RESEARCH DOSSIER
Gradient Boosting for Insurance Loss Prediction
Abstract. Gradient boosting estimates insurance loss functions by sequentially adding weak prediction rules in directions that reduce a chosen empirical risk. The method has become important to actuarial prediction because insured-event data contain variable exposure, large numbers of zero observations, long-tailed payments, correlated rating variables, nonlinear effects, and interactions that conventional additive tariffs may omit. This monograph develops the statistical meaning of functional gradient descent, tree-based second-order updates, and the loss functions appropriate to claim frequency, positive severity, and aggregate loss. It derives exposure-aware Poisson, Gamma, and compound Poisson-Tweedie formulations and clarifies when frequency-severity decomposition agrees with the estimand for pure premium. It then specifies a reproducible validation design emphasizing policy and event grouping, claim development, temporal separation, exposure-weighted calibration, stability, and testing on the deployment population. Four-contract numerical calculations provide a transparent check of deviance, exposure normalization, and the relation between aggregate losses and expected severity; these illustrations are mathematical, not empirical insurer-performance estimates. Finally, the study evaluates the conditions under which boosted trees become interpretable, operationally reliable, and suitable for regulated insurance decisions. Published comparisons support the utility of boosting within their particular motor-insurance datasets and evaluation designs, while leaving general superiority to new, independently validated comparisons. The central conclusion is that a gradient-boosted loss prediction acquires actuarial meaning through an explicit claim definition, insured time at risk, correct statistical objective, disciplined out-of-sample evaluation, and governed use of the output. The work provides a substantive synthesis of established methods without claiming a new algorithm, theory, or empirical discovery.
Keywords: Gradient boosting; functional gradient descent; insurance loss prediction; claim frequency; claim severity; pure premium; exposure offset; Poisson deviance; Gamma regression; Tweedie compound Poisson model; decision trees; early stopping; temporal validation; predictive calibration; model interpretability; actuarial governance.
Introduction
Prediction of an insurance loss is a statistical statement about a contract, exposure period, event definition, information set, and monetary measure. A forecast of reported claims in the next 12 months differs from a forecast of ultimate incurred loss for the same accident year; a prediction of a claim count differs from a probability that at least one claim is filed; and a risk score differs from an expected monetary cost. Each quantity may be useful, but each requires its own data construction and validation. In insurance, the observation is a consequence of risk, coverage, underwriting, reporting, settlement, and time. The statistical method only becomes relevant after those mechanisms are specified.
The foundation of modern gradient boosting lies in the statistical interpretation of boosting by Friedman, Hastie, and Tibshirani (2000), Friedman's (2001) functional gradient formulation, and his (2002) stochastic gradient boosting analysis. Their methods organize prediction as sequential optimization of a loss criterion. In an insurance setting, that criterion can express the probability law of counts, positive claim amounts, or nonnegative aggregate losses. The insurance GLM framework presented by de Jong and Heller (2008) remains indispensable because it gives estimands, exposure adjustment, distributional assumptions, and diagnostics a clear actuarial interpretation. Boosted trees extend the functional shape that may be learned; they do not dispense with these requirements.
The central object of this Research Dossier is gradient-boosted additive trees as predictors of insurance claim frequency, positive severity, and pure premium. Tree ensembles trained by bagging, neural-network architectures, stacking and automated model selection form adjacent, separately analyzable fields. They enter this study where a comparison clarifies boosting mechanics and insurance-specific performance. The discussion takes an international perspective while using several motor-insurance studies as particularly well documented empirical examples. Insurance products differ in legal entitlement, loss emergence, claim frequency, benefit design, and tail behavior; transfer from any particular case study therefore requires independent evidence.
The scope is actuarial prediction and the scientific evaluation of models. A fitted expected loss is an input to pricing, underwriting, reserving, or portfolio planning and is not identical to an authorized premium. The premium can also reflect expenses, capital cost, reinsurance, taxes, competitive conditions, consumer protections, and the applicable legal rating framework. Equally, the model's statistical explanation does not by itself justify a causal interpretation of a risk factor. This distinction determines how predictive success is to be understood.
The evidence method combines original derivations, source-specific review of academic and institutional primary records, a transparent arithmetic demonstration, and comparative assessment of modeling designs. Where journal full texts are restricted, conclusions from those works are limited to the publisher abstracts and bibliographic material actually examined. No private policy data were obtained, no trained model on insurer observations is presented, and no previously published performance improvement is represented as a replicated finding. The contribution lies in analytically relating established boosting methods to the sampling, measurement, validation, and governance conditions of insurance.
Functional Gradient Descent
Insurance estimands and prediction horizons
Consider policy or coverage-period record i, with covariates X_i observable at a specified decision time, positive exposure e_i in a defined time unit, event count N_i, and ultimate aggregate covered loss Y_i. A frequency estimand is lambda(x) = E[N_i | X_i=x, e_i]/e_i under an assumption of proportionality in exposure and stable conditions within that interval. The aggregate expected loss per exposure unit is m(x) = E[Y_i | X_i=x, e_i]/e_i under the analogous proportionality assumption. For a future insured interval with exposure e*, the expected claim count is e* lambda(x*) and the expected total covered loss is e* m(x*). Both proportionality assumptions need testing when contract duration, seasonality, reporting practices, or hazard varies with time.
The observation horizon and the valuation horizon must be recorded separately. A claim in a 2024 accident period may be reported later, reopened, litigated, or settled over years. Training on paid loss at a valuation date predicts paid-to-date loss with development bias unless it is developed or explicitly modeled to the desired maturity. Data already known to a claim handler after notification cannot legitimately enter a pre-inception pricing forecast. The feature cutoff, target window, payment definition, inflation basis, salvage and subrogation treatment, deductible, limit, and currency must be held constant through modeling and validation.
If claim j on policy i has severity Z_ij, then Y_i = sum_{j=1}^{N_i} Z_ij for a consistently defined nonnegative indemnity measure. The conditional per-claim mean E[Z_ij | X_i=x] is a distinct object. The identity E[Y_i|x] = E[N_i|x] E[Z|x] requires that the latter severity mean is appropriately defined as the claim-weighted mean associated with the same risk and event process; it must not be obtained from an arbitrary average of the positive-policy observations. The choice of individual claims, claim periods, or policies as the sampling unit changes both estimates and their uncertainty.
From additive models to stagewise optimization
Let F(x) be a real-valued prediction function, often used as the log of an expected claim rate or loss. Define empirical risk L(F) = sum_i w_i ell(y_i,F(x_i)), where w_i is a justified observation weight and ell is a specified loss function. Gradient boosting constructs F_M(x) = F_0(x) + sum_{m=1}^M nu rho_m h_m(x), where h_m belongs to a restricted class such as small regression trees, rho_m is a fitted step, and nu is a shrinkage coefficient. F_0 is chosen to minimize risk among constant functions. This is function-space optimization: each new tree alters the prediction where the preceding additive model retains systematic error.
Friedman's greedy function approximation establishes the practical connection between steepest descent in the function space and sequentially fitted base learners. At iteration m the negative gradient r_im = -partial ell(y_i,F(x_i))/partial F, evaluated at F_{m-1}(x_i), becomes the target for the next tree. The resulting regression tree approximates the gradient field using partitions of the covariate space. Step size and leafwise adjustments can then be selected to reduce the actual objective rather than treating all gradient values as an ordinary regression response.
A depth-one tree creates a piecewise-constant effect in one variable; greater depth permits interactions within a new tree. Repeated trees construct nonlinear main effects and interactions through additive accumulation. This is a flexible approximation of the conditional predictive function, not a statement that any discovered threshold is a physical or causal boundary. The Elements of Statistical Learning by Hastie, Tibshirani, and Friedman (2009) develops these relationships among tree size, stagewise fitting, shrinkage, and predictive complexity in a common statistical treatment.
Pseudo-residuals, regularization, and convergence
For squared error ell = (y-F)^2/2, the negative gradient is y-F, the familiar residual. For Poisson or Gamma losses on a log-mean scale, the gradient is asymmetric and depends on the current predicted mean. Thus the slogan that boosting repeatedly fits residuals is only literally correct for squared error; more generally it fits loss-specific pseudo-residuals or second-order local approximations. A model chosen with a Poisson loss makes a different scientific commitment from a model chosen with least squares, especially in a portfolio dominated by zero counts.
Shrinkage controls how much each new tree changes F. Subsampling policies or records at each stage introduces another regularization mechanism; Friedman's stochastic boosting study explicitly evaluates subsampling in gradient boosting. Row subsampling is not a license to split multiple observations of the same policy, claim, broker or catastrophe event across training and test sets. The sampling unit for regularization and the unit of external generalization are different design decisions.
Tree depth, minimum leaf exposure, minimum Hessian mass, split penalties, column sampling, and iteration count jointly determine effective complexity. Boosting may fit rare interactions that are scientifically real but statistically weak. Consequently, convergence of the training objective and improvement on the intended deployment population are distinct properties. Early stopping and a genuinely unused test set remain necessary even when gradients and numerical optimization are implemented correctly.
Why insurance-specific loss functions matter
Squared error gives each currency-unit discrepancy a quadratic penalty, often assigning overwhelming leverage to a few large claims. Absolute error targets a conditional median, which for a sufficiently zero-heavy aggregate-loss variable can be exactly zero. The expected claim cost, needed for pure premium, is a mean functional and therefore requires a compatible objective or an explicit method for recovering the mean. A classification objective targets a probability of at least one claim and cannot be converted into the expected number of claims unless the count model or further assumptions supply the missing multiplicity.
Actuarial model design is therefore an alignment of the question, the estimand, and the loss. A zero-heavy nonnegative response may justify the compound Poisson-Gamma representation for certain products, but the same Tweedie family is not a universal representation for liability tail losses, disability durations, or correlated catastrophe accumulations. Yang, Qian, and Zou (2018) study a gradient-tree boosted Tweedie compound Poisson model for insurance pure premium, providing a directly relevant insurance-specific implementation. Their result demonstrates feasibility and case-study predictive gains, not an automatic distributional fit for every line of insurance.
Boosted Trees and Loss Functions
Tree partitions and second-order fitting
A regression tree maps covariates to terminal regions R_k and gives a constant leaf increment a_k in each region. Its splits can capture sharp nonlinearities, conditional thresholds, and interactions, subject to sample support. Modern systems often approximate the change in training loss for a proposed tree with the second-order expansion sum_i [g_i f(x_i) + (1/2) h_i f(x_i)^2] plus a tree-complexity penalty, where g_i and h_i are first and second derivatives of the loss with respect to the current margin. For leaf k, let G_k = sum_{i in R_k} g_i and H_k = sum_{i in R_k} h_i. Under an L2 leaf penalty alpha, the unconstrained quadratic leaf update is a_k = -G_k/(H_k+alpha), provided the curvature and constraints support that update.
The same expression makes the importance of Hessian-based minimum leaf support explicit. In an insurance Poisson model, H_k reflects predicted expected counts, which may be low in a leaf with many policies but little exposure or very low claim rates. A high number of rows is consequently not sufficient to ensure stable claim-frequency estimation. Split gains also depend on the shape of the objective and its curvature. The XGBoost system paper by Chen and Guestrin (2016) describes sparse-aware tree learning, approximate split finding, regularization and system-level efficiencies. The second-order expression is useful for reasoning about optimization; the exact implementation depends on objective, constraints and library version.
LightGBM's original study by Ke and colleagues (2017) introduces gradient-based one-side sampling and exclusive feature bundling to improve tree-building efficiency. CatBoost's ordered boosting approach by Prokhorenkova and colleagues (2018) addresses target leakage induced by categorical encoding and prediction shift through ordered procedures. These developments concern computational and statistical implementations of boosting. Selection among them should be made on an identical insurance estimand, validation protocol, software configuration and resource budget.
Poisson count prediction
Under a conditional Poisson model N_i | X_i,e_i ~ Poisson(mu_i), set mu_i = e_i exp(F(X_i)). The negative log likelihood, up to constants independent of F, is ell_i = e_i exp(F_i) - N_i [log(e_i)+F_i]. Since log(e_i) is fixed, gradients with respect to F_i are g_i = e_i exp(F_i)-N_i and h_i = e_i exp(F_i). The derivatives show directly that time at risk belongs in the mean, not in a label correction applied only after fitting. The Poisson parameter assumes equidispersion conditional on the modeled effects: Var(N_i|X_i,e_i) = E[N_i|X_i,e_i]. Overdispersion and dependence can make this variance assumption false even when the conditional mean is useful.
The Poisson deviance for observed count n and predicted total count mu is D_P(n,mu)=2[n log(n/mu)-(n-mu)], with n log(n/mu) defined to be zero at n=0. A model may be evaluated by total, exposure-weighted, or averaged deviance, but comparisons must use the same normalization. Pooled portfolio calibration additionally requires comparison of sum_i N_i with sum_i mu_i, including by segment and calendar period. A small deviance does not guarantee correct aggregate predicted cost in a changed portfolio, nor does perfect aggregate balance ensure discrimination among risk classes.
The XGBoost objective documentation describes a count:poisson objective and a reg:tweedie objective, including the power parameter. It does not, by listing these objectives, certify a particular exposure-offset interface for every estimator wrapper. The data scientist must inspect whether the API supports base margins or offsets, fixed initial margins, and sample weights as intended; multiplying a claim count by an arbitrary weight is generally not equivalent to adding log exposure to the Poisson linear predictor.
Positive claim severity
For positive severities Z_j, a Gamma mean model with log link s_j = exp(G(X_j)) can use loss ell_j = Z_j/s_j + log s_j, up to factors or data-only terms. Its derivative in G is 1-Z_j/s_j and the observed second derivative is Z_j/s_j. Gamma variance proportional to s_j^2 may be reasonable for some moderate indemnities; it can be inadequate for highly irregular large losses or policy-limit mixtures. A lognormal, inverse-Gaussian, distributional, quantile or specialized heavy-tail model may be preferable depending on what is predicted. All such alternatives require their own calibration and tail checks.
A severity training set formed by positive claims represents the population of reported positive claims at the chosen valuation maturity. If multiple claims arise from one policy, each claim contributes an observation and the estimated severity is claim-weighted; if a mean of total positive-policy payments is trained instead, the response is a conditional aggregate and must not be treated as a per-claim severity. A binary claim indicator plus a positive conditional mean targets expected payment per risk only if the event definition allows at most one claim or count multiplicity is properly incorporated. Henckaerts and colleagues (2021) explicitly address the combination of unbalanced frequency data, varying exposure and scarce, potentially long-tailed severity observations in insurance tariff construction.
Tweedie aggregate losses
A Tweedie exponential dispersion model with power parameter 1<p<2 has nonnegative support with a point mass at zero and a positive continuous component, corresponding to a compound Poisson-Gamma construction. If the aggregate loss per exposure unit R_i=Y_i/e_i has mean m_i=exp(F_i), the actuarially coherent variance scaling under a stationary independent-increments compound model is Var(R_i|X_i,e_i)=phi m_i^p/e_i. Here phi is the per-unit exposure dispersion. The corresponding weighted unit-loss quasi-likelihood has observation weight e_i, with the same goal as applying the appropriate exposure-scaled distribution to Y_i. The exact software objective convention should be tested rather than inferred from its name.
For p in (1,2), the unit-response Tweedie deviance is
D_T(r,m;p)=2{r^(2-p)/[(1-p)(2-p)] - r m^(1-p)/(1-p) + m^(2-p)/(2-p)}.
The limit r=0 is well defined in this range. Suppressing terms independent of m, the per-observation loss for margin F=log m is e[m^(2-p)/(2-p) - r m^(1-p)/(1-p)]. Its first derivative is e[m^(2-p)-r m^(1-p)]; the second derivative is e[(2-p)m^(2-p) - (1-p)r m^(1-p)], positive for r>=0 and 1<p<2. This is an analytically useful form for assessing boosted-tree gradient and curvature calculations. An alternative absolute-total parameterization is possible, but its dispersion and weights must agree with the exposure convention.
If phi is constant, the conditional probability of zero aggregate loss under the Tweedie compound Poisson mechanism is exp[-e m^(2-p)/(phi(2-p))]. The exponent changes with exposure. One can therefore match mean loss while miscalibrating zero frequency and the positive tail if the assumed p, phi, or claim mechanism is wrong. The gradient-tree Tweedie insurance study of Yang, Qian, and Zou (2018) fits the flexible nonlinear mean and estimates power and dispersion through a profile-likelihood procedure. Its method is a specific modeling design that must be replicated with the matching exposure conventions before performance comparisons are accepted.
Alternative targets, constraints, and software
A binary loss occurrence target is suitable when the business decision concerns P(Y>0|X), but claim counts and aggregate costs supply more information for ratemaking. Quantile loss addresses conditional percentiles rather than a mean. Huber-type losses reduce sensitivity to unusually large observations while changing the estimand toward a robust location. A log-transformed severity fitted by squared error estimates E[log Z|X], whose simple exponential is not E[Z|X] without a distributional correction. Even for a correct log-mean model, exposure and contract economics remain independent of a software optimizer's naming conventions.
The public LightGBM parameter reference lists Poisson, Gamma and Tweedie objectives, explicitly identifying Gamma for positive claims and Tweedie for possible total-loss modeling. That reference describes supported options, not insurer-specific predictive validity. Monotonic constraints can encode a substantively justified direction for an effect, but they should not be imposed solely because a variable name looks intuitively ordered. Medical age, vehicle age, policy tenure, risk scores and coverage limits may have nonlinear exposure and selection mechanisms that change the direction of an observed association.
Data encodings also create a statistical problem. Categorical variables such as broker, postal unit, vehicle model, diagnostic code or commercial class can contain many levels, some rare or absent from later periods. Target encoding based on the entire dataset leaks outcome information into validation rows. Ordered or out-of-fold encoding and a defined unseen-category policy are necessary; the advantage of CatBoost's categorical treatment should be evaluated rather than assumed under every portfolio design. Finally, software reproducibility requires recording package versions, loss configuration, split rules, data extraction date, and all seed or determinism settings.
Exposure-Weighted Objectives
Exposure as sampling intensity
Insurance contracts contribute unequal opportunities for insured events. A six-month automobile policy, a two-year policy, and a fleet endorsement may cover different fractions of an actuarial exposure unit. Reporting every row as one policy conceals these differences. Exposure must be defined in relation to the insured event: vehicle-years for vehicle claims, insured-life years for mortality, property-years for relevant property coverage, or an appropriate hazard-specific measure. A contract may change deductibles, coverage, limits, insured value, territory, or risk characteristics during its lifetime. Splitting the record into homogeneous time intervals preserves the interpretation of both the predictors and the denominator.
The proportional-intensity hypothesis is E[N_i|X_i,e_i]=e_i lambda(X_i). It is an assumption about the data-generating process, not a tautology. A newly written policy may exhibit acquisition effects, a catastrophe season may concentrate hazard within a fraction of the year, and policyholders may select different durations in response to impending risk. Where proportionality fails, duration and calendar-time effects, survival or event-history models, or finer exposure splits may be required. Setting e_i=1 for every policy and interpreting the result as an annualized rate is therefore an unsupported modeling shortcut.
An observation with e_i=0 has no log exposure offset; it belongs in a separate quality-control reconciliation unless the zero is a recording convention that can be accurately repaired. Negative exposures, reversal transactions, and amendments should be investigated before model construction. Portfolio predictions use the exposure at the actual decision horizon, including endorsements that alter risk over time. The framework in de Jong and Heller's insurance-data treatment is useful precisely because it requires a statistical response to correspond to the insurer's risk unit.
Offsets and rate-weighted likelihood
For count N_i with known e_i>0 and intensity lambda_i=exp(F_i), the conditional Poisson negative log likelihood is e_i lambda_i-N_i log(lambda_i) plus terms independent of F_i. This can be implemented by predicting the total count with fixed log(e_i) offset or by fitting the annualized rate r_i=N_i/e_i under exposure-weighted Poisson deviance. Multiplying the per-unit rate loss e_i lambda_i-N_i log(lambda_i) by constants common to all observations does not change its minimizer. For appropriate implementations, the two views are algebraically equivalent at the level of the objective.
The equivalence must be verified in the exact model library. In generic tree boosting, a sample weight may scale gradients and Hessians while a base margin may shift the current prediction; an offset that is meant to remain fixed across all boosting iterations is a distinct contract. Some libraries allow initial score vectors but alter their treatment during prediction, refitting or early stopping. A model that sets target N_i/e_i without exposure weights gives an extremely short policy the same influence as a full-year policy; a model that weights N_i rather than its likelihood also changes the stochastic problem. The public XGBoost parameter specification identifies objectives and parameters, while the desired equivalence follows from the derivation above and requires an implementation test.
For Gamma severities, the exposure of the parent policy is generally not an additional weight for every individual positive claim. A policy with many realized claims already generates several positive-severity observations, so blindly adding policy duration weights may count exposure twice for a claim-level conditional-mean estimand. Conversely, when modeling policy-level aggregate loss rates R_i=Y_i/e_i under an exposure-indexed Tweedie variance, e_i supplies the precision weight. A rigorous model specification records the unit of observation, the sampling mechanism, the likelihood normalization, and the prediction returned by the software.
The distinction matters in the presence of policy limits and deductibles. Y_i is the amount actually covered under the defined contract, whereas the latent ground-up loss may differ. Models for changes in deductible or coverage need to distinguish exposure to events from exposure to retained or insured monetary loss. A high excess can generate an observed zero even when an economically material accident occurred. Thus, zero claim payments are a mixture of event absence, coverage restrictions, and reporting or development mechanisms.
The compound model and the zero-claim mass
A compound Poisson representation can be written N_i|X_i,e_i ~ Poisson(e_i lambda_i), Y_i=sum_{j=1}^{N_i} Z_ij, with positive severity marks drawn under a specified dependence structure. If event amounts are conditionally identically distributed with mean s_i and independent of N_i given X_i, then E[Y_i|X_i,e_i]=e_i lambda_i s_i. Relaxing independence changes the decomposition and may require conditioning severity on claim-level circumstances or using a direct aggregate-loss model. The variance of a Poisson sum is e_i lambda_i E[Z^2|X_i] under the independent-mark assumptions; financial tail exposure therefore depends heavily on severity dispersion, even if frequency is stable.
The compound Poisson-Gamma member of the Tweedie family with 1<p<2 gives a parametric way to represent both a point mass at zero and positive loss. Its mean and variance structure permits modeling aggregate expected cost per risk unit with flexible m(X). The mathematical exposition in Taylor and McGuire's Tweedie analysis also illustrates the connection between compound claim processes and actuarial aggregate data, although reserve-development aggregates and contract-level pure premiums are different target populations.
One must not infer that a small Tweedie deviance implies accurate large-loss capital projections. The Tweedie mean regression is a model of conditional expectation under a particular variance law, and rare extremes can dominate tail quantities beyond the resolution of the training sample. Testing predicted zero proportions, mean annual loss, conditional positive payments and upper-tail behavior identifies different failure modes. Losses arising from multiple catastrophe events or correlated insured risks violate an independent-claims account even if the marginal loss distribution appears plausible.
Worked mathematical examples
Consider four hypothetical policies, each with an observed period and consistently defined indemnity claim count and aggregate payments. These numbers have been chosen for transparent arithmetic and are not observations from an insurer or an estimated market benchmark.
| Policy | Exposure in years | Claim count | Aggregate paid loss | Annualized observed count | Annualized observed loss |
|---|---|---|---|---|---|
| A | 1.00 | 0 | 0 | 0.000 | 0.00 |
| B | 0.50 | 1 | 1,500 | 2.000 | 3,000.00 |
| C | 1.50 | 2 | 4,800 | 1.333 | 3,200.00 |
| D | 0.25 | 0 | 0 | 0.000 | 0.00 |
| Total | 3.25 | 3 | 6,300 | — | — |
The portfolio claim frequency is 3/3.25 = 0.9230769 claims per exposure-year. The mean positive-claim amount is 6,300/3 = 2,100, provided the three positive claim payments comprise all claim-level amounts contributing to that total. The pure premium is 6,300/3.25 = 1,938.4615 monetary units per exposure-year. The product 0.9230769 × 2,100 = 1,938.4615, reproducing the observed aggregate loss rate. This calculation is an accounting identity in the sample; it does not establish conditional independence of claim frequency and severity in the underlying risk population.
By contrast, the unweighted average count per policy equals 3/4 = 0.75 and measures a different quantity because policies have unequal durations. The unweighted average of nonzero annualized paid-loss rates equals (3,000+3,200)/2 = 3,100 and exaggerates the aggregate exposure-based pure premium because it conditions on positive losses and gives a short policy the same standing as a long policy. Even the average of all four annualized loss rates, 1,550, differs from the exposure-weighted portfolio rate. These divergences arise before any machine-learning method is applied.
Now assume a constant annual frequency lambda_0=3/3.25 and compare it with illustrative annual predictions (0.4,1.2,1.1,0.4) for policies A–D, respectively. Their predicted total expected counts are, in order, (0.4,0.6,1.65,0.1), while the constant baseline predicts (0.9231,0.4615,1.3846,0.2308). Using the count-level Poisson deviance summed over these four rows yields approximately 3.0173 for the baseline and 1.2911 for the illustrative alternative. The smaller figure is an arithmetic demonstration of the deviance calculation on the displayed records. Because the alternative predictions are hypothetical and assessed on the same four displayed outcomes, the comparison cannot establish out-of-sample advantage. A predictive claim requires independently held-out data, comparable training procedures and uncertainty analysis.
Finally consider an insurer forecasting expected annualized claim cost m=1,000 for a policy lasting e=0.25 year. Its expected aggregate covered loss is e m=250, not 1,000. If a Tweedie model yields a predicted annualized mean m with dispersion phi and power p, the probability of an exact zero for the quarter-year contract depends on the compound-event intensity through e and cannot be copied from the annual policy unchanged. This distinction is especially consequential when many short-duration endorsements are present.
Data integrity, censoring, and claim development
Claims records are generated by business operations as well as by insured hazards. A payment, reserve transaction, recovery, reopening, and claim-status change may each appear as a separate row. Summing such transactions without a transaction taxonomy can duplicate payments, reverse their signs incorrectly, or count reopened incidents as new claims. Policyholder, claim and exposure joins require stable keys and an auditable event definition. Currency conversion and inflation adjustment must specify the index, valuation date, jurisdiction and treatment of exchange-rate effects.
Historical claim severity often has incomplete development. Reporting delay produces left or right truncation relative to extraction rules; settlement produces censoring, and paid-to-date payments are incomplete outcomes for long-tailed liability risks. Development-adjusted ultimate losses are themselves estimated quantities, with estimation error that may correlate with accident period and claim complexity. A boosting model trained on developed amounts should document the development model, its information cutoff, revisions and uncertainty. Randomly splitting later versions of the same claim between training and validation is leakage, even when the record identifiers differ.
An evaluation intended for new-business pricing must use only attributes available before inception. A claims-handling forecast may legitimately use early claim notes and reserve estimates, but then its target, user and decision time have changed. The insurer must also inspect the policy's selection mechanism: quoting, binding, cancellation and renewal may all alter the covariate distribution and the observed count process. This is a foundational source of dataset shift in insurance, and cannot be repaired merely by increasing the number of trees.
Tuning and Early Stopping
Capacity, depth, shrinkage, and regularization
Boosted trees have several interacting complexity controls. Maximum depth or leaf count restricts the order and fineness of interactions. Minimum child weight, minimum samples per leaf or related Hessian thresholds restrict splitting where claims or exposure provide little information. Learning rate reduces the stagewise update; tree count sets the total number of steps. L1/L2 leaf penalties, split gain thresholds, row sampling and feature sampling further alter the fitted function. Their precise names and semantics vary among implementations.
A small learning rate with more boosting rounds can represent an effective optimization path, but the combination must be selected against validation loss. There is no universally optimal tree count, maximum depth, or Tweedie power. Short-tail standardized personal-lines data may support finer partitions than a sparse commercial portfolio with limited exposure in each risk class. Tree depth also changes the interpretive questions: a shallow additive structure is simpler to audit than repeated high-order interactions, but an underspecified structure can systematically misprice known combinations of risk characteristics.
Friedman's stochastic boosting method motivates controlled subsampling as one regularization strategy. XGBoost additionally incorporates explicit penalties and split approximations in its scalable boosting implementation. LightGBM's leaf-wise growth can achieve high flexibility at a given leaf budget, but it changes the risk of creating deep local paths; the LightGBM documentation must be interpreted together with the actual sample size and exposure distribution. CatBoost's categorical handling may reduce particular forms of target-encoding bias while leaving the general need for temporal and portfolio-level validation intact.
For insurance, effective sample size depends on event count and exposure concentration in addition to row count. Suppose a leaf contains 10,000 annual policy exposures but no insured events; it can establish a useful upper bound on frequency under a Poisson assumption. Another leaf with 10,000 rows covering only 100 exposure-years contains much less information. If a third leaf is dominated by one enormous payment, its fitted loss mean may be highly sensitive to one record. These are different estimation problems requiring objective-sensitive regularization and stability diagnostics.
Chronological, grouped, and nested validation
A defensible validation design begins by specifying the prediction decision date. Training data contain only information available by that date; tuning data occur later or represent an appropriately separated risk sample, and an untouched test set estimates final generalization error. In an operational forecasting exercise, chronological separation is usually essential. A random row split may put the same policyholder, insured property, claim, vehicle, event or catastrophe into both training and test samples, rewarding the model for memorizing shared information.
Group keys should reflect the mechanisms by which outcomes are dependent. A property portfolio may require grouping by catastrophe event and geographic cluster, a fleet portfolio by business and vehicle group, and claim-level severity records by policy and event. Where the intended deployment includes repeat renewals, the test design should distinguish generalizing to new periods for known policyholders from generalizing to previously unseen insureds. Both can be relevant, but the former should not stand in for the latter. Geographic holdouts may be necessary to evaluate transportability to a newly entered region, though a purely geographic split can also create distribution shifts beyond the operational use case.
Hyperparameter tuning and early stopping should occur within a training/tuning design insulated from the final test set. A nested resampling scheme gives an outer evaluation of the full model-selection procedure and an inner procedure for selecting hyperparameters, where data volume permits. The final test set should be accessed once for the declared candidate and its prespecified comparisons. Repeated changes to encoding, features, tail treatment, or evaluation metrics after seeing test results consume information from the held-out sample and bias the reported estimate of performance.
Temporal validation must also handle delayed loss maturity. Let a 2024 accident-year test population be evaluated in mid-2025 and again in late-2026. These are not the same realized severity labels for late-reported or unresolved claims. The analyst must either use a sufficiently mature evaluation cohort, explicitly model development, or state that the score concerns an early-valuation target. Out-of-time backtesting should reconstruct the actual information set at each historical decision point rather than reading today's enriched data and pretending those attributes existed years earlier.
Metrics and actuarial calibration
The training loss defines a statistical estimand and should be aligned with the principal assessment metric. For count models, Poisson deviance measures fit to a count mean under a Poisson likelihood and penalizes poor rate estimates. A frequency model's annualized rates should be multiplied by exposure before deviance is calculated on counts. For positive severity, Gamma deviance, log score under an explicit severity density, mean absolute error, or a suitably tailored tail score may be informative depending on the decision. For pure premiums, exposure-weighted Tweedie deviance and aggregate calibration are natural candidates when the compound model is appropriate. A ranking metric alone does not certify monetary calibration.
Proper scoring rules encourage reporting the predictive distribution of a quantity when the full distribution is the object of interest. The strict propriety analysis of Gneiting and Raftery (2007) explains why honest probabilistic forecasts must be evaluated by criteria matched to the forecast object. A deterministic mean forecast under squared-error loss and a full loss distribution evaluated by log score answer distinct questions. In actuarial production, the choice should be explicit, particularly where tail losses, aggregate counts and zero-event probabilities each matter.
Calibration must be examined at both portfolio and segment levels. For expected counts, an overall expected-to-observed ratio compares sums of expected and actual counts, but it may hide opposing errors across young and older vehicles, territories, deductibles, industries or seasons. For loss costs, compare sums of predicted aggregate losses e_i m_i with sums of appropriately valued actual losses Y_i, and report uncertainty alongside them. Exposure-weighted decile plots can make risk differentiation visible, but the deciles must be constructed using predicted risk without contaminating results with future outcomes. Credibility becomes weak in thin tails; a segment-level ratio based on two severe claims is not equivalent to a precisely measured systemic bias.
The publicly reproducible scikit-learn Poisson example contrasts models of claim counts with appropriate distributions and exposure handling. Its companion insurance Tweedie example shows frequency, severity and aggregate objectives on motor-insurance data. These are educational implementations with a particular dataset and pipeline; their scores should not be restated as universal benchmarks for insurers or interpreted as validations of datasets with different contractual meanings.
Model comparison and selection uncertainty
An evaluation set should compare boosting with relevant alternatives under the same target and exposure construction: an intercept-only baseline, a reasonable actuarial GLM, a generalized additive model where appropriate, a tree ensemble, and a suitably structured frequency-severity or direct pure-premium formulation. The baseline is a scientific control. A complex model may lose to an offset-correct GLM when risk information is weak or when the test distribution differs from training. A favorable result against an under-tuned alternative demonstrates less than a comparison in which both models receive proportionate optimization budgets.
Comparisons need uncertainty estimates respecting data dependence. Ordinary row-bootstrap confidence intervals can understate uncertainty when many rows share one catastrophe or policyholder. Group-level or time-block resampling, suitably designed likelihood comparisons and sensitivity analysis across time windows provide more useful evidence. Tail contamination, inflation, reporting lag, and policy cancellations should be prespecified sensitivity dimensions. An apparent 0.5% improvement in average deviance may be operationally fragile when sampling variability and loss development are large. Improvement should be assessed on a meaningful business scale and after recognizing its uncertainty.
Reproducibility and deployment handoff
A reproducible experiment fixes a dataset snapshot, target formula, exposure definition, claim handling rules, feature availability timestamps, extraction code, train-validation-test allocation, software versions, random seeds, preprocessing, parameter search ranges, selection metric and final scoring script. Saving only a fitted model file does not preserve the experiment. A secure deployment also requires versioned transformations for unknown categories, missing values, changing coverages, categorical dictionaries and recalibration, together with a fallback when incoming data breach assumed ranges.
An experiment ledger should record model challengers that were tried and discarded as well as the selected winner. This guards against selective reporting of a favorable validation period. After deployment, monitoring must track the distribution of inputs and exposure, event frequencies, monetary loss calibration, claim maturity, prediction stability and the operational interventions triggered by model output. Retraining and change controls should require new backtests rather than automatically repeating the original tuning recipe on changed data.
Insurance Loss Prediction
Motor, property, health, and commercial settings
Motor insurance commonly offers large policy counts, repeated annual exposures and relatively structured rating attributes. The French motor-liability datasets have supported numerous methodological comparisons; their public existence does not imply that they contain all features needed for an insurer's contemporary underwriting or that the claims remain representative of current inflation and legal rules. French motor third-party liability is one contract and jurisdictional setting. Its statistical findings require direct validation before extension to another country's property or health market.
Property insurance can contain a higher share of severe catastrophe-driven losses, correlated spatial hazards, building attributes, reconstruction-cost inflation and coverage-specific deductibles. A boosted mean-loss model may learn useful nonlinear interaction patterns from exposure and hazard variables, but geographically correlated catastrophes reduce the effective number of independent extreme events. A split between training and test that shares the same storm or wildfire footprint can substantially overstate event-level transferability. The technical response is event-aware validation and, when warranted, explicit catastrophe and spatial models connected to the portfolio's exposure base.
Health insurance may involve repeated encounters, treatment costs, eligibility windows and changes in morbidity associated with service use. The relevant response could be paid plan cost, incurred claim cost or allowed cost; each belongs to a different benefit and contractual definition. Features arising after a patient's claim are inappropriate to a precoverage forecast. Commercial insurance introduces larger insured-specific heterogeneity, limits, deductibles, claims-made triggers, policy layering and long-tailed liability development. The best statistical target may be a frequency–severity decomposition, a direct aggregate response, a censored-loss model, or a hierarchical system with stable portfolio covariates. The gradient booster must be adapted to this target rather than treated as a generic claims-prediction machine.
Frequency-severity and direct pure premium prediction
There are two main pathways to expected aggregate losses per exposure unit. The first estimates frequency lambda(X) and conditional severity s(X) and forms m(X)=lambda(X)s(X), provided the decomposition's conditional definitions make the identity valid. The second directly estimates m(X)=E[Y/e|X] with a compound Poisson-Tweedie or another nonnegative regression objective. In a homogeneous, correctly specified joint distribution these approaches target the same conditional mean. In finite samples with distinct objectives, feature sets, smoothing structures, zero mechanisms and claim dependencies, the fitted predictions can diverge substantially.
The two-part route offers decomposable diagnostics. A rise in pure premium can be attributed to a modeled frequency effect, severity effect or their interaction. Yet severity is trained on the selected population of claims, and large-claim mixtures may require explicit truncation or a flexible marked-event model. A direct Tweedie approach avoids multiplying two separately estimated conditional quantities but entails a particular variance–mean relationship and zero-mass structure. Its power p is a substantive modeling parameter that should be estimated or tuned using training data and evaluated on held-out outcomes.
The study by Yang, Qian, and Zou (2018) on gradient-boosted Tweedie premiums provides published support for the direct approach and its insurance motivation. It does not establish equivalence to every frequency–severity design or dominance across all portfolios. Conversely, Henckaerts and colleagues (2021), in their insurance tariff comparison, examine tree-based methods in tariff estimation while acknowledging the distinction between prediction and usable insurance tariff plans. A deployable tariff may need a clear rate structure, constraints, smoothing or transformation separate from the unconstrained learned surface.
Where claims are censored at policy limits, loss cost modeling must specify whether the estimand is observed insurer payment under that limit or an underlying ground-up severity. Using an uncensored Gamma response for capped claims can falsely suggest a thin distributional tail; a large-loss model based on capped payments will not automatically price a higher limit without contract adjustments. Similarly, a frequency model that ignores changes in deductibles may predict claim reporting or payment incidence rather than physical event incidence. Explicitly separating these objects is an essential actuarial requirement.
Large losses, heterogeneity, and distribution shift
Loss distributions typically contain many zeros and a minority of large observations. While squared-error boosting may concentrate on the largest payments, an aggressively robust loss can suppress costs that are economically essential to an expected indemnity forecast. Winsorization, exclusion of catastrophe amounts or large-loss caps must be identified as changes to the target, followed where necessary by an explicit tail-load or separate severity component. The result should reconcile to the original ultimate covered-cost definition. A common source of pricing distortion is evaluating an attritional model on capped payments and then reporting its predictions as if they were full uncapped insured losses.
Structural changes may affect both the covariates and the conditional loss distribution: climate and catastrophe conditions, judicial and medical cost inflation, technology of vehicles or buildings, fraud schemes, underwriting selection, insured behavior, repair networks and changes in reporting incentives. A monitored model therefore needs a clear retraining trigger and an evaluation protocol comparing the new population with the historical reference. Distribution shift can appear as degraded monetary calibration despite stable conventional discrimination metrics. It may also produce errors concentrated in a newly acquired segment, even when portfolio-wide ratios look acceptable.
High-cardinality identifiers such as broker, employer, medical provider or region deserve attention. They may encode legitimate risk structure, historically contingent operational behavior, or protected-attribute proxies. A variable can improve predictive fit while transporting poorly to a new insurer or period. Out-of-fold target encoding and ordered boosting safeguards address one mechanism of leakage, but governance still must examine whether a particular grouping is explainable, stable, legally permitted and suitable for consumer-facing decisions.
Prediction uncertainty and operational consequences
Point predictions conceal multiple kinds of uncertainty. Parameter and model-selection uncertainty arises from limited data; intrinsic process randomness remains even with a known conditional mean; future portfolio composition and distribution shift introduce further uncertainty. A mean-only gradient booster supplies no automatic confidence interval, predictive interval or capital tail. Bootstrap or resampling methods must respect clustered claims and temporal dependence. A probabilistic model can quantify count and severity distributions only to the extent that its distributional assumptions are supported.
For an individual short-duration policy, the zero outcome may be the most probable even when the expected loss is positive. Forecasting a mean of 2,000 monetary units per year does not mean a claim of that exact amount is likely. For a portfolio, aggregation of risks introduces dependence: regional disasters, supply-chain shocks and systemic claim inflation can make independent-claim simulation materially understate tail outcomes. Expected loss models may contribute to solvency analysis, but risk aggregation and capital requirements demand an appropriately dependent stochastic loss model.
Operational use also creates feedback. A model may influence the price quoted, coverage selected, claim investigations initiated, or business accepted. Those actions alter which risks remain in the subsequent dataset. Updating the model as if its outcomes came from an unchanged random population can reinforce selection bias. Responsible evaluation records the decision policies through which predictive scores enter operations, as well as the outcomes observed after those interventions.
The empirical literature and its limits
The French motor comparative study by Noll, Salzmann, and Wüthrich (2020 revision) compares actuarial predictive methods on a particular public motor-liability claims problem. Its value lies in providing a concrete comparison design; its conclusions are bounded by the sample, features, endpoints and model implementations chosen. The study cannot be cited as a randomized multi-country insurance evaluation. Separate empirical findings must be compared on normalized loss definitions and matched validation methods before quantitative results are pooled or generalized.
Henckaerts, Côté, Antonio, and Verbelen (2021) study tree-based tariff techniques and provide methodological evidence about interpretability and insurance pricing. Yang, Qian, and Zou (2018) examine Tweedie tree boosting for claim cost with a compound-loss objective. These works address overlapping methods, but they do not constitute interchangeable tests of the same algorithm on the same response, exposure and calendar population. A responsible synthesis uses them to motivate testable procedures, not to announce a cross-market accuracy ranking.
The Casualty Actuarial Society's 2026 comprehensive pricing monograph situates nonlinear modeling among a broader set of insurance pricing tasks: feature engineering, interactions, high-cardinality variables, validation and combining different actuarial methods. Its status as a later professional monograph is particularly useful for placing gradient boosting within contemporary insurer workflows, while any specific performance claim still requires its own underlying dataset and experimental result. No numerical gain or regulatory acceptance is inferred merely from the publication of these methods.
Interpretability and Overfitting
Interpreting nonlinear tariffs
A boosted tree ensemble represents a fitted nonlinear mapping from known covariates to a margin and then, under a specified link, to predicted expected claim frequency, severity or loss. An explanatory account must preserve the link and exposure transformation. An increment of 0.1 on a log-rate margin corresponds to multiplication of the rate by exp(0.1), approximately 1.1052, while the effect on an aggregate count also depends on exposure. Showing such a quantity as an absolute currency increment without applying the link would misrepresent the prediction.
Global interpretation asks which variables contribute to prediction across a portfolio and how effects vary by subgroup. Local interpretation asks why a model assigned a particular prediction to a specified record relative to a defined baseline. These are complementary inquiries. An effect plot can expose nonlinearity and interaction, yet its interpretation depends on the distribution of the remaining covariates and the way the plotting method averages them. A model may heavily rely on a proxy variable that has no stable causal relationship with insured loss. The Elements of Statistical Learning provides the broader statistical context for model complexity, fitting and predictive interpretation.
Model reliance, dependence, and explanation
Feature importance derived from split count or split gain measures how the algorithm used variables during fitting; it does not directly measure causal importance or fairness. Permutation importance measures a model's predictive reliance under a given perturbation scheme, but unconditional shuffling can produce implausible predictor combinations where variables are strongly correlated. Alternative conditional importance methods have been analyzed, including correlation-sensitive permutation analysis by Strobl and colleagues (2008) in random forests. The underlying problem of dependent features extends to supervised prediction more broadly, although an RF-specific estimator should not be represented as automatically valid for boosted models.
Lundberg and Lee's (2017) unified additive attribution method motivates SHAP-style explanations. Attribution values depend on the definition of the background distribution and the conditional or interventional interpretation of feature coalitions. A purported individual 'reason' is therefore a decomposition relative to a reference distribution, not a direct physical cause of a loss. Explainers should disclose whether they operate on the raw margin, transformed mean, or final premium after business adjustments; decomposing one and labeling it as another is an interpretive error.
Accumulated local effects, discussed by Apley and Zhu (2020) in their model-visualization research, provide another way to summarize fitted feature influence, often with advantages over partial dependence when variables are correlated. Nonetheless, an ALE plot remains a description of the fitted model under an empirical covariate distribution. It does not establish what would happen if an insurer deliberately changed an insured characteristic, or whether the predictor is permissible in a particular regulatory regime.
Explanations require model-specific and policy-specific validation. A local attribution that changes dramatically with minor, plausible data corrections may be unreliable for a customer appeal. A globally smooth plot may conceal local instability for thinly represented risk classes. Explanatory artifacts must be versioned with the exact fitted model, transformation pipeline, reference population and decision rules. The insurer should be able to reconstruct the score that was actually used in a consequential decision.
Robustness, fairness, and insurance governance
Overfitting appears when a model captures chance irregularities, artifacts of data processing, rare-category encodings, historical operational decisions or information that would have been unavailable at deployment. Some overfitting is revealed by a rising out-of-time loss despite decreasing training loss; other forms remain hidden in average metrics but appear in specific customer groups or catastrophe periods. Stability analysis should therefore combine temporal backtesting, grouped tests, sensitivity to preprocessing, plausible stress scenarios and review of the business meaning of learned interactions.
Fairness in insurance prediction is closely connected to the rating law and contractual principles in the jurisdiction concerned. A risk factor can be predictive and still be prohibited, restricted, operationally indefensible, or a proxy for a protected characteristic. Removal of a named characteristic is insufficient if the remaining variables reproduce its effect. A model review should test disparate calibration and outcome patterns, investigate high-impact proxies, document allowed rating variables and assess legal rules for automated decision-making. Statistical parity is not an automatic actuarial or legal mandate; legitimate risk differentiation, consumer rights, local prohibitions and contractual protections must be analyzed together.
For the European insurance sector, the European Insurance and Occupational Pensions Authority's 2025 AI governance and risk-management opinion clarifies supervisory expectations concerning systems within its stated scope, including data governance, fairness, explainability and human oversight. The scope of that opinion and the EU AI Act's separate legal provisions should be read precisely. In the United States, the National Association of Insurance Commissioners' 2023 AI Systems Model Bulletin gives model guidance on governance, compliance and documentation. Adoption by the NAIC does not itself make that model bulletin binding law in every state. Both instruments are relevant to governance, but their legal force and addressees differ.
The National Institute of Standards and Technology's AI Risk Management Framework, issued in 2023, offers a voluntary general-purpose framework for managing AI risk and designing accountable organizational processes. For insurance loss prediction, this supports a practical division of duties: actuarial staff own the target and loss assumptions, data engineers establish lineage and feature cutoffs, validation staff independently examine performance, legal and compliance staff assess permitted uses, and governance bodies control deployment and remediation. The listed responsibilities are an analytical application of ordinary model-governance practices, not an official checklist quoted from a particular instrument.
Internal analytical synthesis and research agenda
The preceding mathematical and documentary analysis yields five connected conclusions. First, the insurance risk unit precedes the algorithm: a frequency rate, a severity conditional mean and an aggregate pure premium must be defined by claim event, policy duration, coverage and monetary valuation before training begins. Second, the objective carries an actuarial assumption. A Poisson rate objective requires a count-intensity interpretation; a positive Gamma response requires a properly selected severity sample; and a Tweedie loss rate represents a compound nonnegative response with a specific mean–variance relation. Third, predictive validation must recreate the decision environment, including exposure, claim maturity, event dependence and policyholder selection. Fourth, interpretability is meaningful only with the model link, reference population, dependence assumptions and decision context specified. Fifth, production quality is established by monitored calibration, permissible use and traceable governance, alongside statistical performance.
These conclusions are a synthesis of established statistical and actuarial reasoning. They do not assert a newly invented system, novel optimization algorithm, original named doctrine or unreplicated empirical breakthrough. The principal substantive contribution is their integration into an explicit chain of conditions under which a gradient-boosted estimate can be treated as a defensible insurance-loss prediction. If any condition is absent, the interpretation of the result must narrow to the quantity actually fitted and tested.
Several research questions remain open for prospective empirical work. How much of an apparent gradient-boosting advantage survives after equalizing the exposure treatment and development maturity of conventional GLMs? Which losses and rating classes benefit most from joint Tweedie fitting rather than independently fitted frequency and severity? How should event-grouped tests be designed when tail events are too few for stable historical holdout estimates? Which feature-attribution methods remain reliable when rating variables are correlated and policy limits induce target censoring? How rapidly does calibration decay after inflation, underwriting selection or regulatory changes? Answering these questions requires declared datasets, experimental protocols, confidence intervals and independent replication. The present Research Dossier identifies those designs without claiming to have executed them.
Conclusion
Gradient boosting supplies a flexible approach to approximating conditional insurance risk functions. Its distinctive mechanism is sequential functional optimization, commonly implemented through regularized tree partitions, stagewise derivatives, subsampling and early stopping. For actuarial applications, however, the decisive scientific choices concern the event definition, exposure measure, valuation horizon, conditional distribution, observation scheme and validation population. The same set of trees can have different meanings under a Poisson claim-frequency objective, a Gamma positive-severity objective and a Tweedie pure-premium objective. Correctly specifying that meaning is necessary for every comparison and downstream decision.
An exposure-adjusted expected claim count is e lambda(X); an exposure-adjusted expected total loss is e m(X). Under an appropriate frequency–severity decomposition, m(X)=lambda(X)s(X), while a direct aggregate response can be fitted under an adequately validated compound-loss objective. The worked four-policy example demonstrates how a change in exposure denominators affects estimated rates and how Poisson deviance can be computed without confusing an arithmetic illustration with evidence of future predictive accuracy. The example also makes clear why monetary predictions, observed loss rates and actuarially meaningful calibration depend on matching response units.
A credible gradient-boosting study uses independent chronological or operationally equivalent validation, protects against policy/event leakage, records claim-development assumptions, compares relevant actuarial baselines, tests segment calibration, quantifies uncertainty where possible and evaluates distribution shift. Explainability methods help examine the fitted mapping, but legal defensibility and causal interpretation require additional domain-specific analysis. Governance must retain the ability to reconstruct decisions and to correct unstable or unfair use of model outputs.
The reviewed academic and institutional sources support gradient boosting as a method worthy of serious insurance investigation, with documented applications to motor loss data and actuarial tariffs. They justify neither universal superiority over simpler models nor a transfer of performance claims across portfolios and jurisdictions. The scientific result of this monograph is a precise account of the assumptions, numerical relationships, controls and evidentiary conditions needed to evaluate gradient-boosted insurance loss predictions. Further empirical claims are reserved for independently documented experiments. The manuscript constitutes original analytical synthesis of established literature, with no newly instituted authorial theory or doctrine.
References
Apley, Daniel W., and Jingyu Zhu. 2020. “Visualizing the Effects of Predictor Variables in Black Box Supervised Learning Models.” Journal of the Royal Statistical Society: Series B (Statistical Methodology) 82 (4): 1059–1086. https://doi.org/10.1111/rssb.12377.
Chalk, Alan, David Deacon, Montserrat Guillén, and Max Martinelli. 2026. From GLMs to Comprehensive Insurance Pricing: Techniques and Challenges. CAS Monograph No. 16. Arlington, VA: Casualty Actuarial Society. https://www.casact.org/publications-research/publications/flagship-publications/cas-monographs/monograph-no-16.
Chen, Tianqi, and Carlos Guestrin. 2016. “XGBoost: A Scalable Tree Boosting System.” In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 785–794. New York: ACM. https://doi.org/10.1145/2939672.2939785.
de Jong, Piet, and Gillian Z. Heller. 2008. Generalized Linear Models for Insurance Data. Cambridge: Cambridge University Press. https://doi.org/10.1017/CBO9780511755408.
Dutang, Christophe, and CASdatasets contributors. n.d. “French Motor Third-Party Liability Data (
freMTPL).” CASdatasets: R Datasets for Actuarial Science. Project documentation. https://dutangc.github.io/CASdatasets/reference/freMTPL.html.European Insurance and Occupational Pensions Authority (EIOPA). 2025. “Opinion on Artificial Intelligence Governance and Risk Management.” August 6. Frankfurt am Main: EIOPA. https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en.
Friedman, Jerome H. 2001. “Greedy Function Approximation: A Gradient Boosting Machine.” The Annals of Statistics 29 (5): 1189–1232. https://doi.org/10.1214/aos/1013203451.
Friedman, Jerome H. 2002. “Stochastic Gradient Boosting.” Computational Statistics & Data Analysis 38 (4): 367–378. https://doi.org/10.1016/S0167-9473%2801%2900065-2.
Friedman, Jerome H., Trevor Hastie, and Robert Tibshirani. 2000. “Additive Logistic Regression: A Statistical View of Boosting.” The Annals of Statistics 28 (2): 337–407. https://doi.org/10.1214/aos/1016218223.
Gneiting, Tilmann, and Adrian E. Raftery. 2007. “Strictly Proper Scoring Rules, Prediction, and Estimation.” Journal of the American Statistical Association 102 (477): 359–378. https://doi.org/10.1198/016214506000001437.
Hastie, Trevor, Robert Tibshirani, and Jerome Friedman. 2009. The Elements of Statistical Learning: Data Mining, Inference, and Prediction. 2nd ed. New York: Springer. https://doi.org/10.1007/978-0-387-84858-7.
Henckaerts, Roel, Marie-Pier Côté, Katrien Antonio, and Roel Verbelen. 2021. “Boosting Insights in Insurance Tariff Plans with Tree-Based Machine Learning Methods.” North American Actuarial Journal 25 (2): 255–285. https://doi.org/10.1080/10920277.2020.1745656.
Ke, Guolin, et al. 2017. “LightGBM: A Highly Efficient Gradient Boosting Decision Tree.” In Advances in Neural Information Processing Systems 30. https://proceedings.neurips.cc/paper_files/paper/2017/hash/6449f44a102fde848669bdd9eb6b76fa-Abstract.html.
LightGBM Development Team. 2025. LightGBM Documentation: Parameters, version 4.6.0. https://lightgbm.readthedocs.io/en/v4.6.0/Parameters.html.
Lundberg, Scott M., and Su-In Lee. 2017. “A Unified Approach to Interpreting Model Predictions.” In Advances in Neural Information Processing Systems 30. https://proceedings.neurips.cc/paper_files/paper/2017/hash/8a20a8621978632d76c43dfd28b67767-Abstract.html.
National Association of Insurance Commissioners (NAIC). 2023. “NAIC Members Approve Model Bulletin on the Use of Artificial Intelligence Systems by Insurers.” December 4. Washington, DC: NAIC. https://content.naic.org/article/naic-members-approve-model-bulletin-use-ai-insurers.
National Institute of Standards and Technology (NIST). 2023. Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. Gaithersburg, MD: NIST. https://www.nist.gov/itl/ai-risk-management-framework.
Noll, Alexander, Robert Salzmann, and Mario V. Wüthrich. 2020. “Case Study: French Motor Third-Party Liability Claims.” Revised working paper. SSRN preprint no. 3164764. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3164764.
Prokhorenkova, Liudmila, et al. 2018. “CatBoost: Unbiased Boosting with Categorical Features.” In Advances in Neural Information Processing Systems 31. https://proceedings.neurips.cc/paper_files/paper/2018/hash/14491b756b3a51daac41c24863285549-Abstract.html.
Scikit-learn Developers. n.d. “Poisson Regression and Non-Normal Loss.” Scikit-learn User Guide: Examples. https://scikit-learn.org/stable/auto_examples/linear_model/plot_poisson_regression_non_normal_loss.html.
Scikit-learn Developers. n.d. “Tweedie Regression on Insurance Claims.” Scikit-learn User Guide: Examples. https://scikit-learn.org/stable/auto_examples/linear_model/plot_tweedie_regression_insurance_claims.html.
Strobl, Carolin, Anne-Laure Boulesteix, Thomas Kneib, Thomas Augustin, and Achim Zeileis. 2008. “Conditional Variable Importance for Random Forests.” BMC Bioinformatics 9:307. https://doi.org/10.1186/1471-2105-9-307.
Taylor, Greg. 2009. “The Chain Ladder and Tweedie Distributed Claims Data.” Variance 3 (1): 96–104. https://variancejournal.org/article/142013-the-chain-ladder-and-tweedie-distributed-claims-data.
XGBoost Contributors. n.d. “XGBoost Parameters.” XGBoost Documentation. https://xgboost.readthedocs.io/en/stable/parameter.html.
Yang, Yi, Wei Qian, and Hui Zou. 2018. “Insurance Premium Prediction via Gradient Tree-Boosted Tweedie Compound Poisson Models.” Journal of Business & Economic Statistics 36 (3): 456–470. https://doi.org/10.1080/07350015.2016.1200981.