Stochastic programming approaches to pricing in non-life insurance

Transcription

1 Stochastic programming approaches to pricing in non-life insurance Martin Branda Charles University in Prague Department of Probability and Mathematical Statistics 11th International Conference on COMPUTATIONAL MANAGEMENT SCIENCE May, 2014, Lisbon

2 Table of contents 1 Introduction 2 Pricing of non-life insurance contracts 3 Approach based on generalized linear models 4 Optimization models expected value approach 5 Optimization models individual chance constraints 6 Optimization models a collective risk constraint 7 Numerical comparison

3 Introduction Table of contents 1 Introduction 2 Pricing of non-life insurance contracts 3 Approach based on generalized linear models 4 Optimization models expected value approach 5 Optimization models individual chance constraints 6 Optimization models a collective risk constraint 7 Numerical comparison

4 Introduction Multiplicative tariff of rates Motor third party liability (MTPL): Engine volume between 1001 and 1350 ccm, policyholder age 18 30, region over inhabitants: 300 ( ) ( ). Engine volume between 1351 and 1600 ccm, policyholder age over 50, region between and inhabitants: 450 ( ) ( ).

5 Introduction Four methodologies The contribution combines four methodologies: Data-mining data preparation. Mathematical statistics random distribution estimation using generalized linear models. Insurance mathematics pricing of non-life insurance contracts. Operations research mathematical (stochastic) programming approaches to tariff of rates estimation.

6 Introduction Practical experiences More than 4 years of cooperation with Actuarial Department, Head Office of Vienna Insurance Group Czech Republic (VIG CR). VIG CR the most profitable part of Vienna Insurance Group. VIG CR the largest group on the market: 2 universal insurance companies (Kooperativa pojišt ovna, Česká podnikatelská pojišt ovna) and 1 life-oriented (Česká spořitelna). Kooperativa & ČPP MTPL: 2.5 mil. cars from 7 mil. per year (data for more than 10 years)

7 Pricing of non-life insurance contracts Table of contents 1 Introduction 2 Pricing of non-life insurance contracts 3 Approach based on generalized linear models 4 Optimization models expected value approach 5 Optimization models individual chance constraints 6 Optimization models a collective risk constraint 7 Numerical comparison

8 Pricing of non-life insurance contracts Tariff classes/segmentation criteria Tariff of rates based on S + 1 categorical segmentation criteria: i 0 I 0, e.g. tariff classes I 0 = {engine volume up to 1000, up to 1350, up to 1850, up to 2500, over 2500 ccm}, i 1 I 1,..., i S I S, e.g. age I 1 = {18 30, 31 65, 66 and more years} We denote I = (i 0, i 1,..., i S ), I I a tariff class, where I = I 0 I 1 I S denotes all combinations of criteria values. Let W I be the number of contracts (exposures) in I.

9 Pricing of non-life insurance contracts Compound distribution of aggregated losses Aggregated losses over one year for risk cell I L T I = W I w=1 N I,w L I,w, L I,w = X I,n,w, n=1 where all r.v. are assumed to be independent (N I, X I denote independent copies) N I,w is the random number of claims for a contract during one year with the same distribution for all w X I,n,w is the random claims severity with the same distribution for all n and w Well-known formulas for the mean and the variance: µ T I = IE[L T I ] = W I µ I = W I IE[N I ]IE[X I ], (σ T I ) 2 = var(l T I ) = W I σ 2 I = W I (IE[N I ]var(x I ) + (IE[X I ]) 2 var(n I )).

10 Pricing of non-life insurance contracts Multiplicative tariff of rates We assume that the risk (office) premium is composed in a multiplicative way from basic premium levels Pr i0 and nonnegative surcharge coefficients e i1,..., e is, i.e. we obtain the decomposition Pr I = Pr i0 (1 + e i1 ) (1 + e is ). We denote the total premium TP I = W I Pr I for the risk cell I. Example: engine volume between 1001 and 1350 ccm, age 18 30, region over inhabitants: 300 ( ) ( )

11 Pricing of non-life insurance contracts Prescribed loss ratio random constraints Our goal is to find optimal basic premium levels and surcharge coefficients with respect to a prescribed loss ratio LR, ˆ i.e. to fulfill the random constraints L T I TP I ˆ LR for all I I, (1) and/or the random constraint I I LT I I I TP I ˆ LR. (2) The prescribed loss ratio LR ˆ is usually based on a management decision. If LR ˆ = 1, we obtain the netto-premium. It is possible to prescribe a different loss ratio for each tariff cell.

12 Pricing of non-life insurance contracts Sources of risk Two sources of risk for an insurer: 1. Expectation risk: different expected losses for tariff cells. 2. Distributional risk: different shape of the probability distribution of losses, e.g. standard deviation.

13 Pricing of non-life insurance contracts Prescribed loss ratio expected value constraints Usually, the expected value of the loss ratio is bounded IE[L T I ] = IE[L I ] LR ˆ for all I I. (3) TP I Pr I The distributional risk is not taken into account.

14 Pricing of non-life insurance contracts Prescribed loss ratio chance constraints A natural requirement: the inequalities are fulfilled with a prescribed probability leading to individual chance (probabilistic) constraints ( L T ) P I LR ˆ 1 ε, for all I I, (4) TP I where ε (0, 1), usually ε {0.1, 0.05, 0.01}, or a constraint for the whole line of business: ( ) P I I LT I LR ˆ 1 ε. I I TP I Distributional risk allocation to tariff cells will be discussed later.

15 Approach based on generalized linear models Table of contents 1 Introduction 2 Pricing of non-life insurance contracts 3 Approach based on generalized linear models 4 Optimization models expected value approach 5 Optimization models individual chance constraints 6 Optimization models a collective risk constraint 7 Numerical comparison

16 Approach based on generalized linear models Generalized linear models A standard approach based on GLM with the logarithmic link function g(µ) = ln µ without the intercept: Poisson (overdispersed) or Negative-binomial regression the expected number of claims: IE[N I ] = exp{λ i0 + λ i1 + + λ is }, Gamma or Inverse Gaussian regression the expected claim severity: IE[X I ] = exp{γ i0 + γ i1 + + γ is }, where λ i, γ i are the regression coefficients for each I = (i 0, i 1,..., i S ). For the expected loss we obtain IE[L I ] = exp{λ i0 + γ i0 + λ i1 + γ i1 + + λ is + γ is }.

17 Approach based on generalized linear models Generalized linear models The basic premium levels and the surcharge coefficients can be estimated as a product of normalized coefficients Pr i0 = exp{λ i 0 + γ i0 } ˆ LR e is = S min exp(λ i ) i I s s=1 S min exp(γ i ), i I s s=1 exp(λ is ) min is Is exp(λ is ) exp(γ is ) min is Is exp(γ is ) 1, Under this choice, the constraints on loss ratios are fulfilled with respect to the expectations.

18 Approach based on generalized linear models Generalized linear models The GLM approach is highly dependent on using GLM with the logarithmic link function. It can be hardly used if other link functions are used, interaction or other regressors than the segmentation criteria are considered. For the total losses modelling, we can employ generalized linear models with the logarithmic link and a Tweedie distribution for 1 < p < 2, which corresponds to the compound Poisson gamma distributions.

19 Optimization models expected value approach Table of contents 1 Introduction 2 Pricing of non-life insurance contracts 3 Approach based on generalized linear models 4 Optimization models expected value approach 5 Optimization models individual chance constraints 6 Optimization models a collective risk constraint 7 Numerical comparison

20 Optimization models expected value approach Advantages of the optimization approach GLM with other than logarithmic link functions can be used, business requirements on surcharge coefficients can be ensured, total losses can be decomposed and modeled using different models, e.g. for bodily injury and property damage, other modelling techniques than GLM can be used to estimate the distribution of total losses over one year, e.g. generalized additive models, classification and regression trees, not only the expectation of total losses can be taken into account but also the shape of the distribution, costs and loadings (commissions, tax, office expenses, unanticipated losses, cost of reinsurance) can be incorporated when our goal is to optimize the combined ratio instead of the loss ratio, we obtain final office premium as the output,

21 Optimization models expected value approach Total loss decomposition We can assume that L I contains not only losses but also various costs and loadings, thus we can construct the tariff rates with respect to a prescribed combined ratio. For example, the total loss over one year can be composed as follows L I = (1 + vc I ) [ (1 + inf s )L s I + (1 + inf l)l l I ] + fci, where small L s I and large claims L l I are modeled separately, inflation of small claims inf s and large claims inf l, proportional costs vc I and fixed costs fc I are incorporated. We only need estimates of E[L T I ] and var(l T I ) for all I.

22 Optimization models expected value approach Optimization model expected value approach The premium is minimized 1 under the conditions on the prescribed loss ratio and a maximal possible surcharge (r max ): min Pr,e w I Pr i0 (1 + e i1 ) (1 + e is ) I I s.t. ˆ LR Pri0 (1 + e i1 ) (1 + e is ) IE[L i0,i 1,...,i S ], (5) (1 + e i1 ) (1 + e is ) 1 + r max, e i1,..., e is 0, (i 0, i 1,..., i S ) I. This problem is nonlinear nonconvex, thus very difficult to solve. Other constraints can be included. 1 A profitability is ensured by the constraints on the loss ratio. The optimization leads to minimal levels and surcharges.

23 Optimization models expected value approach Optimization model expected value approach Using the logarithmic transformation of the decision variables u i0 = ln(pr i0 ) and u is = ln(1 + e is ) and by setting b i0,i 1,...,i S = ln(ie[l i0,i 1,...,i S ]/ ˆ LR), the problem can be rewritten as a nonlinear convex programming problem: min w I e u i 0 +u i1 + +u is u I I s.t. u i0 + u i1 + + u is b i0,i 1,...,i S, (6) u i1 + + u is ln(1 + r max ), u i1,..., u is 0, (i 0, i 1,..., i S ) I. The problems (5) and (6) are equivalent.

24 Optimization models individual chance constraints Table of contents 1 Introduction 2 Pricing of non-life insurance contracts 3 Approach based on generalized linear models 4 Optimization models expected value approach 5 Optimization models individual chance constraints 6 Optimization models a collective risk constraint 7 Numerical comparison

25 Optimization models individual chance constraints Optimization model individual chance constraints If we prescribe a small probability level ε (0, 1) for violating the loss ratio in each tariff cell, we obtain the following chance constraints P ( L T i 0,i 1,...,i S LR ˆ ) W i0,i 1,...,i S Pr i0 (1 + e i1 ) (1 + e is ) 1 ε, which can be rewritten using the quantile function F 1 L T I of L T I as ˆ LR W i0,i 1,...,i S Pr i0 (1 + e i1 ) (1 + e is ) F 1 L T i 0,i 1,...,i S (1 ε). By setting b I = ln F 1 (1 ε) L T I, W I LR ˆ the formulation (6) can be used.

26 Optimization models individual chance constraints Optimization model individual chance constraints min u w I e u i 0 +u i1 + +u is I I s.t. u i0 + u i1 + + u is b i0,i 1,...,i S, u i1 + + u is ln(1 + r max ), u i1,..., u is 0, (i 0, i 1,..., i S ) I, with b I = ln F 1 (1 ε) L T I. W I LR ˆ

27 Optimization models individual chance constraints Optimization model individual reliability constraints It can be very difficult to compute the quantiles F 1, see, e.g., L T I Withers and Nadarajah (2011). We can employ the one-sided Chebyshev s inequality based on the mean and variance of the compound distribution: P ( L T I TP I ˆ LR ) ( LR ˆ TP I µ T I ) 2 /(σi T ) 2 ε, (7) for LR ˆ TP I µ T I. Chen et al. (2011) showed that the bound is tight for all distributions D with the expected value µ T I and the variance (σi T ) 2 : sup P ( L T I D for ˆ LR TP I µ T I. LR ˆ ) 1 TP I = 1 + ( LR ˆ TP I µ T I ) 2 /(σi T ), 2

28 Optimization models individual chance constraints Optimization model individual reliability constraints The inequality (7) leads to the following constraints, which serve as conservative approximations: 1 ε µ T I + σi T LR ˆ TP I. ε Finally, the constraints can be rewritten as reliability constraints 1 ε σ µ I + I LR ˆ Pr I. (8) ε WI If we set [( ) ] 1 ε b I = ln µ I + σ I / LR ˆ, εw I we can employ the linear programming formulation (6) for rate estimation.

29 Optimization models individual chance constraints Optimization model individual reliability constraints min u w I e u i 0 +u i1 + +u is I I s.t. u i0 + u i1 + + u is b i0,i 1,...,i S, u i1 + + u is ln(1 + r max ), u i1,..., u is 0, (i 0, i 1,..., i S ) I, with [( ) ] 1 ε b I = ln µ I + σ I / LR εw ˆ. I

30 Optimization models a collective risk constraint Table of contents 1 Introduction 2 Pricing of non-life insurance contracts 3 Approach based on generalized linear models 4 Optimization models expected value approach 5 Optimization models individual chance constraints 6 Optimization models a collective risk constraint 7 Numerical comparison

31 Optimization models a collective risk constraint Optimization model a collective risk constraint In the collective risk model, a probability is prescribed for ensuring that the total losses over the whole line of business (LoB) are covered by the premium with a high probability, i.e. P ( L T I I I ) W I Pr I 1 ε. I I

32 Optimization models a collective risk constraint Optimization model a collective risk constraint Zaks et al. (2006) proposed the following program for rate estimation, where the mean square error is minimized under the reformulated collective risk constraint using the Central Limit Theorem: 1 min IE [(L T I W I Pr I ) 2] Pr I r I I I s.t. (9) W I σi 2, I I W I Pr I = I I W I µ I + z 1 ε I I where r I > 0 and z 1 ε denotes the quantile of the Normal distribution. Various premium principles can be obtained by the choice of r I (r I = 1 or r I = W I leading to semi-uniform or uniform risk allocations).

33 Optimization models a collective risk constraint Optimization model a collective risk constraint According to Zaks et al. (2006), Theorem 1, the program has a unique solution ˆPr I = µ I + z 1 ε r I σ rw I, with r = I I r I and σ 2 = I I W I σi 2. If we want to incorporate the prescribed loss ratio LR ˆ for the whole LoB into the rates, we can set [( ) ] r I σ b I = ln µ I + z 1 ε / LR rw ˆ, I within the problem (6).

34 Optimization models a collective risk constraint Optimization model a collective risk constraint min u w I e u i 0 +u i1 + +u is I I s.t. u i0 + u i1 + + u is b i0,i 1,...,i S, u i1 + + u is ln(1 + r max ), u i1,..., u is 0, (i 0, i 1,..., i S ) I, with [( ) ] r I σ b I = ln µ I + z 1 ε / LR rw ˆ. I

35 Numerical comparison Table of contents 1 Introduction 2 Pricing of non-life insurance contracts 3 Approach based on generalized linear models 4 Optimization models expected value approach 5 Optimization models individual chance constraints 6 Optimization models a collective risk constraint 7 Numerical comparison

36 Numerical comparison MTPL segmentation criteria We consider policies with settled claims simulated using characteristics of real MTPL portfolio. The following segmentation variables are used: tariff group: 5 categories (engine volume up to 1000, up to 1350, up to 1850, up to 2500, over 2500 ccm), age: 3 cat. (18-30, 31-65, 66 and more years), region (reg): 4 cat. (over , over , over 5 000, up to inhabitants), gender (gen): 2 cat. (men, women). Many other available indicators related to a driver (marital status, type of licence), vehicle (engine power, mileage, value), policy (duration, no claim discount).

37 Numerical comparison Software SAS Enterprise Guide: SAS GENMOD procedure (SAS/STAT 9.3) generalized linear models SAS OPTMODEL procedure (SAS/OR 9.3) nonlinear convex optimization

38 Numerical comparison Parameter estimates Overd. Poisson Gamma Param. Level Est. Std.Err. Exp Est. Std.Err. Exp TG TG TG TG TG reg reg reg reg age age age gen gen Scale

39 Numerical comparison Employed models GLM The approach based on generalized linear models EV model Deterministic optimization model with expected value constraints SP model (individual) Stochastic programming problem with individual reliability constraints ε = 0.1 SP model (collective) Stochastic programming problem with collective risk constraint ε = 0.1

40 Numerical comparison Multiplicative tariff of rates Individual risk model Collective risk model Parameter GLM EV Exp Exp. 300 Exp. 600 Exp. 60 Exp. 300 Exp. 600 TG TG TG TG TG region region region region age age age gender gender Exposure in thousands

41 Numerical comparison Conclusions (open for discussion) GLM/EV model good start SP model (ind.) appropriate for less segmented portfolios with high exposures of tariff cells SP model (col.) appropriate for heavily segmented portfolios with low exposures of tariff cells

42 Numerical comparison M. Branda (2012). Underwriting risk control in non-life insurance via generalized linear models and stochastic programming. Proceedings of the 30th International Conference on Mathematical Methods in Economics 2012, M. Branda (2014). Optimization approaches to multiplicative tariff of rates estimation in non-life insurance. To appear in Asia-Pacific Journal of Operational Research. L. Chen, S. He, S. Zhang (2011). Tight bounds for some risk measures, with applications to robust portfolio selection. Operations Research, 59(4), J. Nelder, R. Wedderburn (1972). Generalized Linear Models. Journal of the Royal Statistical Society, Series A, 135(3), E. Ohlsson, B. Johansson (2010). Non-Life Insurance Pricing with Generalized Linear Models. Berlin Heidelberg: Springer-Verlag. Ch. Withers, S. Nadarajah (2011). On the compound Poisson-gamma distribution. Kybernetika 47(1), Y. Zaks, E. Frostig, B. Levikson (2006). Optimal pricing of a heterogeneous portfolio for a given risk level. Astin Bulletin 36(1),