RESEARCH DOSSIER
Insurance Model Ensembling and Stacking
Abstract. Insurance prediction joins heterogeneous statistical tasks: estimating claim counts per unit of exposure, conditional claim severity, aggregate insured cost, the probability of a large loss, and complete distributions of uncertain liabilities. Model ensembling combines estimates of the same defined quantity; stacking learns a combination from predictions generated outside the fitting sample of each component. This monograph examines the mathematical foundations, estimation procedures, identification requirements, insurance applications, calibration, generalization behavior, and governance of these methods. It distinguishes bagging and boosting within model families from heterogeneous pooling and cross-validated stacking across generalized linear models, tree ensembles, and other actuarial predictors. The study derives error-covariance conditions for favorable weighting, identifies the distinction between combinations of count intensity and probability of any claim, proves when multiplication of separately averaged frequency and severity differs from averaging pure-premium predictions, and develops a leakage-safe stacking design for grouped, temporally ordered policy and claims data. It further explains why performance improvements depend on the target functional and loss: squared error, Poisson deviance, gamma deviance, a proper distributional score, and financial tail criteria answer different questions. Published actuarial case studies provide evidence of heterogeneous strengths of base learners, while their reported results are explicitly separated from the original analytical derivations here. The principal conclusion is that ensemble quality arises from a valid estimand, complementary predictive errors, independently constructed combination weights, and evaluation on a deployment-relevant distribution. No new named theory or unperformed empirical experiment is asserted.
Keywords: Insurance model ensembling; stacked generalization; stacking; Super Learner; convex combination; prediction averaging; portfolio risk; predictive diversity; generalized linear models; bagging; random forests; gradient boosting; out-of-fold predictions; cross-fitting; model calibration; Poisson deviance; gamma deviance; Tweedie distribution; frequency-severity models; pure premium; claim counts; exposure; insurance pricing; proper scoring rules; distribution shift; cross-validation; model risk governance.
Introduction: Research Object and Evidence Method
A prediction is an assertion about an outcome under a specified information set, population and time horizon. Insurance complicates this ordinary statistical statement because a physical event becomes an insured payment only through coverage terms, exclusions, deductibles, limits, reporting, adjustment, and settlement. Consequently, accurate prediction of an underlying accident rate is not identical to accurate prediction of reported claims, paid losses, ultimate incurred costs or a prospective premium. Each outcome has its own observational denominator, development horizon, and decision relevance. Combining models makes sense only after their outputs have been reconciled to the same target.
This Research Dossier concerns the combination of already constructed prediction procedures. It addresses three analytically distinct operations. First, a collection of fitted predictors can be averaged with predetermined weights. Second, weights or a combination function can be learned using training data designed to preserve out-of-sample prediction status. Third, insurance components such as expected count and conditional severity can be combined into a claim-cost forecast, before or after aggregation across competing model specifications. Each operation has a different estimand and risk of optimistic evaluation.
The seminal stacked generalization study of Wolpert (1992) places predictions from first-level generalizers into a second learning stage trained on held-out responses. Breiman's stacked regressions analysis (1996) studies nonnegative weights estimated from cross-validation predictions. The Super Learner framework of van der Laan, Polley, and Hubbard (2007) uses V-fold cross-validation to determine a weighted combination across candidate learners under a defined loss function. These foundations require careful interpretation: the fitted procedure, the cross-validation population, the class of combination functions, and the loss are all part of the result.
Insurance-specific evidence comes from Statistical Foundations of Actuarial Learning by Wüthrich and Merz (2023), the Casualty Actuarial Society's GLM and pricing monograph by Chalk, Deacon, Guillén, and Martinelli (2026), and the empirical motor study Modelling Motor Insurance Claim Frequency and Severity by Clemente, Guerreiro, and Bravo (2023). The latter comparison reported a gradient-boosting advantage on claim frequency and a GLM advantage on severity for its own study design. That split result is a reason to investigate complementary error structures, not evidence that a particular stacked combination was tested or proved superior by those authors.
The method of the present monograph is a critical synthesis of publisher-verified statistical and actuarial literature, plus original elementary derivations, explicit counterexamples and hypothetical arithmetic illustrations. No insurer microdata have been obtained; no training run, independent empirical replication, confidence interval from observed claims, or comparative new benchmark is claimed. Published experiments are attributed to their original authors and their scope is respected. Accessible publisher abstracts, institutional guidance and open article text were examined, whereas restricted full articles were not represented as fully inspected. Regulatory and actuarial sources are interpreted within their jurisdictions and professional scope, including the ASOP 56 modeling standard and the international IAIS AI supervision paper.
The central research questions are: Which kinds of model diversity make averaging productive? How should combination weights be estimated without reusing fitted responses? What estimands and loss functions are compatible across count, severity, pure-premium and distributional models? How should validation account for temporal, geographic, policyholder and claim-event dependency? Under what conditions does a more flexible meta-learner actually yield a more reliable insurance decision? These questions organize the chapters without instituting a new proprietary theory.
Ensemble Diversity and Insurance Prediction Targets
The predictive object
Let X denote information genuinely available at the decision date; let E denote earned exposure, N the future number of covered claims over the horizon, and S_j the paid or ultimate amount of claim j under the specified policy and accounting definitions. Define total cost Y = sum over j from 1 to N of S_j. A frequency estimator may target lambda(X) = E[N | X,E]/E, whereas a severity estimator may target m(X) = E[S | a defined claim occurs,X]. An aggregate-cost estimator targets mu_Y(X,E) = E[Y | X,E]. A probability-of-claim estimator targets p(X,E) = Pr(N > 0 | X,E). An upper-tail estimator may instead target a quantile of Y or its conditional expected shortfall under a specified convention.
These are different functionals of the underlying conditional distribution. Under a conditional Poisson count model with mean E lambda, Pr(N > 0 | X,E) = 1 - exp(-E lambda). This transformation is nonlinear. Averaging count intensities and converting to probabilities generally produces a different result from converting each count intensity to a probability and then averaging. Both operations can be defensible, but they answer distinct combination questions. Insurance applications must specify whether models represent the same conditional risk with different estimation errors, distinct possible conditional distributions, or separate expert scenarios.
A generalized linear model supplies a structured conditional mean through a link function and a linear predictor, often including log exposure as an offset for counts. A tree-based procedure can represent nonlinear thresholds and interactions. A neural estimator may capture additional representation structure. Their comparison requires identical coverage definitions and prediction-time information. The open-access actuarial treatment of Wüthrich and Merz (2023) provides a useful foundation for distinguishing actuarial learning objectives, while Antonio and Valdez (2012) explain the actuarial setting of prior and posterior risk classification.
Diversity as error structure
Suppose K candidate estimators f_1(X),...,f_K(X) target the same conditional mean mu(X). Write their forecast errors as epsilon_k = f_k(X)-mu(X), evaluated over the deployment distribution and over fitting randomness if relevant. For fixed weights w_k summing to one, the error of the ensemble is sum_k w_k epsilon_k. Its mean squared error can be decomposed into squared weighted bias and the weighted covariance matrix of zero-mean errors. With vector notation,
where b is the vector of mean errors and Sigma is the covariance matrix of centered errors under the evaluation distribution. This is an identity under finite second moments; it is not an estimator of the true error covariance obtained merely from a small training set. A model pair can have excellent individual metrics while sharing correlated residuals, in which case averaging offers limited variance reduction. Two individually modest models can combine beneficially if their errors cancel in relevant regions and their biases do not align.
For two unbiased estimators with error variances v_1,v_2 and covariance c, the variance of weight-w averaging is w^2 v_1 + (1-w)^2 v_2 + 2w(1-w)c. If v_1+v_2-2c > 0, the unconstrained minimum is
When w* lies outside [0,1], the minimum among convex combinations occurs at a boundary. The derivation shows why naive equal weighting is a defensible baseline and sometimes suboptimal, and why seemingly precise covariance-based optimal weights can be extremely unstable when prediction errors are nearly collinear. If base-model errors share a common systematic bias, the variance improvement may coexist with substantial remaining systematic error.
Breiman's bagging predictors study (1996) constructs model diversity by bootstrap perturbations of the sample; Random Forests (Breiman, 2001) additionally introduce randomized tree construction. The article's connection between predictive strength and correlation in a forest illustrates the importance of dependence among component predictors. Friedman (2001) constructs gradient boosting as function optimization, where successive additions are designed in relation to the loss gradient. Bagging and boosting are themselves ensembles, but stacking operates at another architectural level: a random forest or a boosted-tree predictor can be one member of a heterogeneous library containing a GLM.
Structured diversity and shared errors
A library that contains ten minor hyperparameter variations of the same boosted-tree algorithm may have less functional diversity than a small library combining a well-specified exposure-offset GLM, a smooth nonlinear model, and an interaction-sensitive tree method. The number of models is not a mathematical measure of diversification. Relevant diversity must be measured in a held-out population and with respect to the required target and loss.
Insurance creates several recurrent shared-error mechanisms. Claim delays distort the same target used by all learners. Missing coverages or an incorrect exposure denominator enter all training tables. Territorial changes create distribution shift beyond the range of all base models. Deliberate or inadvertent use of post-outcome claim adjustment variables contaminates all predictions, regardless of how sophisticated the stacking algorithm becomes. A combination cannot repair an invalid data-generating definition merely through optimization.
The actuarial importance of coherent model construction is reinforced by CAS Monograph No. 16 (Chalk et al., 2026), which addresses correlated variables, nonlinearities, high-cardinality factors, component blending and practical model validation. A combined estimator should therefore be understood as a defined predictive procedure with provenance, not an informal portfolio of apparently independent numerical opinions.
The unit of independence
Repeated renewals of a policyholder, multiple vehicles of one household, claims from one storm, and development snapshots of the same claim generate dependence. If such linked observations are split across fitting and validation folds, a model can recognize stable customer or event identifiers and appear to generalize to observations that are structurally related to its training sample. The effective unit for independent evaluation might therefore be a policyholder, household, catastrophe event, underwriting year, or chronological block rather than an individual database row.
The assessment begins with an explicit deployment question. Predicting a renewal for a known insured is a different task from predicting a new insured in an unseen territory. The former can legitimately use historical claims available by renewal date; the latter cannot use any future observation from the unseen territory. All base learners and meta-learners must share the information cutoffs required by the intended task. The ASOP 23 data-quality standard (ASB, 2016) and the ASOP 56 modeling guidance (ASB, 2019) establish professional expectations to evaluate data and modeling choices in relation to their intended purpose, although neither specifies a universal machine-learning fold design.
Simple and Weighted Averaging
Arithmetic, convex and unrestricted combinations
For common-scale predictions f_1(x),...,f_K(x), the arithmetic mean is K^{-1} sum_k f_k(x). A fixed convex combination is f_w(x)=sum_k w_k f_k(x), subject to w_k >= 0 and sum_k w_k=1. Convex weighting preserves every point prediction between the smallest and largest component prediction, assuming real-valued outputs on the same scale. The restriction also preserves nonnegative expected loss if every base prediction is nonnegative. This does not imply calibration, unbiasedness or superior performance under every scoring rule.
Unrestricted linear stacking may use negative weights or a nonunit sum and can extrapolate outside the component range. Extrapolation sometimes corrects systematic errors, but it can produce negative fitted expected claims or unstable premiums where each individual base model remains positive. Practical insurance stacking therefore often uses a positive link at the meta-layer, a convex weight constraint, or direct prediction on a transformed scale followed by a carefully assessed retransformation. These choices impose different functional classes and should be compared as distinct candidates.
The historical Stacked Regressions study by Breiman (1996) is particularly important here: it estimates weights from cross-validation predictions with nonnegativity constraints. The later Super Learner algorithm (van der Laan et al., 2007) generalizes the combination strategy by choosing a weighted library under the chosen cross-validated risk. The distinction between precommitted and learned weights is practical as well as statistical. When weights are learned, the fitting procedure acquires another source of estimation uncertainty and another layer of tuning.
What is being averaged?
For count frequency, one may average nonnegative rate estimates lambda_k(x) and apply exposure e afterward: lambda_w(x)=sum_k w_k lambda_k(x), and E[N|x,e] is predicted as e lambda_w(x). Averaging the mean counts e lambda_k is equivalent when each component uses the same exposure e. It is not equivalent when components silently use inconsistent exposure bases. If some models predict per-trip rates, some annualized rates and some policy-period counts, they must be converted to a common denominator and time horizon before combining.
For binary claim occurrence, the relevant object is p_k(x) in [0,1]. A convex combination of such probabilities lies in [0,1], but this geometric property is weaker than probabilistic calibration. As a hypothetical example, models predicting 0.10 and 0.30 for an event combine at 0.20 under equal weights. If the models instead provide Poisson intensities of 0.10 and 0.30 per exposure unit, the equal mean intensity is 0.20 and the associated any-claim probability over one unit is 1-exp(-0.20), approximately 0.1813. Averaging the two individual probabilities produces [1-exp(-0.10)+1-exp(-0.30)]/2, approximately 0.1771. The difference is Jensen's inequality for the concave map lambda to 1-exp(-lambda). It is a direct algebraic consequence of choosing a combination order.
For claim severity, raw amounts and logarithms of predicted amounts should not be treated as interchangeable. Averaging log-scale predictions and exponentiating yields a geometric rather than arithmetic mean, and the former generally differs from the conditional expected severity on the monetary scale. A predictive model of log severity also requires attention to transformation bias and the conditional error distribution. If the target is expected insurer cash outflow, model averaging should be defined on expected paid or ultimate dollar costs after consistent coverage transformations. If the target is an entire conditional distribution, the mixture of distributions is an appropriate distinct construction.
Mixtures of distributions and averages of parameters
Let F_k(y|x) be a predictive cumulative distribution for a common outcome. A linear opinion pool F_w(y|x)=sum_k w_k F_k(y|x) with convex weights is itself a distribution. If each distribution has finite expectation, the mixture mean equals the weighted mean of component means. But the variance of the mixture includes both within-model variance and variability of component means:
This is the law of total variance for a mixture index. The second term recognizes between-model variation. Replacing a mixture by a single distribution with averaged parameters will generally discard or alter this term. In particular, mixing Poisson distributions with distinct intensities is not the same distribution as a Poisson with the average intensity, because the mixture variance is the mean intensity plus the variance of the mixing intensities. This difference matters for tail risk and capital calculations.
A mixture should also not be mistaken for an empirically identified posterior distribution over models. A weight chosen to minimize point-prediction loss is not automatically a Bayesian model probability. A distributional mixture with fitted weights must be evaluated using an appropriate strictly proper score if its asserted output is a predictive distribution. The theoretical basis for that choice is supplied by Gneiting and Raftery (2007), who analyze the properties of scoring rules including the logarithmic score and continuous ranked probability score.
Loss functions define the combination
For an observed target y, ordinary squared error L(y,mu)=(y-mu)^2 elicits the conditional mean under finite second moments. Absolute error targets a conditional median under the usual qualifications. For Poisson counts, the unscaled Poisson deviance per observation is
with the zero-count contribution interpreted by continuity as 2mu. Prediction mu must be positive and the same exposure unit must be used in fitting and scoring. For positive gamma severity y>0 and predicted mean mu>0, a common gamma deviance is
These expressions are losses for different statistical tasks; they should not be pooled into an arbitrary sum without defining a decision criterion and meaningful scale. When data represent claim counts, Poisson deviance measures mean-count fit under a Poisson likelihood criterion, although conditional overdispersion may necessitate additional diagnostic models. Gamma deviance addresses positive severity data, not zero-claim policy outcomes. For mixed zeros and positive continuous costs, a compound Poisson-gamma or Tweedie construction can be used when its assumptions are appropriate.
The foundational Tweedie claims dispersion study of Smyth and Jørgensen (2002) explains the importance of modeling dispersion as well as mean claim cost. Dunn and Smyth's Tweedie density evaluation study (2005) addresses numerical evaluation of exponential-dispersion densities. A Tweedie mean score or likelihood should never be described as a fully validated tail model merely because it aggregates zeros and positive losses. Heavy tail behavior, coverage limits and dependence require additional study.
Learning weights from historical errors
The target weights of a squared-error convex stack solve min_w sum_i a_i (y_i - sum_k w_k z_ik)^2 with nonnegative weights summing to one. Here z_ik are forecasts for observation i generated from base models that did not use its outcome during fitting, and a_i are explicitly justified evaluation weights. This is a constrained quadratic optimization. The matrix Z of forecasts may be nearly rank deficient if base predictions are highly correlated. In such cases, fitted weights can change dramatically after small data perturbations even when ensemble predictions change little. Weight instability is itself informative about identifiability of the component contribution.
A ridge penalty or preference for equal weights can stabilize learned combinations, but the penalty must be trained inside the permitted validation design. Alternative weight estimation using logistic, Poisson or other losses changes the estimand and constraint space. An optimizer that minimizes squared policy-level paid dollars may prioritize a few extreme claims, whereas a count-deviance optimizer may prioritize overall frequency fit; neither automatically minimizes a portfolio's capital quantile or a coverage-adjusted combined ratio.
The ASOP 25 credibility procedures (ASB, 2013) illustrate an adjacent but distinct actuarial practice: blending subject experience with relevant experience based on predictive value and professional judgment. Credibility weighting is an established actuarial framework addressing the reliability of experience; stacking learns a combination of forecasts using a defined held-out loss. They can coexist, for example by using a credibility-adjusted estimate as one candidate learner, but they should not be represented as mathematically equivalent procedures.
Prediction stability and nonnegativity
Convex predictions preserve numerical bounds established by the component point forecasts, which can aid operational review. However, extremely large positive severity estimates are not capped by taking their convex average unless another model supplies lower values, and a fixed combination can still understate unusual tail events if all components were trained on truncated or selectively settled claims. Negative meta-weights can amplify volatility. A positive meta-link may restore nonnegativity but also change whether the ensemble mean is the arithmetic sum of component means.
A suitable evaluation table therefore records for each proposed combination: its target; prediction-time information; scale of its input columns; training loss; constraints and penalty; cross-fitting method; treatment of exposure and zeros; out-of-sample diagnostics; and the economic decision for which the prediction will be used. These are substantive design choices whose omission often produces apparently favorable yet irreproducible ensemble results.
Out-of-Fold Stacking
Two learning levels and the information boundary
The defining operation in stacking is a second-stage learner trained from base-learner forecasts that are generated without fitting the base learner on the target observation. Let D={(x_i,y_i,e_i,g_i,t_i)} be training records including exposures e_i, group identifiers g_i and information cutoff times t_i. Construct a set of K base learning algorithms A_1,...,A_K and partition the available fitting observations into V folds I_1,...,I_V. For each k and v, fit model A_k to D excluding I_v, including within-fold feature preparation and parameter selection that would otherwise learn from target values. Predict for i in I_v and record z_ik=f_{k,-v}(x_i). Every observation now has one out-of-fold prediction per base learner, on the appropriate output scale.
Fit a meta-learner h to (z_i1,...,z_iK; y_i) under a loss and constraints appropriate to the target. Optionally provide genuine prediction-time covariates in a documented passthrough architecture. After meta-fitting, refit the K base procedures on the full development set (using only permitted information), and predict new x by h(f_1,D(x),...,f_K,D(x)). This full-data refitting step distinguishes the deployed base estimators from the fold-specific estimators used to construct the meta-training matrix. It also introduces a small distribution mismatch: meta-fitting saw predictions from models trained on roughly (V-1)/V of development data, while deployment uses models trained on the whole development set. That mismatch should be evaluated rather than denied.
This procedure follows the substantive construction of Wolpert's stacked generalization (1992) and van der Laan et al.'s Super Learner (2007). The implementation guidance for scikit-learn's StackingRegressor similarly distinguishes full-data-fitted base estimators used for prediction from cross-validated base predictions used to fit the final estimator; its warning about a prefit stack trained on the same training data is a useful operational illustration.
The out-of-fold property is not automatic independence
A held-out row may still leak information through preprocessing. A mean target encoding computed once over the complete dataset, a normalization calibrated using test outcomes, a feature-selection procedure performed before folds, and oversampling that places near-duplicate claim records across folds can invalidate the effective out-of-fold boundary. If the target is ultimate claims, development information unavailable at quote time may leak into contemporaneous rating variables, even when the train/test split is formally disjoint.
Leakage-safe implementation therefore requires both observational separation and chronological information controls. Every supervised transform, target encoding, model selection rule and imputation choice that uses target information must be fit inside the fold's training portion. Unsupervised transformations can also leak deployment-distribution knowledge if fitted on a future validation period, particularly when they choose categories, rare-level thresholds or vocabulary using future samples. The analyst should document what information would actually have been available at the intended operational date.
Policies create a second challenge: correlated rows. If the task is predicting outcomes for new policyholders, all historical policy terms for one holder should ordinarily remain within one fold. If the task is predicting next-year costs for returning policyholders, past terms may be legitimate training input, but future renewals cannot precede training cutoffs. When a catastrophe affects many claims, splitting event-linked records among folds can obscure generalization to a future distinct catastrophe. The relevant split unit follows the deployment estimand, not a default random-number generator.
Temporal and grouped cross-fitting
For a chronological forecast, let training window T_v contain only information available before validation interval V_v. Construct z_ik using base models trained on T_v and predict records in V_v. When claims settle after occurrence, a gap between the training cutoff and validation interval may be necessary to avoid post-cutoff label or feature information. An expanding-window evaluation can combine successive historical origins, but the early windows may be small and distribution shift may render later windows more relevant. All candidate learners must use the same permitted historical exposures, not different effective future knowledge.
A V-fold design based on exchangeable observations serves a different inferential target from a rolling-origin design. Published work on cross-validation and its estimand by Bates, Hastie, and Tibshirani (2024) shows that cross-validation error generally estimates a particular training-procedure performance rather than a guarantee about the single fitted model at hand; familiar variance intervals can miscover because errors across folds are dependent. That finding is particularly relevant when estimating a small gain from stacking and reporting it as a stable deployment improvement.
Cross-fitting grouped data is constrained by available fold sizes. If claim severities are rare and extreme, the number of independent high-severity event groups may be much smaller than the number of rows. A small holdout containing a single catastrophe cannot reliably establish tail improvement; a fold with no large event may understate the dispersion in severity. The analyst should describe the distribution of exposure, count, loss, segments, years and catastrophes across folds. Stratifying by outcome can be sensible in some exchangeable problems but may itself use future event labels in a way that is inconsistent with realistic time-series deployment.
Nested selection and an untouched test period
Suppose an analyst tries eight candidate libraries, five meta-learners, four validation schemes and several losses, then picks the combination with the smallest cross-validation score. The chosen result inherits selection optimism even if every candidate used valid out-of-fold training. A genuinely independent final evaluation set, separated before the entire search, is therefore necessary for an unbiased assessment under the intended evaluation distribution; if the test is reused for adjustment, it becomes part of development.
Nested cross-validation can estimate the behavior of the complete model-selection procedure when repeated outer folds are feasible. In each outer training portion, library selection, preprocessing, hyperparameter tuning and construction of the out-of-fold stack occur entirely within that portion. The outer validation portion is then predicted once by the complete fitted stack. This is computationally demanding because each base and meta learner is repeatedly refit, yet the method aligns the reported statistic with the procedure that will be used operationally.
An evaluation design should further distinguish pointwise out-of-sample scores from decision experiments. If predicted expected losses will determine premiums, the evaluation must consider changing customer selection and demand under the resulting prices. Historical claim outcomes under an existing underwriting policy do not automatically identify claims under a radically altered tariff or portfolio. A stack that improves historical mean error can induce adverse economic consequences when implemented outside its observed risk mix.
Why training the meta-learner on fitted predictions fails
Consider an aggressively interpolating base learner that outputs y_i on every development observation, yet has weak prediction capability for future insureds. If its fitted predictions, rather than genuinely held-out predictions, enter the meta-training matrix, the meta-learner observes an apparently perfect predictor and may assign it nearly all weight. The validation error used to train that layer collapses, while the deployment error remains. This is target leakage through the level structure of the stack. Out-of-fold predictions remove this specific direct leakage route, although they do not cure sampling bias, temporal shift, or tuning based on the final test.
The implementation consequence is operationally precise: a stack's provenance should preserve every fold assignment, each base learner's training membership and preprocessing pipeline, the stored out-of-fold prediction matrix, the meta loss and constraints, retraining procedure, independent test period, and final model artifacts. A report that records only the list of algorithms and resulting ensemble weights cannot substantiate whether stacking was scientifically implemented.
Super Learner and the limits of guarantees
Super Learner optimizes a library under a cross-validated risk criterion and is often discussed in relation to asymptotic oracle behavior under specified regularity assumptions. The foundational 2007 Super Learner paper establishes the algorithm and its theoretical setting, not an unconditional finite-sample assertion that any learned mixture dominates all candidates in every dataset, score, subgroup, year or jurisdiction. Library size, estimand definition, losses, fold construction and stability determine what a useful empirical claim can support.
In insurance applications, the best practical meta-model can be a constrained linear combiner rather than a highly flexible neural network. A nonlinear combiner may exploit genuine regional differences in model reliability, but it can also overfit the out-of-fold matrix, particularly when the number of effectively independent risk clusters is small. A model search should therefore compare the simplest credible combination and each individual base learner against the more complex stack in the same untouched evaluation design.
Blended Actuarial Components
Frequency, severity and the conditional mean
For each policy or homogeneous exposure group, the frequency-severity representation uses a claim count N and individual insured losses S_1,...,S_N. The conditional expected aggregate payment satisfies
If, conditional on X, expected claim severity per occurrence has the same value m(X) independently of the realized count and all relevant claim-selection effects, then E[Y|X]=E[N|X]m(X). For exposure e and Poisson frequency with intensity lambda(X), this becomes e lambda(X)m(X). More generally, if the expected severity per occurrence varies with the count or hidden catastrophe state, multiplication of marginal count expectation and an unconditional average severity may miss conditional dependence. An exact frequency-severity decomposition must state the conditioning and definition of the severity estimand.
Insurance practice makes that distinction consequential. Count and severity can be connected through weather severity, inflation, litigation environment, repair complexity, contractual limits and claims-reporting thresholds. A count model trained on all insured policy terms and a severity model trained only on observed claims also describe differently selected samples. Claims with zero payments, closed-without-payment status or delayed reporting may be missing from the severity training data for contractual rather than physical reasons. Combining their forecasts without reconciling populations risks inconsistency even when each model performs well within its own training sample.
The insurance source literature supports separate estimation while documenting its limits. Clemente, Guerreiro, and Bravo (2023) found in their motor frequency-severity comparison that gradient boosting and the standard GLM had different relative performance for count and severity tasks. Smyth and Jørgensen's compound Poisson claims model (2002) supplies a distributional alternative that connects aggregate loss to a count process and gamma severity assumptions. These published studies do not establish that the same choice of component-level stacking weights transfers unchanged to pure-premium loss.
Two aggregation orders
Suppose model k predicts count intensity lambda_k(x) and severity m_k(x), with positive values and common exposure units. A combined expected aggregate loss may be constructed by averaging model-specific pure premiums:
Alternatively, analysts may average frequency and severity separately, using the same weights in this simple illustration, and multiply:
Write E_w for the average over the discrete model index. Then
The covariance here is between model-specific predictions for the same insured, calculated with combination weights. It is conceptually distinct from covariance between actual claim frequency and actual severity. If models with high predicted count intensity also tend to predict high severity, averaging the paired products yields a higher value than multiplying their separate averages. If the models disagree in opposite directions, the sign reverses. Therefore a single notion of 'blending the two components' does not specify one unique premium.
For a transparent hypothetical pair, let lambda_1=0.05, m_1=2,000, lambda_2=0.10, m_2=1,000, each with weight one half and one exposure-year. The two model-specific pure premiums are both 100, so mu_A=100. The separately averaged intensity is 0.075 and severity 1,500, producing mu_B=112.50. The difference -12.50 equals the weighted covariance of the component predictions. The inputs are invented solely to demonstrate the identity; this is not an observed insurance result. It follows that selecting weights based on count deviance and gamma deviance separately does not guarantee optimal combined pure-premium error.
Direct aggregate-cost and component ensembles
A direct pure-premium model estimates E[Y|X,e] on the same aggregate-cost target and can be a candidate alongside a frequency-severity product. The direct model may naturally account for some dependence in aggregate outcomes, while the component approach can exploit more numerous count observations and a claim-level severity sample. Their relative advantages are empirical questions. A defensible heterogeneous library may include a Poisson GLM multiplied by gamma severity GLM, a boosted count model multiplied by another severity estimator, a compound Tweedie aggregate-cost model, and a direct aggregate-cost learner. The predictions enter a meta-layer only after being converted to comparable gross or net insured costs, monetary currency, exposure and horizon.
Wüthrich and Merz's actuarial learning reference (2023) and Chalk et al.'s practical pricing monograph (2026) supply the relevant connection between statistically different learners and the production of insurance rating structures. Actuarial model choice should preserve the available policy terms, claims definitions, censoring or development information, and permissible rating factors. A component blend that uses predictive variables forbidden under the applicable regulatory framework cannot become acceptable through averaging with a lawful component.
Normalization, earned exposure and offsets
Prediction units must remain explicit. A rate lambda of 0.08 claims per policy-year predicts a count of 0.02 on three months of exposure, ignoring other seasonality and assuming a homogeneous intensity. If one base learner returns annual rate 0.08 and another directly predicts count 0.02 for the quarter, averaging these raw outputs produces a quantity with no meaningful unit. Either annualize the count or convert the rate to expected count over the same horizon. When the model permits seasonality, the adjustment should use the integral of time-varying intensity rather than merely multiplying by calendar duration.
Likewise, expected severity per paid claim, per reported claim and per incurred covered event are different conditional measures. Reinsurance recovery, policy limits and deductibles can alter the transformation from gross physical costs to insurer cash flow. If the objective is net claims cost after a treaty, a gross-cost ensemble may be used only in combination with a coherent treatment of net recoveries, attachment limits, reinstatements and coverage definitions. Stacking can be applied at the gross stage or to consistent net predictions, but the evaluation must score the final net estimand if that is what the decision requires.
Reserving and development dependencies
Reserving ensembles differ from new-business pricing ensembles. The target may be outstanding ultimate loss or the future cash flow from claims that already occurred. Candidate forecasts can come from development triangles, granular claim-level models, case-reserve adjusted projections and statistical learning systems. Their training data frequently include overlapping accident years, valuation dates and repeated development states of one claim. Random rowwise folds can place later development of the same claim in the fit and earlier development in validation, creating hindsight leakage.
A valid reserving evaluation must freeze the information known at each historic valuation date and compare predictions with outcomes subsequently revealed by a specified maturity cutoff. The choice of outcome is material: ultimate losses may not yet be fully known, and historical paid claims can be sensitive to inflation, settlement programs, litigation and changes in claims operations. A weight that minimizes early development error may be less appropriate for mature ultimate liability or short-term liquidity. The ASOP 56 guidance on model output validation (ASB, 2019) provides a professional standard for assessing intended use, testing, model risk and controls rather than a universal algorithm for choosing a reserving ensemble.
Product and jurisdictional distinctions
For automobile insurance, observation units may be insured vehicles or policy periods, and ratings must account for earned exposure and policy-level renewal structure. For life insurance, mortality or lapse predictions can include time-varying hazard models and censoring; an ensemble should respect survival probabilities and the dependence of cash flows on contract duration. For health insurance, severity and utilization are affected by benefit design, provider networks and claim adjudication, while regulatory rules may restrict available variables and rating differentiation. For property catastrophe insurance, event clustering and spatial dependency dominate ordinary cross-sectional measures of effective sample size.
The general scientific architecture is transferable across products; exact distributions, legal constraints, validation periods and loss functions are not. International application requires a declared insurance product, jurisdiction, pricing or valuation purpose, and information set. The authorial synthesis developed here treats model combination as an operation over contractually defined conditional outcomes rather than over generic numerical outputs. This interpretation follows from established mathematics and insurance practice; it is not proposed as a new original theory.
Calibration of Combined Models
Calibration belongs to an identified forecast target
A model may rank risks correctly and still misstate expected claims in every group. For a claim-frequency model, a straightforward global calibration check compares observed claims with predicted claims after accounting for exposure and comparable development: sum_i N_i versus sum_i e_i lambda_hat_i. Groupwise calibration examines this relationship within relevant classes, horizons and underwriting segments. A ratio of observed to expected counts near one across a complete portfolio can conceal systematic overpricing or underpricing of subgroups.
Calibration of an event probability p means, in the ideal population, E[Y|p(X)=p]=p for a binary Y, interpreted using appropriate conditioning and finite-sample diagnostics. For aggregate loss mean mu(X), a related conditional calibration condition is E[Y|mu(X)]=mu(X). These conditions are target-specific. Calibrating expected count does not automatically calibrate the probability of at least one claim or the full loss distribution.
Mixture forecasts introduce an additional layer: even if each component is individually calibrated within a historical population, the mixture may not be fully conditionally calibrated under a shifted portfolio. Conversely, a weighted sum can have sound global aggregate calibration while misranking risk or mischaracterizing tails. A combined forecast should be assessed on both aggregate exposure-weighted totals and relevant disaggregations, with uncertainty made explicit.
Proper scores and sharpness
A statistical evaluation should identify what the score rewards. Gneiting and Raftery's proper scoring-rule analysis (2007) distinguishes calibration and sharpness in probabilistic forecasting and provides a framework for honest predictive distributions. The log score assesses density placed on the realized outcome; the continuous ranked probability score compares a complete predictive cumulative distribution with observed outcomes. Neither is interchangeable with point squared error. A model can win on mean prediction and lose on calibrated distributions, especially when rare large losses matter.
For an event probability p and binary outcome y, the Brier loss (p-y)^2 is proper. Its average value supports comparison of probabilistic forecasts, subject to matching sampling, prevalence and evaluation weights. With continuous severity or aggregate insured cost, the selected score should reflect whether a point mean, a quantile, a predictive distribution or an economic utility is the stated object. If insurers care about tail capital, a mean-loss scoring result alone is insufficient to support a capital claim.
A combined distribution can be evaluated with a proper distributional score and tested for probability integral transform behavior when the outcome and regularity conditions allow. Insured costs often contain a probability mass at zero and may include discrete limits or contractual point masses; standard continuous PIT diagnostics must then be adapted for mixed outcomes. Tail assessments must consider exceedance counts, magnitude, clustering and the adequacy of simulated or scenario-based losses. Strong overall distribution scores need not imply precise inference in a very rare tail.
Calibration may require a separate stage
A positive multiplicative recalibration of a mean count forecast can take the form mu_new(x)=c mu_old(x), where c is estimated on a disjoint calibration period to target a specified observed-to-expected ratio. If a single portfolio-wide factor is used, it can correct a global level but leaves relative segment errors unchanged. A richer calibration curve may correct nonlinear bias but can itself overfit small segments. Recalibration belongs within the training or validation design and must never use the untouched test sample whose performance is being reported.
Stacking and calibration are related but separate operations. A meta-model trained under a proper mean loss may improve both fit and calibration, yet this is not guaranteed. A model could be selected for ranking, then calibrated in a subsequent stage. The full pipeline, including recalibration, must be cross-fitted or evaluated externally to avoid optimistic reporting. Model comparison should include the original component forecasts, their direct average, the fitted stack, and any recalibrated versions as different pre-specified candidate procedures.
Exposure-weighted portfolio assessment
Insurance loss prediction often has uneven exposure. If the response is a claim frequency N_i/e_i, giving every short-lived policy equal weight with a full-year policy is not the same estimation criterion as using an exposure-weighted rate loss. For a Poisson model of count N_i with mean e_i lambda_i, a likelihood-based criterion naturally accounts for expected count over exposure. Similarly, evaluating aggregate insured cost per contract versus cost per unit of exposure changes the emphasis of the objective.
Evaluation weights may also reflect decision importance, but they cannot be chosen arbitrarily after looking at the test errors. Premium-weighted, exposure-weighted and policyholder-weighted evaluations represent different social and financial summaries. In a catastrophe-exposed portfolio, weighting by insured value may dominate all small risks even when risk frequency is rare. The chosen operational loss should therefore be explicit, with secondary diagnostics showing which parts of the portfolio drive the advantage.
Heterogeneous improvements and false universal claims
Clemente, Guerreiro, and Bravo (2023) provide a useful empirical counterexample to an unconditional machine-learning dominance claim. Their motor claims comparison reports lower test Poisson deviance for gradient boosting in frequency estimation but lower gamma deviance for the standard GLM in severity estimation. Their insurer data were proprietary and the study's metric and sample determine the result; this manuscript does not re-estimate their figures or extrapolate them to every market. The proper inference is methodological: when alternative estimators differ in strengths across target functionals, model combination may deserve a controlled evaluation. Whether combining them helps remains an independent empirical question.
Predictions can also vary in calibration by risk subgroup, period or claim size. The simple blended model that gives the best overall test score may perform poorly for a small yet financially material region. Consequently, both aggregate and conditional calibration analyses should be recorded, accompanied by practical confidence or uncertainty bounds estimated using methods compatible with dependence. The warning in Bates et al. (2024) concerning conventional cross-validation variance estimates is directly relevant when a claimed difference in ensemble scores is small.
Numerical stability and computational evaluation
Deviance losses involve logarithms, zero outcomes and lower bounds on predicted means. Predictions approaching zero require numerical safeguards consistent with the likelihood and model definition; arbitrary clipping can change test performance and must be documented. Tweedie densities require specialized numerical evaluation, as discussed by Dunn and Smyth (2005). A workflow that silently changes the loss, clamps predictions, drops extreme claims or scales target values differently across learners can create apparently compelling gains that are computational artifacts.
Reproducible evaluation records the model versions, feature transformations, random seeds, precise loss implementation, exposure units, zero conventions, fold memberships, calibration method, model-selection search and versioned data. For insurance this record must also preserve valuation dates, inflation and currency adjustments, contract boundaries, gross/net loss choice and benefit transformations. A forecast assessment is scientifically meaningful only in relation to these declarations.
Generalization, Model Risk, and Governance
The evaluation distribution is part of the scientific claim
Every forecast assessment is conditional on a population and time horizon. A claim that an ensemble has lower expected loss should specify an evaluation distribution Q for future records, outcomes, exposures and dependencies. An empirical score on historical observations sampled from P identifies performance under Q only to the extent that P and Q are aligned or a justified method transports results between them. In insurance, changing coverages, repair inflation, extreme weather, claims management, demographic mix, legal environments, fraud patterns and rate actions can change the joint distribution.
Prediction averaging can reduce estimation variance without correcting a common structural bias. For example, if every base model assumes a stable frequency rate when actual frequency is increasing due to changing hazard exposure, all members may underestimate the same future year. A stack trained to emphasize whichever model was most successful in past years may exacerbate such underprediction if it assigns small weight to a model that is more conservative under the new conditions. This illustrates the limitations of historical fit as evidence of future robustness.
A deployment-oriented validation strategy can include successive time-origin backtests, geographic transfer tests, product-cohort holdouts, policyholder-group holdouts, rare-event stress scenarios, and scenario-based reweighting of material risk classes. These tests answer different questions. A territorial transfer test examines a particular geographic shift, while a rolling-origin backtest addresses temporal stability. Neither alone establishes robust performance against all unobserved sources of change.
Error correlation is state-dependent
The error-covariance formula developed earlier holds relative to a specified distribution. In practice, two estimators may be negatively correlated in ordinary claim periods but simultaneously underpredict catastrophe losses. Unconditional residual covariance can then overstate diversification precisely where an insurer needs reliable estimates. A richer diagnostic calculates error dependence within prespecified regimes such as catastrophe exposure, low-exposure policyholders, high repair-inflation periods or large-loss bands, if sample sizes permit meaningful interpretation.
There is a fundamental distinction between increasing model heterogeneity and securing source diversity. A GLM and random forest trained on the same incomplete loss ledger share missing claims, outlier miscoding, policy-form changes and historical survivorship. Their algorithmic difference does not eliminate common data risk. A well-governed ensemble separates uncertainty due to estimation, candidate specification, observational limitations and future environmental change. Internal variation across model predictions is only one imperfect indicator of total uncertainty.
Model governance is a system-level obligation
An ensemble consists of base algorithms, feature processors, a combination rule, training and update logic, score transformations, and potentially calibration and business overrides. The whole system must be documented, owned, monitored and capable of being independently reproduced. Its component-level interpretability does not automatically establish system-level transparency. For example, assigning weights to recognizable GLMs and trees may seem transparent, but an undisclosed dynamic selection rule can make the final tariff difficult to explain.
For applicable U.S. actuarial professional work, the ASOP 56 modeling standard requires attention to intended purpose, assumptions, model structure, testing, output validation, risk mitigation, governance and documentation. The ASOP 23 data guidance concerns selection and review of data in actuarial assignments. Their professional applicability depends on the actuary and assignment; they are not international statutes applicable to all insurers.
International supervisory material supplies complementary expectations. The IAIS 2025 AI supervision application paper addresses risk-based supervision and proportionality, governance, robustness, transparency, fairness and redress. EIOPA's 2025 AI governance opinion clarifies supervisory interpretation of existing insurance-sector governance principles within its stated scope and does not itself create a universal new statutory requirement. The U.S. National Association of Insurance Commissioners adopted a model AI bulletin in December 2023 for use by states; the model bulletin's adoption by the association should not be confused with identical binding enactment in every U.S. jurisdiction. The relevance of each source to an insurer depends on its jurisdiction, business line, function and legal status.
These materials support a common governance question: can the insurer explain how the combined system was developed, what it estimates, why its results are fit for the decision, what records allow inspection and how errors are detected or corrected? The answer needs source-level evidence, a stable model registry and documented human responsibility. A favorable average score is only one part of that evidence.
Fairness, rating variables and accountability
Ensembling may obscure the influence of protected or restricted characteristics when multiple components encode proxy information. A feature rejected in one learner can reappear through an external score, geographic aggregation, history-based claim indicator or residual-correction layer. Auditing only the final meta-weights misses these pathways. A comprehensive review evaluates legal data eligibility, input provenance, segment-specific error, pricing effects and the downstream decision process.
Fairness has jurisdiction-specific meanings and cannot be reduced to a universal statistical parity criterion. Risk-based pricing, anti-discrimination rules, consumer protection, insurance availability and pooling objectives interact differently across systems and products. The IAIS insurance AI supervision paper frames fairness and redress as part of a broader governance setting, while EIOPA's insurance AI risk opinion discusses explainability, oversight and proportionate controls in the EU supervisory context. These sources justify undertaking legal and social outcome evaluation without pretending that predictive accuracy alone settles the relevant normative issues.
A stack also complicates recourse. If a consumer contests an insurance decision, it may be insufficient to explain that three base models contributed 20, 30 and 50 percent. The complete explanation may require documenting the meaning of the target, the input information, the effect of material variables, the applicable contract and the human or institutional review path. Weight transparency and decision transparency are distinct objectives.
Operational monitoring and the replacement decision
Once deployed, the ensemble should be monitored for data drift, prediction drift, calibration deterioration, score changes, segment-specific anomalies, claim-development differences and operational failures. Monitoring should compare the combined forecast with individual learners and with a simple benchmark. If the stack's performance deteriorates while a stable GLM remains adequate, a documented fallback procedure can reduce operational harm. Conversely, an apparent improvement from replacing the model must be evaluated against costs of governance, explainability, maintenance and impact on insureds.
Recalibration or weight updates alter the complete decision procedure and deserve versioning. A monthly re-estimation schedule may chase noise when claim feedback arrives with long delays; an annual update can be too slow under rapid distribution shift. The cadence depends on product, observation delay and economic materiality. Testing should distinguish a model failure from claims delays, changes in coverage or a portfolio distribution change. A governance record that treats every observed-to-expected deviation as model misspecification can lead to unstable repetitive adjustments.
Analytical Examples and Verification
Example: covariance and model averaging
Consider two hypothetical unbiased predictors of the same conditional mean with equal error variances v_1=v_2=9 and covariance c=3. Equal averaging has variance (9+9+2*3)/4=6, compared with variance 9 for either predictor. If instead c=9, equal averaging has variance 9, reflecting perfectly shared error fluctuation. If c=-3, equal averaging has variance 3. These are direct calculations from the two-predictor covariance identity. They illustrate the dependence mechanism and are not empirical estimates for insurance business.
If the first predictor has variance 4, the second variance 9, and covariance 1, the theoretical unconstrained optimum weight on the first predictor is (9-1)/(4+9-2)=8/11, approximately 0.7273. The resulting variance is (4*9-1)/(4+9-2)=35/11, approximately 3.1818. Equal averaging yields (4+9+2)/4=3.75. A real analyst does not know these true population quantities and must estimate them; the gain from an estimated optimal weight can disappear when covariance estimates are noisy or when next year's error structure differs.
Example: probability transformations
For hypothetical one-year count intensities 0.1 and 0.3 under two conditional Poisson models, equal average intensity is 0.2. The derived any-claim event probability is 1-exp(-0.2), or about 0.181269. The mean of model-specific any-claim probabilities is one half of [(1-exp(-0.1))+(1-exp(-0.3))], or about 0.177172. The first is the event probability under a single Poisson model with averaged intensity; the second is the mixture probability under a 50:50 selection between the two Poisson models. Neither is intrinsically the correct description without stating which statistical object the combination represents.
The probability mixture also illustrates tail implications. If the underlying conditional count distribution is a mixture of Poisson models with distinct rates, its conditional variance exceeds its conditional mean by the variance of the random rate under the mixing weights. Treating the same prediction as a single Poisson with average rate removes this extra dispersion. Even when the two formulations agree on expected counts, they differ in event-count uncertainty.
Example: component-order counterexample
Let two component models, evaluated at one hypothetical insured with one year of exposure, predict respectively (frequency, severity)=(0.05,2,000) and (0.10,1,000). Averaging the two products produces 100 monetary units of expected cost. Multiplying the separately averaged frequency and severity produces 112.50. This is a 12.5 monetary-unit discrepancy created solely by aggregation order and model-specific co-movement. A meta-learner targeting claim counts cannot resolve which pure-premium construction is economically preferred without an appropriate pure-premium evaluation.
If the count and severity models are drawn from two different collections with independently fitted weights, the distinction becomes more pronounced: each component library has its own weighting operator, which need not correspond to a coherent joint distribution of frequency and severity. The appropriate method follows the intended aggregate-cost estimand and available information, not a presumption that component-wise optimization composes automatically into optimal total-cost prediction.
Example: leakage through an interpolating component
Suppose a training sample contains 1,000 observations and a base predictor memorizes all of them. Its in-sample squared error is zero. If a meta-learner is trained on those in-sample predictions, it may assign maximal weight to this component. Yet the component might have expected squared error 100 on unseen observations. If a second stable component has expected out-of-sample squared error 25, the in-sample stack could select the worse predictor. With genuinely out-of-fold predictions, the memorizing component is instead assessed on data it did not fit; the original inflated advantage disappears to the extent that the folds represent the operational prediction task.
The numbers in this example are hypothetical by construction; they demonstrate the direction of a failure mode rather than the performance of any specific machine-learning algorithm. Out-of-fold training solves the in-sample comparison problem, while further controls remain necessary for hyperparameter search, linked policyholders, chronological cutoffs and independent test evaluation.
Example: a minimal leakage-safe study protocol
A reproducible future insurer study could use a portfolio containing policy-period records with stable customer identifiers, contract details, earned exposure and reported or ultimate loss outcomes. The investigator would first define the valuation date, outcome development horizon and target. A chronological final test year would then be reserved and untouched. Within earlier development years, grouped or rolling-origin folds would be constructed and the complete training transforms fitted inside each fold. Every candidate learner would generate out-of-fold predictions for each eligible fold. The meta-combiner would be fit to those predictions under a prespecified actuarial loss, and all tuning would occur inside the development set.
For an expected-count target, the investigator could compare an exposure-offset Poisson GLM, a random forest or boosted-tree procedure designed for the target, a fixed convex average, and a nonnegative cross-validated stack. For expected aggregate cost, additional appropriately linked count-severity and direct pure-premium candidates could be considered. The final holdout would report mean count deviance or aggregate-cost loss according to the chosen estimand, observed-to-expected calibration, relevant segment and year diagnostics, and explicit uncertainty from linked observations. Computational cost, interpretability, permitted variables, claims delay and the effects on premiums would be assessed separately.
This is a research protocol, not an experiment conducted for the present dossier. It specifies the evidence needed to substantiate a future claim that a particular ensemble improves insurance forecasts. The open questions include how effective sample size changes with event clusters, whether dynamic weights remain stable across market shifts, and whether gains in expected-loss accuracy translate into better long-horizon insured outcomes.
Internal mathematical and bibliographic checks
The identities in the monograph can be independently checked from their defining assumptions. The optimal two-predictor weight follows by differentiating the quadratic w^2 v_1+(1-w)^2 v_2+2w(1-w)c and setting its derivative to zero. The variance of a predictive mixture follows by conditioning on the discrete component index. The discrepancy between paired-product and product-of-averages follows by expanding the weighted covariance. The Poisson probability comparison follows by applying the nonlinear map 1-exp(-lambda) before or after averaging. Their proofs are elementary and conditional, with finite second moments where needed and explicitly stated model assumptions.
Publication-based claims were separated from these derived identities. The actuarial motor study is cited for its reported within-study frequency/severity comparison, the stacking papers for their published algorithms, the scoring-rule paper for its theory, and official insurance guidance for its jurisdiction-scoped expectations. No unverified new performance percentages, new insurance datasets, or reconstructed significance tests are introduced. All hypothetical arithmetic is marked accordingly.
Research Synthesis and Implications
The combined literature establishes that a meaningful insurance ensemble is jointly determined by a target, library, training design, combination rule, loss and evaluation population. Its effectiveness arises when complementary component errors outweigh weight-estimation noise and shared structural error, measured under a deployment-relevant test. These conclusions follow from the error-covariance identity and from the design of stacked generalization; they do not require an original named doctrine.
Several implications deserve particular emphasis. First, diversification of algorithms is less informative than diversification of the errors relevant to a stated target. Second, prediction-time validity must be maintained across both levels of the stack; ordinary k-fold labels do not establish this property when rows share policyholders, events, development periods or future information. Third, the choice of where to combine estimates changes the mathematical object: averaging count rates differs from averaging event probabilities; averaging paired pure premiums differs from multiplying averaged components; mixing distributions differs from averaging their parameters. Fourth, a scoring rule identifies the question answered by model comparison. An apparent improvement in mean cost forecast cannot be assumed to imply a superior claim-frequency model, calibrated full distribution or accurate capital tail.
A further insurance-specific implication concerns governance. The ensemble is a system consisting of all constituent models, transforms and decision rules. Supervisory and actuarial expectations about documentation, transparency, data quality, validation and fairness apply to that system within the relevant professional or jurisdictional scope. Combining models may improve predictive performance while increasing dependencies and organizational complexity. The practical value of the combination is therefore an evidence-based decision incorporating reproducibility and institutional consequences.
The research agenda arising from this synthesis includes temporally adaptive but stable weighting under long-tailed settlement delays; dynamic mixtures that preserve coherent pure-premium and count-distribution identities; event-cluster-aware estimation of effective out-of-sample precision; evaluation of the economic consequences of price-induced selection; and methods for auditing changes in fairness or coverage access after ensemble updates. These are research questions and possible studies, not introduced theories or promises of performance.
Conclusion
Insurance model ensembling and stacking provide established methods for integrating predictions from alternative statistical procedures, including generalized linear models, random forests, boosted trees and actuarial frequency-severity structures. Averaging can reduce error variance when component errors are imperfectly correlated, and cross-validated stacking can learn a combination from out-of-fold predictions. Both benefits depend on correctly defined targets and appropriately independent evaluation.
The analysis demonstrates that insurance prediction exposes several especially important conditions: exposure normalization and contractual outcome definitions must be consistent; observational and claims-development dependence must determine validation splits; the full meta-learning and tuning procedure requires evaluation that is independent of its own selection; loss functions must match counts, severity, pure premium or distributions; and calibration and governance must be assessed after combination as well as before it. Elementary covariance, mixture and frequency-severity identities establish why an ensemble can have different predictions and risk implications depending on the order and scale of its operations.
The substantive result is a defensible research interpretation of model combination as a fully specified predictive and institutional process. Its value must be established against individual learners, simple averages, and credible actuarial benchmarks within the same prospective task. The original contribution of this dossier is the integrated analytical treatment of error dependence, two-stage validation, actuarial aggregation order, loss-specific calibration, and jurisdiction-sensitive model governance. It makes no claim to an original statistical theory, externally replicated insurer benchmark, universal predictive superiority or public release.
References
Actuarial Standards Board. (2013). Actuarial Standard of Practice No. 25: Credibility Procedures (revised December 2013; effective May 1, 2014). ASOP 25.
Actuarial Standards Board. (2016). Actuarial Standard of Practice No. 23: Data Quality (revised December 2016). ASOP 23.
Actuarial Standards Board. (2019). Actuarial Standard of Practice No. 56: Modeling (adopted December 2019; effective October 1, 2020). ASOP 56.
Antonio, K., & Valdez, E. A. (2012). Statistical concepts of a priori and a posteriori risk classification in insurance. AStA Advances in Statistical Analysis, 96, 187–224. https://doi.org/10.1007/s10182-011-0152-7.
Bates, S., Hastie, T., & Tibshirani, R. (2024). Cross-validation: What does it estimate and how well does it do it? Journal of the American Statistical Association, 119(546), 1434–1445. https://doi.org/10.1080/01621459.2023.2197686.
Breiman, L. (1996a). Bagging predictors. Machine Learning, 24, 123–140. https://doi.org/10.1007/BF00058655.
Breiman, L. (1996b). Stacked regressions. Machine Learning, 24, 49–64. https://doi.org/10.1023/A:1018046112532.
Breiman, L. (2001). Random forests. Machine Learning, 45, 5–32. https://doi.org/10.1023/A:1010933404324.
Chalk, A., Deacon, D., Guillen, M., & Martinelli, M. (2026). From GLMs to Comprehensive Insurance Pricing: Techniques and Challenges. Casualty Actuarial Society, CAS Monograph No. 16. CAS monograph.
Clemente, C., Guerreiro, G. R., & Bravo, J. M. (2023). Modelling motor insurance claim frequency and severity using gradient boosting. Risks, 11(9), 163. https://doi.org/10.3390/risks11090163.
Dunn, P. K., & Smyth, G. K. (2005). Series evaluation of Tweedie exponential dispersion model densities. Statistics and Computing, 15, 267–280. https://doi.org/10.1007/s11222-005-4070-y.
European Insurance and Occupational Pensions Authority. (2025). Opinion on Artificial Intelligence Governance and Risk Management (August 6). Official EIOPA opinion.
Friedman, J. H. (2001). Greedy function approximation: A gradient boosting machine. The Annals of Statistics, 29(5), 1189–1232. https://doi.org/10.1214/aos/1013203451.
Gneiting, T., & Raftery, A. E. (2007). Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association, 102(477), 359–378. https://doi.org/10.1198/016214506000001437.
International Association of Insurance Supervisors. (2025). Application Paper on the Supervision of Artificial Intelligence (July 2). Official IAIS publication.
National Association of Insurance Commissioners. (2023). NAIC Members Approve Model Bulletin on Use of AI by Insurers (December 4). Official NAIC announcement.
Scikit-learn developers. (Documentation accessed November 14, 2026). StackingRegressor: Stacked Generalization for Regression. Official API documentation.
Smyth, G. K., & Jørgensen, B. (2002). Fitting Tweedie's compound Poisson model to insurance claims data: Dispersion modelling. ASTIN Bulletin: The Journal of the International Actuarial Association, 32(1), 143–157. https://doi.org/10.2143/AST.32.1.1020.
van der Laan, M. J., Polley, E. C., & Hubbard, A. E. (2007). Super Learner. Statistical Applications in Genetics and Molecular Biology, 6(1), Article 25. https://doi.org/10.2202/1544-6115.1309.
Wolpert, D. H. (1992). Stacked generalization. Neural Networks, 5(2), 241–259. https://doi.org/10.1016/S0893-6080%2805%2980023-1.
Wüthrich, M. V., & Merz, M. (2023). Statistical Foundations of Actuarial Learning and its Applications. Springer Actuarial. Springer, Cham. https://doi.org/10.1007/978-3-031-12409-9.