Optimal (Re)insurance Contract Design

Part IV: Empirical Reinsurance Models and Applications

to Agricultural Sector
C HENGGUO W ENG

Department of Statistics and Actuarial Science

University of Waterloo

READI Research Workshop - Module 2

Feb 22, 2019

C. Weng (c2weng@uwaterloo.ca) – p. 1/??

 Outline

• Section IV-1. Empirical Reinsurance Models

• Section IV-2. Optimal Reinsurance Analysis from a Crop

Insurer’s Perspective

C. Weng (c2weng@uwaterloo.ca) – p. 2/??

 • Section IV-1.
Empirical Reinsurance Models

C. Weng (c2weng@uwaterloo.ca) – p. 3/??

 Motivation
• The general reinsurance model can be formulated as:


 minf
 ρ(X, f )
s.t. 0 ≤ f (x) ≤ x, for all x ≥ 0,



Π(f (X)) ≤ π.

 Often difficult to solve (due to infinite dimension)

 In practice, the distribution of the underlying risk X is estimated from the

observed data {x1 , · · · , xN }.

• Empirical-based reinsurance model:

 exploits the observed data directly

C. Weng (c2weng@uwaterloo.ca) – p. 4/??

 Problem Formulation
• Collect samples x := (x1 , x2 , · · · , xN ) corresponding to the underlying risk X

• Introduce decision variable f := (f1 , f2 , · · · , fN ) where each fi represents the

reinsurance indemnification for the loss amount xi , i = 1, 2, · · · , N .

• Define empirical-based estimates as:

ρ(X, f ) → \
ρ(x, f)
0 ≤ f (x) ≤ x, for all x ≥ 0 → 0 ≤ fi ≤ xi , i = 1, 2, · · · , N,
Π(f ) ≤ π → d) ≤ π
Π(f

C. Weng (c2weng@uwaterloo.ca) – p. 5/??

 Problem Formulation (Cont’d)
• General reinsurance
 model

 minf ρ(x, f )

s.t. 0 ≤ f (x) ≤ x, for all x ≥ 0,



Π(f ) ≤ π.
• Empirical reinsurance
 model:

 min f
\
ρ(x, f)

s.t. 0 ≤ fi ≤ xi , i = 1, 2, · · · , N,


 d) ≤ π.
Π(f
• Many empirical reinsurance model can be cast as Second-Order Conic (SOC)
programming:
 A wide class of optimization problems
 Efficient softwares are available for solving SOC programming: CVX (Grant
and Boyd, 2008)
• In my thesis, I investigated the variance, CTE, and VaR minimizing models.

C. Weng (c2weng@uwaterloo.ca) – p. 6/??

 Variance Minimization
• Consider the following
 variance-minimizing model:
 min Var(T ) = Var(X − f (X) + Π(f ))
f f
 s.t. 0 ≤ f (x) ≤ x, Π(f ) ≤ π.
• Empirical version of the goal function:
XN
1 
2
\
Var(Tf ) = ¯
xi − fi ) − (x̄ − f ,
N − 1 i=1
where x̄ denote the mean of x, and f¯ denote the mean of f .
• Empirical version of the constraints:
d) ≤ π.
0 ≤ fi ≤ xi , i = 1, 2, · · · , N, and Π(f
• Empirical variance
 minimizingP model:
N  ¯

2

 minf ∈RN
 i=1 xi − fi ) − (x̄ − f
subject to 0 ≤ fi ≤ xi , i = 1, 2, · · · , N.


 d) ≤ π.
Π(f

C. Weng (c2weng@uwaterloo.ca) – p. 7/??

 Reinsurance Premium Budget Constraint
• P1. Expectation principle: Π(f ) = (1 + θ)E[f ] with θ > 0.

d) ≤ π ⇐⇒ (1 + θ)f¯ ≤ π,
Π(f

p
• P2. Standard deviation principle: Π(f ) = E[f ] + β Var[f ], where β > 0.
"N #1/2
X
2
d) ≤ π ⇐⇒ f¯ + √ β
Π(f ¯
fi − f ≤ π.
N −1 i=1

• Many other principles are investigated.

C. Weng (c2weng@uwaterloo.ca) – p. 8/??

 CTE Minimization
• CTE minimizing model:
  
 minf CTEα (Tf ) = CTEα X − f (X) + Π[f (X)]
 s.t. 0 ≤ f (x) ≤ x, Π[f (X)] ≤ π,

• Reformulated model:
   
 1
 min(ξ,f ) Gα (ξ, f ) = ξ + E X − f (X) + Π(f (X)) − ξ .
α +

 s.t. 0 ≤ f (x) ≤ x, Π(f (X)) ≤ π.

• Empirical model:

 XN   

 min \ 1 d) − ξ
(ξ,f ) Gα (ξ, f ) = ξ + xi − fi + Π(f ,
αN i=1 +


 s.t. d) ≤ π.
0 ≤ fi ≤ xi , i = 1, 2, · · · , N, Π(f

C. Weng (c2weng@uwaterloo.ca) – p. 9/??

 CTE Minimization (Cont’d)
• Empirical model:
   + 
 P
 min(ξ,f ) G\
α (ξ, f ) ≡ ξ + 1
αN
N
i=1 xi − fi + d) − ξ
Π(f ,

 s.t. d) ≤ π.
0 ≤ fi ≤ xi , i = 1, 2, · · · , N, Π(f

• Empirical CTE-minimizing model:

 PN

 min ξ + 1
i=1 zi ,


(ξ,f ,z) αN

 s.t. d) ≤ π,
0 ≤ fi ≤ xi , Π(f

 z ≥ 0, z ≥ [) − fi − ξ + xi ,
Π(f

 i i

 i = 1, 2, · · · , N.

• The above empirical model can be cast as Second-order conic programming for as
many as TEN reinsurance premium principles.

C. Weng (c2weng@uwaterloo.ca) – p. 10/??

 Example for CTE Minimization
• use samples from an exponential loss distribution with mean 1000

• assume the standard deviation principle with loading factor β = 0.2

• consider different levels of the reinsurance premium budget

• the solutions f ∗ are illustrated by their scatter plots over sample x

C. Weng (c2weng@uwaterloo.ca) – p. 11/??

 CTE min & Expectation Principle (1/2)
6000 6000 6000

5000 5000 5000

4000 4000 4000

3000 3000 3000

2000 2000 2000

1000 1000 1000

0 0 0
0 1000 2000 3000 4000 5000 6000 0 1000 2000 3000 4000 5000 6000 0 1000 2000 3000 4000 5000 6000

1) π = 200 2) π = 400 3) π = 600

6000

5000 6000 6000

5000 5000
4000

4000 4000
3000

3000 3000

2000
2000 2000

1000
1000 1000

0
0 1000 2000 3000 4000 5000 6000 0 0
0 1000 2000 3000 4000 5000 6000 0 1000 2000 3000 4000 5000 6000

4) π = 800 5) π = 1000 6) π = 2000

C. Weng (c2weng@uwaterloo.ca) – p. 12/??

 CTE min & Std Principle (1/2)

2000 2000

1500 1500

1000 1000

500 500

0 0
0 1000 2000 3000 4000 5000 6000 0 1000 2000 3000 4000 5000 6000

1) π = 80 2) π = 100
6000 6000

5000 5000

4000 4000

3000 3000

2000 2000

1000 1000

0 0
0 1000 2000 3000 4000 5000 6000 0 1000 2000 3000 4000 5000 6000

3) π = 150 4) π = 200

C. Weng (c2weng@uwaterloo.ca) – p. 13/??

 CTE min & Std Principle (2/2)
6000 6000

5000 5000

4000 4000

3000 3000

2000 2000

1000 1000

0 0
0 1000 2000 3000 4000 5000 6000 0 1000 2000 3000 4000 5000 6000

5) π = 400 6) π = 600
6000 6000

5000 5000

4000 4000

3000 3000

2000 2000

1000 1000

0 0
0 1000 2000 3000 4000 5000 6000 0 1000 2000 3000 4000 5000 6000

7) π = 1000 8) π = 1500

C. Weng (c2weng@uwaterloo.ca) – p. 14/??

 Objective & Analysis Method
• Objective: analyze the stability and consistency of the empirical solutions.

 stability: generate the same functional form for independent random samples
from the same loss distribution and over the same set of parameter values.

 consistency: converge to the true optimal ceded loss function as we increase

the sample size.

• Method: numerical experiments with the CTE minimization model under

 expectation reinsurance premium principle
 standard deviation reinsurance premium principle

C. Weng (c2weng@uwaterloo.ca) – p. 15/??

 Expectation Principle
• Numerical setting:

(i) X ∼ Exp(µ) with mean µ = 1000,

(ii) expectation premium principle with loading factor θ = 0.2

(iii) reinsurance premium budget π = 200.

(iv) α = 0.05.

• Solution to the theoretical model:

 Case 2 of Part 1 in Chapter 4 is applicable

 a stop-loss reinsurance f ∗ (X − d∗ )+ with d∗ = µ ln(6) = 1791.76.

C. Weng (c2weng@uwaterloo.ca) – p. 16/??

 Expectation Principle
• Numerical Experiment:
 For each empirical solution {x}, {f }, we fit a change-loss function
f (x) = c(x − d)+ .
ˆ:
 Admissibility of (ĉ, d)

ˆ + | < ε, for i = 1, 2, . . . , N.
|f(i) − ĉ(x(i) − d)

 N = 150 to N = 390 in multiple of 30.

 For each sample size N , we replicate the random samples M = 1000 times
independently to obtain 1000 independent estimates of ĉ and dˆ using ε = 0.1.
ˆ we also keep track of the proportion of admissibility
 Of the fitted pair ĉ and d,
and we report only the mean and standard errors of the admissible fitted ĉ and
ˆ
d.

C. Weng (c2weng@uwaterloo.ca) – p. 17/??

 Expectation Principle
• Numerical Experiment (Cont’d): (expectation principle)

N Admissibility mean of ĉ mean of dˆ mean of x

150 100% 1.00 (≈ 0) 1,802.22 (8.88) 1,000.10 (2.62)
180 100% 1.00 (≈ 0) 1,801.09 (7.98) 1,000.23 (2.35)
210 100% 1.00 (≈ 0) 1,799.05 (7.23) 1,000.16 (2.16)
240 100% 1.00 (≈ 0) 1,794.71 (6.78) 9,991.87 (2.04)
270 100% 1.00 (≈ 0) 1,791.93 (6.40) 9,986.34 (1.92)
300 100% 1.00 (≈ 0) 1,789.18 (6.17) 9,984.83 (1.84)
330 100% 1.00 (≈ 0) 1,788.93 (5.81) 9,987.20 (1.74)
360 100% 1.00 (≈ 0) 1,791.40 (5.69) 9,989.56 (1.67)
390 100% 1.00 (≈ 0) 1,791.99 (5.37) 9,995.38 (1.60)

C. Weng (c2weng@uwaterloo.ca) – p. 18/??

 Expectation Principle

1.000000000

0.999999998

0.999999996

0.999999994

0.999999992
values

0.999999990

0.999999988

0.999999986

0.999999984

0.999999982
N=150 N=180 N=210 N=240 N=270 N=300 N=330 N=360 N=390
sample size
C. Weng (c2weng@uwaterloo.ca) – p. 19/??
 Std Principle
• Numerical setting:
(i) X ∼ Exp(µ) with mean µ = 1000,
(ii) standard premium principle with loading factor β = 0.2
(iii) reinsurance premium budget π = 100.
(iv) α = 0.05.
• Numerical Experiment:
 For each empirical solution {x}, {f }, we fit a capped change-loss function
f (x) = min{c(x − d)+ , m}, and determine its admissibility by an error tolerance ε = 0.1 in a
similar way.

 Consider nine different sample sizes.

 For each sample size N , we replicate the random samples M = 1000 times independently to
obtain 1000 independent estimates of ĉ, dˆ and m̂.

 Of the fitted pair ĉ, dˆ and m̂, we also keep track of the proportion of admissibility and we report
only the mean and standard errors of the admissible ones.

C. Weng (c2weng@uwaterloo.ca) – p. 20/??

 Std Principle
• Numerical Experiment (Cont’d): (Std principle)

N Admissibility mean of ĉ mean of dˆ mean of m̂ mean of x

150 89.7% 1.00 (≈ 0) 2,694.13 (11.45) 1,502.85 (6.90) 1,002.62 (2.75)
180 93.8% 1.00 (≈ 0) 2,686.45 (10.04) 1,505.36 (6.51) 1,001.31 (2.42)
210 95.8% 1.00 (≈ 0) 2,686.67 (9.17) 1,509.18 (5.99) 1,001.31 (2.21)
240 97.1% 1.00 (≈ 0) 2,679.99 (8.51) 1,509.82 (5.94) 999.59 (2.06)
270 97.9% 1.00 (≈ 0) 2,677.78 (7.97) 1,507.24 (5.71) 999.08 (1.95)
300 98.8% 1.00 (≈ 0) 2,676.32 (7.58) 1,507.23 (5.36) 998.94 (1.84)
330 98.9% 1.00 (≈ 0) 2,673.58 (7.17) 1,499.80 (4.97) 999.17 (1.76)
360 99.8% 1.00 (≈ 0) 2,672.89 (6.61) 1,504.78 (4.71) 1,000.20 (1.65)
390 99.9% 1.00 (≈ 0) 2,677.89 (6.55) 1,501.04 (4.80) 999.56 (1.60)

C. Weng (c2weng@uwaterloo.ca) – p. 21/??

 Std Principle

0.99999

0.99998

0.99997

0.99996
values

0.99995

0.99994

0.99993

N=150 N=180 N=210 N=240 N=270 N=300 N=330 N=360 N=390

sample size
C. Weng (c2weng@uwaterloo.ca) – p. 22/??
 • Section IV-2.
Optimal Reinsurance Analysis
from a Crop Insurer’s
Perspective

C. Weng (c2weng@uwaterloo.ca) – p. 23/??

 Thank You

Chengguo Weng

C. Weng (c2weng@uwaterloo.ca) – p. 24/??