1 Risk Modeling of Multi-year, Multi-line Reinsurance Using Copulas by Ping Wang St John’s University, New York on CICIRM 2011 at Beijing, China.

1

Risk Modeling of Multi-year, Multi-line Reinsurance Using

Copulas

by Ping Wang

St John’s University, New York

on CICIRM 2011 at Beijing, China

2

Agenda Today

• Multi-year, multi-line reinsurance

• A Framework Using Copulas to model time dependence

• Application using real data

• Concluding remarks

• Q & A

3

Multi-year, multi-linereinsurance policies

• Cover losses arising from multiple lines of business over multiple years (3 or 5 most common)

• Stop-loss type, commonly. Reinsurer pays claims only if the accumulated losses from several business lines over an extended period exceed a fairly high threshold.

• Reduced volatility compared to separate coverage

4

Difficulty Facing Actuaries

• Simultaneous modeling dependence – Across time, and– Across business lines (e.g., workers

compensation and commercial multiple perils)

5

Modeling Product Risk With Copula

• Assume independence between business lines

• Model time-dependence of each line using copula

• Simulate the distribution of future accumulated losses

• Estimate the payoff of multi-year, multi-line reinsurance

6

Marginal Distribution

• Suppose that there are Ti years data for a business line of the ith primary insurer

• Univariate marginal distribution functions

• Fit with Gamma, normal, lognormal, t-dist’n

iiTiiii

YYYY ,,,,321Y

ititititititit yPyYy ,PProbP

7

Modeling Time Dependencies Using Copulas

• With Copula C, the joint distribution function of Yi can be expressed as

• The log-likelihood of ith primary insurer is

• where c(.) is the probability density function corresponding to the copula function

• Predictive distribution is obtained based on the results of maximum likelihood estimation

ii iTiiTii PPyy ,,C,,P 11

i

i

iTii

T

tititi PPPθyl ,,,cln),p(ln 21

1

8

Estimate Product Risk

• Simulation of joint distribution of each business line over multiple years

• Calculate the policy payoff • Analyze the risk using VaR and CTE

9

Real Data

• Loss ratios of workers compensation (WC) and commercial multiple perils (CMP)

• 32 primary insurers• Task: based on the loss history of 5 years,

fit the multivariate distribution, simulate the future losses, then model the risk of the reinsurance policy that covers accumulated losses of both lines over next three years.

10

Correlations across Time: WC

• Loss ratios among years are not independent.

WC04 WC03 WC02 WC01 WC00

WC04 .6483(<.0001)

.6640(<.0001)

.4611(.0079)

.6128(.0002)

WC03 .6586(<.0001)

.3132(.0809)

.3398(.0571)

WC02 .6144(.0002)

.3796(.0321)

WC01 .5617(.0008)

Reported are the value of Pearson correlations and corresponding p-values.

11

Correlations across Time: CMP

CMP04 CMP03 CMP02 CMP01 CMP00

CMP04 .4771(.0058)

.3327(.0628)

.3200(.0742)

.3661(.0394)

CMP03 .4999(.0036)

.1510(.4093)

.1225(.5041)

CMP02 .4212(.0164)

.2571(.1554)

CMP01 .3589(.0437)

Reported are the value of Pearson correlations and corresponding p-values.

12

Relationship between WC & CMP

• Correlation coefficient: 0.1510

Scatter plot of WC vs CMP loss ratio

0

20

40

60

80

100

120

140

0 20 40 60 80 100 120CMP loss ratio

WC

lo

ss r

atio

13

Fitted Marginal Distribution

WC loss ratio CMP loss ratio

Distribution AIC K-S stat* AIC K-S stat

Lognormal 2176.4276 0.0383 2087.3092 0.0538

Gamma 2176.0656 0.0399 2087.6407 0.0709

t-dist’n 2588.6599 0.2707 2411.356 0.2561

*: kolmogorov-Smirnov test statistic

14

t-copula

• t-copula:

• where Gr is CDF of t-distribution function and

m

i irrmrrm

uguupuu

11-

1-1

1-1

))(G(

1)(G),...,(G),,(c T

2)(

1

2/12/

11

||)2

()(

)2

(),;(

mr

m rrr

mr

rp

tΣtΣ

tT

15

Different “correlation matrices”

55

234

23

22

32

432

1

1

1

1

1

X

AR

551

1

1

1

1

X

EX

551...00

............

0...10

0...01

X

I

16

Maximum Likelihood Estimation

• Parameters to be estimated: – of copula: in correlation matrix Σ and

degrees of freedom r– of marginal distribution, e.g. shape and scale

parameters for Gamma

17

MLE Results: WC

t-copula + Gamma margin t-copula + lognormal margin

parameter estimate StdError p-value estimate StdError p-value

0.6443 0.09136 <0.0001 0.6634 .0900 <0.0001

Shape/mu 10.6546 1.9740 <0.0001 4.1954 0.0455 <0.0001

Scale/sigma 6.6438 1.2528 <0.0001 0.3235 0.0310 <0.0001

DF r 4.2362 0.2704 <0.0001 4.2519 0.2704 <0.0001

AIC 999.77 1000.52

18

MLE Results: CMP

t-copula + Gamma margin t-copula + lognormal margin

parameter estimate StdError p-value estimate StdError p-value

0.4339 0.0925 <0.0001 0.4493 .0947 <0.0001

Shape/mu 11.4205 1.6132 <0.0001 3.9882 0.0296 <0.0001

Scale/sigma 4.9811 0.7206 <0.0001 0.3083 0.0222 <0.0001

DF r 4.2524 0.2703 <0.0001 4.2641 0.2703 <0.0001

AIC 979.07 981.18

19

Simulation and Analysis

• Based on the multivariate distribution of the loss ratio for business lines (WC, CMP separately) for the primary insurer

• Simulate the multivariate variables and

• The overall loss across two lines over three years is

• Where P denotes the annual premium• Payment on the reinsurance policy after deductible D

3

1,, )(

ttTtCtTtW YPXPlossTotal

),,( ,1, tTiTi xx ),,( ,1, tTiTi yy

0,)(max

3

1,, DYPXP

ttTtCtTtW

20

Histogram of Total Loss Using Different Assumptions

21

VaR and CTE of Total Loss (in millions)Using Different Assumptions

• Of 10,000 simulations of Total Loss

• Based on temporal independent loss ratios 196 are greater than the threshold; the reinsurer expects claims at a frequency of one in about fifty years, with average claims of $24.50 million.

• Based on copula dependence the frequency of claims is about 5% (495 of 10,000), or one in twenty years, and the average claims $41.71 million.

VaR and CTE of Total Loss (in millions)Using Different Assumptions

Copula dependence Independence

Percentage (%)

VaR CTE VaR CTE

99.5 698.080 732.394 631.948 655.245

99 660.840 704.613 610.872 637.249

95 595.420 637.094 568.998 595.911

90 563.536 607.559 545.428 576.016

22

Remarks

• Copulas – can use information developed over time to

better fit the multi-year claims experience– Can use information from similar risk classes

23

Thank You!

Questions and comments?