Masteruppsats i matematisk statistikMaster Thesis in Mathematical Statistics
The impact of the underlyinginterest rate process- When calculating the best estimate of liabilities
Johan Dellner
Matematiska institutionen
Masteruppsats 2014:3
Försäkringsmatematik
Maj 2014
www.math.su.se
Matematisk statistik
Matematiska institutionen
Stockholms universitet
106 91 Stockholm
Mathematical StatisticsStockholm UniversityMaster Thesis 2014:3
http://www.math.su.se
The impact of the underlying interest rate process- When calculating the best estimate of liabilities
Johan Dellner∗
May 2014
Abstract
Solvency II requires a stochastic valuation for most products withguarantees. This is done in order to determine the time value ofoptions and guarantees (TVOG), which is a part of the best esti-mate of liabilities. One way to determine the TVOG is by project-ing a large number of economic scenarios in a financial projectionmodel. This paper aims to explain and examine the impact of us-ing the following underlying interest rate processes: Hull and White(HW); Cox–Ingersoll–Ross (CIR); and Libor Market Model (LMM)when generating the economic scenarios used for the valuation. Thesethree processes are all used in the insurance industry and fulfill themarket consistency and risk neutrality required by EIOPA under Sol-vency II; while HW allows for negative interest rates; the CIR andLMM does not. The differences in their distributions yield differentresults when the TVOG is determined; which indicates the impor-tance of using the appropriate model as well as understanding it. TheTVOG is significantly different for products with a guaranteed rateof 0 % due to the allowance of negative rates in HW. This masterthesis is limited to a simple guarantee product and a sample of 80model points, applying different guaranteed rates to examine how thedistributions of the generated scenario impacts the outcome.
∗Postal address: Mathematical Statistics, Stockholm University, SE-106 91, Sweden.E-mail: [emailprotected]. Supervisor: Dmitrii Silvestrov.
3
Acknowledgement
I would like to take this opportunity to thank all of my former colleagues at Towers
Watson for the help and inspiration during the progress of writing this paper, as well as
the usage of its software. I’m also grateful for my supervisor professor Dmitrii Silvestrov
for his support and valuable feedback during the duration of this work.
4
[This page is intentionally left blank.]
5
Contents
1 INTRODUCTION ....................................................................................................................... 6
2 INTEREST RATE MODELS ..................................................................................................... 7
2.1 Vasicek model .......................................................................................................................... 7 2.2 Cox-Ingersoll-Ross model ........................................................................................................ 9 2.3 Hull and White model ............................................................................................................. 10 2.4 Libor Market Model ................................................................................................................ 12
3 MARKET DATA ...................................................................................................................... 14
3.1 Yield Curve............................................................................................................................. 14 3.2 Swaption volatilities ................................................................................................................ 14
4 ANALYSIS OF SCENARIO FILES ......................................................................................... 15
4.1 Risk neutrality and market consistency .................................................................................. 15 4.2 Is the yield curve replicated? .................................................................................................. 15 4.3 Is the swaption prices replicated? .......................................................................................... 17 4.4 The distributions ..................................................................................................................... 18
5 THE GUARANTEE PRODUCT............................................................................................... 21
5.1 Assumptions ........................................................................................................................... 21 5.2 Model points ........................................................................................................................... 21 5.3 Portfolio investment ................................................................................................................ 21
6 RESULT .................................................................................................................................. 22
6.1 Guaranteed rate of 0% ........................................................................................................... 22 6.2 Guaranteed rate of 2% ........................................................................................................... 23 6.3 Guaranteed rate of 4% ........................................................................................................... 23
7 DISCUSSION AND CONCLUSION ........................................................................................ 25
8 REFERENCES ........................................................................................................................ 27
6
1 Introduction
The upcoming European regulation Solvency II, as described in full detail in EIOPA’s
technical documentation (2013), increases the overall effort insurance companies’ needs
to put into their regulatory reporting; in many areas this will also force companies to
perform calculations which haven’t be required before. This master thesis will focus on
one of these new “areas”; the calculations of the time value of options and guarantees
(TVOG). The purpose of the TVOG is to reflect the value of the uncertainty of the
obligations the insurance company has taken by promising future guaranteed amounts to
their policyholders; guarantees which create an option value due to the asymmetry in the
contracts. A typical example is an insurance contract with profit participation. If the
insurance company earns an investment return which is in excess of the guaranteed
amount, the policyholder will get a discretionary amount in addition to the guaranteed.
However, if the investment return is insufficient to cover the guaranteed amount; the
insurance company has to cover the loss. This creates the asymmetry which creates the
option value, Hull (2006) describes option values in great detail in his book Options,
Futures, and Other Derivatives.
Frasca and LaSorella (2009) describe one of the most common ways to determine the
TVOG in their article Embedded Value: Practice and Theory. It is done by projecting a
large number of economic scenarios in a financial projection model. This is done by
calculating the best estimate of liabilities for each of the economic scenarios and then
taking the average of it, this is followed by calculating the best estimate of liabilities
under the deterministic certain equivalent scenario, also known as the “central scenario”.
By subtracting the best estimate of liabilities under this certain equivalent scenario from
the averaged amount of the stochastic results, one would get the TVOG.
Many articles, an example is Li and Zhao (2006), have been written discussing the
difficulties of replicating market prices; complex models have been introduced to capture
the behavior of the market. The complexity of the models increases the overall effort
required for model calibrations, so a more complex model might not always be
preferable.
There exists a very limited amount of articles (no articles have been found during the
duration of this work) which examines the prospective impact of the underlying interest
rates models for life insurance contracts, i.e. the valuation of the TVOG. This work will
focus on this area and give an answer the question of if the choice of interest rate model
impacts the valuation of the best estimate of liability. The work is structured with six
district sections as follow; an overview of the properties of the different interest rate
models, a short description of the market data used in the calibrations, an analysis of the
resulting scenario files once they have been generated, a descriptions of the guaranteed
product which has been modelled, the results of the valuation and finally the conclusions
of the findings.
7
2 Interest rate models
There are several different providers of economics scenario generators (ESG) in Europe
as analyzed by InsuranceERM (2013); these generators are in many cases based on
different underlying interest rate models. The two main requirements for scenarios used
to determine the TVOG is that they are market consistent and risk neutral. In short,
market consistent refers to reproduce any market price observed in the market; risk
neutral refers to being arbitrage free by fulfilling the a martingale test (1=1). The
following underlying interest rate processes: Hull and White (HW); Cox–Ingersoll–Ross
(CIR); and Libor Market Model (LMM) all fulfill these demands to certain extent. All
processes are used in the insurance industry, even though LMM tend to be the most
common one based the data collected by InsuranceERM (2013). It’s clear from the
structure of the interest processes that the distribution will be very different; while HW
allows for negative interest rates; the CIR and LMM does not. There are modifications of
LMM which allows for negative interest rates, but the process used here does not. The
CIR uses a fix volatility term; HW and LMM allows for fluctuation of the volatility term
structure. This makes it significantly more difficult to replicate the markets option prices
surface using CIR, compared to HW and LMM. There are other versions of CIR where it
allows for a fluctuation of the volatility term; thus those have been considered to be out of
scope for this work.
2.1 Vasicek model
Both the CIR and the HW models are extensions of the Vasicek model which was
introduced by Oldrich Vasicek in 1977 and has an Ornstein-Uhlenbeck process as its
base. This is a short rate model following the dynamics given by the stochastic
differential equation:
( ) ( ( )) ( ) [1]
Where:
( ) = Short rate
= Short rate reversion parameter
= Long term mean
= Volatility
( ) = Wiener process
The Vasicek model is known as being the first model to capture the mean reversion
characteristics. This can be described in the notation above as moving towards the long
term mean of over time, the speed of this reversion is given by the short rate reversion
parameter of . It’s shown below that the variance of the random process ( ) is
independent of , but not from , which is intuitive since only carries information of
the long term mean.
8
Ito’s Lemma states that for any transformed random process f(r(t), t), where f(r, t) is a
function, which has continuous derivatives, second in r and first in t, and r(t) is the
random process given by the stochastic differential equation [1]:
( ( ) ) [ ( ( )) ( ( ) )
( ) ( ( ) )
( ( ) )
( ) ]
( ( ) )
( ) ( ) [2]
By setting ( ( ) ) to:
( ( ) ) ( ( ) ) [3]
We get the following derivatives:
( ( ) )
( ) [4],
( ( ) )
( ( ) ) [5],
( ( ) )
( ) [6]
Using these results of [4], [5] and [6] in [2] yields:
( ( ) ) [ ( ( )) ( ( ) ) ] ( )
( ) [7]
If we now would integrate [7] between 0 and t:
( ( ) ) ( ( ) ) ∫ ( )
[8]
Rewriting [8]:
( ) ( ) ( ) ∫ ( )
[9]
It’s now straightforward to calculate the expected value and the variance of the Vasicek
model:
[ ( )] ( ) ( ) [10]
[ ( )] [(∫ ( )
)
] [ ]
∫
( )
( )
[11]
A observation of the results tells us that if ( ) , then the expected value of ( ) is
equal to . As well as that an increase of would decrease the variance.
9
2.2 Cox-Ingersoll-Ross model
Cos-Ingersoll-Ross (CIR) in a one factor interest rate model and was introduced by John
C. Cox, Jonathan E. Ingersoll and Stephen A. Ross in 1985. The model is a modification
of the Vasicek model, with the following short rate dynamics given by the stochastic
differential equation:
( ) ( ( )) √ ( ) ( ) [12]
Where:
( ) = Short rate
= Short rate reversion parameter
= Long term mean
= Volatility
( ) = Wiener process
The square root of the short rate guarantees the short rate will never turn negative.
The expected value and variance of ( ) can be determined very similar to the way it was
done for the Vasicek model. The only change one would need to do in equation [7] is to
replace with √ ( ), it would then yield:
( ( ) ) √ ( ) ( ) [13]
If we now would integrate [13] between 0 and t:
( ( ) ) ( ( ) ) ∫ √ ( ) ( )
[14]
Rewriting [14]:
( ) ( ) ( ) ∫ √ ( ) ( )
[15]
One can now calculate the expected value and the variance of the CIR model:
[ ( )] ( ) ( ) [16]
[ ( )] [(∫ √ ( ) ( )
)
]
[ ] ∫ [ ( )]
[ [ ]]
∫ ( ( ) ( ))
10
∫ ( ( ) ( ))
( ( )
( ))
( )
( )
( )
( ) [17]
The expected value is the same as in the Vasicek model, but the variance has changed due
to the square root of ( ) in [12]. In the Vasicek model the variance of the random
process was independent of , this is no longer the case for the CIR-process. The reason
for this dependency is related to the dependence of ( ) in the variance, since the two
factors becomes more integrated in CIR than in Vasicek.
In order to fit the market value of zero coupon bonds, Brigo and Mercurio (2001)
introduced an external deterministic time-dependent shift, ( ) of ( ), such that:
( ) ( ) ( ) [18]
It is in fact this ( ) that is used in the ESG. Once the parameters have been calibrated
in the ESG with market data, the scenarios can be generated through the short rate
dynamics described above.
The scenarios used in this work have been calibrated and generated with Towers
Watson’s MoSes ESG.
2.3 Hull and White model
There are several different types of Hull and White models (HW) the description here will
be limited to the one-factor HW (also known as the extended Vasicek model), the model
was introduced by John Hull and Alan White in 1993. The model follows to the short rate
dynamics given by the stochastic differential equation:
( ) ( ( ) ( )) ( ) ( ) [19]
Where:
( ) = Short rate
= Short rate reversion parameter
( ) = Long term mean
( ) = Volatility
( ) = Wiener process
11
The short rate reversion parameter is adjusting the speed of reversion towards the long
term mean; naturally, the long term mean is set by the initial yield curve. For the volatility
a piecewise volatility is calibrated in the economic scenario generator (ESG) with the
following structure:
( )
{
Similar to the calculations of the Vasicek and CIR model, one can relatively easy
calculate the expected value and the variance of the HW model by applying Ito’s lemma
to [18]:
( ( ) ) [ ( ( ) ( )) ( ( ) )
( ) ( ( ) )
( )
( ( ) )
( ) ]
( ) ( ( ) )
( ) ( ) [20]
By setting ( ( ) ) to:
( ( ) ) ( ( ) ( )) [21]
We get the following derivatives:
( ( ) )
( ) [22],
( ( ) )
( ( ) ( )) [23],
( ( ) )
( ) [24]
Using these results of [22], [23] and [24] in [20] yields:
( ( ) ) [ ( ( ) ( )) ( ( ) ( )) ] ( ) ( ) ( ) [25]
If we now would integrate [25] between 0 and t:
( ( ) ( )) ( ( ) ( )) ∫ ( ) ( )
[26]
Rewriting [26]:
( ) ( ) ( ) ( ) ∫ ( ) ( )
[27]
The expected value is straight forward:
12
[ ( )] ( ) ( ) ( ) [28]
The variance:
[ ( )] [(∫ ( ) ( )
)
] [ ]
∫ ( )
[29]
A notable observation is that in comparison to CIR, the variance is independent of ( ) which correspond the in [17].
Once the parameters have been calibrated in the ESG with market data, the scenarios can
be generated through the short rate dynamics described above.
The scenarios used in this work have been calibrated and generated with Towers
Watson’s MoSes ESG.
2.4 Libor Market Model
Brigo and Mercurio (2001) explores the Libor Market Model (LMM) as it is very
different in comparison to the above mentioned models of Vasicek, HW and CIR; LMM
captures the dynamics of the entire yield curve by using building blocks of forward rates.
They define the standard lognormal LMM model as:
( ) ( ( ) ) ( ) ( ) ( ) [30]
Where:
( ) = Forward rates
( ) = The drift of forward rate
( ) = Volatility
( ) = Wiener process
In practice, the standard lognormal LMM fails to capture the volatility smile in
derivatives market; however, by adjusting the volatility within the LMM to the SABR
volatility model this can be accomplished. Hagan and Lesniewski (2008) derive the
SABR-LMM by first defining the SABR volatility model using the following system of
stochastic differential equations:
( ) ( ) ( ) ( ) [31]
( ) ( ) ( ) [32]
Where:
( ) = Wiener process with correlation coefficient -1 < p < 1 with ( ).
= A constant parameter, > 0
13
= A constant parameter, 1 > b > 0
This yields the SABR-LMM; which is the one used in the scenario generation;
( ) ( ( ) ( ) ) ( )
( ) ∑
( ) [33]
( ) ( ( ) ( ) ) ( ) ∑
( ) [34]
Where:
= The drift of forward rate in SABR-LMM
= The drift of the volatility in SABR-LMM
= A time independent parameter
= A time independent parameter
( ) = A time dependent exponential function
The complexity of the structure in comparison to the Vasicek, CIR and HW makes it
significantly more difficult to derive the expected value. Hence, the derivation is therefore
left out of this work. The reason for this increased complexity is mainly caused by the
fact that SABR-LMM isn’t a short rate model, but a model of the entire yield curve at
every point t. The other models described here only model the short rate at every point t.
The scenarios used in this work have been calibrated and generated with an under
development version of Towers Watson’s STAR RN ESG.
14
3 Market data
The market data used in the scenario files is based on data from Bloomberg as of 2013-
06-30.
3.1 Yield Curve
The yield curve has been derived out of Swedish (SEK) swap rates up until the last liquid
point; the last liquid point has been defined to be 10 years. After the last liquid point the
Smith-Wilson extrapolation method has been applied using a convergence period of 10
years with an ultimate forward rate of 4.2%.
3.2 Swaption volatilities
Swedish (SEK) swaption volatilities are the following.
0,00%
0,50%
1,00%
1,50%
2,00%
2,50%
3,00%
3,50%
4,00%
Years
Yield Curve
1 2 3 4 5 6 7 8 9 10 15 20 25 30
1 30.30% 34.60% 36.05% 34.75% 33.35% 31.55% 30.00% 28.85% 27.60% 26.60% 22.58% 21.34% 20.89% 20.59%
2 32.10% 33.50% 32.70% 31.60% 30.80% 29.80% 29.00% 28.30% 27.50% 26.60% 23.47% 22.55% 22.23% 21.98%
3 33.90% 31.80% 30.40% 29.35% 28.60% 28.00% 27.60% 27.00% 26.60% 26.10% 23.73% 23.02% 22.85% 22.67%
4 31.45% 29.90% 28.70% 27.70% 26.95% 26.60% 26.10% 25.80% 25.60% 25.40% 23.58% 23.19% 23.28% 23.01%
5 28.75% 27.70% 26.80% 26.15% 25.60% 25.30% 25.00% 24.90% 24.80% 24.70% 23.23% 22.98% 23.07% 22.89%
7 25.60% 25.60% 25.20% 24.40% 23.60% 23.70% 23.90% 24.10% 24.30% 24.40% 23.67% 22.99% 22.99% 22.79%
10 23.60% 22.80% 22.50% 22.50% 22.40% 22.40% 22.40% 22.40% 22.40% 22.40% 21.39% 21.39% 20.91% 20.42%
15 21.86% 21.74% 22.02% 22.60% 22.98% 23.17% 23.05% 22.96% 22.90% 22.81% 21.54% 20.71% 19.65% 18.79%
20 22.43% 23.86% 23.75% 23.85% 23.75% 23.75% 23.88% 23.49% 23.23% 22.90% 21.20% 18.98% 17.82% 16.96%
25 23.80% 23.95% 23.94% 24.05% 23.75% 23.55% 23.88% 23.33% 22.88% 22.35% 19.27% 17.25% 16.19% 15.42%
30 22.33% 21.93% 21.54% 21.24% 20.86% 20.48% 20.72% 20.20% 19.78% 19.37% 16.86% 15.32% 14.45% 14.26%
Volatility Surface (%) - Market
Option Expiry (Years)Swap Tenor (Years)
15
4 Analysis of scenario files
The scenario sets used in this work consists of 3 000 scenarios. For the three different
interest rate models described, one set of 3 000 scenarios have been created.
4.1 Risk neutrality and market consistency
The scenario sets must satisfy a number of conditions in order to be risk neutral and
market consistent, a few of those question one need to ask are the following;
Are all investment strategies “in average” equivalent to a risk-free investment?
Is the yield curve derived from the average of discount factor equal to the market
yield curve?
Do the scenarios replicate swaption prices?
Do the scenarios replicate index asset-option prices?
Is the correlation reflected correctly?
Do the scenarios reflect the markets expectations (e.g. inflation)?
In addition to these conditions other validations and checks are performed to confirm that
the scenario set reflects the market (graphical inspections of distributions etc.).
Since only the interest rate models are included in this work and more precisely only the
short rate, this will now be limited to look at the average of discount factors, swaption
prices and distributions.
4.2 Is the yield curve replicated?
The implied yield curve from the different scenario sets is calculated by averaging the
discount factors in the scenario sets. Once the discount factors have been determined the
spot rates can be calculated.
( ) [ ( ) ], ( ) ( )
Where;
( ) Discount factor at time t and interation i in a scenario set.
( ) The implied spot rate at time t.
Note that this is not the same as taking the average of the spot rates in the scenario set.
16
Figure 2: The differences between the implied yield curve and the market yield curve used in the calibrations.
As illustrated in Figure 1 and Figure 2; the differences between the yield curves are no
more than 2 bps, which should be considered to be low.
-0,10%
-0,05%
0,00%
0,05%
0,10%
Differences from Market
Implied LMM
Implied HW
Implied CIR
0,00%
0,50%
1,00%
1,50%
2,00%
2,50%
3,00%
3,50%
4,00%
Yield Curves
Market
Implied LMM
Implied HW
Implied CIR
Figure 1: The implied yield curve using the different interest models together with the market yield curve; which was used during the calibration processes. It difficult by a visual inspection to notice any differences; they are all very close to one another.
17
4.3 Are the swaption prices replicated?
The swaption prices are numerically calculated based on the scenario set and these prices
are compared to the market prices. As illustrated in Figure 3, Figure 4 and Figure 5; the
differences are significant and as expected the CIR-process fails to capture the full
volatility surface. The reason for this is mostly due to the fact that the CIR uses a single
volatility factor, which makes it impossible for CIR to match the entire volatility surface.
LMM has an absolute total error of 7.5 while HW has an absolute total error of 8.4. The
reasoning around why this is would mostly be in regards to the structure of LMM, which
is a lot more complicated than the one around HW. This more advanced structure enables
the model to capture more of the surface.
Figure 3: Total error (%) between Market price and Scenario price for CIR.
Figure 4: Total error (%) between Market price and Scenario price for HW.
Figure 5: Total error (%) between Market price and Scenario price for LMM.
1 2 3 4 5 6 7 8 9 10 15 20 25 30
1 -10.62% -30.79% -39.47% -41.39% -41.80% -40.23% -37.94% -35.55% -31.98% -28.41% -7.20% 7.53% 17.40% 24.50%
2 -34.94% -42.13% -43.33% -42.40% -40.65% -37.36% -33.57% -29.14% -23.95% -18.05% 10.27% 28.07% 37.97% 43.46%
3 -46.72% -46.04% -44.44% -41.46% -37.61% -32.91% -27.56% -21.14% -15.08% -8.57% 23.00% 40.48% 47.93% 50.64%
4 -49.89% -48.45% -45.17% -40.14% -34.08% -27.58% -20.03% -12.72% -5.70% 1.11% 33.50% 48.42% 52.15% 53.17%
5 -50.05% -47.82% -43.19% -37.03% -29.23% -21.07% -12.43% -4.41% 3.30% 10.49% 42.74% 55.48% 56.99% 55.32%
7 -48.01% -46.03% -40.13% -30.91% -19.37% -9.79% -0.75% 7.52% 14.81% 21.50% 48.67% 60.50% 58.48% 53.43%
10 -47.41% -42.70% -34.62% -24.12% -11.41% 1.16% 12.81% 23.06% 31.96% 39.56% 68.40% 70.91% 67.47% 60.85%
15 -51.61% -45.62% -35.08% -22.28% -8.51% 4.03% 16.12% 26.49% 35.06% 42.33% 65.96% 67.27% 63.05% 56.80%
20 -56.79% -52.06% -38.22% -21.85% -6.45% 6.65% 17.15% 28.16% 36.93% 44.55% 64.58% 71.23% 65.05% 58.02%
25 -61.80% -53.10% -36.26% -19.16% -2.92% 10.96% 20.36% 31.64% 40.70% 48.68% 74.30% 76.76% 68.29% 60.33%
30 -62.07% -49.48% -27.57% -6.86% 12.05% 28.71% 39.12% 51.60% 61.37% 69.14% 90.90% 87.25% 76.00% 61.85%
Total error (%) - Market price vs Scenario price
Option Expiry (Years)Swap Tenor (Years)
1 2 3 4 5 6 7 8 9 10 15 20 25 30
1 39.64% 8.15% -6.03% -10.32% -12.68% -12.55% -11.85% -11.45% -9.85% -8.60% -1.70% -2.16% -4.20% -5.74%
2 17.00% 3.07% -1.77% -4.13% -6.08% -6.34% -6.51% -6.25% -5.37% -3.93% 1.16% 0.08% -2.05% -3.56%
3 -0.40% -0.50% -1.15% -1.60% -2.00% -2.30% -2.58% -1.99% -2.03% -1.61% 1.81% 0.53% -1.84% -3.43%
4 -2.99% -2.43% -1.74% -0.73% -0.02% -0.09% 0.48% 0.31% -0.19% -0.64% 1.46% -0.75% -3.97% -5.08%
5 -1.70% -0.83% 0.42% 1.21% 2.28% 2.39% 2.43% 1.69% 0.98% 0.32% 1.88% -0.42% -3.31% -4.49%
7 -2.89% -4.10% -3.10% -0.82% 1.42% -0.02% -1.83% -3.55% -5.19% -6.39% -7.03% -6.86% -8.75% -9.52%
10 0.39% 2.51% 2.69% 1.67% 1.18% 0.34% -0.43% -1.13% -1.77% -2.36% -0.40% -2.16% -1.42% -0.43%
15 6.91% 6.83% 4.95% 1.89% -0.19% -1.43% -1.39% -1.41% -1.59% -1.60% 2.14% 4.48% 8.35% 11.58%
20 -1.29% -7.04% -6.98% -7.66% -7.59% -7.88% -8.63% -7.53% -6.87% -5.89% -0.32% 8.99% 14.01% 17.69%
25 -10.23% -10.90% -11.02% -11.52% -10.71% -10.24% -11.52% -9.84% -8.47% -6.71% 5.56% 15.41% 20.46% 23.91%
30 -11.16% -9.82% -8.43% -7.40% -5.97% -4.48% -5.59% -3.50% -1.73% 0.08% 12.95% 22.44% 28.04% 28.16%
Total error (%) - Market price vs Scenario price
Option Expiry (Years)Swap Tenor (Years)
1 2 3 4 5 6 7 8 9 10 15 20 25 30
1 8.72% 0.27% -4.11% -4.62% -5.38% -4.16% -2.87% -2.42% -1.12% -0.39% 3.97% 1.54% -2.16% -5.54%
2 4.18% 0.57% -0.71% -1.96% -3.57% -3.83% -4.44% -5.05% -5.08% -4.49% -2.56% -5.41% -9.05% -12.23%
3 -3.87% -0.56% -0.04% -0.07% -0.74% -1.67% -3.06% -3.63% -4.74% -5.31% -5.24% -8.24% -11.99% -15.16%
4 -4.35% -1.86% -0.10% 1.02% 1.02% -0.37% -1.16% -2.50% -4.09% -5.55% -7.04% -10.73% -14.84% -17.49%
5 -1.34% 1.54% 2.99% 2.94% 2.51% 1.13% -0.15% -2.08% -3.86% -5.46% -6.96% -10.46% -14.34% -17.27%
7 6.22% 4.51% 4.14% 4.84% 5.56% 2.56% -0.51% -3.25% -5.68% -7.48% -9.91% -11.29% -14.65% -17.73%
10 5.30% 7.24% 6.64% 4.73% 3.37% 1.71% 0.37% -0.76% -1.68% -2.49% -1.86% -5.47% -7.73% -9.72%
15 7.89% 7.49% 4.88% 1.34% -1.08% -2.63% -3.02% -3.46% -4.14% -4.71% -3.94% -5.90% -6.33% -6.36%
20 5.38% -1.59% -2.39% -4.00% -4.70% -5.75% -7.30% -6.92% -6.92% -6.69% -6.80% -3.49% -2.38% -1.73%
25 -1.51% -3.16% -4.01% -5.41% -5.52% -6.30% -9.11% -8.96% -9.06% -8.71% -2.99% 2.28% 3.92% 4.53%
30 -0.15% -1.39% -2.13% -3.04% -3.52% -3.89% -6.62% -5.89% -5.42% -4.69% 3.08% 8.11% 9.59% 7.06%
Total error (%) - Market price vs Scenario price
Option Expiry (Years)Swap Tenor (Years)
18
4.4 The distributions
As a result of the differences in the model structure; the distribution are looks very
different from each other.
4.4.1 Short rate
The distributions of the short rate are different between the models; both CIR
and LMM results in very high short rate towards the end of the projection. HW
on the other hand ends up relatively low rates. Another observation is that the
HW distribution includes negative values; while LMM and CIR doesn’t.
Figure 6: Percentile graph of the short rate for CIR.
Figure 7: Percentile graph of the short rate for HW.
19
Figure 8: Percentile graph of the short rate for LMM.
4.4.2 Short rate index
The non-negative rates for CIR and LMM prevents them to end up with an
index below 1. This is particularly notable for LMM where there is a sharp edge
by 1. The HW index has a wider spread than LMM and CIR.
The total number of scenarios is 3 000 per set and there are scenarios with a
short rate index above 10 after 30 years for all of these sets, those values are not
shown in the figures below.
50
100
150
200
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5 10
Nu
mb
er
of
sce
nar
ios
CIR - Short rate index after 30 years
Figure 9: A histogram of the short rate index after 30 years with 3 000 scenarios.
20
10
20
30
40
50
60
70
80
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5 10
Nu
mb
er
of
sce
nar
ios
HW - Short rate index after 30 years
Figure 10: A histogram of the short rate index after 30 years with 3 000 scenarios.
20
40
60
80
100
120
140
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5 10
Nu
mb
er
of
sce
nar
ios
LMM - Short rate index after 30 years
Figure 11: A histogram of the short rate index after 30 years with 3 000 scenarios.
21
5 The guarantee product
The policyholder will receive a lump sum payment upon retirement and if the
policyholder dies before retirement the lump sum payment occurs at the time of death. In
both cases the payment will be the maximum of the fund value and the guaranteed value.
If the policyholder surrenders before retirement, he will only receive the fund value.
5.1 Assumptions
For simplification; the assumption is made that all expenses are perfectly matched with
premium fees taken directly on the premium paid. It’s the net premium paid (after fees)
which is added to the guaranteed value and the fund value. In addition to the premium
fees, there’s also an annual guarantee charge of 0.2% on the fund value.
5.1.1 Guarantee rate: The model will perform three runs with three different
net guarantee rates; 0%, 2% and 4%.
5.1.2 For mortality M90 is assumed:
( ) ( )
Where:
, and {
5.1.3 Paid-up: The model assumes a 2% flat annual paid-up rate.
5.1.4 Surrender: The model assumes a 1% flat annual surrender rate.
5.2 Model points
The portfolio consists of 80 model points; 40 males and 40 females. They are evenly
distributed between the ages of 21 and 60; and are all assumed to retire at the age of 65.
All policyholders just entered the contract and will pay an annual net premium of 1000
SEK (after fees) as long as they remain in the premium paying state.
5.3 Portfolio investment
The entire fund value is assumed to be invested in short duration bonds and credited
interest based on the short rate.
22
6 Result
The time value of options and guarantees is calculated as the difference between the best
estimate of the liabilities using a central scenario and the mean of the stochastic set of
scenarios. It has been modeled in the financial projection software MoSes, according the
assumptions presented in Chapter 5.
6.1 Guaranteed rate of 0%
In the central scenario with a guaranteed rate of 0%; the maximum of the fund value and
guaranteed value will always be equal to the fund value. This is not the case for the
stochastic scenario set due to the volatility underlying the scenarios. However, since
LMM and CIR doesn’t allow for negative interest rates the option values generated for
those are low. The reason they end up with an option value is due to the guarantee charge
on the fund value, if the guarantee charge would have been zero; the option value would
have been zero. HW on the other hand allows for negative interest rates; and this yields a
significantly higher option value compared to both LMM and CIR.
Figure 9 illustrates the differences between the different models graphically. The
downward slope of the curve is caused by two main factors; policyholders entering the
contract at a high age have paid fewer premiums by the time of retirement; secondly, the
volatility between the different outcomes in the scenario set grows larger over time.
Females live longer, by the assumptions made in M90, which in average generates a
higher guarantee; this explains that females have a slightly higher TVOG.
1 000
2 000
3 000
4 000
5 000
6 000
21 26 31 36 41 46 51 56
SEK
Age of Entry
Time value of option and guarantees for one policy
LMM - Males
LMM - Females
HW - Males
HW - Females
CIR - Males
CIR - Females
Figure 12: An illustration of how the TVOG varies between different ages and models with a guaranteed rate of 0%. For one female policyholder who enters at the age of 21; the TVOG value for her would almost be 5 000 SEK with the HW-model.
23
6.2 Guaranteed rate of 2%
With a guaranteed rate of the 2%; the value of the guarantees in central scenario is 0
(similar to 6.1). As illustrated in Figure 10, the HW model gives the highest TVOG. The
TVOG calculated using the LMM model generates a lower value than HW for the
policyholders entering at a low age, while the value for those entering at a higher age is
more similar.
The CIR model generates a value close to 0; it should be noted that the volatility
underlying the scenarios is very different from HW and LMM due to the limitation with
only one volatility parameter (see section 4.2). This limitation could be causing the low
TVOG.
5 000
10 000
15 000
20 000
21 26 31 36 41 46 51 56
SEK
Age of Entry
Time value of option and guarantees for one policy
LMM - Males
LMM - Females
HW - Males
HW - Females
CIR - Males
CIR - Females
Figure 13: An illustration of how the TVOG varies between different ages and models with a guaranteed rate of 2%. For one female policyholder who enters at the age of 21; the TVOG value for her would be 18 000 SEK with the HW-model.
6.3 Guaranteed rate of 4%
In the central scenario; with a guaranteed rate of 4%; the maximum of the fund value and
guaranteed value will not always be equal to the fund value. The HW model generates the
largest TVOG; the values are closer to LMM for the policyholders entering the policy at a
higher age. For CIR the values increased compared to the previous guaranteed rates of
0% and 2% (see 6.1 and 6.2); however, the limitations of CIR volatility match could
potentially still be the main source of the differences.
24
10 000
20 000
30 000
40 000
50 000
21 26 31 36 41 46 51 56
SEK
Age of Entry
Time value of option and guarantees for one policy
LMM - Males
LMM - Females
HW - Males
HW - Females
CIR - Males
CIR - Females
Figure 14: An illustration of how the TVOG varies between different ages and models with a guaranteed rate of 4%. For one female policyholder who enters at the age of 21; the TVOG value for her would almost be 47 000 SEK with the HW-model.
25
7 Discussion and conclusion
The main findings in this master thesis, commented in more details are the following:
The choice of interest rate model used for the valuation impacts the TVOG.
The differences between the interest rate models seem to the most significant for
products with a guaranteed rate of 0% due to the allowance of negative rates in
HW.
The interest rate models fail to fully replicate the market prices, especially the
CIR.
An initial inspection of the numbers indicates that the choice of interest rate model has a
significant impact on the TVOG. The differences could potentially cause management
decisions which wouldn’t have been taken if another interest model would have been
used; giving the management a different view of the liabilities. It’s important to
understand the differences and be aware of its limitations when applying them; an interest
rate model used properly for one purpose doesn’t necessarily need to be ideal for another.
This work is limited to a very particular guaranteed product, three different interest rate
models with certain calibrations modifications and market data from a specific date. The
differences illustrated in Chapter 6 can be a result of these conditions; in this case HW
always ends up with a higher option value than both LMM and CIR. The conclusion
shouldn’t be that HW will generate a higher TVOG; it’s rather the fact that these different
models will result in different option values and this will directly have an impact on the
balance sheet.
This work has been limited to a simple guaranteed product with a single payment; it’s to
be expected that products with other structures could have a very different outcome. If
one would consider any ratcheting product where the guaranteed amount increases
depending on the market’s development; the path of the fund value up until retirement
would have a crucial impact of the TVOG. Another example of these path dependencies
is by increasing the number of payments from a single payment; if one would pay a
discretionary benefit at the age of 65; the fund value might not be able to cover the
guarantees at the age of 70 due to the payment at 65.
In general; equities have a higher volatility than bonds; and the volatility is the main
source of any financial option value. Therefore an obvious limitation in this work to
assume 100% short duration bonds when investigating the sensitivities of the TVOG.
Nevertheless, it’s still interest rate models that’s underlying any index asset model which
is typically added to model equity returns and the cash flows will always be discounted
by the short rate.
It’s noted that the CIR model is using a single volatility parameter and therefore fails to
capture the full volatility surface of the market. As a result; it’s very difficult to draw any
conclusion in regards to the differences in the generated TVOG using CIR compared to
HW and LMM; other than the fact that using CIR with a single volatility parameter is too
simplistic to capture the TVOG. CIR could potentially be used to capture it by using
different sets of stochastic scenarios; depending on the duration of guarantees; no further
26
analysis has been done in this respect. Another option would be to use another modified
version of CIR which allows for a fluctuating volatility term.
The figures (Figure 9, 10 and 11) illustrating the short rate index explains the outcome of
the modeled guaranteed product very well. It’s possible to directly spot the result of the
non-negative rates in CIR and LMM; all values are above 1. On the other hand, HW has
several index values below 1; this could be linked the Figure 12 where a guaranteed rate
of 0% has been applied. In short and slightly simplified, the policyholder is guaranteed to
end up with an index value of 1 and any values below 1 will create an option value. This
is no longer the case with the guaranteed rate of 2%; instead (still simplified) the
policyholder is guaranteed ( ) ( ) . The value of 1.8 explains the
low option value for CIR with the guaranteed rate of 2%; one could see in Figure 9 that
CIR got almost its entire distribution above 1.8. For the guaranteed rate of 4% we end up
with a simplified guaranteed ( ) ( ) and all distribution are
covering that value and as it’s calculated all models ends up with a significant option
value.
Finally, the focus in this work has been to present and illustrate that the choice of interest
model is important and will affect the results. It’s left for the reader to select which model
would suit ones purpose. Further research studies could potentially be done to conclude
the most appropriate model for this specific product. Additionally, it would be interesting
to look into the impact for other guaranteed products, e.g. ratcheting products.
27
8 References
[1] Hagan, P., and Lesniewski, A.: LIBOR market model with SABR style stochastic volatility, draft (2008). [2] EIOPA: Technical Specifications for the Solvency II valuation and Solvency Capital Requirements calculations (Part I), EIOPA-DOC-12/362
[3] Towers Watson: MoSes Economic Scenario Generator 3.1.1 – Technical Documentation, Towers Watson (2011). [4] Brigo, D., and Mercurio F.: Interest Rate Modeling – Theory and Practice, Springer (2001). [5] Hull, J. C.: Options, Futures, and Other Derivatives, Sixth Edition, Prentice Hall (2006). [6] Wu, L.: Interest Rate Modeling – Theory and Practice, CRC Press (2009).
[7] Andersson, G.: Livförsäkringsmatematik, Svenska Försäkringsföreningen (2005).
[8] Market Data taken from Bloomberg
[9] Lindgren, B. W.: Statistical Theory, CRC Press (1993).
[10] Frasca, R., and LaSorella K.: Embedded Value: Practice and Theory, Society of Actuaries (2009). [11] The InsuranceERM guide to ESG producers, InsuranceERM (2013), published 21 May 2013 in Software IT. [12] Vasicek, O.: An Equilibrium Characterisation of the Term Structure, Journal of Financial Economics 5 (1977). [13] Hull, J. and White, A.: One factor interest rate models and the valuation of interest rate derivative securities," Journal of Financial and Quantitative Analysis, Vol 28, No 2 (1993). [14] Li, H. and Zhao, F.: Unspanned Stochastic Volatility: Evidence from Hedging Interest Rate Derivatives, Journal of Finance 61 (2006).