Joint Distributional Copula Regression for Insurance Claim Frequency and Severity via Boosting

1 Introduction

• Actuarial distributional regression remains relatively limited, and separate frequency and severity models can overlook dependence between the outcomes.The paper notes that pricing and risk management require understanding claim variability and tail risk as well as expected claims.Source: §1 · p. 2
• The proposed GAMLSS-based approach links parameters of both marginal distributions and the copula dependence parameter to structured additive predictors.Component-wise boosting selects base learners and regularizes effects through small step lengths and early stopping.Source: §1 · p. 3
• The framework is evaluated on simulated and telematics data for effect and dependence recovery, variable selection, predictive performance, and model interpretation.Its additive predictors are presented as directly inspectable, supporting transparency and auditability in regulated actuarial settings.Source: §1 · p. 3

2.1 Model structure

• The framework represents a bivariate conditional distribution by combining parametric marginal distributions with a copula that separates margins from dependence.With a discrete margin, the copula representation still defines a valid joint distribution, although it is unique only over the ranges of the marginal CDFs.Source: §2.1 · p. 4
• Each marginal and dependence parameter has a structured additive predictor, and the covariate subset may differ across parameters.Link functions keep parameters within their admissible ranges; considered dependence structures include Gumbel, Frank, and Gaussian copulas.Source: §2.1 · pp. 4–5

2.2 Application to discrete–continuous responses

• For a discrete–continuous response, the joint likelihood takes a finite difference over the discrete margin and differentiates with respect to the continuous margin.The resulting expression applies to copulas with a partial derivative in the continuous margin’s argument.Source: §2.2 · pp. 5–6

3.1 Boosting algorithm for joint distributional copula regression

• The boosting procedure minimizes empirical risk using negative log-likelihood and parameter-specific base learners fitted to loss gradients.At each iteration, gradients are evaluated at the previous predictor estimates and serve as pseudo-observations for fitting candidate learners.Source: §3.1 · p. 7
• Noncyclic boosting updates only the parameter-specific predictor whose best-fitting base learner gives the lowest candidate empirical loss.This across-parameter selection step also yields inherent variable selection; the algorithm is outlined in Figure 1.Source: Figure 1 · p. 7, §3.1 · p. 8
• The stopping iteration is selected where validation-set out-of-bag risk reaches its minimum.The dataset is split into training and validation sets to assess predictive risk during fitting.Source: §3.1 · p. 8

3.2 Negative gradients of the implemented loss function

• For each additive predictor, the negative gradient of the negative log-likelihood is the corresponding partial derivative of the log-likelihood.The method derives these gradients for marginal and copula predictors in the discrete–continuous setting.Source: §3.2 · pp. 8–9

4 Simulated experiment

• The simulation compares prediction, effect and dependence recovery, and variable selection against separate-response benchmarks that assume independence.Source: §4 · p. 9

4.1 Data generation

• The data-generating process links Poisson and Gamma margins with a Frank copula, using seven informative covariates among p = 10, 100, or 1000 predictors.Each setting uses 2,000 observations split into training, tuning, and test sets, with 10 independent datasets.Source: §4.1 · pp. 9–10

4.2 Estimation and evaluation setup

• B-spline base learners allow linear and nonlinear effects, while out-of-bag risk on the tuning sample selects the stopping iteration.The comparisons use negative log-likelihood and energy score; independent models’ joint negative log-likelihood is the sum of marginal values.Source: §4.2 · p. 10

4.3 Results

• Across all three dimensional settings, Frank copula boosting has the lowest negative log-likelihood and energy score, followed by univariate boosting.The univariate MLE model could not be estimated at p = 1000.Source: Table 1 · p. 11, §4.3 · p. 11
• For p = 10, the correctly specified Frank copula has the best predictive accuracy, while Gaussian and Gumbel alternatives outperform independence benchmarks on negative log-likelihood.Their energy scores remain competitive with the univariate boosting and MLE models.Source: Table 2 · p. 11, §4.3 · pp. 11–12
• Estimated effects generally track the true effects across parameters, though some nonlinear estimates are attenuated and the X1 curve flattens at higher values.Linear effects for X2, X3, and X4 are recovered with high accuracy; Figure 2 contrasts true effects with estimates across 10 replications.Source: Figure 2 · p. 12, §4.3 · pp. 12–13
• The fitted conditional Gamma distributions reflect covariate-linked changes in both mean and spread, including nonlinear dispersion patterns for X5 and increasing spread with X6.Figure 3 displays conditional CDFs and quantiles as selected covariates vary.Source: Figure 3 · p. 13, §4.3 · pp. 13–14
• Across p = 10, 100, and 1000, informative covariates rank among the most frequently selected, including when p equals the 1,000 training observations.Some noise variables enter the top 10, but their selection rates remain markedly lower than those of informative variables.Source: Figure 4 · p. 14, §4.3 · p. 14

5.1 Data and Model Specification

• The analysis uses synthetic telematics data with 100,000 observations and 52 variables, Poisson and Gamma margins, and non-cyclic boosting.The candidate set contains five categorical and 22 numerical covariates; the frequency predictor also includes log policy exposure as a fixed offset.Source: §5.1 · p. 15, §5 · p. 15

5.2 Empirical Results

• All three copula boosting models achieve smaller negative log-likelihood and energy score values than the univariate benchmarks, with Frank strongest on both metrics.The Gaussian copula ranks next and the Gumbel copula last. Figure 9 shows joint-probability values concentrated around 0.35 for claim-free policies and mostly near one for claim-positive policies.Source: Table 3 · p. 16, §5.2 · pp. 16, 19, Figure 9 · p. 19
• Under the Frank copula, estimated Kendall’s τ varies substantially across policyholder risk profiles from −0.157 to 0.72, indicating both negative and positive dependence whose strength differs markedly.The 2–3-hour continuous-driving proportion shows a modest declining trend, with τ ≈0.58–0.61 across its domain.Source: §5.2 · pp. 16–17
• Total.miles.driven is informative across all four predictors and is selected most often; other frequently selected covariates vary by parameter.For frequency, Total.miles.driven and Brake.total are retained; mileage, credit score, and braking are prominent in severity predictors, while mileage and 2–3-hour driving proportions lead for dependence.Source: §5.2 · pp. 16–17, Figure 5 · p. 17
• Estimated expected claim counts increase with total mileage and the braking index, with mileage’s effect rising sharply at low levels and more gradually thereafter.These are partial effects on the Poisson mean predictor.Source: §5.2 · p. 18, Figure 7 · p. 18
• Severity quantiles rise and spread with mileage, while higher credit scores shift them downward and compress their spread; braking particularly widens the upper tail.The conditional Gamma plots show covariate effects on the fitted distribution, not only its center.Source: §5.2 · pp. 18–19, Figure 8 · p. 18

6 Concluding Remarks

• The framework jointly models claim frequency and severity through parameter-specific additive predictors and a covariate-dependent copula.Noncyclic model-based boosting produces sparse predictors and performs variable selection separately for each marginal and dependence parameter.Source: §6 · p. 20
• Simulation and synthetic telematics analyses show improved out-of-sample predictive performance over univariate benchmarks while maintaining interpretability.Partial-effect plots and conditional distributional summaries illustrate covariate associations with parameters and implied predictive distributions.Source: §6 · p. 20
• The fitted dependence structure helps explain how driving-related characteristics relate to the conditional association between claim frequency and severity.The paper describes the framework as promising for insurance settings requiring accurate forecasts and transparent explanations.Source: §6 · p. 20

Appendix A. Poisson–Gamma responses with a Gaussian copula

• The Gaussian copula links a Poisson claim-count margin and a Gamma severity margin, with latent correlation ρ as its dependence parameter.The Gamma parameterization uses mean severity µ and dispersion σ.Source: Appendix A · p. 21
• For the discrete Poisson margin, the likelihood uses its probability mass, CDF, and left limit alongside the continuous Gamma density and CDF.The Gamma shape and scale are given as 1/σ^2 and σ^2µ, respectively.Source: Appendix A · p. 21
• The mixed discrete–continuous log-likelihood combines the marginal quantities with the Gaussian copula derivative, and its second logarithm term vanishes when claim count is zero.The full likelihood is formed from the individual observation contributions.Source: Appendix A · p. 22

Appendix B. Noncyclic boosting algorithm for distributional copula regression

• At each iteration, the algorithm computes parameter-specific negative gradients and fits candidate base learners for the predictors.Base learners are defined for covariates and model parameters, with initial predictors set using suitable offsets.Source: Appendix B · p. 23
• The algorithm chooses the parameter-specific predictor with the smallest candidate loss, then updates that predictor using its best-fitting base learner.This selection across parameters is the noncyclic step.Source: Appendix B · p. 23
• After updates, the procedure computes out-of-bag risk and uses it to determine the optimal stopping iteration.The tuning sample is out of bag relative to the training sample.Source: Appendix B · p. 23

• Table 4 is titled as a listing of variables in the synthetic telematics dataset.Source: Table 4 · p. 24
• The seven day-of-week proportions form one group, while weekday/weekend proportions form another; each group sums to one, or 100%.Source: Appendix C · p. 24
• The passage identifies six acceleration and braking threshold labels—06, 08, 09, 11, 12, and 14—and lists turning-intensity labels separately.Source: Appendix C · p. 24
• The reported turning-intensity categories are five labeled levels, 08, 09, 10, 11, and 12, distinct from the separately listed acceleration and braking thresholds.Source: Appendix C · p. 24