A Multivariate Analysis of Intercompany Loss Triangles

Transcription

1 A Multivariate Analysis of Intercompany Loss Triangles Peng Shi School of Business University of Wisconsin-Madison ASTIN Colloquium May 21-24, 2013 Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 1 / 30

2 Outline Introduction: background and motivation Modeling Data distribution Bayesian hierarchical model Model assessment Data analysis NAIC schedule P Estimation and inference Predictive applications Reserving variability Reinsurance example Concluding remarks Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 2 / 30

3 Background Introduction Loss reserves: the technical provisions to support outstanding liabilities of a property casualty insurer Reserves represent the largest balance sheet liability Loss reserving is a classic actuarial problem Improper reserving could be detrimental Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 3 / 30

4 Background Introduction An example of run-off triangle Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 4 / 30

5 Introduction Literature Univariate loss reserving Chain ladder and others See Taylor (2000) and Wüthrich and Merz (2008) Multivariate loss reserving Naive approach: the silo method Additivity issue (see Ajne (1994)) attracts more attention recently EU capital adequacy regime: Solvency II CAS loss reserve dependency working party Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 5 / 30

6 Introduction Literature Recent literature emphasizes dependencies among triangles Distribution-Free approach Multivariate chain-ladder, e.g. Braun (2004), Merz and Wüthrich (2008), Zhang (2010) Multivariate additive model, e.g. Hess et al. (2006), Merz and Wüthrich (2009) Parametric approach Parametric distributions, e.g. Shi et al. (2012) Copula approach, e.g. Shi and Frees (2011), de Jong (2012) Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 6 / 30

7 Introduction Motivation Prediction of insurance liabilities often requires aggregating the experiences of loss payment from multiple insurers to borrow information for lines of business from other insurers to identify industry-wide under- or over-reserving problem to predict outstanding liabilities for a reinsurer The resulting dataset of intercompany loss triangles displays a multilevel structure of claim development a portfolio consists of a group of insurers each insurer several lines of business and each line various cohorts of claims Our goal is to propose a Bayesian hierarchical model to analyze intercompany claim triangles, accommodating association within and between insurers Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 7 / 30

8 Modeling Some Notations Index n = 1,, N indicates the n-th insurer; l = 1,, L indicates the l-th line of business i = 1,, I indicates the i-th accident year j = 1,, J(= I) indicates the j-th valuation point t j Variable of interest X n,l,i (t j ) denotes incremental paid losses ω n,l,i denotes the exposure in accident year i We normalize incremental payments by Y n,l,i (t j ) = X n,l,i (t j )/ω n,l,i Index set {(i, j) : i + j I + 1} divides data into two subset D I : information available by calendar year I + 1 DI c : future payments in years t = i + j > I + 1 Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 8 / 30

9 Modeling Data Distribution - Marginal Assume a parametric distribution for Y n,l,i (t j ), for example: Y n,l,i (t j ) F n,l ( ; η n,l,i,j, φ n,l ) η n,l,i,j determines the location such that η n,l,i,j = g(µ n,l,i (t j )) Vector φ n,l summarizes additional parameters Alternative models for location Parametric with two factors g(µ n,l,i (t j )) = δ n,l + α n,l,i + ζ n,l,j Semiparametric regression g(µ n,l,i (t j )) = δ n,l + α n,l,i + s n,l (t j ) Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 9 / 30

10 Modeling Data Distribution - Joint Dependency among multiple lines is accommodated by a copula Distribution of ( Y n,1,i (t j ),, Y n,l,i (t j ) ) has the following copula representation F n (y n,1,i,j,, y n,l,i,j ) =Prob(Y n,1,i (t j ) y n,1,i,j,, Y n,l,i (t j ) y n,l,i,j ) =H n (F n,1 (y n,1,i,j ; η n,1,i,j, φ n,1 ),, F n,l (y n,l,i,j ; η n,l,i,j, φ n,l ); ρ n ) Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 10 / 30

11 Modeling Multilevel Structure Consider a Bayesian hierarchical model allows insurers to learn from each other provides predictive distribution f (y Dc I y D I ) = f (y Dc I Θ)f (Θ y D I )dθ For instance g(µ n,l,i (t j )) = δ n,l + α n,l,i + ζ n,l,j δ n,l N(0, σδ 2 [l]) for n = 1,, N α n,l,i N(0, σ 2 α[l, i]) for n = 1,, N σ 2 α[l, i] IG(ψ α [l], ψ α [l]) for i = 1,, I Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 11 / 30

12 Model Assessment Modeling For training data, we consider logarithm of the pseudo-marginal likelihood (LPML) statistic Denote y n,i,j = (y n,1,i,j,, y n,l,i,j ) Define CPO n,i,j = f (y n,i,j y D n,i,j ) = f (y n,i,j Θ)f (Θ y D n,i,j )dθ Calculate CPO M = log CPO n,i,j For validation data, we consider L-criterion L measure = 1 S b S N L r=b+1 n=1 l=1 {(i,j):i+j>i+1} ([y n,l,i,j ] r y n,l,i,j ) 2 Can be evaluated using the basis of either paid losses or loss ratio Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 12 / 30

13 Data Analysis Data NAIC Schedule P Data preparation Triangles constructed from Schedule P of year 1997 Future payments in lower triangles from year Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 13 / 30

14 Data Analysis Data Consider a hypothetical portfolio Five insurers Each with personal and commercial auto lines Triangles of paid losses Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 14 / 30

15 Model Specification Sampling distribution Data Analysis Lognormal model for incremental payment, i.e. Y n,l,i (t j ) LN(η n,l,i (t j ), σ 2 n,l) Parametric regression for personal auto Penalized regression spline for commercial auto s n,l (t j ; θ n,l ) = β n,l t j + K λ n,l,k t j ν k 3 ν 1 < ν 2 < < ν K are fixed knots, could be the k/(k + 1)th sample quantile of covariate t j K k=1 λ2 n,l,k < τ (τ is a constant) to penalize the roughness of the fit Frank copula to join multiple lines H n (u, v; ρ n ) = 1 ( ln 1 + (e ρnu 1)(e ρnv ) 1) ρ n e ρn 1 k=1 Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 15 / 30

16 Data Analysis Inference Vague priors are used in the inference Run 50,000 MCMC iterations in two parallel chains First 40,000 iterations in each chain discarded as burn-in sample Some selected results Left panel: σ 2 in log-normal Right panel: ρ in Frank copula Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 16 / 30

17 Model Comparison Data Analysis Consider three models Model 1: Assume independence among business lines and no learning across insurers Model 2: Allow for dependence among business lines within each insurer but no learning across insurers Model 3: Allow for dependence among business lines within each insurer and information sharing between insurers LPML L-Measure Amount Ratio Model e Model e Model e Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 17 / 30

18 Predictive Applications Reserving Variability Define reserves as R = g(y Dc I ) Quantities of interest could be: accident year reserves calendar year reserves reserves by business line firm-level reserves. Bayesian model provides predictive distributions of Y Dc I, and thus of reserves R Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 18 / 30

19 Reserving Variability Predictive Applications Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 19 / 30

20 Reserving Variability Predictive Applications Distribution-free approaches rely on conditional mean squared error of prediction (MSEP) MSEP R DI = E [(R R ] B ) 2 D I Given R B = E[E(R Θ) D I ] = E(R D I ) MSEP R DI = Var(R D I ) = E [Var(R Θ) D I ] + Var [E(R Θ) D I ] variability = process variance + estimation error (PV) (ER) Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 20 / 30

21 Reserving Variability Predictive Applications ER Accident Year Calendar Year PV MSEP ER PV MSEP ,001 1,844 3, , ,358 1,008, ,544 5,002 8, , , , ,511 9,883 14, , , , ,314 20,405 26, , , , ,482 42,333 51, ,037 49,870 61, ,676 92, , ,904 25,770 34, , , , ,263 13,017 19, , , , ,246 6,208 10, , ,187 1,217, ,051 1,976 3,635 Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 21 / 30

22 Predictive Applications Reinsurance Example Consider two types of reinsurance contract Quota share reinsurance: share risk proportionally Excess-of-loss reinsurance: share risk above threshold Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 22 / 30

23 Predictive Applications Reinsurance Example Risk capital for the hypothetical reinsurance portfolio Value-at-Risk (VaR): VaR(p) = Q R (p) Conditional Tail Expectation (CTE): CTE(p) = E[R R > Q R (p)] VaR CTE 90% 95% 99% 90% 95% 99% Quota = ,662,280 22,293,481 23,500,626 22,505,966 23,066,293 24,175,243 Quota = ,441,520 14,862,321 15,667,084 15,003,977 15,377,529 16,116,829 Quota = ,220,760 7,431,160 7,833,542 7,501,989 7,688,764 8,058,414 Retention = 1 27,085,174 27,971,135 29,646,489 28,230,301 28,989,703 30,506,736 Retention = 5 22,581,757 23,273,150 24,809,406 23,551,479 24,202,636 25,621,111 Retention = 10 15,205,267 15,922,672 17,681,267 16,526,250 17,525,540 21,394,198 Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 23 / 30

24 Concluding Remarks Summary Several features of our approach Both parametric and semi-parametric formulations Copula model to associate business lines A hierarchical structure to allow for learning across insurers Predictions are allowed at different levels of interest Future research To compare with classical multilevel modeling To incorporate collateral information Firm-level heterogeneity Look at triangles by state Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 24 / 30

25 Copula Appendix A copula is a multivariate distribution function with uniform marginals. Let U 1,..., U T be T uniform random variables on (0,1). Their distribution function C(u 1,..., u T ) = Pr(U 1 u 1,..., U T u T ) For general applications, consider arbitrary marginal distributions F 1 (y 1 ),..., F T (y T ). Define a multivariate distribution function using the copula such that F(y 1,..., y T ) = C(F 1 (y 1 ),..., F T (y T )) Sklar(1959) established the converse: any multivariate distribution function F can be written in the form of the above equation, i.e., using a copula representation Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 25 / 30

26 Appendix Penalized Regression Spline g(µ n,l,i (t j )) = δ n,l + α n,l,i + β n,l t j + Γ j γ n,l Denote γ n,l = (γ n,l,1,, γ n,l,k ) Define Γ = Γ K Λ 1/2 K and Γ j is the jth row of Γ t 1 ν 1 3 t 1 ν 2 3 t 1 ν K 3 t 2 ν 1 3 t 2 ν 2 3 t 2 ν K 3 Γ K = t J ν 1 3 t J ν 2 3 t J ν K 3 0 ν 1 ν 2 3 ν 1 ν K 3 ν 2 ν ν 2 ν K 3 Λ K = ν K ν 1 3 ν K ν Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 26 / 30

27 Data Exploration Appendix Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 27 / 30

28 Appendix Model Specification Some examples on priors δ n,l N(0, σδ 2 [l]) for n = 1,, N σ 2 δ [l] IG(10 4, 10 4 ) for l = 1,, L σ 2 n,l IG(ψ[l], ψ[l]) for n = 1,, N ψ[l] Gamma(10 4, 10 4 ) for l = 1,, L θ n Uniform( 100, 100) for n = 1,, N Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 28 / 30

29 Convergence Appendix Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 29 / 30

30 Copula Validation Appendix Use K -plot to validate the Frank copula Based on function K (w) = w φ(w) φ (w) Visualize parametric and non-parametric estimates Parametric: φ(w) = ln[(e ρw 1)/(e ρ 1)] for Frank Non-parametric: use pseudo-observations W s = Ĥ(U s, V s ) Peng Shi (Wisconsin School of Business) Intercompany Loss Triangles 2013 ASTIN Colloquium 30 / 30