Hedging and Risk Management of Variable Annuity Portfolios ......Variable Annuity (VA) A variable...
Transcript of Hedging and Risk Management of Variable Annuity Portfolios ......Variable Annuity (VA) A variable...
Hedging and Risk Management of Variable Annuity Portfolios
A Functional Data Approach
X. Sheldon Lin
Department of Statistical Sciences
University of Toronto
Joint work with Guojun Gan, University of Connecticut
Based on a paper that has been accepted by Insurance: Mathematics and Economics
WatRISQ Seminar, March 17, 2015
Outline
Introduction and Motivation
A Brief Introduction of Variable Annuity (VA)
Issues and Challenges when Risk Managing a VA Portfolio
Nested Simulation
Our Approach: Clustering and Universal Kriging
Numerical Illustration
Concluding Remarks and Selected References
Variable Annuity (VA)
◮ A variable annuity is a tax-deferred retirement saving plan (deferred
annuity).
◮ It has two phases: an accumulation phase and a payout phase:
◮ During the accumulation phase, the annuitant pays
premium(s) into the contract and accumulates assets;◮ During the payout phase, the annuitant has options to receive
a lump sum payment or to convert the account value into a
fixed immediate annuity.
◮ It is wrapped with guaranteed death benefits and/or guaranteed living
benefits.
Embedded Guarantees
A variable annuity is wrapped with one or more of two types of guarantees:
◮ The guaranteed minimum death benefit (GMDB);
◮ The guaranteed minimum living benefit (GMLB), which is further divided
into
◮ The guaranteed minimum accumulation benefit (GMAB)◮ The guaranteed minimum income benefit (GMIB)◮ The guaranteed minimum withdrawal benefit (GMWB)
Guaranteed Minimum Death Benefits (GMDB)
◮ A GMDB provides a level of protection against the loss in the
subaccount(s) in the event of death.
◮ Common types of GMDBs are ‘Return of Premium’, ‘Annual Rachet’ and
and ‘Annual Roll-up’
◮ ‘Return of Premium’ guarantees the beneficiary of the variable
annuity the greater of (a) the subaccount value or (b) the total
premiums paid in the past;◮ ‘Annual Rachet’ is an enhanced guarantee that pays the
greater of (a), (b) above or (c) the account value on a prior
contract anniversary date.◮ ‘Annual Roll-up’ is a rising floor GMDB, meaning that the
benefit level increases annually at a specified rate of interest.
‘Annual Roll-up’ is often combined with ‘Annual Rachet’.
Guaranteed Minimum Living Benefits (GMLB)
◮ A GMLB is typically offered as a rider.
◮ A GMAB usually is a ‘Return of Premium’ guarantee (or its enhanced
version).
The annuitant will receive a lump sum payment after a specified period
(8-10 years).
◮ A GMIB guarantees the greater of the subaccount value or a minimum
payout base (the total premiums paid with annual credited interest rate
(3%-5%)) and a guaranteed annuitization rate after a waiting period (7+
years).
◮ A GMWB guarantees a systematic withdrawal of a certain percentage
(5%-7%) of the total premiums paid annually until the total premiums
are exhausted, regardless of the market performance.
Essentially the VA guarantees are (path dependent) put options written on the
sub-account value.
As a result, single contracts can be valuated using option pricing theory.
Issues and Challenges when Risk Managing a VA Portfolio
◮ In-house hedging program: construction of an investment portfolio to
dynamically hedge the VA portfolio.
◮ Calculation of required capitals, Regulatory Capital (RC) and Economic
Capital (EC), to meet the regulatory requirements.
◮ The payoff function of the guarantee in most contracts is path dependent
that results in no closed-form formulas for the liability value of the
guarantees.
◮ The number of contracts is very large, often in the range of 100,000 to
1,000,000.
◮ The portfolio is highly non-homogeneous: each contract is different in
terms of age, gender, time to maturity, guarantee type, account value,
etc.
◮ In practice, insurance companies follow a market-to-model approach and
use the so-called nested simulation (to be described).
Nested Simulation
◮ Nested simulation is a two-level simulation procedure: outer loop
simulation and inner loop simulation.
◮ The outer loop involves projecting the VA liabilities along real world
scenarios. At each node of an outer loop scenario, quantities of interest
(the fair market value, greeks) are valuated at each node, using a large
number of simulated risk-neutral paths (the inner loop).
◮ Those are used to help construct a hedging portfolio and to determine
required capitals for the future mismatches between the VA portfolio and
the hedging portfolio.
Nested Simulation: Graphical Illustration
Figure: Nested simulation has two loops: the outer loop and the inner
loop.
Nested Simulation: Challenges
◮ The computation of nested simulation of a large VA portfolio is highly
intensive and often prohibitive.
◮ For example, if we use 1000 real world scenarios in the outer loop and
1000 risk-neutral paths in the inner loop, and project the liabilities or a
quantity of interest at yearly steps for 30 years, then the total number of
projections for each contract is
1000× 1000× 30× 31/2 = 4.65× 108,
which is a very big number. For a portfolio of 100,000 contracts, the
number of projections would be 4.65× 1013.
Our Approach
◮ Select a small number of representative VA policies from the portfolio.
◮ Evaluate the representative VA policies using a nested simulation model.
◮ Build a universal kriging model to evaluate all the VA policies in the
portfolio without simulation.
Selection of Representative VA Policies by Clustering
◮ Let X = {x1, x2, · · · , xn} denote the portfolio of VA contracts, where n is
the number of VA contracts and xi represents the i-th VA contract.
◮ Without loss of generality, we assume that every VA contract is
characterized by d attributes (e.g., gender, age, time to maturity,
guarantee type, etc.) and that the first d1 attributes are numeric and the
remaining d2 = d − d1 attributes are categorical.
◮ Define the distance/metrics between two VA policies x and y as
D(x, y) =
√
√
√
√
d1∑
h=1
wh(xh − yh)2 +
d∑
h=d1+1
whδ(xh, yh),
where xh and yh are the hth component of x and y respectively, wh > 0 is
a weight assigned to the hth component, and δ(·, ·) is the simple
matching distance defined as
δ(xh, yh) =
{
0, if xh = yh,
1, if xh 6= yh.
Choice of Weights wh
◮ For the numeric attributes, assume the form
wh =1
R2h
,
where R2h measures the degree of variability of the data of the h-th
attribute.
◮ The most sensible choice for R2h is the corresponding sample
variance:
R2h =
1
n − 1
n∑
i=1
(xih − x̄h)2,
where x̄h is the sample mean.◮ Another choice is to use the range of the data set:
Rh = max1≤i≤n
xih − min1≤i≤n
xih.
However, this choice would give a small weight to a numeric
variable with a large extreme value.
◮ For the categorical attributes, a common weight, say α, is assigned (e.g.
α = 0.5, 1, or 2).
The K-prototypes (Modified K-Mean) Clustering Algorithm
◮ Initialize cluster centers (representative policies)
selecting k distinct contracts from portfolio X randomly: Suppose
µ(0)1 ,µ
(0)2 , · · · , µ
(0)k .
◮ Update cluster memberships
Updated the cluster memberships γ1, γ2, · · · , γn as follows:
γi = argmin1≤j≤k
D(xi ,µ(0)j ).
◮ Update cluster centers
µ(1)jh =
1
|Cj |
∑
x∈Cj
xh, h = 1, 2, · · · , d1,
µ(1)jh = modeh(Cj), h = d1 + 1, · · · , d .
◮ Repeat the above steps until the memberships of each cluster is stabilized.
◮ The k distinct representative VA contracts Z = {z1, z2, . . ., zk} are the
VA contracts that are the closest to each of the cluster centers.
The Universal Kriging Model
◮ Let Z = {z1, z2, . . ., zk} now be the k distinct representative VA
contracts.
◮ Let Xzj (t) be a quantity of interest, say the dollar Delta, at time t of the
representative contract zj along an outer loop scenario obtained by a
nested simulation model.
◮ The dollar Delta at time t of an arbitrary contract xi in the VA portfolio
is expressed linearly as
X̂xi (t) = λtri XZ (t) =
k∑
j=1
λijXzj (t),
where λi = (λi1, λi2, . . . , λik)
◮ The key is how to obtain the optimal value of λij ! This is an optimal
projection (in math) or process regression (in statistics) problem.
Universal Kriging◮ It is a method in spatial functional data analysis
◮ Let {Xx(t)}, x ∈ P, where x represents a ‘location’ or a VA contract of
the VA portfolio P in our case.
◮ All the stochastic processes {Xx(t), x ∈ P} are assumed to have the
following stationary properties:
E [Xx(t)] = m(t)
Var[Xx(t)] = σ2(t)
Cov(Xxi (t),Xxj (t)) = C(h; t),
where h is the distance between xi and xj .
◮ C(h; t) is often expressed in terms of the so-called semivariogram function
γ(h; t) =1
2Var[Xxi (t)−Xxj (t)] = σ2(t)− C(h; t).
◮ The accumulative semivariogram function
γ(h) =
∫ T
0
γ(h; t)dt
plays an important role in obtaining optimal values.
Determining the Semivariogram Function γ(h)
◮ Choose the spherical function of the form:
γ(h; a, b, c) =
0, if h = 0;
a + b[
3h2c
− 12
(
hc
)3]
, if 0 < h ≤ c;
a + b, if h > c.
◮ Estimate parameters a, b and c by minimizing the L2 distance between
the spherical function and the empirical semivariogram function:
γ̂(h) =1
2|N(h)|
∑
(x,y)∈N(h)
S∑
s=1
(Xx(ts)−Xy(ts))2 .
Here S is the number of discretization steps, and
N(h) = {(x, y) : D(x, y) = h, x ∈ Z , y ∈ Z},
where |N(h)| is the number of elements in N(h).
The Optimal Solution
The optimal values λi = (λi1, λi2, . . . , λik) is the solution of the system of
linear equations:(
A(Z ) B(Z )
Btr (Z ) 0
)
·
(
λi
vi
)
=
(
A(Z , xi )
B(xi )tr
)
.
Here, A(Z ) is a k × k matrix defined as
A(Z ) =
γ(D(z1, z1)) γ(D(z1, z2)) · · · γ(D(z1, zk))
γ(D(z2, z1)) γ(D(z2, z2)) · · · γ(D(z2, zk))...
.... . .
...
γ(D(zk , z1)) γ(D(zk , z2)) · · · γ(D(zk , zk))
.
The Optimal Solution
A(Z , xi ) =
γ(D(z1, xi ))
γ(D(z2, xi ))...
γ(D(zk , xi ))
.
B(Z ) =
f1(z1) f2(z1) · · · fq(z1)
f1(z2) f2(z2) · · · fq(z2)...
.... . .
...
f1(zk) f2(zk) · · · fq(zk)
,
where f1(·), f2(·), . . ., fq(·) are q basis functions linking to key features of a VA
contract.
B(xi ) =(
f1(xi ) f2(xi ) · · · fq(xi ))
.
Numerical Illustration: A Synthetic VA Portfolio
100,000 VA policies with a total account value of 25.552 billion dollars are
randomly generated.
Attribute Values Distribution
Guarantee type {GMDB, GMDB + GMWB} 50%, 50%
Gender {Male, Female} 50%, 50%
Age {20, 21, 22, . . ., 60} Uniform
Account value [10000, 500000] Uniform
GMWB withdrawal rate {0.04, 0.05, 0.06, 0.07, 0.08} Uniform
Maturity {10, 11, 12, . . ., 25} Uniform
Model Specifications
◮ The underlying equity model: a regime switching lognormal model with
two regimes.
◮ 30 year projection period with annual time step.
◮ Projecting dollar Deltas of the portfolio along each of 1,000 outer loop
scenarios.
◮ The dollar Deltas of each of representative VA policies are calculated
based on 1,000 inner loop risk neutral paths.
Model Specifications
◮ We use 6 basis functions (i.e., q = 6) defined as follows:
f1(x) = 1, f2(x) =
{
0, if x is male;
1, if x is female,
f3(x) =
{
0, if x contains GMDB only;
1, if x contains GMDB and GMWB,
f4(x) = normalized age of x,
f5(x) = normalized guaranteed withdrawal rate of x,
f6(x) = normalized maturity of x.
◮ All contracts are scaled to have the same account value to ensure the
weak stationary conditions. We then scale the dollar Deltas estimated by
the UK method back to reflect the account value of the original contracts.
Testing
◮ We test the UK method with the target number of representative
contracts being 550, 1,100 and 2,200, respectively. These numbers
represents roughly 0.5%, 1%, and 2% of the total number of contracts in
the VA portfolio.
◮ For illustration purposes, we randomly select 5 outer loop real world
scenarios and calculate the annual dollar Deltas along the scenarios for
each of the representative contracts using the nested simulation model.
◮ We then estimate the annual dollar Deltas of every scaled contract in the
portfolio using the UK method.
◮ For comparison, we also calculate the annual dollar Deltas of every scaled
contract using nested simulation (extremely time consuming).
Estimated Dollar Delta vs Actual Dollar Delta (550
clusters)
0 5 10 15 20 25−8
−7
−6
−5
−4
−3
−2
−1
0x 10
9
Year
Dol
lar
Del
ta
NS1UKFD1NS2UKFD2NS3UKFD3NS4UKFD4NS5UKFD5
Figure: The annual dollar Delta along five outer loop scenarios estimated
by the UK method with 512 representative contracts.
Estimated Dollar Delta vs Actual Dollar Delta (550
clusters)
0 5 10 15 20 250
0.5
1
1.5
2
2.5x 10
4
Year
MS
E
RW1RW2RW3RW4RW5
(a)
−2.5 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5
x 104
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5x 10
4
(b)
Figure: Figure (a) shows the mean squared errors (MSE) at each point in
time along the five outer loop scenarios. Figure (b) shows the histogram
of the differences of dollar Deltas at time 0.
Estimated Dollar Delta vs Actual Dollar Delta (550
clusters)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
x 105
−2.5
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5x 10
4
Account Value
Diff
eren
ces
of D
olla
r D
elta
s
(a)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
x 105
−0.05
−0.04
−0.03
−0.02
−0.01
0
0.01
0.02
0.03
0.04
0.05
Account Value
Nor
mal
ized
Diff
eren
ces
of D
olla
r D
elta
s
(b)
Figure: Figure (a) is the scatter plot between the account values and the
differences of the dollar Deltas calculated by the simulation model and
those estimated by the UK model.
Figure (b) is the scatter plot between the account values and the
normalized differences of the dollar Deltas.
Estimated Dollar Delta vs Actual Dollar Delta (1100
clusters)
0 5 10 15 20 25−8
−7
−6
−5
−4
−3
−2
−1
0x 10
9
Year
Dol
lar
Del
ta
NS1UKFD1NS2UKFD2NS3UKFD3NS4UKFD4NS5UKFD5
Figure: The annual dollar Delta along five outer loop scenarios estimated
by the UK method with 987 representative contracts.
Estimated Dollar Delta vs Actual Dollar Delta (1100
clusters)
0 5 10 15 20 250
2000
4000
6000
8000
10000
12000
14000
16000
18000
Year
MS
E
RW1RW2RW3RW4RW5
(a)
−2.5 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5
x 104
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5x 10
4
(b)
Figure: Figure (a) shows the mean squared errors (MSE) at each point in
time along the five outer loop scenarios. Figure (b) shows the histogram
of the differences of dollar Deltas at time 0.
Estimated Dollar Delta vs Actual Dollar Delta (2200
clusters)
0 5 10 15 20 25−8
−7
−6
−5
−4
−3
−2
−1
0x 10
9
Year
Dol
lar
Del
ta
NS1UKFD1NS2UKFD2NS3UKFD3NS4UKFD4NS5UKFD5
Figure: The annual dollar Delta along five outer loop scenarios estimated
by the UK method with 1845 representative contracts.
Estimated Dollar Delta vs Actual Dollar Delta (2200
clusters)
0 5 10 15 20 250
2000
4000
6000
8000
10000
12000
14000
Year
MS
E
RW1RW2RW3RW4RW5
(a)
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
x 104
0
1
2
3
4
5
6
7x 10
4
(b)
Figure: Figure (a) shows the mean squared errors (MSE) at each point in
time along the five outer loop scenarios. Figure (b) shows the histogram
of the differences of dollar Deltas at time 0.
Mean Absolute Percentage Error
Number of ContractsMean Absolute Percentage Error
RW1 RW2 RW3 RW4 RW5
512 4.49% 8.75% 10.17% 5.42% 4.18%
987 2.82% 5.75% 6.67% 3.52% 2.62%
1,845 1.72% 3.31% 3.90% 2.07% 1.59%
Table: The mean absolute percentage errors of annual dollar Deltas
estimated by the UK method and those calculated by the simulation
model. RW1 to RW5 denote the five outer loop real world scenarios.
Computing Times
Number of Contracts Entire Portfolio
512 987 1845 100,000
k-prototypes 4.52 3.83 4.22 NA
Nested simulation 166.50 317.75 580.18 32520.68
UKFD 48.16 100.71 199.05 NA
Total 219.18 422.29 783.45 32520.68
Table: Computing times used by the clustering algorithm, the nested
simulation model, and the UKFD method. The numbers are in seconds.
Concluding Remarks and Ongoing/Future Work
◮ The UK method performs well in terms of accuracy and speed.
◮ Although we used dollar Deltas to illustrate the effectiveness of the
proposed approach, it can be applied to other quantities of interest such
as dollar Rho, reserves and risk measures.
◮ In our test cases, we used annual time steps in the nested simulation
model. The method can be also applied to cases when monthly steps or
variable-time steps are used. When the method is applied to monthly
data, the gain in computation time will be even more significant because
the UK method’s performance does not depend on the number of time
steps.
Ongoing/Future Work
◮ Better clustering/sampling methods.
◮ Utilize time-dependent variogram function for more accurate projection.
◮ Application to dynamic hedging via meta-modelling.
◮ Risk capital calculation: VaR and CVaR.
Selected References
Bauer, D., A. Reuss and D. Singer (2012). “On the calculation of the
solvency capital requirement based on nested simulations”, ASTIN
Bulletin, 42, 453–499.
Caballero, W., R. Giraldo and J. Mateu (2013). “A universal kriging
approach for spatial functional data”, Stochastic Environmental Research
and Risk Assessment, 27, 1553–1563.
Gan, G. (2013). “Application of data clustering and machine learning in
variable annuity valuation”, Insurance: Mathematics and Economics, 5,
795–801.
Gan, G. and X.S. Lin (2015). “Valuation of large variable annuity
portfolios under nested simulation: a functional data approach”,
Insurance: Mathematics and Economics, accepted.
Ramsay, J.O. and B.W. Silverman (2005). Functional Data Analysis, 2nd
Edition, Springer Series in Statistics, New York, NY.
Questions?
Thank you for listening. Questions?