Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Decision Trees and Random Forests in Insurance Prediction

Abstract. Decision trees and random forests transform heterogeneous insurance records into locally adaptive predictions by recursively partitioning risk covariates and aggregating predictions from randomized trees. Their actuarial value depends on specifying the insured event, observation window, exposure denominator, settlement maturity, predictive loss function, and decision for which a forecast will be used. This monograph analyzes the mathematical construction of classification and regression trees, bootstrap aggregation, random-subspace forests, insurance-specific frequency and severity targets, loss-based evaluation, interpretation of variable importance, and external validity. It distinguishes empirical evidence from mathematical implications and illustrative calculations. The central analytic finding is that tree-based flexibility can help identify risk interactions neglected by restrictive parametric specifications, while the validity of a premium or underwriting decision still turns on data-generating processes, tail coverage, calibration, fairness constraints, and independent validation. Insurance applications are evaluated against credible generalized linear and distributional benchmarks. Motor-insurance evidence, cross-product transfer limits, correlated exposures, right censoring, correlated features, claims leakage, and post-deployment drift are examined. This research is a critical synthesis of established literature and explicit actuarial reasoning; it proposes no new named theory, new estimator, or empirical performance claim.

Keywords: Decision trees; classification and regression trees; CART; bagging; random forests; insurance prediction; claim frequency; claim severity; pure premium; exposure; out-of-bag estimation; variable importance; insurance ratemaking; motor insurance; calibration; predictive validation; distribution shift.

Introduction

Insurance prediction concerns future events under contracts that specify coverage, limits, deductibles, claims conditions and the financial meaning of losses. A model may forecast whether a claim occurs, the count of covered events per exposure period, the cost of an individual reported claim, the ultimate amount of an incurred event, the aggregate loss cost of a policy, or the time until a claim settles. These response variables are different scientific objects. Even when the same records and algorithm are used, they lead to different likelihoods, evaluation metrics and deployment decisions. Model selection therefore begins with the estimand, the population, and the date at which the prediction is made.

A recursive tree is attractive because it expresses interactions directly in conditional partitions. A vehicle characteristic may predict little on average but considerably more among a particular age, territory or mileage segment. A tree can represent that conditional pattern without the analyst explicitly adding every cross-product. The foundational CART analysis of risk partitions by Breiman et al. (1984) supplies the statistical grammar; Breiman's later random forests formulation (2001) formalizes the use of randomized tree ensembles.

This flexibility operates within actuarial constraints. A partition with attractive retrospective fit may have few earned policy-years, may mix immature and developed claims, and may exploit information recorded only after the decision date. An accurate claim/no-claim classifier can still produce a poorly calibrated pure premium because the cost conditional on a claim is misspecified. An insurer also makes decisions under supervisory, consumer, reserving and governance requirements that a training loss does not encode automatically.

The study combines primary algorithmic papers, original actuarial case studies, published methodological critiques, official software documentation and an actuarial ratemaking reference. Primary-source scope is indicated in the evidence record. Mathematical statements are derived from definitions or established results; the numerical illustrations later in the monograph are expressly hypothetical and do not measure the performance of an insurer or an identified commercial model. International application is addressed through transferability and institutional heterogeneity, rather than by assuming one jurisdiction's tariff rules apply globally.

Scientific questions and evidentiary distinctions

The first question concerns what trees learn about conditional insurance risk and which data conditions make those patterns reproducible. The second examines when bagging and feature randomness improve generalization, including their implications for interpretability and stability. The third concerns what is established by published insurance examples and what remains untested in any particular portfolio. Throughout, a theorem about an algorithm, a benchmark result on one dataset, a prospective actuarial application, and an operational control are treated as distinct forms of knowledge.

The actuarial question is prospective. The CAS ratemaking statement links property-casualty rates to anticipated costs of transferred risk, while explicitly having a particular U.S. reference status. It is useful for clarifying why a training-set score cannot alone establish an adequate future rate. Legal adoption, fairness constraints and documentation duties must be established for the actual jurisdiction and product.

Tree-Based Risk Partitions

Observation, exposure and target construction

Let an insured unit i have covariates X_i measured at an underwriting cut-off, exposure t_i in insured-years, count N_i of covered occurrences during that period, and ultimate covered aggregate amount Y_i. For a prospective pricing forecast, natural targets include frequency lambda(x) = E[N_i | X_i=x,t_i]/t_i, expected severity m(x) per covered occurrence and pure premium q(x) = E[Y_i | X_i=x,t_i]/t_i. The duration t_i is essential: one claim during a month and one claim during a full year represent very different observed rates.

When conditional claim severities have a common mean m(x) per event and claim counting and severity mechanisms meet the necessary conditional expectations, E[Y_i | x,t_i] = E[N_i | x,t_i] m(x). This expectation identity need not assume fully independent severities if the event-sum conditional expectation is properly specified. It fails as a shortcut if the severity sample refers only to paid or closed claims while the count target includes reported or incurred events at a different valuation date. Deductibles, limits, expense loadings and payment timing can also break an apparently simple equivalence between recorded claims and prospective covered cost.

The grain of the dataset must be explicit. Policy-level annual records, policy-period transactions, vehicle-months, insured objects, individual claims and individual payments are not exchangeable units. A claim-severity training set has an event-conditioned selection mechanism and often multiple events per contract. Multiple years for one insured and multiple policies for one catastrophe are correlated. A conventional IID analysis that ignores this structure understates uncertainty and can make validation optimistic. The data dictionary should identify the contractual period, earned exposure, valuation date, amount basis, revision sequence and censoring rules.

Recursive binary partitions

A regression tree constructs regions R_1,...,R_K by selecting splits that reduce a node's weighted empirical loss. For squared-error regression a leaf prediction is the weighted within-leaf mean: a_R = sum(i in R) w_i y_i / sum(i in R) w_i, provided the denominator is positive. The chosen split minimizes the combined weighted residual sum of squares across the two child nodes. For binary classification, a split can maximize reduction in Gini impurity or entropy; leaf class frequencies may be used as preliminary event probability estimates. Both procedures are dependent on how class weights and observation weights are specified.

The same y variable can be framed differently: an observed annual count N, a rate N/t, or an event indicator 1(N>0). For rare claims and short periods, an indicator may approximate frequency informally, but it discards multiple-claim counts. A squared-error tree fitted to raw counts without accounting for variable exposure can preferentially sort policy duration rather than underlying intensity. Modeling the rate N/t with exposure weights, using a count-appropriate split objective, or fitting an exposure-aware statistical benchmark addresses different aspects of that problem; the weighting and objective should be documented explicitly. A mechanical replacement of a Poisson offset by an extra duration covariate is not generally equivalent.

CART's original treatment of pruning and prediction (Breiman et al., 1984) also shows why terminal-node complexity belongs to model design. Deep partitions isolate local combinations and expose small samples to noise. Cost-complexity pruning balances training improvement against leaf proliferation, whereas constraints on depth, minimum leaf exposure and minimum claim count act during construction. These alternatives may create substantially different extrapolations at the margins.

From within-leaf fit to future risk

A tree defines piecewise-constant predictions. Two policies on opposite sides of a split threshold can receive materially different fitted outcomes even when their covariates are almost equal. The estimated discontinuity reflects the partition, not proof of an underlying physical discontinuity in insured risk. Conversely, when the true conditional function is smooth, large leaves introduce approximation error. The choice of tree complexity involves a bias-variance tradeoff and reflects sample size, noise level, target distribution and the decision value of local accuracy.

A single large loss can dominate squared-error reduction in severity models. The appropriate response may be a robust objective, a separate large-loss layer, conditional quantile analysis, or a distributional model rather than deleting legitimate claims. The Tweedie density study by Dunn and Smyth (2005) illustrates why distributional modeling involves more than an ordinary squared-error fit, particularly when aggregate insurance cost combines zeros and continuous positive amounts. Choice of objective should correspond to the estimand and risk measure.

Operational feature availability

Only information available at the prescribed prediction time may enter a valid deployment model. A motor underwriting model cannot use a future at-fault determination, later litigation costs, final settlement amount or a claims-handling code created after policy inception. Even seemingly innocuous features, such as the number of document revisions or a missingness flag, may disclose downstream processing outcomes. A temporal feature inventory should record first-observed time, revision time, processing owner and whether availability depends on the future event being forecast.

Record linkage matters equally. If accident attributes from the same event appear in training and test with altered identifiers, a random split permits cross-event memorization. For frequency prediction the unit of independence may be a driver, vehicle, contract, household or fleet, depending on deployment. For catastrophe loss the event grouping may span insurers and geographic zones. The correct sampling unit follows the prediction problem, not the database's convenient primary key.

Formal exposure weighting, policy aggregation, and split choice

An exposure-adjusted empirical tree estimator can be written as the solution to an explicitly weighted mean-loss problem. Suppose y_i=N_i/t_i is an observed claim rate and the squared-error weight is w_i=t_i. Then the fitted rate in a leaf R is sum(i in R) t_i(N_i/t_i)/sum(i in R)t_i = sum(i in R)N_i/sum(i in R)t_i. This exact algebraic identity shows why exposure weighting returns the crude aggregate frequency for that node, provided all t_i are positive and the claim counts are coherently defined. It says nothing, by itself, about how effective the tree's split selection will be under a Poisson count variance or in the presence of temporal dependence.

A Poisson exposure model instead posits N_i conditional on x_i and t_i with mean t_i lambda(x_i), and its likelihood weights claim counts and exposure in a particular way. A regression tree trained on per-exposure rates with squared loss may select different splits from a tree optimized using a Poisson deviance. The estimation objective must therefore remain visible in documentation: the same tabular predictors can generate distinct partitions even when their leaf means eventually use equivalent exposure-normalized ratios. The distinction is essential when small exposures produce very volatile annualized rates.

For a direct aggregate-cost tree, an analogous leaf mean sum Y_i/sum t_i is the exposure-weighted average of Y_i/t_i. A separate conditional claim-severity tree usually has claims rather than policy-years as observations, and its denominator is the number or weight of eligible positive-cost events. It is erroneous to carry the denominator of the count model into the severity model without defining the event unit. Repeated payments for a single accident can likewise multiply the apparent sample size unless consolidated to the intended individual claim or ultimate-event observation.

A practical split audit records, for the parent and each child leaf, policy count, distinct insured count, earned exposure, event count, positive-event count, aggregate covered loss, loss maturity, and the single-largest-loss share. It also records the proposed split threshold and weighted loss improvement. This makes visible whether a candidate division represents a reproducible conditional interaction or is mainly a response to one unusually large incurred case. The statistical foundations of classification and regression tree pruning (Breiman et al., 1984) support comparing complexity against generalization rather than treating all training improvements as information.

Sampling selection at underwriting and claims stages

Training records are generated by the insurer's history of decisions. Earlier underwriting may have excluded certain industries, hazards, buildings or vehicles. A new model fitted only on accepted risks consequently estimates experience within that accepted distribution. If it is offered to a broader applicant population, it extrapolates beyond observed coverage and may also misinterpret the relationship between prices and risk selection. A portfolio-scale approval policy can create its own sample-selection mechanism.

Claims records add further selection. Some insured events are never reported, small claims may be absorbed by deductibles, reopened cases can create duplicate rows, and fraud investigations may change observed labels. For underwriting prediction, these are properties of the outcome-measurement process, even when they arise after the original risk period. A tree can discover them statistically, but it cannot identify the unobserved counterfactual claims that would have occurred under another coverage, reporting or settlement regime without additional identifying assumptions.

A coherent causal interpretation therefore requires a separate design and evidence base. The causal forest analysis by Wager and Athey (2018) treats treatment-effect heterogeneity under potential-outcome assumptions; its estimand is not the same as a standard insurer's conditional probability of a future claim. When defining the underwriting target, the analyst should decide which policy decision and insured population are actually being forecast before borrowing claims data from another stage of the insurance process.

Bootstrap Aggregation

Sampling with replacement and aggregation

Breiman's bagging predictors paper (1996) proposed building multiple predictors from bootstrap resamples and aggregating their output. Let D denote a learning sample and D_b* an independently drawn bootstrap sample of size n from D. If T(D_b*,x) is a numerical tree prediction, the bagged prediction is f_B(x) = (1/B) sum(b=1 to B) T(D_b*,x). For a class label, voting aggregates categorical decisions; averaged class probabilities, when available, require calibration assessment rather than being assumed to be calibrated.

Sampling the same n records with replacement leaves approximately a fraction (1-1/n)^n of the original records unseen by any one tree, approaching e^(-1) or 36.8 percent as n grows. The complementary fraction of distinct observed records approaches 63.2 percent. These are expectations under row bootstrap sampling and not a count of unique insureds, accident events, or geographical clusters. In portfolio data a resample of rows can include correlated observations belonging to identical risk entities.

The mechanism of bagging is especially relevant to unstable estimators. Small data perturbations can alter early tree splits and create different downstream partitions. Averaging these predictions reduces variation, albeit subject to correlation among trees. Breiman's 1996 paper reports improvements across certain empirical and simulated tasks, not a universal dominance guarantee. Stable low-variance estimators may gain less, and aggregation can introduce its own bias for certain targets.

Variance and correlated predictions

For a simplified exchangeable collection of B tree predictions at a fixed x, each with variance sigma^2 and pairwise correlation rho, the variance of their average is sigma^2 [rho + (1-rho)/B]. This is an algebraic identity under equal variance and equal pairwise correlation, not an empirical insurance parameter estimate. As B grows, the independent component vanishes while the correlated component remains. Randomization can be useful precisely when it decreases dependence enough to compensate for any weakening of individual trees.

In insurance, large shared events make the analogy more consequential: many tree fits may differ while all reflect the same unmodeled catastrophe event or retrospective reporting convention. Increasing the number of trees does not create independent catastrophic experience. The uncertainty represented by bootstrap perturbations is narrower than uncertainty about future climate, inflation, legal coverage or a shifted underwriting population.

Out-of-bag estimation and its limits

For each observation, an out-of-bag prediction uses only the subset of trees trained without that observation. This is convenient internal validation when the sampling units behave sufficiently like independent draws from the relevant population. OOB can estimate predictive error at low computational cost and can support exploratory variable importance. It is not automatically an honest estimate for a later policy year, another country, a new insurer, or a portfolio with duplicated claim events.

If a motor portfolio contains multiple exposure rows from a driver, withholding just one row while other rows remain in a tree does not reproduce prediction for an entirely unseen driver. Similarly, development payments on one claim cannot be separated across training and test without contaminating settlement forecasts. For an out-of-time prediction problem, OOB is a model-development diagnostic; a separate chronological validation is decisive for prospective use.

Resampling design for insurance populations

Row bootstrap, policyholder bootstrap, vehicle bootstrap, accident-event bootstrap and calendar-block bootstrap answer different statistical questions. Cluster resampling can preserve within-unit dependencies, although it changes the distribution of unique observations and the effective sample size. Sampling by calendar block addresses temporal co-movement but has limited meaning when only a few independent periods exist. Sampling of weather and catastrophe exposures requires a separate view of hazard intensity, spatial dependency and contractual aggregation.

Weighted resampling also matters. Oversampling rare claims can make a model learn a useful decision boundary, but it changes the effective class prior unless corrected in a probabilistic evaluation. Severity models conditioned on positive claims are, by design, trained on a selected sample; their deployment must account for the probability of entering that sample. Each design should preserve a clear distinction between frequency weighting, exposure weighting, cost weighting and operational loss weighting.

Dependence, bootstrap distributions, and uncertainty statements

Bootstrap resampling supplies a distribution over fitted predictors conditional on the recorded sample. It does not directly describe every source of uncertainty in a future insurance portfolio. A year with a large windstorm can appear repeatedly in many resamples, and a year without such a storm cannot generate unobserved catastrophe scenarios. Replicated exposure rows also share underwriting and claims context. The relevant probability model for total insurer loss depends on these common shocks, and a random row bootstrap can hide them.

This distinction can be made algebraically. For B equally variable tree errors with pairwise correlation rho, the variance of the average is sigma^2(rho + (1-rho)/B). A higher positive rho produces a larger irreducible component when B increases. In a setting of common reporting inflation, duplicated source fields or strong shared hazards, rho may be high even though individual tree structures look different. The formula is a stylized exchangeable covariance result; actual forest trees do not generally have exactly constant correlation, and no numeric rho is assigned to insurance in this monograph.

What an analyst calls 'uncertainty' must be declared explicitly. There is Monte Carlo variation from using a finite number of randomized trees. There is sampling uncertainty arising from the observed policy and event cohort. There is label uncertainty from incomplete claim development. There is distribution shift from new policies, geography or future inflation. There is contractual uncertainty about claims definitions and covered-loss valuation. Different procedures measure different components; adding trees mainly addresses Monte Carlo variability, and even that effect depends on how the ensemble is built.

The out-of-bag error is particularly prone to overinterpretation. Each omitted row is predicted by trees trained without that row, yet the same accident event, insured enterprise or catastrophe may have contributed other correlated rows. A genuinely new event or new region cannot be emulated by merely withholding one member of a connected event group. Moreover, repeated selection of tuning parameters against the same OOB score weakens its role as a neutral final assessment. Internal OOB curves should be labeled 'development diagnostics' unless a research design justifies stronger population inference.

The bagging predictors study (Breiman, 1996) provides the general variance-stabilization motivation, while the random forests study (Breiman, 2001) makes ensemble correlation a central theoretical concern. The sound actuarial extension is a hierarchy of tests: resampling diagnostics for training stability, grouped holdout evaluation for cross-entity generalization, chronological testing for prospective risk, and stress scenarios for outcomes outside ordinary historical support. These evaluations answer related but distinct questions.

Random Forest Construction

Randomness in features as well as observations

A random forest extends tree aggregation by adding randomness in split-variable selection. At each candidate node, an implementation samples a subset of predictors and searches for a split within that subset. This decorrelates trees that would otherwise repeatedly select one powerful predictor at the root. Breiman (2001) analyzes the relation between tree strength and forest correlation; Liaw and Wiener (2002) describe the corresponding randomForest software implementation.

For a regression forest, f_RF(x) usually averages predictions from B randomized trees. For classification, the aggregation may be voting or class-probability averaging depending on the estimator and use case. These are constructional conventions rather than actuarial promises. Software versions, treatment of missing values, class weighting, resampling, and splitting criteria can change predictions materially while preserving the label 'random forest.'

Tuning dimensions and interactions

The central controls include the number of trees, candidate variables at a split, maximum depth, minimum leaf size, split loss, sampling fraction, bootstrap choice, feature encoding, and randomness seed. Probst, Wright, and Boulesteix (2019) systematically examine these dimensions and tuning strategies. More trees usually stabilize Monte Carlo variation of an ensemble but do not correct leakage, bad labels, missing hazard periods, or model-class misspecification.

Minimum leaf size is especially important for claims: a leaf holding hundreds of zero-claim policies but almost no actual claims can give a superficially stable frequency estimate with wide event-rate uncertainty. An analogous severity leaf may contain numerous small claims and one exceptional settlement that controls its mean. For reporting, leaf records, total exposure, event counts, paid-to-ultimate maturity and large-loss concentration are more informative than leaf record count alone.

A tuning search must be nested inside the development data. If one chooses hyperparameters by observing the final holdout score, the holdout becomes training information. Repeated experimentation can overfit a fixed test partition without any gradient descent. A defensible procedure specifies the candidate families, tuning budget, resampling units, selection criterion and untouched evaluation window before reporting comparative results.

Algorithm and insurance loss are separate specifications

An implementation may support squared-error, absolute-error or Poisson-deviance splitting. The scikit-learn RandomForestRegressor reference documents these software-dependent choices and minimum leaf controls. However, availability of a Poisson criterion does not by itself solve exposure offsets, claims-development truncation, or exceedance severity. The end-to-end specification must state whether a prediction represents conditional annual count, raw count over the observed exposure, cost per policy, cost per earned-year or an event-time hazard.

Monotonicity and continuity constraints may also matter in rating. Ordinary unconstrained trees can create counterintuitive rate relativities when local data are sparse; smooth additive models or constrained approaches may be easier to explain for some variables. Since rating coherence and competitive advantage are decision criteria, the best statistical score need not be the only admissible design.

Correlation, margins and extrapolation

Forests average observations associated with terminal regions. This grants local adaptivity but weakens extrapolation beyond the observed support of covariates. When accident repair costs rise beyond the historical range, a conventional forest trained on earlier amounts cannot reliably extrapolate an inflation trend simply by averaging historical leaf responses. Explicit inflation adjustments, forecasted exposure mix and exposure to new hazards may be required.

Correlation among features presents a second challenge. If territory, garage location and income-related proxies encode overlapping information, split selection may distribute predictive credit across substitutes. The resulting variable importance pattern depends on which features were made available and how correlated predictors compete for splits. Conditional predictive association must be kept distinct from an intrinsic characteristic of the insured and from a causal interpretation.

Finite-ensemble behavior and structural generalization

A forest is often called nonparametric because it does not impose a fixed additive linear predictor. Yet its finite implementation is highly structured. Split rules select axis-aligned thresholds, minimum leaf sizes determine local pooling, candidate-feature restrictions affect competition among predictors, and averaging attenuates extreme local estimates. These construction choices represent a particular inductive bias. They enable interactions, but their existence does not guarantee the model will discover the important ones in sparse claims experience.

Two problems can occur simultaneously. A dominant predictor may cause many bootstrap trees to split similarly, leaving high forest correlation. A truly important interaction involving a rare risk may be available in only a small number of terminal nodes and appear unstable across samples. Reducing candidate features per split can help diversify trees but may also weaken the ability of an individual tree to find a rare, informative split. The methodological overview by Probst, Wright, and Boulesteix (2019) supports treating tuning as a coupled design problem instead of assigning a universal best value to any single parameter.

A forest can also have difficult boundary behavior. Its standard regression predictions are constructed from observed responses in leaves, so a conventional averaging ensemble usually remains tied to the range of its training targets. Suppose claim repair inflation causes average settled costs to double in the following period while covariates remain unchanged. Historical response averaging alone will not establish the new level. Explicit nominal-to-constant-currency normalization, credible future inflation assumptions, or recalibration to recent developed claims are necessary components of the insurance forecast.

The model's effective complexity is not the number of trees by itself. A thousand very shallow trees may be much more constrained than a hundred almost unpruned trees. Strong dependence among trees can make a large ensemble behave like a far smaller set of genuinely independent predictors. The validation record should therefore report sensitivity to depth, leaf exposure, feature subsampling and random seed, as well as an empirical forest-size stabilization curve. Computational convergence of the ensemble is a diagnostic; generalization to the next risk year remains a separate empirical proposition.

The software implementation must be part of the scientific record. The RandomForestRegressor documentation illustrates specific supported criteria and hyperparameters for one implementation. Corresponding implementations in another library or version can differ in the treatment of missing values, categorical variables, bootstrap sampling and tree growth. Reproducibility requires recording the actual version and preprocessing pipeline used to produce every reported score, rather than assuming the name 'random forest' specifies a complete computational experiment.

Frequency and Severity Prediction

Frequency, probability and count

The three common responses for a covered-event process are P(N>0 | X,t), E[N | X,t] and a full count distribution P(N=k | X,t). A forest trained as a binary classifier targets the first, whereas a count regression targets the second. The first probability equals one minus the probability of zero claims; it equals expected count only in special approximations. With increasing frequency and multiple possible events, that approximation becomes poor.

A Poisson baseline with conditional mean E[N | X,t] = t exp(g(X)) provides a transparent intensity specification. A negative-binomial benchmark can address overdispersion. A forest may learn flexible interactions in g-like structure, yet it must account for varying t. An exposure-weighted regression on N/t is an actuarial approximation under a mean-based objective; its equivalence to a count likelihood depends on variance assumptions and weighting. A benchmark comparison should disclose those differences.

Frequency performance must be assessed by exposure and by business segments, not simply averaged per policy record. If an insurer has short-term contracts and annual policies, unweighted row-level loss can give disproportionate influence to short terms. A stated weighted deviance or appropriately normalized aggregate residual makes the prediction target auditable.

Positive severity and ultimate development

Claim severity can refer to paid amount, incurred amount, settlement cost, ultimate amount, ground-up economic loss, or insurer net amount after deductible and limits. They are distinct variables with different availability and different tail properties. At a valuation date, open claims have incomplete payments and may have reserves correlated with claim complexity. A severity model trained only on settled claims inherits settlement selection; excluding open claims does not make the missing ultimate payments disappear.

For heavy-tailed severity, terminal-node means can be unstable. A forest's averaging across trees reduces split variability but does not automatically repair underrepresentation of rare catastrophes or latent severe litigation. Tail measures require tail-focused evaluation and may require separate parametric tails or stress scenarios. Mean severity may be the primary premium target, while a high conditional quantile can be relevant to claims operations or large-loss monitoring. These are different decisions with different appropriate losses.

The quantile regression forests method of Meinshausen (2006) uses forest-derived observation weights to approximate conditional quantiles. It demonstrates how an ensemble can support more than an average, but the resulting high quantile still depends on observed tail data and the sampling regime. One cannot infer extreme capital adequacy from an apparently smooth estimated 99th percentile obtained from a short claims history.

Pure premium and multiplicative decomposition

For the policy contract, aggregate covered loss Y is the sum of individual insured event losses. The relevant predictive mean per unit exposure is q(x) = E[Y | x,t]/t. A two-part estimate q_hat(x) = lambda_hat(x) m_hat(x) can help diagnose whether a deterioration comes from counts or cost per event. It is correctly interpreted only when both components refer to coherent contract limits, valuation assumptions, event units and future covariates.

A direct pure-premium model predicts aggregate cost, often dominated by zeros and an asymmetric positive tail. The statistical family can be a Tweedie compound-like distribution, a two-part construction, or a flexible regressor with exposure-aware loss. The Tweedie computation paper (Dunn and Smyth, 2005) documents density evaluation issues in this class, while the actuarial literature gives concrete out-of-sample comparisons across pure-premium and component models.

Aggregate pure premium remains a conditional expectation; it is not the final retail price. Expenses, taxes, margins, reinsurance costs, policy terms and jurisdictional constraints require separate treatment. A model cannot infer a lawful rate merely by minimizing empirical loss. The CAS cost-estimation principles provide a useful actuarial conceptual distinction within their stated regulatory context.

Insurance loss functions and scoring

Evaluation should target the same scientific quantity that the model predicts. For occurrence probabilities, the Brier score is the mean squared difference between a quoted probability and a realized binary outcome; log loss evaluates a different penalty profile. For probabilistic forecasts, Gneiting and Raftery (2007) explain strictly proper scoring rules, which encourage distributional reporting that matches a forecaster's probabilities. Discrimination statistics such as AUC can be valuable but do not establish calibration or appropriate monetary costs.

For annual counts, Poisson deviance or a corresponding likelihood can be more informative than plain classification accuracy. For positive severity, gamma deviance, robust absolute loss and tail-stratified error answer different questions. For mixed-zero aggregate cost, Tweedie deviance may reflect the intended mean-variance family more closely than RMSE, but it is sensitive to the chosen power parameter. A full distributional target can be assessed with a proper score suited to that family. These comparisons require strictly held-out predictions and prespecified definitions of exposure weighting.

Metric selection has economic consequences. An underwriting threshold maximizing accuracy may reject many true negative no-claim policies and have unacceptable commercial value. A ranking score may be informative for prioritizing audits but poorly calibrated for premium construction. A mean-square-optimal model may devote disproportionate weight to a small number of largest losses. These are not contradictions: different metrics elicit different functionals and encode different operational costs.

Temporal, clustered and geographic validation

For prospective insurance, a realistic test trains on information available by a cut-off and evaluates policies underwritten in later periods, with claims matured sufficiently for the target. A separate validation set may be needed for tuning. An intervening gap can prevent leakage through delayed claims, reserving updates and rolling-window features. The time split must respect both policy start date and loss valuation date.

Clustered structures require grouped separation where the deployment question concerns new units. A driver or household appearing across years can make a random row split easier than a new-customer underwriting problem. Flood events, storms, cyber outbreaks and fleet operations create shared outcomes across many records. A geographic holdout can test transfer to an unfamiliar region; it is not interchangeable with the next-year holdout in the same territory.

A published nonparametric method on one exposure population is evidence for a particular data-generating setting. Internal model success cannot be assumed to transfer across national motor markets with different deductibles, roads, driving behavior, claims practices, inflation, liability rules and distribution channels. External validation should measure calibration-in-the-large, segment drift, variable availability and economic outcomes, with explicit recalibration if permitted.

Severity tails, accumulation, and decision-sensitive forecasting

Expected claims cost is only one functional of the future aggregate-loss distribution. A business line with many ordinary claims and very rare extreme losses may have a stable median and highly uncertain mean. An ensemble selected by average squared prediction error can react strongly to large observed losses, yet few such observations are available to reveal the frequency and dependence of extreme events. Separating the ordinary attritional component from an explicitly modeled catastrophic or exceptional tail may help make these uncertainty sources auditable. Such a separation is a research design choice that requires evidence; it is not a universal prescription.

Consider a severity sample with 99 independent claims of 1,000 currency units and a single valid claim of 901,000. Its empirical average is 10,000 per claim. Removing the single largest claim reduces the sample average for the remaining 99 to 1,000, a tenfold change. The arithmetic illustrates the sensitivity of mean loss to a rare large event, not a factual statement about any insurance population. A bootstrap forest fitted to the full sample may repeat the same large claim across many training replicates. That does not create additional distinct extreme-loss observations, and the apparent stability of aggregate predictions can be misleading if the reporting system or future hazard differs.

Claim limits and reinsurance attachments also alter the estimand. A capped insured loss, ground-up property damage, indemnity net of deductible, and insurer retention after a treaty are different target variables. A random forest trained on one layer cannot be relabeled as a forecast for another without applying the correct contract transforms and checking the distribution of exceedances. Catastrophic accumulation across policies requires aggregation under shared events. Averaging contract-level expected loss is sufficient to form the expectation of a sum under linearity, but it is insufficient to determine the tail probability of that sum when losses are dependent.

Risk-sensitive scoring should disclose whether the research targets a mean, a conditional quantile, an upper-tail expected shortfall, a large-loss classification probability or a policy-level ranking. For quantiles, the quantile regression forests article (Meinshausen, 2006) offers an established extension. For a forecast distribution, a proper scoring rule is preferable to ad hoc manipulation of one favorable summary statistic; Gneiting and Raftery (2007) provide the general elicitation rationale. Metrics should be applied to the same valid holdout losses under the same valuation and exposure conventions.

Temporal cost inflation and claims settlement development are especially important in long-tail business. Some old claims become fully observed years after underwriting. An apparent prediction improvement may disappear when the target is updated to ultimate settlement amounts. A valid study should publish the maturity horizon and any development-adjustment methodology, and it should distinguish model error from subsequently updated labels. Without those disclosures, a comparison of flexible algorithms can confound improvements in feature learning with differences in label quality.

Variable Importance and Stability

What an importance score measures

An impurity-based importance measure accumulates reductions in the training split criterion credited to a variable. A permutation importance compares model performance before and after a variable is shuffled in a validation sample. Neither is a universal measure of causal effect, and neither identifies a variable's unchanging moral or regulatory status. Both are relative to the model, dataset, feature set and metric.

Strobl et al. (2007) demonstrate biases in random forest selection and importance measures associated with predictors having different possible splitting structures. An insurance variable with many categories can receive apparent explanatory prominence because it offers more candidate partitions, particularly when estimates are trained and evaluated under the same data. Unchecked importance can therefore misdirect investigation of underwriting factors.

Correlated predictors and conditional importance

With strongly correlated features, marginal permutation breaks the dependence structure and creates unrealistic covariate combinations. Shuffling vehicle use while holding related mileage measures fixed may construct combinations that almost never arise in business data. A model might place reliance on one proxy in one fit and another in the next fit while collective predictive performance stays stable.

Conditional variable importance (Strobl et al., 2008) was developed to address this issue by respecting relevant dependence in the permutation mechanism. The resulting conditional and marginal importances answer different questions. Neither transforms observational association into an intervention effect. Reporting should identify the specific estimand of importance, the correlated-variable groups and the reference sample.

Stability, calibration and segment analysis

A usable risk partition should survive plausible perturbations of data, period and modeling choices. Stability can be examined by repeating fits across seeds and time slices, comparing rank order, group-level mean predictions, relative rates and concentration of predictions at terminal extremes. A stable average forest can nevertheless have unstable local explanations. Conversely, importance rankings may vary while the decision-level projected claims cost remains sufficiently stable.

Insurance validation should be segmented by exposure, claim occurrence, major coverage, territory, policy tenure and protected or restricted attributes where lawful audit access is available. A favorable global score may mask substantial underprediction for a small class or a growing hazardous exposure. Segment materiality must be assessed in claim amounts and exposure volume as well as statistical significance. Multiple-comparison searches across numerous groups should be interpreted cautiously.

Calibration-in-the-large compares total predicted and observed insured losses on a coherent maturity basis. Calibration curves compare predicted and realized frequencies or average costs across risk groups. Yet a model can be calibrated on observed historical mix and fail after the insured population changes. Recalibration and drift surveillance need governance over which coefficients, base rates or tree structures may be changed and how such changes are evaluated.

Interpretability as a layered requirement

A single shallow decision tree can often be rendered as a collection of rules; a forest of hundreds of deep trees cannot ordinarily be given the same complete direct reading. Local surrogate rules and partial dependence displays summarize the fitted predictor from selected viewpoints. Their quality depends on approximation accuracy and the degree to which the displayed covariate combinations are supported by actual observations.

Interpretability in insurance has several audiences: actuaries analyzing expected costs, underwriters making decisions, model validators checking failure modes, regulators reviewing permitted classifications, and policyholders receiving explanations of outcomes. A variable's numerical rank may be insufficient for all of them. Each requires a truthful mapping among the prediction target, key data inputs, policy rules, uncertainty, and actual authority for the decision.

The transparency program is strongest when it includes dataset provenance, versioned feature logic, tree settings, cross-time metrics, subgroup checks, explanation fidelity, override procedures and documented human escalation. This is a practical synthesis of established governance considerations, not a claim that any single algorithm creates those duties or that one jurisdiction mandates the same control.

Prediction, fairness and causality

Historical claims do not record a neutral natural risk in isolation. Coverage selection, reporting incentives, underwriting exclusion, geography, repair supply chains, litigation and settlement practices affect the observed label. A model trained on these outcomes may learn differences in reporting and treatment as well as insured hazards. Fairness evaluation must therefore specify whether the objective concerns expected covered cost, allowed rating variables, access, consumer outcomes or a constrained allocation.

Predictive trees are also distinct from causal forests. Wager and Athey (2018) develop methods for heterogeneous causal effects under explicit potential-outcome assumptions. A conventional random forest that assigns higher claims probability to one risk factor does not establish what would happen if the insurer changed that factor, adjusted the premium, or denied coverage. Pricing decisions can influence take-up and claim behavior, shifting the subsequent training population.

Variable importance audits, correlated proxies, and drift

The scientific target of an importance audit must be stated precisely. If the question asks which features help a frozen model predict in a new portfolio, out-of-time permutation loss measures a deterioration in that portfolio's forecast score under the chosen perturbation. If the question asks whether an underwriting characteristic intrinsically causes additional covered loss, such permutation evidence is inadequate. If the question asks which risk factors entered a given deployed decision, global importance is an incomplete account of that individual decision.

A variable also changes meaning when its data pipeline changes. 'Annual mileage' might refer to a customer's estimate, a telematics reconstruction, an odometer reading or an updated administrative proxy. The label can remain the same while its measurement error and access conditions shift. A year-to-year change in importance may therefore reflect altered data capture rather than altered driving hazard. Versioned feature lineage is essential to interpret the stability of an insured-risk model.

Strobl et al. (2007) establish why split-based forest importance can favor variables with many available split points. The later conditional importance study by Strobl et al. (2008) addresses dependence among predictors that distorts unconditional permutation interpretations. These are methodological results about an estimator and its perturbation schemes. Their insurance implication is that one should examine the distribution of variable levels, feature cardinality, missingness, dependence and validation metrics before interpreting a ranking.

An operational audit can report multiple complementary views: impurity importance used during fitting, loss-based permutation importance on a frozen future sample, grouped permutation effects for correlated variables, and stability across rolling time windows. The audited importance should be paired with calibration and counterfactual feature availability, and the proposed business action must be specified separately. A high importance score can justify closer measurement study, while a low importance score need not establish that a variable is irrelevant to fairness, contractual coverage or an important small subpopulation.

Drift surveillance requires a defensible baseline. For each monitoring interval, report the number of new contracts, earned exposures, realized and mature claim events, segmentation by relevant hazard and coverage, feature availability and changes in the price or risk-selection regime. Calibration residuals, missingness changes and performance confidence limits should be examined together. Automated alerts should not be interpreted as proof of a particular cause: repair-price inflation, altered reporting practice and a genuine change in accident frequency can produce overlapping symptoms.

Motor Insurance Evidence and Limits

French third-party liability evidence

Motor insurance provides a valuable empirical setting because exposure and policy characteristics are commonly recorded at the contract level, while claims are sparse and heterogeneous. The French motor third-party liability case study by Noll, Salzmann, and Wuthrich (revised 2020) compares a classical generalized linear model with regression-tree, boosting-machine and neural-network approaches for claim frequency. The paper reports that the simple GLM did not capture some covariate interactions as well as the alternative approaches in its studied setting.

This evidence supports the proposition that flexibility can be relevant to frequency forecasting. It does not establish that random forests outperform every GLM, boosting model or neural network in motor portfolios worldwide. A regression tree in the paper must not be relabeled a random forest; likewise, an interaction improvement does not automatically establish regulatory appropriateness, stable tail prediction, or causal interpretability.

The practical implication is that the benchmark GLM should be credible, not deliberately weak. A rich GLM with selected interactions, splines, offset treatment and regularization may close a portion of the gap. Comparison across model families should include tuning fairness, data-availability constraints, equivalent held-out periods and the same economically meaningful scoring rule.

Insurance tariff studies with random forests

A more directly relevant actuarial investigation is the tree-based tariff study by Henckaerts, Cote, Antonio, and Verbelen (2020 online). Its publisher record expressly describes development of frequency and severity tariffs using a progression from regression trees to random forests and boosted trees. It emphasizes excess-zero counts, variable exposure, scarce long-tailed severity observations, transparent tariff construction and tuning through cross-validation.

The study establishes a concrete actuarial research program in which ensemble performance is assessed in relation to the distinct response types that constitute a tariff. It provides a reason to test forests as serious contenders, not an unconditional result that forests are the universally optimal rating technology. Reported models, datasets and losses must be read in their exact experimental context; an approved deployment in another insurer requires its own validation.

Lessons from homeowners insurance

The multi-peril homeowners analysis by Frees, Meyers, and Cummings (2012) compares direct pure-premium and component models, evaluates out-of-sample performance, and examines dependencies across perils. Its primary contribution here is methodological: claims costs are sparse, multi-peril dependence matters, and model comparisons require more than one global fit statistic. The paper should not be cited as evidence of an RF-specific improvement it did not estimate.

This distinction matters for transferability. Motor frequency evidence concerns collision, liability or other contract-defined events; property insurance may include correlated catastrophes and a different severity tail. Homeowners, medical, commercial liability and life insurance each require target-specific handling. A method supported for one policy-count forecast cannot be presumed adequate for long-tail claims reserving or an event-time survival model.

Censoring, duration and alternative forest families

For claim closure duration, lapse, mortality or time-to-event analysis, binary trees trained on an observed 'closed' flag can induce bias when records have unequal follow-up times. The random survival forests paper (Ishwaran et al., 2008) explicitly develops forest procedures for right-censored survival responses. That literature illustrates the principle of matching splitting and estimation methods to the observational mechanism.

Survival forests do not automatically solve reporting delay, left truncation, competing risks, selective claim closure or jurisdiction-specific settlement effects. Those features require correct risk sets, event definitions, censoring assumptions and data histories. In every insurance product, the data-generating process precedes algorithm choice.

International research boundaries

The geographical label 'international' describes the scientific applicability of the questions, not the representativeness of one French or U.S. dataset. Country-specific motor exposure rules, claim classifications, repair inflation, tort structures and insurance supervision alter both predictors and outcomes. A future comparative study should record target compatibility, contract basis, feature reproducibility, privacy constraints, historical catastrophe load and local economic cost measure before interpreting cross-country performance differences.

An insurer adopting forests should preserve auditability of training jurisdiction, geography, time window, target maturity, contractual scope and software configuration. It should be possible to explain when the model has moved outside its observed support. This is especially important when relying on external research as rationale for internal development: empirical literature supplies hypotheses and methods, while local validation establishes fitness for the exact portfolio.

Critical synthesis of the evidence

The literature provides convergent support for three qualified conclusions. Tree-based splits can discover conditional interactions that the analyst has not explicitly encoded; aggregating randomized trees can reduce sensitivity to individual tree partitions; and claim targets require exposure, heavy-tail, dependence and time-aware evaluation. The strength of each conclusion is different. The first follows from the function class and selected empirical studies, the second from ensemble mathematics and general random forest research, and the third from actuarial data construction and evaluation requirements.

There is no defensible universal numerical uplift to assign to random forests in insurance. The French case study is not a dedicated RF experiment, the tariff study has a specific design and data scope, and the homeowners study examines multiple statistical strategies rather than offering a global winner. Original synthesis therefore centers on the fit between objective, observation process and decision, with no newly invented predictive architecture or model proposed.

Conditions for international comparative inference

A credible comparative study across insurance markets needs a protocol that separates statistical method from product and jurisdiction. The study should first harmonize a target such as one-year covered claim count per earned vehicle-year and explicitly identify where it cannot be made equivalent. A third-party liability claim in one jurisdiction may involve a different compensation system, deductible, reporting obligation or litigation process from a superficially similar claim elsewhere. Comparable names do not guarantee comparable random variables.

The evidence base for motor forecasting demonstrates why this distinction matters. The French motor frequency case study (Noll, Salzmann, and Wuthrich, 2020 revision) tests one explicit claims-frequency problem; it neither provides a global motor benchmark nor documents universal random forest superiority. The actuarial tree-based tariff study (Henckaerts et al., online 2020, journal 2021) explicitly develops frequency and severity tariff components and compares tree-based model families. Its substantive results depend on the underlying insurance sample and chosen objective.

Cross-country benchmarking should establish at least the following comparability conditions in narrative documentation: a common event definition, consistent exposure measurement, aligned policy limits and deductibles, compatible historical periods, observed valuation maturity, harmonized or explicitly stratified cost inflation, the same permissible feature set, equivalent training and tuning budgets, and evaluation on comparable prospective calendar intervals. Contract and legal differences that prevent equivalence should be reported as study strata, not silently averaged away.

A useful experimental design trains each candidate method on a distinct country's eligible development data, tunes within that country's historical window, and evaluates it in a genuinely later period. A second transport experiment can apply the trained model to another country after documenting feature translation, omitted risk factors and any allowed recalibration. The first tests local out-of-time performance; the second tests cross-domain generalization. A third experiment can evaluate performance after a new insured population or policy rule changes, making selection effects visible. It would be misleading to combine these fundamentally different experiments into a single unqualified ranking.

The multi-peril homeowners case (Frees, Meyers, and Cummings, 2012) adds a distinct lesson about dependencies and alternative out-of-sample scoring measures, although it investigates homeowners rather than French motor liability. Together these papers motivate research designs that preserve the insurance mechanism and the evaluation context. They do not establish any new theoretical doctrine or a general coefficient of improvement; the original contribution of this dossier is a documented synthesis of their methodological implications for a common evaluation problem.

Technical Appendix: Transparent Illustrations and Audit Protocol

Hypothetical exposure and pure-premium calculation

Consider three invented motor segments for one annual observation basis. The amounts are mathematical examples, not observed data and not estimates from a fitted tree. Group A has 500 earned vehicle-years, 40 claims, and total covered incurred cost of 180,000 currency units; Group B has 300 earned vehicle-years, 45 claims, and cost of 315,000; Group C has 200 earned vehicle-years, 40 claims, and cost of 440,000. Every group uses the same ultimate-cost definition and valuation assumptions.

Segment Earned vehicle-years Covered claim count Aggregate cost Claims per vehicle-year Cost per claim Cost per vehicle-year
A 500 40 180,000 0.080 4,500 360
B 300 45 315,000 0.150 7,000 1,050
C 200 40 440,000 0.200 11,000 2,200
Total 1,000 125 935,000 0.125 7,480 935

The total cost per earned-year equals 935,000/1,000 = 935. Using frequency times severity yields 0.125 x 7,480 = 935. Segment calculations give 0.08 x 4,500 = 360; 0.15 x 7,000 = 1,050; and 0.20 x 11,000 = 2,200. The exposure-weighted mean of segment costs, (500 x 360 + 300 x 1,050 + 200 x 2,200)/1,000, also equals 935. These identities hold for this internally consistent invented example; they do not prove that a fitted forest will preserve the same relationship out of sample.

If a regression tree places a new observation into Group C, its historical within-leaf mean may be 2,200 cost units per vehicle-year. A forest averages many such data-dependent leaf estimates. Neither prediction is automatically a legally permitted premium, an inflation-adjusted future ultimate cost, or a confidence interval. The calculation illustrates why denominators must be preserved and why an apparently successful event classifier is insufficient to estimate aggregate expected costs.

Bagging and effective sample calculation

For n independent records, the probability that a given record is absent from one size-n bootstrap is (1 - 1/n)^n. With n = 1,000, this is approximately 0.3677, so a tree has roughly 368 out-of-bag records in expectation, including approximately 632 distinct in-bag records. Actual realizations vary. The estimate does not apply as a count of independent policies if claims, monthly transactions and payments refer repeatedly to a smaller number of economic exposures.

The forest-averaging formula also clarifies diminishing returns. Suppose individual tree errors have common variance 100 and pairwise correlation 0.20. With 10 trees, the variance of the mean error under the exchangeable covariance assumptions is 100[0.20 + 0.80/10] = 28. With 100 trees it is 20.8, and the large-ensemble limit is 20. This is an illustrative covariance calculation, not a measured insurance forest. It shows why another order of magnitude in tree count can leave substantial common error.

Minimum reproducibility record

A reproducible insurance forest evaluation identifies dataset owner and permission, observation unit, exact prediction date, training start and end, data freeze, contract and claim definitions, exposure basis, known-maturity boundary, event and insured grouping, data exclusions, preprocessing steps, feature timestamps and encodings, random seeds, candidate models, tuning ranges, split construction, validation periods, metrics and decision thresholds. The model artifact should include library versions, trained parameter values, input schema and a fixed mapping from raw data to deployment features.

The empirical results should provide the number of policies or exposure units, earned exposure, total event count, positive-cost count, significant tail concentration and sample sizes for each independent test segment. Report the benchmark fit and the forest under identical forecast horizons, along with confidence or resampling uncertainty suitable to the actual dependence structure. Avoid manufacturing confidence intervals from unclustered row resampling when the sample contains shared catastrophic events.

A validation report separates discrimination, calibration, cost accuracy and business utility. For pricing, it can include expected and actual aggregate cost by independent risk band, with credibility intervals and sample sizes. For fraud investigation it may use workload-constrained precision and recall; for claim closure a censoring-aware survival metric; for large losses a tail-focused criterion. Every metric must be linked to an estimand and implementation decision rather than used as an interchangeable headline score.

Source-to-claim research audit

The foundational source analysis establishes algorithm construction and mathematical interpretation; the actuarial source analysis provides concrete examples of exposure-aware insurance modeling; the critical literature establishes limits of variable importance and inference. The random forest hyperparameter review supplies methodological context, the bias analysis of forest importance supports warnings about categorical feature competition, and the proper scoring rule literature grounds probabilistic evaluation. None of these sources reports the numerical hypothetical table above, whose arithmetic is independently inspectable.

This document deliberately does not claim a new fitted insurance dataset, a systematic meta-analysis, full restricted-book access, reproduction of proprietary insurance tariff experiments or performance forecasts for a specific insurer. It establishes a source-checked methodological synthesis with transparent reasoning, falsifiable validation requirements and a reproducible numerical illustration. Stronger empirical claims require data access and predeclared comparisons.

Prepublication reproduction and falsification checklist

A final research package should give another qualified analyst enough information to falsify its conclusions under a declared data contract. In the present documentary study, the source metadata, exact claims, mathematical derivations, and hypothetical computations are exposed for inspection. No real insurer observations are furnished, so no empirical out-of-sample accuracy statement is made. The inferential scope of the manuscript is consequently narrower than that of a fitted multi-country insurance model, and its conclusions identify the conditions that such a model would have to satisfy.

For a future portfolio experiment, the first reproducibility object would be a frozen cohort manifest identifying policy inception and expiration dates, risk and policy units, claim events, transaction versions, and cutoff dates. A second would be a target manifest describing reporting, ultimate adjustment and contract transforms. The third would specify training and evaluation partitions, including the rationale for any entity, event, time or geographic grouping. A fourth would record all preprocessing, feature timestamps, model hyperparameters and software versions. A fifth would contain model outputs paired with a separately frozen holdout outcome file and complete score calculations.

The empirical null comparisons should include transparent actuarial baselines, a naive recent-history baseline, a single suitably pruned tree, bagged trees and random forests; optional other model families can then be compared under a common tuning budget. All are assessed on the same predeclared estimand and evaluation sample. Incremental benefit should be reported in the original units of the decision, such as expected claims cost per earned-year or deviance relative to exposure, alongside the statistical score. When a difference is small or dependent on one tail event, the result should state that sensitivity directly.

The governance layer should track the researcher's choices and the insurer's subsequent operational choices separately. A high-quality forecast can still be misused if its input cutoff moves, its contractual loss layer changes or its predictive risk class is presented as a causal explanation. The appropriate documentation makes the scope of inference clear and preserves a path for independent reassessment. These conditions are integral to disciplined insurance science and can be implemented without positing a new named theory.

Conclusion

Decision trees identify interactions by subdividing covariate space, and random forests stabilize predictions by averaging trees randomized through sampling and feature selection. These advantages belong to the statistical architecture of the methods. Their realized value in insurance depends on disciplined alignment of the learning task with contractual risk, exposure, loss maturity, valuation basis, observation dependence and the prospective decision to be made.

The substantive research establishes five findings. First, a claims frequency rate, an event probability, a positive severity mean and a policy pure premium are different prediction targets and must be evaluated separately. Second, forests reduce the instability of individual tree partitions but retain correlated modeling errors and do not generate experience with unobserved hazards. Third, variable importance and local explanations measure model-dependent associations and require careful treatment of correlations, categorical split opportunities and causal ambiguity. Fourth, credible out-of-time, grouped and segment-level comparisons with well-designed actuarial baselines are indispensable to performance claims. Fifth, cross-product and international transfer require compatibility of contracts, observation mechanisms, hazard distributions, legal rules and policy decisions.

Published motor-insurance and tariff studies justify serious consideration of tree-based learning, especially where nonlinear interactions complicate conventional specifications. They do not supply a universal ranking or a transferable numerical accuracy premium for random forests. The decisive research standard is an auditable alignment among target, dataset, algorithm, loss function, independent validation and operational decision. Future empirical studies can test this alignment across insurers and jurisdictions using comparable outcome definitions, transparent uncertainty analysis and controlled changes to both portfolio mix and claim-development assumptions.

References