RESEARCH DOSSIER
Explainability Methods for Insurance Predictive Models
Abstract. An explanation of an insurance predictive model is a specified statistical account of how a fitted prediction function behaves with respect to selected features, reference distributions, perturbations, populations, and outputs. This monograph investigates the foundational methods of global and local explanation, their mathematical objects, their reliability when features are dependent, and their application to insurance frequency, severity, aggregate cost, underwriting, claim management, fraud screening, and survival-related prediction. The principal finding is that an explanation has determinate meaning only after its estimand, output scale, reference population, feature intervention or conditioning mechanism, and validation criterion are fixed. Partial dependence, individual conditional expectation, accumulated local effects, permutation reliance, local surrogate methods, Shapley allocations, and contrastive counterfactuals answer related but materially different questions. In insurance datasets, correlated risk attributes, exposure offsets, rare claims, heavy tails, administrative selection, temporal drift, and pricing transformations can make superficially similar explanations disagree legitimately. A mathematical example proves the divergence between conditional and interventional Shapley allocations for a perfectly dependent feature pair; a second derivation clarifies additive explanation on the log scale and its nonadditive translation to monetary insurance outcomes. The monograph compares empirical fidelity, statistical uncertainty, actionable relevance, and communication needs, and establishes a reproducible analytical review procedure as an ordinary synthesis of documented methods. It distinguishes predictive association from causal effects and source-grounded explanation from an unwarranted justification of an adverse insurance decision. The resulting account is international and method-centered; supervisory documents are interpreted within their own authority and scope. No new named theory, empirical field result, or externally peer-reviewed finding is claimed.
Keywords: insurance predictive modeling; explainable artificial intelligence; actuarial analytics; model interpretation; partial dependence; individual conditional expectation; accumulated local effects; permutation importance; LIME; SHAP; Shapley values; counterfactual explanations; correlated predictors; predictive calibration; model fidelity; uncertainty; insurance pricing; claims; governance.
Introduction
Insurance prediction converts information about exposures, contracts, events, and outcomes into quantities used for risk classification and operational decisions. Predicted claim counts, positive claim sizes, expected aggregate costs, lapse probabilities, mortality or morbidity probabilities, and claim-review scores all arise from different sampling structures and different loss functions. A single feature-importance chart cannot by itself express their distinct statistical meanings. The open-access actuarial foundations assembled by Wüthrich and Merz (2023) in Statistical Foundations of Actuarial Learning establish the importance of aligning statistical procedures with insurance response variables and sampling mechanisms. The insurance tariff study by Henckaerts and colleagues (2020) in Boosting Insights in Insurance Tariff Plans develops tree-based modeling for claim frequency and severity while addressing exposure variation, zero-heavy claims, and long-tailed outcomes.
Explainability is particularly consequential because many insurance predictions enter processes involving financial protection, eligibility, price, compensation, fraud review, and communication to affected parties. The evidence for insurance-sector demand is not confined to generic claims about artificial intelligence. Owens and colleagues (2022) conducted a systematic review of XAI in insurance, screening 419 research articles and analyzing 103 within the stated 2000–2021 period; they found XAI concentrated in claims management, underwriting, and actuarial pricing. Those counts describe the review's own search design and historical window, rather than the total present-day literature.
An explanation should carry a complete declaration of its object. Consider an insurer that produces a predicted annual claim cost of $620 for an insured risk. The explanation might seek the fitted function's dependence on annual mileage, the difference from a particular reference premium, the local role of vehicle age, or the operational cause of a price change following a renewal. These are different questions. A statement such as "mileage contributed $70" is incomplete until the analyst identifies the baseline, the response scale, whether the amount is an additive allocation or a finite prediction contrast, and whether other features were allowed to move jointly with mileage.
The present research follows a methodological sequence: define what is being explained; investigate global function behavior; examine local explanations; identify statistical nonidentifiability under dependence; establish quality criteria; and translate findings into insurance communication. Foundational technical statements are traced to the originating papers; institution-specific statements are attributed to official supervisory materials. Mathematical examples below are original analytic demonstrations, not fitted estimates from any insurer's confidential data. The interpretation remains bounded by the sources actually inspected, the assumptions made explicit, and the absence of original observational datasets.
The substantive thesis is that faithful explainability is a relationship among a fitted predictor, a defined reference distribution, an explanation functional, and an intended decision question. Fidelity to one such object does not automatically demonstrate validity for another. The many Shapley values problem, examined by Sundararajan and Najmi (2020), exemplifies why a computationally exact allocation can remain contingent on the chosen explanation game. This finding organizes the analysis without introducing a new named theory.
Explanatory Objectives
The statistical object of explanation
Let observations be indexed by , with recorded response , predictors , exposure where relevant, and observation time . A learning algorithm fitted to training data yields . An explanation of is a functional , where is a specified reference distribution, is a possible individual risk, describes feature subsets or perturbations, and is an output scale, contrast, or predictive loss when needed. The inclusion of these arguments forces the analyst to state what a number or plot represents.
The fitted predictor is itself an estimator. For a frequency model, may estimate expected claims per unit exposure, whereas represents the expected count over exposure . For a severity model, may estimate a conditional mean given a positive observed claim; a pure-premium model estimates an expected aggregate loss per policy or per exposure. A logistic classifier can output either a score, log odds, or calibrated probability. Changes of scale alter attribution magnitudes and often change feature rankings.
Three different targets are commonly conflated. An explanation of the prediction function describes how a locked model maps inputs to outputs. An explanation of a data association describes patterns in the joint distribution of observed predictors and losses. A causal explanation requires a counterfactual intervention structure and sufficient assumptions to identify its effect. These objects may correspond under restrictive circumstances, but the equality must be argued rather than assumed.
Write for a frozen learned model and for a chosen population. One global explanation is a dependence summary , another is performance loss under a disturbance , and a local explanation is anchored to a policy at . Moving from one insurer's portfolio to another changes . Training the same architecture on a new calendar period changes . A change in insured exposure, claim definition, inflation adjustment, or deductible structure changes . None of these transformations is a neutral formatting operation.
Rudin (2019), in interpretable models for high-stakes decisions, argues that interpretable model design itself deserves priority in consequential settings. This argument does not imply that every general linear model is substantively transparent or that every complex model is necessarily opaque. Interpretability depends also on feature definitions, preprocessing, hierarchy, interactions, and the audience's knowledge. A regression coefficient presented without its link function, baseline category, smoothing parameter, or exposure offset can be less informative than a carefully validated explanation of a tree ensemble. The substantive comparison is between actual models and actual explanation tasks.
Insurance targets and observation mechanisms
Insurance targets need explicit temporal and contractual definitions. For claim frequency, a typical counting process specification is as a working model, with adjustments needed for overdispersion or excess zeros. The annualized rate and the count mean answer different questions. If a short-duration policy has exposure , the contribution of exposure to expected count is typically mechanical. Treating that offset as a discretionary underwriting factor leads to an explanation of accounting time rather than insured risk.
In positive-claim severity, one often examines conditional on claim occurrence and recorded coverage terms. The target could be incurred loss, paid loss, ultimate loss, capped amount, or net loss after a deductible. Paid-claim sampling can favor settled claims, so a severity explanation computed on settled payments may partly explain claim closure and payment practices rather than latent damage cost. Claim development, large-loss censoring, reserve changes, and litigation timing should therefore appear in the outcome specification.
For pure premium, the most familiar illustrative compound expression is if the conditional mean of each claim's severity under is exchangeable across its within-policy count and the aggregation assumptions are satisfied. When event severity depends on the number of claims or other latent exposures, the general identity is and the simple product requires additional restrictions. Explaining frequency and severity separately yields useful operational decomposition, but the allocation for their product need not equal the sum of their separate feature attributions.
Fraud detection illustrates a different measurement problem. Confirmed fraud labels often require investigation; investigation selection can depend on an earlier score. A predictive explanation of a model fitted to investigated claims can therefore describe the historical selection mechanism as well as alleged fraud. Rejection, referral, confirmed fraud, and actual fraudulent intent are distinct labels. The interpretation needs to name the recorded outcome and the observation protocol before claiming any connection to wrongdoing.
Time-to-event insurance models similarly distinguish hazard, survival function, cumulative incidence, and predicted event probability over a finite horizon. Censoring assumptions, competing risks, and age or duration time scales alter explanation meanings. A hazard-model explanation on the log hazard scale is not numerically interchangeable with an explanation of five-year event probability.
The CAS authors Chalk, Deacon, Guillen, and Martinelli (Monograph No. 16) emphasize correlation, nonlinearities, and blended tariff models. Their domain concerns illustrate why explanation requires both predictive sophistication and sound actuarial definitions. The monograph provides case studies; it does not establish that any single interpretation technique always performs best.
Explanation questions, audiences, and causal scope
A global question can ask whether a model relies heavily on a rating feature across the target portfolio. A local question can ask how the fitted model arrives at the risk score for a single insured. A sensitivity question asks how the prediction would change if one feature value were replaced. A stability question asks whether a similar explanation appears across fit seeds, plausible perturbations, claim years, and near-optimal fitted models. A communication question asks which verified model facts will help an underwriter, actuary, reviewer, consumer, or supervisor understand a decision.
These questions impose different evidence standards. A policyholder may need an accurate description of the important declared attributes and the terms of a decision. A model validator may need sampling distributions, test losses, training versus holdout comparisons, and perturbation support diagnostics. A regulator may need records of data provenance, model versions, governance steps, and evidence of consumer impacts. Explanations should be faithful to their assigned scope; they should not be used as moral approval of a model whose data or outcomes fail independent review.
There is also a distinction between explanation and justification. Consider a high insurance quote. A valid local attribution can establish that a model assigns a large contribution to a geographic risk indicator under a specified background distribution. Whether using that indicator is statistically warranted, legally permissible, equitable under the applicable rating regime, or consistent with a firm's coverage goals requires further evidence. Consequently, the output of an attribution algorithm can support a review while remaining unable to resolve the entire underwriting or legal decision.
The international insurance supervisory perspective is documented in the IAIS (2025) Application Paper on AI supervision. Its Section 5 discusses explaining outcomes and adapting explanations to intended recipients. The paper gives application guidance in relation to existing Insurance Core Principles; its own opening provisions state that it introduces no new requirements. Its relevance here is conceptual and procedural, not a claim of universally binding legal force.
Global Importance and Partial Dependence
Performance-based importance
Global importance is sometimes treated as an intrinsic ranking of predictors. A more precise definition makes it relative to the fitted model, the test population, and an evaluation loss. Let be a proper or otherwise appropriate loss and . Define marginal permutation reliance for feature by
where is drawn independently from the marginal distribution of . An estimate is formed by shuffling that column on an evaluation sample and measuring the change in loss. If a performance score is maximized instead, its sign convention reverses. The well-established model reliance and model-class reliance framework of Fisher, Rudin, and Dominici (2019) makes clear that importance can depend on a selected good-fitting model and on the full set of well-performing alternatives.
For insurance, the evaluation loss must preserve the outcome. A binary claim-event classifier may be assessed by log loss or Brier score, whereas a count model may use Poisson deviance and a pure-premium model an appropriate likelihood or loss sensitive to zeros and tails. Gneiting and Raftery's (2007) strictly proper scoring-rule analysis supplies a general statistical basis for assessing probabilistic forecasts. The chosen loss is a normative part of the importance question: a ranking under monetary square error can differ from one under claim occurrence log loss. A result should state which prediction target and score it uses.
A useful estimator computes the original held-out loss, repeats the permutation with independent random seeds, and reports the mean change and dispersion. Test-set temporal alignment is essential. A time split that respects claim occurrence and development is often more informative than a random split mixing historically related claims or renewals. Weighted losses can target policies, exposure units, total monetary loss, or another clearly declared population; weights transform the importance estimand.
A feature with low permutation importance in one fitted predictor can still be a strong predictor in an alternative fitted model. Two closely related rating variables can substitute for each other: perturbing either individually leaves the other available. In the permutation-importance documentation, scikit-learn explicitly notes that correlated predictors can each appear unimportant despite predictive utility. The appropriate conclusion is about the fitted predictor's current reliance under the specified perturbation, not about the independent scientific relevance of the variable.
Permutation methods also reveal a problem that rank stability alone cannot solve: a random replacement can produce combinations rarely or never observed in underwriting data. An age-dependent feature and a driver-license tenure measure are an elementary example. A combination with sixteen years of licensed driving experience at age nineteen might be impossible in the relevant jurisdiction. The model's output on such inputs is an extrapolation, so the measured importance may reflect unstable off-support behavior. Hooker and Mentch's (2019) critique of permuting features formalizes this general concern; their cited work is an author preprint in the evidence record.
A conditional permutation version replaces from rather than the unconditional marginal. It better preserves observed dependence, while testing a different question: the additional information uniquely contributed beyond the other features. Strobl and colleagues (2008) examine conditional variable importance in random forests. Under near-perfect collinearity, conditional importance can be close to zero because little independent variation remains. Such a result is a property of conditional information content and should not be paraphrased as "the variable does not matter" in an economic or causal sense.
Partial dependence and individual curves
A partial dependence function for a feature subset is
Its empirical form averages the fitted predictor over held observations of the remaining features. The idea appears in the interpretation of gradient boosted tree models in Friedman's (2001) gradient boosting and model interpretation. Its meaning is a model sensitivity average under a deliberately constructed hybrid distribution: set the selected features to , retain the other observed components, and average the resulting predictions. It does not automatically represent the expected observed outcome among contracts whose selected feature naturally equals .
This distinction is critical when predictors are dependent. If age, vehicle value, and tenure co-vary, replacing age while preserving the other components creates feature combinations far from the historical support. A smooth PDP can still be numerically stable under an extrapolating model, providing a false impression of empirical certainty. Accordingly, a PDP should accompany a depiction of observed support, relevant variable combinations, and a statement of whether the hybrid values are realistic for the product.
The individual conditional expectation curve for observation is . Goldstein and colleagues (2015) in individual conditional expectation plots show how heterogeneous profiles and interactions become visible where their mean PDP obscures them. For an additive model with no relevant interaction, ICE curves have similar shape apart from vertical offsets. When the curves diverge or cross, feature effects depend on other characteristics. An average curve does not make that heterogeneity disappear.
Centering ICE curves at a reference produces . In a pricing model this can compare modeled changes in rate for two vehicle ages while holding each risk's other coordinates fixed, but it cannot by itself say how that policyholder's price will change next year. The realized renewal will also alter time, underlying loss costs, inflation, legal rating restrictions, and operational rate factors. The distinction between a controlled model contrast and actual renewal dynamics must be retained.
ICE and PDP require choices about grid density, tail handling, categories, interaction subsets, and response scale. For a feature highly concentrated at one value, plotting hundreds of evenly spaced points can devote much of the figure to unsupported regions. Quantile grids and rug plots better convey data density. When examining age or exposure, contractual minimums and valid intervals should be enforced; when examining a binary protected or legally restricted factor, counterfactual replacement demands separate legal and causal interpretation.
Accumulated local effects
Apley and Zhu (2020) introduced accumulated local effects as an alternative to partial dependence, especially under correlated inputs. For a continuous predictor , a population first-order ALE effect can be expressed, under differentiability and regularity, as a centered cumulative integral of the conditional local derivative:
An implementable bin estimator forms prediction differences between the lower and upper boundaries of observed intervals for records inside each interval, cumulates the average differences, and centers the resulting function. By relying on locally observed combinations, ALE generally reduces the gross extrapolation associated with naive PDP substitutions. The method remains a description of the fitted model under a particular observational distribution; a conditional derivative does not by itself identify a causal effect.
Suppose annual mileage and commuting frequency are positively associated. A PDP may repeatedly combine high mileage with records declaring no commuting, even if such records are uncommon. ALE instead examines local mileage changes among contracts currently in mileage neighborhoods. The resulting graph answers a different question and may have a different shape. Neither is an automatic winner; each must be matched to the analyst's desired perturbation and support assumptions.
Quantile binning provides enough records in dense areas and fewer unsupported stretches, but bin-width choice alters resolution and bias. Insurers should test alternative bin counts, temporal folds, exposure-weighted variants when substantively appropriate, and sparse segments. At portfolio extremes, local sample counts and interval-specific uncertainty should be reported. For categorical or ordinal attributes, ALE requires adapted contrasts, and the ordering of categories may introduce interpretation conventions.
Global summaries and interactions
A global summary of absolute local attributions, such as , measures average attribution magnitude under a chosen background and explanation game. It is not automatically the same as permutation reliance , because one is a decomposition of predictions and the other quantifies loss sensitivity. Neither is equal to a causal effect merely because it is large. Likewise, a tree's split count or impurity reduction measures its internal construction and can favor high-cardinality predictors; the scale of the plotted importance is part of the explanation.
Interaction analysis examines whether the effect of one characteristic changes with another. An additive function has no mathematical interaction on that declared scale, whereas the response-scale function has multiplicative interactions in absolute response changes. Hence a log-link GLM can be additive on the log rate scale and nonadditive in dollars. A global interaction summary is valid only with the link and prediction target stated.
Tree-specific global interpretation can aggregate consistent local quantities. Lundberg and colleagues (2020) discuss global understanding from tree explanations, combining local contributions, feature interactions, and monitoring. Averages over a stable and relevant portfolio can reveal rare but substantial localized effects. However, a summary across heterogeneous insurance products can conceal opposite-sign patterns and distort the prevalence of the risk conditions that generate them.
A global analysis thus should report the reference portfolio, target variable, response scale, performance measure, dependence treatment, grouping rules, and relevant stratifications. A rank without these attributes is a visual description with an indeterminate statistical object. When alternative reference populations yield different rankings, the disagreement is an empirical property that requires explanation, not a reason to choose the most convenient figure.
Local Feature Attribution
Local surrogates and LIME
A local explanation seeks to clarify the behavior of a fixed predictive function near a particular input . A widely used approach constructs a simpler approximation in a local neighborhood. Ribeiro, Singh, and Guestrin (2016) introduced LIME in explaining predictions of classifiers, fitting an interpretable surrogate that approximately follows a black-box model in the vicinity of an instance. A general statement of its objective is
where weights locality, restricts interpretation complexity, measures local agreement, and penalizes complexity. The resulting explanation approximates the fitted function under a chosen sampling and distance construction. It does not establish a causal effect.
In a motor insurance example, consider a gradient-boosted model predicting claim occurrence. A local surrogate might assign coefficients to recent claim history, vehicle age, mileage class, and geographic area. These coefficients summarize the model's behavior over selected artificial variations near the focal contract. They need not match coefficients from a global logistic regression on claims, and they need not reproduce Shapley allocations. Even two local fits for the same frozen contract can disagree when sampling, discretization, or feature encoding changes.
Locality is determined by the distance metric, proximity kernel, sampling law, encoding, and sparsity constraint. Standardizing mileage while leaving a high-cardinality categorical geography feature represented by dummies can distort effective distances. Independently perturbing dummy coordinates can create impossible policy records. A local surrogate may achieve high numerical fit to synthetic model scores but have limited validity for actual insurable risks. Local predictive agreement should therefore be evaluated on feasible inputs and, where possible, on a relevant local holdout set.
The minimum record for a credible local surrogate identifies the frozen prediction and scale, the focal risk, its feasible neighborhood, the training and evaluation procedure for the surrogate, a fidelity score, the local sample size, uncertainty across resamples, and the exact code and preprocessing version. Where importance rankings vary sharply across neighborhoods, the proper result is an instability finding. A single attractive bar chart is insufficient.
Attribution and model provenance must also travel together. A feature may receive a small coefficient in one insurer's underwriting-year model and a larger one after claim-cost inflation or changing underwriting experience. An explanation generated against model version cannot be presented as the explanation of version . The archive must preserve the decision model, explanatory background dataset or its approved reproducible substitute, feature mappings, and explanation algorithm version.
Shapley allocations and their reference games
Shapley-value explanations distribute a prediction contrast among features using marginal contributions averaged over feature coalitions. Let index features and let specify a characteristic value for subset . The Shapley attribution of feature is
Under the conventional game conditions, efficiency gives . Lundberg and Lee (2017) present the SHAP additive attribution framework, relating model-specific methods to a shared additive approach. The allocation is a mathematical answer for a defined game; the game itself is selected by the analyst.
Two important games use different information. In an interventional or marginal reference construction, , with omitted features drawn from a background marginal law. In an observational conditional construction, , with missing coordinates integrated over their conditional distribution. The marginal construction can create unrealistic combinations when features depend on each other. The conditional construction can assign predictive information to a variable absent from the explicit functional formula because that variable reveals correlated inputs.
Aas, Jullum, and Løland (2021) investigate Shapley explanations with dependent features, proposing dependence-aware approximations. Sundararajan and Najmi (2020) demonstrate in The Many Shapley Values why competing choices of characteristic function can yield different valid Shapley allocations. The critical question is therefore which absent-feature game answers the insurance analyst's intended question.
Dependency patterns are common in insurance: driver age and driving tenure; construction type and building value; catastrophe zone and property coordinates; insured amount and replacement cost; annual mileage and vehicle use; occupation and industrial classification. A Shapley report should say whether it distributes functional contributions while breaking dependencies, distributes informational relevance under the observed joint law, or uses a third explicitly specified grouping or partitioning procedure.
Reference populations matter independently of the allocation algorithm. The background may comprise all in-force policies, quoted risks, claims submitted in the last year, an approved comparison cohort, or a specific product segment. A young driver's predicted claim frequency can be above the whole-portfolio average and below the average for a high-mileage novice-driver cohort. Attribution signs can reverse because the baseline changed, without any difference in the individual prediction.
Computing exact Shapley allocations generally entails evaluating many subsets. Implementations sample coalitions, estimate conditional distributions, or exploit model-specific structure. Covert and Lee (2021), in improving KernelSHAP estimation, analyze estimation and convergence of practical regression-based Shapley calculations. A technical report must distinguish Monte Carlo error in estimating the attribution, sampling error from the background data, uncertainty from learned conditional distributions, and statistical variation in the original fitted predictor.
The frequent statement that a SHAP bar shows how much a feature "caused" a predicted risk is too strong. Attribution to a model contrast is not intervention identification. A large allocation for area code, prior claim history, or a medical classification does not demonstrate that changing the code or classification would change physical risk by the same amount. Causal conclusions require causal structure and assumptions external to the Shapley game.
Tree-specific explanation, interaction, and output scale
Tree ensembles admit algorithms that exploit splitting structure. Lundberg and colleagues (2020) present tree-specific explanations and interactions, combining local attributions with global analysis. Computational exactness for a particular tree game does not settle the meaning of the background distribution, the reasonableness of conditional independence assumptions, or the substantive interpretation of correlated risk factors.
An insurance frequency predictor may be additive on the log mean scale. If and , then
Consequently, additive log-score contributions become multiplicative factors on the rate scale. In general, . Independently exponentiating log-scale attributions and calling their sum a dollar premium allocation is invalid. One can define a separate Shapley game on the rate or dollar scale, but its attributions need not be algebraic transformations of log-scale allocations.
For a binary model, additive contributions to log odds are not percentage-point changes in predicted claim probability. The same log-odds contrast generates different probability differences depending on the starting probability. For catastrophe severity, a contribution to predicted log ultimate loss is not an absolute increment in loss dollars. Every explanation must carry the correct unit and link function.
Global aggregation raises an additional issue. The arithmetic mean of signed local contributions can vanish when positive and negative values cancel, whereas the mean absolute allocation remains large. Such summaries measure distinct quantities. A distribution of signed effects, a mean absolute value, a tail summary, and a segment comparison can together be informative, provided the portfolio weights and reference game are declared.
An end-to-end insurance decision may contain preprocessing, embeddings, predictive ensembles, calibration, rating rules, expense loadings, minimum premiums, rate caps, discounts, and regulatory adjustments. Explaining a tree model operating upstream of these steps describes that model layer. A complete explanation of a final quoted premium must additionally trace its downstream transformations. In particular, a 15% modeled claim-cost difference need not imply a 15% premium difference after expenses, loading, taxes, or capping.
Counterfactual contrast and feasible action
A counterfactual model explanation searches for an alternative input producing a specified model outcome. One formulation is
where restricts admissible alternatives and is a desired prediction. Wachter, Mittelstadt, and Russell (2018) present counterfactual explanations of automated decisions as contrastive information that can be communicated without full disclosure of a black-box model. The output is a prediction contrast under the specified alternative, not automatically a feasible recommendation or a guarantee.
In insurance, date of birth is immutable, and a record's historical claim cannot simply be undone. Actual mileage or safety equipment may be changeable, but the pricing consequence depends on verification, product terms, eligibility, and applicable rate rules. A property owner may retrofit a roof but cannot alter the building's construction year. A counterfactual whose proposed changes violate such constraints is mathematically possible in input space while operationally meaningless.
Karimi and colleagues (2022), in their survey of algorithmic recourse, distinguish contrastive explanations from feasible and consequential recommendations. Four objects should be documented separately: the change in a model input, the feasible real-world action, the contractual or operational response, and the causal effect of the action on future insured losses. Even when all four are available, their quantitative effects can differ.
For claim triage, a model may predict a different review score if a missing document is present. The insured may submit the document, while a valid coverage determination still requires its own process. For telematics pricing, a change in driving behavior may affect future risk and the model may be retrained after many policyholders respond. Static contrasts are insufficient to estimate the feedback. Where a communication invites a customer to act, it should be based on an authorized and verifiable action pathway, not on the nearest point in arbitrary Euclidean feature space.
Explanation Stability and Fidelity
Statistical and computational uncertainty
An explanation is estimated from data, a fitted function, and an algorithm. Different stages generate distinct uncertainties. Training-sample uncertainty changes the predictive model itself. Sampling uncertainty in the background dataset changes the reference distribution. Monte Carlo error changes approximated model-agnostic attribution values. A learned conditional generator creates additional estimation and specification error. Temporal changes alter the target portfolio. The resulting spread cannot be summarized faithfully by one generic confidence interval without identifying which uncertainty sources the interval represents.
Let an explanation estimator be , where is the training sample, the background sample, and the algorithm's random seed or Monte Carlo draws. A useful conceptual decomposition distinguishes
when these expectations exist and the conditional construction is appropriate. The first term includes variation conditional on a fitted-model training set; the second represents variation from refitting across training samples. This is the law of total variance, not a proposed new statistical principle. It shows why a tightly converged SHAP estimator for one locked model does not establish the stability of explanations across plausible fitted models.
A practical repeated analysis may involve resampled or temporally refitted models, independent explanation backgrounds, and computational seeds. The resulting three-way array can estimate dispersion attributable to each design choice. One need not compute every possible combination if cost prohibits it, but the report should make clear whether uncertainty estimates concern the original function, the empirical population, or only repeated numerical approximations.
Training resamples require attention to dependence. Random bootstrap resampling of individual claim records can create leakage when claims share policyholders, households, vehicles, disasters, or calendar events. Blocked resampling by policy, insured, claim occurrence group, or catastrophe event is more appropriate where those units define dependence. Spatial resampling may be needed for regional hazard features; time-based splits for changing claim practices. A narrow confidence interval from a design that breaks the sampling structure is misleading.
For expensive local explanations, approximate uncertainty can be assessed through repeated seeds and backgrounds, while full model refitting may be limited to selected representative records. Convergence diagnostics should report the estimation target, algorithmic budget, interval construction, and retained background size. A confidence interval for an individual feature attribution must not be presented as a claim that the feature causally alters insured loss within the stated interval.
Global rankings have their own uncertainty. The probability that a feature ranks first across resamples differs from a confidence interval for its mean importance. Two features may have overlapping reliance distributions while a deterministic ordering is still produced by their point estimates. Rank agreement, correlation of estimated importance vectors, and sign consistency of local effects provide complementary diagnostics. Very small differences in importance need no categorical substantive interpretation.
For a portfolio with rare large claims, predictive loss estimates may have high variance, especially if squared monetary loss is used. A single high-severity event can dominate a permutation-importance estimate. Report tail sensitivity, examine capped and uncapped losses for a scientifically justified purpose, and assess whether the metric is aligned with the insurer's actual target. Repeated permutations reduce shuffling noise but do not cure scarce outcome information.
Fidelity testing and adversarial stress
Fidelity means agreement between an explanation and the specified object it purports to represent. A local surrogate's fidelity can be assessed by predictive error on a feasible neighborhood around its target instance. A Shapley implementation's computational fidelity includes the additive reconstruction check and convergence toward the declared coalition game. A global PDP's numerical fidelity can be assessed by reproducibility of the plotted averages under stable software and sampling choices. These checks answer different questions and should be recorded separately.
For a surrogate , an elementary local error functional is
where is the declared feasible local evaluation law. If model outputs are probabilities, a probability-scale error is appropriate for the chosen purpose, while class-label agreement can hide large disagreements in confidence. If outputs are monetary losses, scale and heteroscedasticity must be considered. Evaluating on the very synthetic samples used to fit the surrogate gives an optimistic view of generalization to new local perturbations. Distinct fitting and evaluation perturbations can help.
A complete fidelity exercise may compare several neighborhoods: the algorithm's default kernel, a conditional data-based kernel, a physically feasible neighborhood, and a local empirical holdout. Fidelity under one neighborhood does not establish fidelity under all others. A model may be perfectly linear in an unsupported synthetic region and highly nonlinear within real policy combinations, or vice versa. This makes support diagnostics an integral part of explanation evaluation.
SHAP's efficiency identity provides a useful numerical check:
for output scale and a declared tolerance . Passing this check verifies additive reconstruction under a chosen formulation; it does not prove that the baseline, game, conditioning, or feature set is scientifically appropriate. A reconstruction residual can arise from stochastic approximation, omitted interactions, omitted columns, calibration transformations, or implementation mismatch.
Perturbation methods also face deliberate manipulation. Slack and colleagues (2020) show in adversarial attacks on LIME and SHAP that a model can be constructed to behave benignly on the synthetic distribution sampled by an explainer while pursuing a different decision rule on the actual data distribution. This research establishes a concrete failure mode of certain post hoc explanation pipelines. It does not imply that every explanation or insurance deployment is adversarial. The appropriate safeguard is independent evaluation of explanation fidelity on relevant observations and feasible shifts, not merely faithful reproduction of the explainer's own synthetic samples.
Explanation invariance deserves direct tests. Two functionally equivalent encodings of a rating variable may lead to a different attribution partition; one-hot and grouped representation can redistribute attributions. Monotone transformations of model output can change the allocation game. A model may remain predictively similar after a refit while local attributions change because correlated variables exchange explanatory weight. An explanation system should disclose which invariances it promises and which transformations fundamentally change the estimand.
There is a boundary between an explanation metric and the validity of the model. Calibration evaluates whether predicted event frequencies match observed events across defined groups. Discrimination evaluates ranking ability. Predictive scoring rules evaluate probabilistic accuracy. Local explanation fidelity evaluates agreement with the model in a specific neighborhood. A model may be well calibrated and difficult to interpret locally, or have a highly faithful explanation and poor claim prediction. These are independent axes of evaluation.
Gneiting and Raftery (2007) develop strictly proper scoring rules for forecast evaluation. A scoring rule can evaluate the predictive distribution; it cannot verify that an attribution is causal or ethically appropriate. For insurance models, explanation QA should therefore be attached to, rather than substituted for, ordinary actuarial model validation.
Model-class uncertainty, drift, and reproducibility
When two predictive models achieve similar holdout loss, they can rely on different features. Fisher, Rudin, and Dominici (2019) examine variable reliance across model classes, providing a formal foundation for studying reliance among acceptable models. This is especially relevant in correlated insurance feature sets, where different predictive representations can retain similar performance.
An insurer should investigate whether a purportedly essential factor is important in most well-performing models or only one selected model. A feature that dominates a single boosted-tree run but disappears under equally accurate alternative models has a fragile role in model-based interpretation. Conversely, a feature retaining similar reliance under diverse validated architectures may support a stronger descriptive conclusion, subject to the reference-population and loss assumptions.
Temporal drift requires two distinct comparisons. A fixed model can be explained across successive portfolios to show composition change: versus . A current and past model can be explained against a fixed reference cohort to show model change: versus . Comparing with changes both arguments and does not identify which caused the movement in the explanation. This separation is a straightforward application of controlled comparison, not a new formal theory.
Consider a property insurer whose hurricane exposure portfolio shifts inland while the underlying storm model is recalibrated. A decline in average coastal-distance attribution can reflect fewer coastal buildings, a changed fitted relation, an altered catastrophe hazard prior, or a different attribution background. Recording those factors permits an interpretable historical analysis. Using only a before-and-after feature-importance chart does not.
Versioned reproducibility records at minimum the target definition and claim development cutoff, extraction date, eligible population, feature dictionary, transformations and encodings, model weights, training interval, background source and sample seed, code and package versions, link/output scale, explanation game, evaluation loss, and results. Where data protection limits retaining raw policy records, the archive can preserve secure references, model artifacts, approved sufficient aggregates, privacy-protective validation summaries, and immutable hashes, subject to applicable governance controls.
A source claim also has provenance. A paper can establish a proposed algorithm's properties under assumptions, not prove that an insurer's deployment satisfies them. This monograph therefore assigns original method claims to their publications, distinguishes official supervisory statements from applicable law, and identifies its synthetic calculations as analytic examples. Those are the same boundaries an institutional insurance explanation process should maintain.
Insurance Communication and Review
Audience-specific communication
A useful explanation is a complete statement aimed at a defined question and audience. For an actuarial pricing team, relevant outputs may include loss ratio validation, exposure-weighted distributions, feature influence on frequency and severity, cross-year stability, and the response-scale interpretation of relativities. For underwriting reviewers, a focal prediction contrast, valid input dictionary, and explanation stability under feasible nearby records may be most helpful. For a claims reviewer, the explanation must distinguish predictive triage from contractual adjudication. For a consumer, concise reasons for a material insurance decision should track actual inputs, product rules, permitted evidence, and available review channels.
These audiences require a shared technical record while receiving appropriately structured communication. One layer should identify the prediction object and actual decision component. Another should state the baseline and attributed contrasts. A third should record validation and any instability or data limitation that materially affects interpretation. Merely presenting a visual ranking leaves uncertainty about what the score means or what is within the person's control.
The IAIS (2025) AI supervision application paper discusses the need for explainability appropriate to different stakeholders and the context of insurance supervision. Its publication is a supervisory application paper, not a newly created treaty or directly binding world regulation. The analysis here uses it to identify supervision-relevant communication concerns without projecting it onto every legal system.
The European Insurance and Occupational Pensions Authority's (2025) opinion on AI governance addresses national supervisory authorities and risk-proportionate governance of insurance AI systems within its stated scope. The opinion is distinguishable from the AI Act, from primary European legislation, and from national implementation. Its regulatory relevance supports governance context; it does not make a specific attribution algorithm universally mandatory.
In the United States, the National Association of Insurance Commissioners adopted a model bulletin on insurer AI systems on December 4, 2023. It concerns insurers' AI governance programs and regulatory examination documentation for AI affecting consumers. A model bulletin is a model supervisory instrument; the legal effect in a state depends on that state's adoption, enforcement framework, and existing law. None of these documents alone demonstrates that a particular insurer has met all applicable obligations.
A communication should not confuse model explanation with a completed legal or conduct review. An insurer can faithfully explain why an algorithm raised a claims-review priority while still needing to assess whether the data, claim procedures, and final decision were appropriate. A rate explanation may be numerically exact for the risk-cost model yet omit premium-loading decisions applied downstream. A consumer-facing statement should identify that boundary in ordinary language.
Worked insurance explanation dossiers
The following examples specify analytic records that an insurer could reproduce in a controlled test environment. They contain synthetic values and illustrative assumptions, not claims about actual firms.
Motor insurance frequency and pure-premium example
Suppose a pricing system estimates annualized motor claim frequency using a log-link model and estimates severity separately. A policy has years of exposure, a predicted annualized frequency , and conditional mean severity $6,000$ dollars. Under the stated simple compound assumptions, the expected count during the observed exposure is and expected total claim cost for this exposure period is . The annualized pure premium is . A statement that the model predicts in claims "for the half-year policy" would mix annualized and observed-exposure targets.
The explanation record identifies , , , and whether rate relativities are explained on log or dollar scale. It uses a defined reference portfolio, produces ICE curves for mileage, ALE for a potentially correlated driver-tenure attribute, and a set of local SHAP attributions for the risk-cost layer. It reports that a response-scale premium difference can include interactions not visible in separate component log explanations. If feature groupings differ between rate engine and explanatory software, they must be reconciled before communication.
To test claim relevance, the investigator compares model performance under Poisson-style frequency deviance and an appropriate severity loss, checks out-of-time calibration for claims, computes explanation stability across seasonal training cohorts, and evaluates support constraints for age and tenure. The consumer-facing statement would then identify the material rating inputs and explain that quoted premium also contains nonclaim expense and other contract factors. The example shows how a single coherent record can preserve different actuarial and communication layers.
Property catastrophe scoring example
Imagine an insurer using a property claim probability model with elevation, distance to coast, construction type, insured value, and local hazard score. Elevation and hazard score correlate geographically. A large marginal SHAP attribution to elevation may reflect explicit model dependence, while conditional Shapley values may allocate the same score partly to hazard because location predicts both. Both calculations must name the game. An ALE analysis may remain closer to observed feasible combinations, but it also represents model derivatives under the observed portfolio and its sampling limitations.
For rare catastrophic events, event-level blocked validation is important. Randomly splitting individual properties from the same catastrophe between training and testing can overstate model generalization and make the resulting explanations appear more reliable than they are. The valid analysis explicitly distinguishes event-conditional severity, annual event frequency, building insured value, deductibles, and net covered loss. Explaining insured loss without separating claim frequency and contractual coverage limits can obscure which part of the model responds to physical hazard.
If the aim is to identify a risk-reduction action, a hypothetical roof retrofit requires engineering and policy-term verification. An optimization that simply lowers a nonmodifiable construction-year field is not actionable recourse. A valid communication could compare model outcomes for an authorized, documented retrofit scenario and state that its modeled difference is contingent on verification and the rating framework.
Claims fraud-screening example
Consider a model that assigns a score of to a claim and refers it for specialist review. A local surrogate explanation indicates that high claimed cost, document irregularity, and prior referral codes are important in the chosen synthetic neighborhood. The first essential check is the score's target: whether is a calibrated probability of confirmed fraud or merely a ranking score learned from investigated-claim outcomes. If prior referral codes were themselves produced by earlier fraud models, the variables can capture historical selection.
The explanation is fit against a restricted feasible neighborhood rather than arbitrary combinations of claim status and completion dates. The investigator checks whether the explanation remains similar when referral history is excluded, a matched cohort is used, or a different conditional generator is applied. The downstream human process examines primary documents and coverage conditions independently. A consumer notice about additional verification should not call the model score evidence of guilt.
This example also illustrates the importance of precise probability language. A displayed has no reliable probabilistic interpretation in an uncalibrated ranking system. Explanation fidelity and calibration should both be tested. Neither algorithmic explanation nor well-calibrated probability establishes an individualized legal determination.
Health or life insurance event horizon example
A predictive model estimates probability of a covered event within one, three, and five years, accounting for censoring. A feature can have different SHAP contributions at each horizon when explanations are computed on horizon-specific probability scales. A feature whose effect on hazard is positive in an early interval can be neutral later; an average importance rank does not communicate this time dependence. A technically valid report records censoring assumptions, exposure or follow-up period, competing risks when applicable, and the horizon-specific reference population.
For sensitive data, the model explanation must be connected to the legitimate authorized analysis and privacy controls. The claim that a feature is correlated with higher predicted loss does not establish medical causation. The methods in this monograph support model interpretation, while product legality, permitted risk factors, consumer protection, and data governance require distinct assessments.
Comparative methods and quality conditions
The following methodological comparison names the question each technique estimates, its principal dependence issue, and a relevant quality check. The entries represent an analytical synthesis of the originating methods, rather than rankings derived from a benchmark experiment.
| Method | Statistical object and intended use | Dependence and data concern | Primary validation |
|---|---|---|---|
| Marginal permutation importance | Out-of-sample performance change when a feature is scrambled | May create implausible feature combinations and understate substitution among correlated factors | Holdout metric, repeated permutations, support assessment |
| Conditional permutation importance | Incremental performance information after conditioning on remaining features | Conditional generator must be estimated; near redundancy can erase unique relevance | Conditional fit, alternative generators, repeatable losses |
| Partial dependence | Average frozen-model response to setting feature values across reference records | Marginal substitution may leave observed joint support | Density/rug, support filtering, stratification, ICE comparison |
| Individual conditional expectation | Focal-model response curve for each reference record | Curves can traverse unrealistic combinations | Feasibility, curve heterogeneity, outlier checks |
| Accumulated local effects | Centered accumulation of local fitted-model changes under observed conditional neighborhoods | Sparse bins, variable dependence and conditional sampling affect estimates | Bin counts, alternative binning, temporal stability |
| LIME/local surrogate | Sparse local approximation to a frozen model | Perturbation distribution, locality kernel and encoding determine target | Feasible local holdout fidelity, seed sensitivity |
| Marginal SHAP | Additive attribution under a marginal feature-replacement game | Off-support feature coalitions; reference-population changes | Additive reconstruction, backgrounds, model version |
| Conditional SHAP | Additive allocation under an observational conditional game | Conditional distribution estimation and redundant information | Conditional fit, game comparison, uncertainty |
| Counterfactual explanation | Prediction contrast under feasible input alterations | Nearest mathematical change may be impossible or nonactionable | Feasibility constraints, model robustness, authorized action pathway |
These methods can be combined constructively. For example, permutation reliance can establish whether a variable affects held-out predictive performance, ALE can summarize average local behavior, ICE can reveal interactions, and a local SHAP explanation can attribute a focal difference from a declared baseline. Agreement across techniques can support a robust description when the estimands genuinely overlap. Disagreement can be informative when techniques differ in conditioning, scale, target population, or performance loss.
A complete individual quality record can be summarized by a series of checks expressed as questions. Is the target a count, claim amount, annualized rate, probability, hazard, or downstream operational decision? Is the model frozen and reproducible? Is the background population relevant? Does the perturbation mechanism remain within feasible insurance records? Does the figure use the correct scale, exposure denominator, and time horizon? Are correlated features handled deliberately? Are numerical approximations stable? Are results consistent across credible fitted models and time periods? Does the proposed interpretation exceed what the underlying method proves? This sequence is a synthesis of existing methodological controls and introduces no new named governance doctrine.
The required empirical follow-through depends on the setting. For catastrophe risk, stratified event-level testing is prominent. For motor claim frequency, exposure and seasonality matter. For a claim triage system, label selection and human review are essential. For long-duration life and health outcomes, censoring and the horizon are decisive. An explanation system that prints identical generic warnings for all products can satisfy documentation form while missing the statistical problem specific to each target.
Research priorities and interpretive boundaries
Three methodological gaps merit particular attention. First, explanations of correlated insurance variables remain sensitive to the reference game and conditional generator. Benchmarking algorithms without comparing their estimands can conceal this distinction. Research should evaluate support-conditioned alternatives under realistic actuarial dependency structures and report where features are observationally redundant.
Second, operational premiums and claims outcomes are the outputs of pipelines, not isolated predictors. Research can compare explanations of underlying expected loss with explanations of the complete price or decision transformation, testing which audience questions each representation can answer. Model explanation on a log link, pure-premium explanation in dollars, and final premium reason codes must be aligned before they are communicated as a single story.
Third, stability across model versions, insured populations, and event periods is less visible than stability across computational seeds. In actuarial experience analysis, long-tailed claims and changing exposure composition can make model reliance time-dependent. Scientific comparisons should separate model changes from population shifts and use realistic blocked holdouts.
There is a wider challenge of evaluation incentives. An explanation can be concise, visually persuasive, and yet unfaithful to the model's relevant behavior. Conversely, a technically faithful allocation can be incomprehensible or irrelevant to a consumer asking about a concrete contract action. Future work should jointly assess statistical fidelity, distributional plausibility, audience comprehension, and whether communicated decisions accurately reflect the entire authorized insurance process. These are linked evaluation aims, not interchangeable metrics.
The source-critical contribution of this monograph lies in comparing distinct explanation estimands under actuarial observation structures and deriving their communication consequences. The algebraic Shapley comparison and frequency-severity interaction example establish specific analytical results; neither establishes empirical superiority of a method. No independent insurer dataset, randomized intervention, new algorithm, or formal systematic evidence review was conducted. Assertions about effectiveness remain bounded by the investigated source literature and the definitions used here.
Conclusion
Explanation methods in insurance predictive modeling operate on fitted functions, background distributions, model outputs, and carefully chosen counterfactual or conditional constructions. The first requirement of scientific validity is to identify these objects. A bar in a local attribution chart, a rank from permutation importance, and the slope of an ALE curve cannot be understood independently of the question and probability structure they encode.
The fundamental methods have complementary strengths. Permutation importance examines performance reliance; partial dependence and ICE describe fitted-model response to selected feature settings; ALE summarizes local changes across observed neighborhoods; LIME approximates the local model with an interpretable surrogate; Shapley methods allocate a prediction contrast within a specified game; counterfactual methods compare model outcomes under alternative records. These methods differ in sampling, dependence assumptions, extrapolation, and interpretation. Their differences are scientifically consequential in insurance portfolios with correlated predictors and complex contractual outputs.
The exact two-feature demonstration proved that conditional and marginal Shapley attribution can yield different allocations even with no numerical approximation error. The frequency-severity example showed that pure-premium interactions can generate dollar allocations different from separate component contrasts. The log-link derivation established why additive attribution in transformed model space becomes multiplicative on the insurance-rate scale. Together, these results identify common routes to erroneous monetary or causal communication and offer directly checkable safeguards.
Explanation fidelity, predictive validity, causal identification, and legal or operational justification require their respective evidence. Their coexistence in a high-stakes insurance process is valuable when the model and observation mechanism are documented. An explanation capable of reconstructing a score can still be unstable across insured populations; a well-calibrated risk model can still use nonactionable or legally restricted variables; a feasible model contrast may fail to describe the causal consequences of a real-world intervention. The research therefore supports source-grounded explanations that state exactly what they establish.
For international insurance science, explanation quality rests on explicit prediction targets, sampling frames, reference distributions, link functions, dependence treatment, uncertainty, temporal validity, and the actual decision layer. Institutional communications can use the same verified technical facts at different levels of detail. The IAIS, EIOPA, and NAIC documents surveyed here provide context within their own regulatory and supervisory scopes; they do not replace jurisdiction-specific obligations.
The final scientific conclusion is that explainability in insurance is a disciplined practice of defining and validating statistical comparisons. A complete explanation makes the fitted model's behavior intelligible under declared assumptions and preserves the boundary between predictive description, causal inference, and insurance decision-making. The present dossier provides an original analytical synthesis of established methods and exact illustrative derivations, without claiming a new named theory or externally validated empirical finding.
References
Friedman, Jerome H. (2001). “Greedy Function Approximation: A Gradient Boosting Machine.” Annals of Statistics 29(5):1189–1232. Full source.
Goldstein, Alex, Adam Kapelner, Justin Bleich, and Emil Pitkin (2015). “Peeking Inside the Black Box: Visualizing Statistical Learning with Plots of Individual Conditional Expectation.” Journal of Computational and Graphical Statistics 24(1):44–65. Full source.
Apley, Daniel W., and Jingyu Zhu (2020). “Visualizing the Effects of Predictor Variables in Black Box Supervised Learning Models.” Journal of the Royal Statistical Society: Series B 82(4):1059–1086. Full source.
Fisher, Aaron, Cynthia Rudin, and Francesca Dominici (2019). “All Models Are Wrong, but Many Are Useful: Learning a Variable's Importance by Studying an Entire Class of Prediction Models Simultaneously.” Journal of Machine Learning Research 20(177):1–81. Full source.
Ribeiro, Marco Tulio, Sameer Singh, and Carlos Guestrin (2016). “Why Should I Trust You?: Explaining the Predictions of Any Classifier.” Proceedings of ACM SIGKDD 2016:1135–1144. Full source.
Lundberg, Scott M., and Su-In Lee (2017). “A Unified Approach to Interpreting Model Predictions.” Advances in Neural Information Processing Systems 30. Full source.
Lundberg, Scott M., et al. (2020). “From Local Explanations to Global Understanding with Explainable AI for Trees.” Nature Machine Intelligence 2:56–67. Full source.
Aas, Kjersti, Martin Jullum, and Anders Løland (2021). “Explaining Individual Predictions When Features Are Dependent: More Accurate Approximations to Shapley Values.” Artificial Intelligence 298:103502. Full source.
Strobl, Carolin, Anne-Laure Boulesteix, Thomas Kneib, Thomas Augustin, and Achim Zeileis (2008). “Conditional Variable Importance for Random Forests.” BMC Bioinformatics 9:307. Full source.
Hooker, Giles, and Lucas Mentch (2019). “Please Stop Permuting Features: An Explanation and Alternatives.” arXiv preprint 1905.03151. Full source.
Sundararajan, Mukund, and Amir Najmi (2020). “The Many Shapley Values for Model Explanation.” Proceedings of Machine Learning Research 119:9269–9278. Full source.
Covert, Ian, and Su-In Lee (2021). “Improving KernelSHAP: Practical Shapley Value Estimation Using Linear Regression.” Proceedings of Machine Learning Research 130:3457–3465. Full source.
Rudin, Cynthia (2019). “Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead.” Nature Machine Intelligence 1:206–215. Full source.
Slack, Dylan, Sophie Hilgard, Emily Jia, Sameer Singh, and Himabindu Lakkaraju (2020). “Fooling LIME and SHAP: Adversarial Attacks on Post Hoc Explanation Methods.” Proceedings of AAAI/ACM AIES 2020. Full source.
Wachter, Sandra, Brent Mittelstadt, and Chris Russell (2018). “Counterfactual Explanations without Opening the Black Box: Automated Decisions and the GDPR.” Harvard Journal of Law & Technology 31(2); arXiv 1711.00399. Full source.
Karimi, Amir-Hossein, Gilles Barthe, Bernhard Schölkopf, and Isabel Valera (2022). “A Survey of Algorithmic Recourse: Contrastive Explanations and Consequential Recommendations.” ACM Computing Surveys 55(5):Article 95. Full source.
Gneiting, Tilmann, and Adrian E. Raftery (2007). “Strictly Proper Scoring Rules, Prediction, and Estimation.” Journal of the American Statistical Association 102(477):359–378. Full source.
Wüthrich, Mario V., and Michael Merz (2023). “Statistical Foundations of Actuarial Learning and Its Applications.” Springer Actuarial, Springer Nature. Full source.
Owens, Emer, Barry Sheehan, Martin Mullins, Martin Cunneen, Juliane Ressel, and German Castignani (2022). “Explainable Artificial Intelligence (XAI) in Insurance.” Risks 10(12):230. Full source.
Henckaerts, Roel, Marie-Pier Côté, Katrien Antonio, and Roel Verbelen (2020). “Boosting Insights in Insurance Tariff Plans with Tree-Based Machine Learning Methods.” North American Actuarial Journal 25(2):255–285 (published online 2020). Full source.
Chalk, Alan, David Deacon, Montserrat Guillen, and Max Martinelli (2026). “From GLMs to Comprehensive Insurance Pricing: Techniques and Challenges.” Casualty Actuarial Society Monograph No. 16. Full source.
International Association of Insurance Supervisors (2025). “Application Paper on the Supervision of Artificial Intelligence.” IAIS, 2 July 2025, 43 pages. Full source.
European Insurance and Occupational Pensions Authority (2025). “Opinion on Artificial Intelligence Governance and Risk Management.” EIOPA, 6 August 2025. Full source.
National Association of Insurance Commissioners (2023). “Model Bulletin: Use of Artificial Intelligence Systems by Insurers.” NAIC, adopted 4 December 2023, 9 pages. Full source.
scikit-learn developers (2026). “Permutation Feature Importance.” scikit-learn User Guide (online documentation; accessed 15 November 2026). Full source.