provide legitimate information; if the application is a new unidentified customer, those records are unavailable. Both evaluations are useful, but they estimate different transport tasks. The claimed predictive validity must name which one the experiment addresses.
Suppose the category taxonomy changes at date t*. A code named C001 before t* could refer to a different class after t*. Retaining the string while silently changing its definition creates an unobserved structural break. Conversely, replacing several old levels with a new standardized level changes apparent cardinality and can move observations between groups without risk changing. The analyst should preserve a bitemporal view: what category was assigned when the event occurred, and what mapping was known when the forecast was made. A retrospective corrected category may be helpful for descriptive analysis, while an as-of predictor needs the historically available mapping.
A category-level effect can absorb unmeasured factors that are not portable. For example, a repair-facility identifier may correlate with claim severity because some facilities receive more complex cases. The association can be strong even if changing the facility would not reproduce the risk effect in the same insured population. The feature then reflects allocation and selection mechanisms as well as repair behavior. It should be used with a clear predictive purpose and without a causal interpretation unsupported by research design.
Rare-level thresholds and category life cycles
A simple rule such as "group all categories with fewer than ten observations" ignores exposure differences, changing market size, and the outcome's rarity. A more informative threshold considers effective experience and stability under historical resampling. For a Poisson rate, the standard deviation of the direct estimate is approximately sqrt(lambda_k/E_k) under the conditional model. Relative uncertainty is approximately 1/sqrt(E_k lambda_k), illustrating that expected event count, rather than raw rows, governs precision. When expected frequency is small, a category with hundreds of insured units may still have inadequate experience for a stable independent risk estimate.
This calculation is conditional on the Poisson assumption and on reliable exposure. Overdispersion, catastrophe clustering, changes in deductibles, and duplicate event records can produce greater uncertainty. A portfolio with regionally concentrated properties may have numerous policy-years but only a small effective number of independent storm events. The denominator should therefore be interpreted with the dependence structure in view, and category-level uncertainty should be reported rather than hidden behind a deterministic score.
Categories pass through a life cycle. A new commercial activity code begins with no local outcome history; after several years it accumulates claims experience; later it may be merged, deprecated, or reclassified. A static training dictionary treats these states as interchangeable strings. A better audit stores each level's creation date, training exposure, event count, last observed date, number of coding revisions, and parent classification. These values can support a documented fallback decision and alert model owners when a frequently predicted level has insufficient historical support.
Rare-category grouping is potentially useful at three distinct points. Source-system aggregation replaces multiple raw codes with a stable broader concept before model development. Statistical pooling retains individual level identities but shrinks their estimates toward a baseline. Deployment fallback handles levels absent from the fitted training dictionary. These operations solve different problems. Aggregation changes the scientific feature definition, pooling changes the estimator, and fallback specifies behavior outside the training support. Their quality must be evaluated separately.
Partial pooling can also be examined through the derivative of the posterior mean with respect to an observed category rate. In the Poisson-Gamma example it is w_k=E_k/(E_k+beta). The derivative approaches zero as E_k approaches zero and approaches one as exposure becomes large. This provides an intuitive measure of how strongly a transformed level inherits its own observed experience. It does not guarantee that two levels are socially or legally appropriate to compare. The choice of reference population determines what the estimate means.
For a new level with E_k=0, the formula returns mu_0. That property is useful but reveals a substantive choice: the "global" prior may be dominated by common risks unrelated to the new one. A stable hierarchy can instead produce a baseline for the closest approved parent group, and uncertainty can accompany that estimate. If a category has no plausible peer group, the insurer may prefer a conservative generic score and operational review until evidence accumulates. Such decisions should be supported by documented governance and tested for effects on future exposures.
Detailed leakage counterexample and failure analysis
Consider a purely illustrative population in which each training policy has a unique, arbitrary identifier, and each Y_i is an independent Bernoulli(0.05) claim indicator. By construction, identifier and future outcome are independent in the target population. Under the naive encoder T_{C_i}=Y_i for all training observations because every category is a singleton. A flexible learner trained on T sees perfect target agreement. The resulting training loss may be close to zero even though the original category contains exactly zero legitimate predictive information about a new policy's outcome.
Now suppose the encoder is fitted using only a training partition and the learner evaluated on a disjoint test partition containing entirely new identifiers. Every test identifier is unseen and must receive a common prior. Its best calibrated constant prediction under the stipulated data-generating process is 0.05; arbitrary apparent in-sample discrimination vanishes. The diagnostic difference is explained by feature construction, not by sudden deterioration in the true relationship between an identifier and claim probability.
The argument becomes subtler if identifiers repeat. With two rows per identifier, leave-one-out encoding for the first record uses the second record's outcome. If the second record occurs after the first record's quotation date, the procedure leaks future information even though no row directly encodes its own target. If the two rows correspond to the same claim recorded in different accounting systems, the leakage becomes more severe because the supposedly independent observation is a duplicate. Grouping and temporal boundaries are therefore logical requirements of the experiment, not optional aesthetic refinements.
A random out-of-fold encoder reduces self-inclusion by withholding an observation's own fold. Nevertheless, labels associated with other folds may arrive after the predicted event, and records from the same insured asset may span training and validation. To emulate an actual historical decision, the encoder must be fitted on a date-constrained history, with the appropriate entity restriction and a defensible label-availability rule. The cross-fitting behavior documented by scikit-learn solves a defined cross-validation construction problem, while chronology and claim-development must be addressed by the researcher's split design.
There is also a contamination pathway through unsupervised statistics. Suppose a policy is quoted in January, but a "category frequency" variable is calculated from a file containing all policies sold through December. This frequency is derived without claim outcomes, yet it uses future portfolio composition. If future market participation was influenced by the insurer's own risk selection, the encoded variable can reflect information and decisions unavailable at the original quotation. A correct chronology therefore applies to every learned preprocessing stage, not solely to a target-mean encoder.
Leakage can be investigated with a set of explicit unit tests. Verify that each scoring date precedes the last source observation used for the feature, that training rows exclude the held-out entity where the task requires new-entity evaluation, that duplicate insured events cannot cross a split unnoticed, and that fitted category priors are stored with dates. Create an adversarial new-category test in which all test identifiers are unseen and confirm that fallback values are deterministic under the saved model. Run a synthetic no-signal simulation of the type described above: a pipeline that claims predictive improvement from arbitrary singleton codes under proper holdout should trigger a methodological investigation.
Frequency, severity, and distributional objectives
The objective function determines the statistical meaning of an engineered feature. A Poisson frequency model may express log E[N_i|X_i]=log(e_i)+eta(X_i), with exposure offset log(e_i). A smoothed target statistic expressed as log of a category rate can enter eta, but small or zero observed rates require explicit lower bounds or a posterior estimate that remains positive. The resulting coefficient can be interpreted as modifying expected frequency only under the model's assumptions and with appropriate other predictors.
For severity, define S_i as the positive cost of a reported and covered claim. A category target constructed from claims may use total severity divided by count, but the distribution can have heavy tails and may be distorted by limits, deductibles, salvage, subrogation, or settlement practice. Transforming severity to logarithms changes the estimand: an average log cost is related to a geometric mean, and a naive inverse transformation generally does not recover arithmetic expected cost. A model builder must state whether the desired prediction is expected payment, a conditional quantile, or a full distribution. Extreme claims may require separate treatment rather than one stable scalar level estimate.
In aggregate pure premium, the conditional expectation of total loss per unit exposure is central. A target code from aggregate loss divided by exposure can reflect both claim frequency and average claim cost. It may be useful for a direct Tweedie-type model, but it loses the distinction between a high-frequency, low-severity group and a low-frequency, high-severity group with the same expected total. Those groups may require different deductibles, risk controls, capital assumptions, or claims capacity. The single scalar is a deliberately compressed summary of the target, and its information loss should be recognized.
For time-to-event or lapse prediction, a category estimate constructed from complete observed durations can be biased when censoring differs by group. Category-level survival experience should respect censoring, left truncation, and time-dependent covariates. A count of "lapses per observed policy" may conflate genuine lapse intensity with observation window length. The category representation belongs to the same statistical model family as the estimand; it cannot be carried unexamined across outcomes simply because the raw level strings match.
Where claim outcomes mature gradually, a leakage-safe engineering analysis may require a separate historical triangle of available values at each valuation date. For example, a severity encoder on reported incurred cost at twelve months of development is a different quantity from an encoder on payments recorded at sixty months. If the study compares ultimate cost, the investigator must describe how the outcome was estimated, what mature periods trained the model, and whether future development factors were themselves estimated using the holdout. The same information boundary applies recursively to every intermediate reserving calculation that feeds a predictive feature.
Comparative case analysis across insurance operations
In automobile insurance, vehicle make, model, trim, repair network, and local territory can generate large category dictionaries. A VIN-derived feature may appear highly granular yet repeat across only a small number of vehicles in certain portfolios. Domain grouping by mechanical or safety attributes can provide transportability to new model years; a target statistic for the exact trim level relies on sufficient stable past exposure. Claims involving new technologies may mature under different repair economics, so historical cost codes can drift even when physical model names remain the same.
In commercial property insurance, detailed industry activity and insured-asset identifiers create many levels. A business classification may be recorded at the parent enterprise, site, and insured-location level, with different levels of exposure and control. Using the most detailed code can generate apparent predictive power from a small number of unusually costly claims, whereas pooling within a stable activity taxonomy can reduce volatility. Location and occupation are often correlated; an ablation study should determine whether the granular activity code adds independent prospective information after the model already uses hazard and building variables.
In workers' compensation or employer liability, employer identifiers, occupational classification, and injury descriptors may carry distinct meanings. An employer's own historical losses can legitimately inform experience rating in some institutional settings, yet a research test of prediction for a new employer should not import employer-specific future outcomes. Small businesses with brief coverage generate sparse denominators, and classification changes may arise from audits rather than real changes in work performed. A comparison of category encodings should preserve the operational distinction between prospective risk characteristics and retrospectively audited exposure.
In health and life insurance, fine-grained provider, diagnostic, occupational, or administrative categories may involve sensitive personal information and specific restrictions. Statistical validity and authorized use require separate examination. The EU General Data Protection Regulation and relevant insurance law can affect whether a data feature may be collected, repurposed, or used for decisions. Even where data are accessible for operational analytics, the permitted predictive purpose may be narrower. Governance cannot be retrofitted from an accurate AUC score.
In catastrophe and agricultural insurance, detailed geographic grids and crop/soil classifications may show many levels but few independent catastrophic events. The exposure count can be large while event-level effective sample size remains small. A model that treats policy outcomes as conditionally independent may report unjustifiably narrow category uncertainty after a single storm or drought. The scientifically relevant validation may involve leaving whole events, seasons, or regions out of development. A high-cardinality spatial encoding should be tested against physically grounded hazards, stable mapping, and the possibility that future catastrophe conditions differ from the observed periods.
These examples support a general comparative conclusion: the appropriate method depends on what a category records, the event process, available independent exposure, the forecast horizon, and the insurer's legal and operational role. The same name "target encoding" can represent meaningfully different statistical procedures across insurance lines.
A reproducible comparison specification
For a future empirical study, define a predeclared portfolio and time window, then construct a common baseline dataset using an explicit versioned category dictionary. Specify one outcome at a time, with exposure and known-at-date fields. Retain an untouched chronological evaluation cohort, and if the deployment concerns new entities, hold out entire relevant groups. Establish clear outcome development rules for all methods. This removes several common sources of unintentional procedural advantage.
Fit a benchmark family that includes a no-category baseline, a domain-grouped feature, a sparse one-hot representation, an unsupervised frequency or hashing representation, an out-of-fold smoothed target encoder, and a hierarchical or native categorical estimator where practical. Each candidate should receive a comparable tuning budget and the same labeled development data. A method that uses more historical data must be identified as a different information scenario rather than credited with superior feature engineering by default.
Choose hyperparameters within nested development windows. Report predictive scores on the original scale of the estimand; add calibration curves, residual checks, and uncertainty intervals computed with an appropriate dependence structure. For financial outcomes, assess both aggregate adequacy and segment-specific error. A small shift in the mean loss score can conceal extreme variability at the category level. Include the number of independent loss events where catastrophe concentration is material.
Separate four coverage conditions in the test report: common old levels, rare old levels, entirely unseen levels, and recoded or merged levels. Report the percentage of exposure, the percentage of predicted loss, and the frequency of operational fallback within each condition. These fractions often matter more to actual deployment than the nominal count of one-hot columns. Compare how many categories can receive an independently stable effect at a declared uncertainty threshold.
Attribution to the engineered category should be assessed through matched ablation and retuned baselines. If adding a granular level improves performance only when the data include an imperfectly separated entity or future reporting variable, the reported gain lacks the intended prospective meaning. If it improves performance robustly across historical years, lines, and segments, confidence is stronger but still conditioned on observed distribution shift. These research criteria are consistent with the general statistical learning account and the model-testing expectations contained in ASOP No. 56.
A complete published experimental report should supply the metric definition, numerator and exposure denominator, train and test date boundaries, model and encoder hyperparameters, effective training category experience, unseen-level strategy, subgroup results, confidence intervals or comparable uncertainty summaries, and links to reproducible code where permissions allow. If private policy records cannot be released, a synthetic example, source code, dependency lockfile, and independently checkable calculation can still expose the estimation mechanism without making fabricated claims about real portfolios.
Monitoring, correction, and the limits of historical evidence
High-cardinality features require continued monitoring because the feature map is partly a statistical estimate of the past. Track the volume of new levels, share of exposure flowing through fallback categories, changes in level frequencies, rolling loss experience, source taxonomy revisions, and calibration by major and rare groups. Monitor preprocessing errors separately from model prediction errors: a category-key mismatch may be an administrative failure even when the mathematical estimator remains correct.
A drift signal should prompt a question about cause. Has the insured portfolio changed, has the hazard changed, have claim reporting practices changed, or has a source taxonomy changed? Different causes require different actions. A data-mapping correction may repair a code immediately, while a genuine structural risk change may require new modeling evidence. Retraining simply because a population-level metric changed can reproduce corrupted source definitions and obscure the reason for deterioration.
For categories containing very few insured entities, operational explainability may conflict with privacy. Publishing detailed empirical claims counts and loss rates can reveal confidential experience indirectly. Documentation can preserve formulas, pooling structures, independent validation results, and access-controlled evidence without revealing individual policyholders' identities. Privacy requirements may also constrain which source attributes are legally available for retrospective reproduction.
The strongest evidence of a categorical representation's value is repeated performance on future data under the exact information, exposure, and legal conditions of the intended use. The weakest is a high training fit achieved after target-informed preprocessing of an entire dataset. Between these endpoints lie many valid but limited evaluations. The scientific responsibility is to identify what each result establishes and how far it can be transported.
Conclusion
High-cardinality categorical feature engineering in insurance is the statistical construction of usable risk information from numerous, often sparse, evolving category levels under explicit exposure, time, and governance constraints. The representation problem begins with the meaning of a category and the target to be predicted, continues through statistical pooling and leakage-safe estimation, and ends only when prospective performance, operational fallback, and permissible use have been assessed. Simple unsmoothed target averages can memorize outcomes; shrinkage reduces extreme sampling noise without erasing every possible structural bias; cross-fitting curbs self-inclusion but does not itself reconstruct historical time; and hierarchical, hashed, grouped, and neural representations each trade off transportability, capacity, interpretability, and maintenance cost.
The mathematical illustrations verify how exposure and smoothing jointly determine an encoded rate and why individual outcomes can directly contaminate naive encodings. The primary studies establish plausible and sometimes competitive methods, including the insurance-specific mixed-effects neural approach of Avanzi et al., while the standards and regulatory materials identify separate professional and institutional responsibilities. The defensible evaluation unit is the complete, documented, time-consistent feature-construction pipeline and its deployment population. That conclusion provides a scientific basis for comparative research and internal insurance model review without constituting professional actuarial, insurance, legal, or personalized advice.
References
Avanzi, Benjamin, Greg Taylor, Melantha Wang, and Bernard Wong. 2024. "Machine Learning with High-Cardinality Categorical Features in Actuarial Applications." ASTIN Bulletin: The Journal of the IAA 54 (2): 213–238. https://doi.org/10.1017/asb.2024.7.
Actuarial Standards Board. 2016. "Actuarial Standard of Practice No. 23: Data Quality." Revised December 2016, effective April 30, 2017. Actuarial Standards Board.
Actuarial Standards Board. 2019. "Actuarial Standard of Practice No. 56: Modeling." Adopted December 2019, effective October 1, 2020. Actuarial Standards Board.
Chalk, Alan, David Deacon, Montserrat Guillen, and Max Martinelli. 2026. "From GLMs to Comprehensive Insurance Pricing: Techniques and Challenges." CAS Monograph Series 16. Casualty Actuarial Society.
Chen, Tianqi, and Carlos Guestrin. 2016. "XGBoost: A Scalable Tree Boosting System." Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 785–794. https://doi.org/10.1145/2939672.2939785.
Efron, Bradley, and Carl Morris. 1973. "Stein's Estimation Rule and Its Competitors: An Empirical Bayes Approach." Journal of the American Statistical Association 68 (341): 117–130. https://doi.org/10.1080/01621459.1973.10481350.
European Insurance and Occupational Pensions Authority. 2025. "Opinion on Artificial Intelligence Governance and Risk Management." EIOPA-BoS-25-360, August 6, 2025. EIOPA.
European Parliament and Council. 2016. "Regulation (EU) 2016/679, General Data Protection Regulation." Official Journal of the European Union L 119, May 4, 2016, 1–88.
European Parliament and Council. 2024. "Regulation (EU) 2024/1689, Artificial Intelligence Act." Official Journal of the European Union, July 12, 2024.
Friedman, Jerome H. 2001. "Greedy Function Approximation: A Gradient Boosting Machine." Annals of Statistics 29 (5): 1189–1232. https://doi.org/10.1214/aos/1013203451.
Goldburd, Mark, Anand Khare, Dan Tevet, and Dmitriy Guller. 2025. "Generalized Linear Models for Insurance Rating." Second edition, revised 2025. CAS Monograph Series 5. Casualty Actuarial Society.
Hastie, Trevor, Robert Tibshirani, and Jerome Friedman. 2009. "The Elements of Statistical Learning: Data Mining, Inference, and Prediction." Second edition. Springer Series in Statistics. Springer. https://doi.org/10.1007/978-0-387-84858-7.
Kuznetsov, Vitaly, and Mehryar Mohri. 2017. "Generalization Bounds for Non-Stationary Mixing Processes." Machine Learning 106: 93–117. https://doi.org/10.1007/s10994-016-5588-2.
Micci-Barreca, Daniele. 2001. "A Preprocessing Scheme for High-Cardinality Categorical Attributes in Classification and Prediction Problems." SIGKDD Explorations Newsletter 3 (1): 27–32. https://doi.org/10.1145/507533.507538.
National Association of Insurance Commissioners. 2023. "Model Bulletin: Use of Artificial Intelligence Systems by Insurers." Adopted December 4, 2023. NAIC.
Pargent, Florian, Florian Pfisterer, Janek Thomas, and Bernd Bischl. 2022. "Regularized Target Encoding Outperforms Traditional Methods in Supervised Machine Learning with High Cardinality Features." Computational Statistics 37: 2671–2692. https://doi.org/10.1007/s00180-022-01207-6.
Prokhorenkova, Liudmila, Gleb Gusev, Aleksandr Vorobev, Anna Veronika Dorogush, and Andrey Gulin. 2018. "CatBoost: Unbiased Boosting with Categorical Features." Advances in Neural Information Processing Systems 31: 6638–6648.
scikit-learn Developers. 2026. "TargetEncoder API Reference." Scikit-learn documentation, accessed November 14, 2026.
scikit-learn Developers. 2026. "Target Encoder's Internal Cross Fitting." Scikit-learn example documentation, accessed November 14, 2026.
scikit-learn Developers. 2026. "GroupKFold API Reference." Scikit-learn documentation, accessed November 14, 2026.
Strobl, Carolin, Anne-Laure Boulesteix, Achim Zeileis, and Torsten Hothorn. 2007. "Bias in Random Forest Variable Importance Measures: Illustrations, Sources and a Solution." BMC Bioinformatics 8: 25. https://doi.org/10.1186/1471-2105-8-25.
Tibshirani, Robert. 1996. "Regression Shrinkage and Selection via the Lasso." Journal of the Royal Statistical Society, Series B (Methodological) 58 (1): 267–288. https://doi.org/10.1111/j.2517-6161.1996.tb02080.x.
Weinberger, Kilian Q., Anirban Dasgupta, John Langford, Alexander J. Smola, and Josh Attenberg. 2009. "Feature Hashing for Large Scale Multitask Learning." Proceedings of the 26th International Conference on Machine Learning, 1113–1120. https://doi.org/10.1145/1553374.1553516.
Luteijn, Johannes Michiel, Jacky Tam, Rebecca Denis, Fiona Fan, Jaskaran Minhas, Dhanesh Kadan Puthanveedu, and Abigail Takyi. 2026. "A Framework for Applying Data Science Techniques to Health and Care Actuarial Projects." Cambridge University Press, published 2026; consulted for its sections on leakage-safe encoding and embeddings.