Stochastic Modeling - Financial Reporting - Record

_________________________________
*
Copyright © 2003, Society of Actuaries
RECORD, Volume 29, No. 1*
Washington, D.C. Spring Meeting
May 29–30, 2003
Session 68TS
Stochastic Modeling in the Financial Reporting World
Track: Financial Reporting
Moderator: ROBERT W. WILSON
Panelists: RONALD J. HARASYM
Summary: The risks of variable annuities and the coming of international
accounting standards have brought stochastic modeling out of the back rooms of
investment quantitative analysts and into the everyday world of risk management
and financial reporting. Panelists discuss the basics of stochastic approaches to
setting reserves and embedded values. Attendees leave with a greater knowledge
of how the coming stochastic approaches to valuation will impact financial
reporting.
MR. ROBERT W. WILSON: Ronald Harasym is assistant vice-president of Financial
Risk Management in the Corporate Risk Office with Sun Life Financial in Toronto. In
this capacity, he is responsible for monitoring, quantifying and managing specific
capital market risks for the world-wide operations of Sun Life Financial. He is a
Fellow of the Canadian Institute of Actuaries and a Fellow of the Society of
Actuaries. He is also a Chartered Financial Analyst. Ron is a graduate of the
University of Toronto, with an MBA from the Rothman School of Management. He
has fourteen years of experience in the insurance industry, and over the past eight
years, Ron has worked extensively on the quantification and hedging of embedded
option risk in the United Kingdom and the United States. He's a member of the
Society of Actuaries' Course Seven Education and Examination Committee and he
frequently lectures in the Department of Statistics at the University of Toronto.
MR. RONALD J. HARASYM: I first started working on stochastic modeling back in
1996 where I was a one-person shop looking at embedded asset and liability
options. At the time, the embedded options were not being properly priced—–they
were more or less being given away for free in Canada, the United States and the

Stochastic Modeling in the Financial Reporting World 2
U.K. I started applying stochastic models to estimate the cost of embedded options,
but not to the liking of the marketing people when I started implying that the
embedded options were not without cost to the company. Basically, it started off
fairly small, looking at investment performance guarantees on segregated (separate
account) funds that were in the pipeline, various book value wrappers, as well as
pricing embedded options that were on the asset side.
In 1998, I was asked to look at the risk inherent in guaranteed annuity options on
deferred annuities in our U.K. business. This was an industry-wide problem. The
risk was similar to the risk inherent in guaranteed minimum income benefit riders
that were (and still are) offered on variable annuities in the United States. The
difficulty was that no one was looking at the risk of the embedded option and
stochastically modeling it. The U.K. was a very deterministic modeling environment.
My objective was to model the embedded guaranteed annuity option and to
determine an economic risk profile of the guarantee. Ultimately, we used the tool to
hedge the risk and evaluate trade-offs between various investment strategies. From
this perspective, the use of stochastic modeling worked quite well. My presentation
today will describe how stochastic modeling has moved out from the back rooms of
offices and into the highly visible financial reporting world.
I want to give you a high-level view of stochastic modeling in a generic framework.
It's a framework that I go through when performing any stochastic modeling task.
People tend to think of a stochastic model as a single model, but in fact there are
often several other processes behind it, such as random number generation and
economic scenario generation. For demonstration purposes, I'll use a guaranteed
minimum income benefit (GMIB) rider. I'll show the modeling results and some
sensitivity testing, comment briefly on reserve and capital relief, and then share
some final thoughts.
A few years ago, there was quite a discussion on the discussion boards of the
Canadian Institute of Actuaries as to a definition of stochastic modeling. The word
"stochastic," courtesy of dictionary.com, is derived from the Greek word
stokhastikos, which means, in short, "to guess at." You might want to call it
something like "intelligent guessing." A stochastic model by definition has one
random variable in it and specifically deals with time-variable interaction. The
stochastic model itself doesn't have to be a simulation; it just has to have a random
variable in it. What we typically do is repeat those random elements, ending up
with a series of results. This is what we refer to as a "Monte Carlo" simulation.
A stochastic simulation is an imitation and simplification of a real world system.
People tend to think of a stochastic model as the real world, but let's face it, it's
just another tool in your tool kit. One of the advantages of stochastic modeling is
that if you're going to either offer products that have embedded options or
purchase assets that have embedded options, you can try to price the embedded
options and assess the risk. Then once you've priced the risk, or at least quantified
the risk, you can work the theoretical cost into the pricing. In the past, the majority

of these embedded options were never priced appropriately, often being given away
for free, which has come back to haunt numerous insurance companies.
Stochastic simulation is a useful tool for forecasting purposes. There are also
advantages from a financial reporting perspective. You can forecast where you will
be at future points in time. One benefit is that you end up with distributions of
results. As long as you have a robust model and a robust framework around it, you
can project financial statements. If it's not required yet, it may be required soon;
there are certainly advantages to performing stochastic modeling over not doing it.
From a risk management perspective, some people say that if you can't quantify
the risk, then it's not a risk. Or they think that what you don't know won't hurt you.
But in many products, there are significant embedded options. Unless you quantify
these, you don't really know what's coming down the pipeline.
Another use of stochastic modeling is for the simulation of very complex systems
where a simple closed-form solution does not exist. For example, if you don't have
a simple formula or equation to price your embedded option, then you can perform
a stochastic simulation and derive a theoretical or an expected answer. However,
keep in mind that stochastic modeling is part art, part science and part judgment.
You have to use common sense.
Stochastic simulation is not a magical solution. You need to perform reality checks
during the modeling process and understand the limitations of the model. What you
get out is only as good as what you put in. If the stochastic model is weak and
doesn't capture the appropriate variable interactions, then don't expect the
interactions to fall out afterward.
In a stochastic simulation, complex situations with long time frames can be
compressed into a more manageable package. You can get a better understanding
of the dynamics of your product. You can test policyholder behavior to see where
you get hurt the most. People tend to think about the downside risk in stochastic
modeling. However, there's also upside risk. The advantage of stochastic modeling
is that the whole distribution of risk can to be quantified and examined.
Stochastic simulation is preferred over deterministic modeling when regulations
provide real economic incentives, such as significant reserve or capital relief, for
performing stochastic simulation. When risk is modeled deterministically or if you
model risks independently, then you're not going to pick- -up the benefit of
diversification. Modern investment portfolio theory says that for any risks that have
a correlation of less than positive one, by combining these risks, there will be some
benefit from diversification. When you model risk stochastically and inter-
dependently, the end result is a superior risk assessment with the quantification of
the benefit from diversification. There are some exceptions when dealing with non-
linear risks such as GMIBs. Finally, my favorite benefit from stochastic modeling is
that you can watch your company fail over and over again as you simulate those

embedded liability options that many people originally deemed to be worthless!
There are limitations to stochastic modeling. It requires an enormous investment in
time and expertise. It is technically challenging and computationally demanding.
Often, reliance is placed (and companies become dependent) on a few "good"
people. Frequently, stochastic models become a "closed shop." It seems as if those
individuals who have all the knowledge don't want to share it, don't want to
document and don't want to explain their work to others. So you end up with some
very challenging situations.
Another limitation of stochastic simulation is that the use of thousands of scenarios
may create a false sense of precision. Some people tend to look at a stochastic
simulation and automatically think it has both accuracy and precision, whereas in
reality, it's only as good as what you've put into it. If your model is specified
incorrectly, then your risk assessment will be incorrect and you will be hedging the
wrong target. In the end, the stochastic model will become a disadvantage rather
than a competitive advantage.
The results of a stochastic simulation can be difficult to interpret, especially to
senior management, who might be used to seeing only one number. Often the
results can be presented in many different formats, such as scatter plots,
histograms, conditional tail expectations and so on. Effective communication of the
stochastic results is a challenge. Some people tend to use jargon that is not readily
understandable. I find that in communicating stochastic results, you have to keep it
very simple and distill the information down to a few basic points.
You wouldn't want to perform stochastic modeling for everything; it wouldn't make
sense. There are a number of situations where stochastic modeling is preferred
over deterministic modeling, such as when you have skewed or discontinuous
distributions or cost functions. Examples of those would be when modeling
investment guarantees on segregated funds in Canada, guaranteed minimum death
benefits (GMDBs) and GMIBs in the United States, or any situation where there's
significant volatility or sensitivity to the initial starting conditions. A good example
would be if you have interest rate options, and the current market conditions are
materially different than they were in the past.
Where there's path dependence, such as on guaranteed annuity options in the U.K.
or GMIBs in the United States, the level of policyholder annuitization will depend
upon what path or what level interest rates have been over the period of time being
projected. You can capture this relationship in a stochastic model; it may not
capture as well in a deterministic framework. Finally, cases where volatility or
skewness of the variables is likely to change over time can also be incorporated
into a stochastic simulation.
Now I am going to work through the steps that were followed in order to perform
the stochastic simulation of a GMIB rider on a variable annuity. A generic

framework, from a process perspective, is shown in Chart 1. There really isn't a
starting or an ending point. It's a constantly evolving process; you're constantly
feeding back and looping through it. In the flow chart, the rectangular boxes
indicate processes or models, while the parallelograms indicate input/output
(typically data).
Within this framework, model and data validation are very important steps. Some
people take the historical data "as is." Yet I've spoken to other people who say they
spend an enormous amount of time going through the data, particularly if they're
trying to fit fund returns to a benchmark. When measuring fund returns, statutory
holidays in Canada, the United States or the U.K. don't always line up, so there
may be data lags. Sometimes the fund values are not correctly updated. The
numbers that get fed through to you could be incorrect. People should be prepared,
on a cost-benefit basis, to spend time performing data validation. Finally, once you
have calibrated your models, perform back-testing to check that the outputs are
consistent with the inputs.
Chart 2 provides a verbal interpretation of the steps that you'd go through, from
beginning to end. There are a few points to keep in mind. No one model fits all
solutions. You have to learn to walk before you can run. Keeping it simple is
probably the best thing. There always seems to be a tendency to go the more
complex route, whereas the 80/20 rule seems to apply—80 percent of the benefit is
picked up in 20 percent of the effort. You can certainly get a great pick-up initially
modeling the additional components, but going too far into the details may not be
worth it. Finally, always strive toward actionable results.
I now want to talk briefly about random number generation. It's often overlooked
during the model development phase. The objective is simple enough—to produce
numbers that are uniformly distributed between 0 and 1. You probably never
thought that random numbers could be such an exciting field, but it is a
fundamental building block of any stochastic simulation. People rely on random
number generators within software, but often they have no idea how robust the
random number generation process is. If you look on the Internet, you'll find many
references for random number generators that are available. Many academic sites
indicate tests that you should run your random number generator through.
However, no random number generator will satisfy all tests. They will repeat sooner
or later.
A good practice is for your company to adopt a standard random number generator
that's used for all stochastic modeling purposes. Then as other stochastic models
are developed, they can rely on the standard random number generator. Also, it
can take some of the mystery out of trying to debug unusual results. I have seen
cases where the random numbers don't repeat very often. What ends up happening
is that you run 1,000 scenarios, and after 200 or 300 scenarios, the scenarios start
to repeat or there's something in there that's cycling. You thought you ran 1,000
scenarios, but actually you ran 250 scenarios four times.

An economic scenario generator is another important part of a stochastic simulation
that contains capital market-related guarantees. Here, again, it makes sense to
have a common generator within your company. There are many to choose from.
You need to determine whether you require an economic or a statistical model.
Calibration is always an issue. A desirable characteristic to look for is that the
scenario generator is an integrated model. In other words, it not only models
equity returns, but also interest rate yield curves or fixed income returns as well as
inflation and currency, all in an internally consistent fashion. A few models I've seen
don't model the relationship between interest rates and equity, but I think that's
changing. People now recognize that if you're going to model GMIBs, for example,
you have to capture the joint interest rate and equity market return relationship in
your model. Another desirable characteristic that you want to have in your scenario
generator is a component approach. Instead of having to re-run the whole model
for just one part, it's flexible enough so you can run only the pieces that you need.
I now plan to use a GMIB rider as an example, because it is a fairly simple product
(in concept at least). With this product, a person deposits $100. There's a
guaranteed account and a market value account. The guaranteed account
accumulates at a "roll-up" rate. In this case I've assumed 5 percent per annum.
The market value account is driven by the market value of the funds (accounts)
that the person invests in.
Let's assume that the nature of the situation, which is not unlike companies in the
industry right now, is that we have a block of contract holders with an overall
guaranteed account value of $1.4 billion while the market value is equal to $1
billion. So the guaranteed income benefit option is "in-the-money." The options,
due to assumed exercise constraints, are three to four years from being able to be
exercised. A key thing to remember about the GMIB is that there's a dual impact.
You have an interest rate guarantee tied with a mortality guarantee. On one hand,
you have a conservative interest rate set fairly low, maybe 3 or 4 percent on the
annuity. (It was conservative when it was priced, but unfortunately it's not any
more.) On the other hand, you have very aggressive roll-up rates. You're
guaranteeing 5 percent, but the only way the person can realize that is by taking
the guaranteed annuity.
I generated equity returns using regime switching log-normal model with two
regimes. Fixed income returns were modeled using a Cox-Ingersoll-Ross model.
Historical correlations were used. The economic scenario generator was calibrated
using maximum likelihood estimation. There are calibration issues that you have to
worry about that include limited, often inconsistent, data. The Choleski
decomposition methodology was used to generate correlated returns. If you don't
use time series of equal length, your model may eventually fail because you won't
properly satisfy the requirements for the Choleski decomposition. There are various
"fixes" for some of these problems, but they can become complex.
One logical question is, how often do you recalibrate the economic scenario

generator? If you are presenting results quarterly and are also recalibrating
quarterly, then changes in the results will be driven by the recalibration as well as
by changes in the market condition. You can actually be making the presentation of
your results more complicated if you recalibrate more often.
In my example, I ran 1,000 scenarios using a monthly frequency and a 35-year
projection horizon. Chart 3 is a scatter plot of the present value of GMIB cash flows
as a function of the average interest rate per scenario. In looking at this chart, the
benefit is certainly under water—that is, it's "in-the-money." You can see there's a
limited upside, given the modeling assumptions that went in. Note that whether
interest rates are high or low doesn't make too much difference.
Chart 4 shows the present value of GMIB cash flows as a function of the average
equity return per scenario. It's the same scatter points, but plotted in a different X-
Y plane. Here, you also see there's a limited upside. It should be, because it's the
same results being plotted as in Chart 3; if it were otherwise there would be a
problem. That is one of the difficulties with stochastic modeling. Sometimes you're
dealing with so much data and while condensing it down, it's easy to make
mistakes. Note that as the equity return decreases, from the upper right to the
lower left, there are a number of scenarios where the end result could be classified
as catastrophic in nature.
The conditional tail expectation (CTE) is a measure of downside risk. It's defined as
the average of outcomes that exceed a specified percentile. In other words, if you
want to calculate CTE (90 percent), you order the results from 1 to 100 (if you had
100 observations) and average the ten worst results. It's considered a more robust
measure than a percentile. When you're dealing with embedded options where you
have skewed distributions, the CTE picks up on the catastrophic events in the tail.
In some cases, the CTE is modified. For example, if you have events where you
don't allow the upside to be averaged in, you just floor it out at zero.
Selected CTE measures for our GMIB rider example are shown in Chart 5. By
definition, CTE (0 percent) is equivalent to the average result. The results are
presented as negative present values of the GMIB rider cash flows, so on average
the present value is minus $43 million. Recall that this option is "in the money"—
that is, the option has value to the contract holder. What was thought to be a free
option a few years ago, or given away for virtually nothing, is now seemingly very
expensive.
The percentile and CTE curves are plotted together in Chart 6. I originally also
plotted a modified conditional expectation measure where I zeroed out the
favorable events. Unfortunately, because this product was so under water, the
modified conditional tail was almost identical to the true conditional tail, so it didn't
really capture the point that I wanted to make. Usually, by zeroing out the positive
events, the modified conditional tail will result in a larger number. However, as you
move into the extreme tail, they eventually converge.

In Chart 6, the present value of GMIB cash flows is presented by percentile on the
left side and by CTE on the right side. One observation is that the CTE measure
looks considerably smoother. That's because you're averaging events, so there's
more information contained in there. Another observation is that the percentile
curves have various crossover points.
I performed some sensitivity testing on the model. The base equity return was
assumed to equal 8 percent. I reset the model, recalibrated it such that the average
equity return was equal to 6 percent, and reran the economic scenarios. Under a
lower equity market return, from a percentile and a CTE perspective, the results
have deteriorated. I also independently cut the lapse rate in half and the cost of the
rider skyrocketed.
There are a number of benefits from performing sensitivity testing. Aside from
quantifying an impact of changing assumptions, it is useful for validation of the
model. If you sensitivity test an assumption and the results don't bear out what you
expected, then you can do two things: you can conclude that it's insensitive, or you
can go back and check it. I don't think enough time is spent on the model
validation step. Sensitivity testing, if the model is correctly specified and the
interaction correctly modeled, allows one to direct more attention to the
assumptions to which the results are most sensitive. If you're wrestling with an
assumption for a long period of time and your results aren't sensitive to that
assumption, then you are wasting your time.
When performing stochastic modeling, the analysis-of-change step is very
important. Often there are a number of parameter changes. It is useful to construct
a build from one point in time to the other. Aside from getting a better
understanding of the model, you're confirming that the model is responding in a
way that you believe it ought to be. One is able to gain a better understanding of
the dynamics underlying the stochastic simulation by performing sensitivity testing.
One method of presenting sensitivity testing results is shown in Charts 7 and 8. On
the very left side is the base case. The line going across represents the CTE (70
percent) level for the base case. I picked CTE (70 percent) level merely for a
reference point; it could have been any other CTE level. By looking at the results in
this fashion, you can see how the various CTE levels change when you adjust the
various assumptions.
Based upon the second column in Chart 8, changing the rider premium charge by
10 basis points (which may not be possible on existing business) doesn't make a
big difference. If you're thinking about changing the pricing spread that you use in
annuities at payout time by 10 basis points, it doesn't make a significant difference
either. If you alter your mortality assumptions, in this situation, the GMIB rider is
so far under water that you're just tinkering on the edges.
As far as the lapse rates are concerned, you can see that if by some miraculous

chance we have double the lapse rate, the results are more favorable and
compressed from a risk perspective. On the other hand, if the lapse rates get cut in
half, for example if the contract holders suddenly understand that the GMIB rider
contains a valuable benefit, then you can see that the risk gets significantly worse,
especially when measured at the CTE (90 percent+) level. The discovery of
embedded options is interesting because they can be discovered in one of two
ways. Either you find them, or someone else will find them for you.
Normally the policyholder annuitization rate is a critical assumption; that is the
take-up rate, or the rate at which they annuitize. This assumption probably would
have been more sensitive if it had been more "at the money." Still, the outcome of
increased annuitization is negative.
Sensitivity to the investment assumptions is presented in Chart 7. I've scaled the
charts the same between the investment and the liability assumptions so you can
make a comparative analysis. What does not show up when you look at the
assumptions independently are the multiple variable interactions that occur in
reality. For example, there's often a weak negative correlation between equity
market returns and interest rates. When the equity market goes down, it's not good
for this product, because the market value of the contract holder's account has
fallen, while over time the contract holder's guaranteed account is steadily
increasing at 5 percent per annum. When interest rates fall, then the current pricing
rate for the annuity is lower and it's starting to hit the guaranteed rate, so that's
not good. With the GMIB, when you start looking at the various jointly determined
risk numbers, you quickly realize that you cannot just add the risks. For example, a
minus 25 percent equity market shock combined with a minus 200 basis point
shock in the interest rate market can't be taken in isolation and added. Due to the
non-linear risk profile of the GMIB rider, one needs to test the income and equity
return assumptions jointly.
One of the advantages of performing stochastic modeling is the potential for
reserve and capital relief. In the Canadian environment, the use of stochastic
approaches is favored over deterministic approaches. Stochastic techniques are also
useful in quantifying risk and can be used for risk management and hedging
purposes.
Keep in mind that no one model fits all. In our company, we've tried to adopt a
standardized scenario generator for various stochastic modeling purposes so we can
focus on the specification of the asset-liability model itself and the end results as
opposed to re-building the scenario generation process. However, you still end up
calibrating the model differently for Canadian versus U.S. requirements, or for
whether it's a pricing exercise or a risk-management exercise.
In stochastic modeling, you want to cultivate best practices. You want to avoid
"groupthink," where you have a group of people who think that what they're doing
is absolutely correct. You end up with less communication going out and very poor

judgment being made. It becomes not just a black box of a model, but also a black
hole of a department.
Keep it simple; keep it practical. Don't use a sledgehammer to crack a walnut. I've
heard many analogies describing stochastic modeling, but the point is that you
don't want to "over-do" the situation. Any type of model that you put together
should be suitable for the situation.
Focus on accuracy first and precision second. I tend to round the numbers so as not
to overly imply any precision, because to some extent you're misleading yourself.
Add complexity on a cost/benefit basis. Some people want to include everything in
the model and then it takes forever for the model to run. You want to keep it
simple.
You also want to perform reality checks along the way. Stand back and assess the
situation. For example, take a look at the simulated output and generate summary
statistics. Make sure that what is coming out was what you thought ought to be
coming out. You might find a problem. On the other hand, you might find that you
need significantly more scenarios than you ever thought to get a degree of
confidence that you are comfortable with. There are formulas for calculating
confidence intervals around the percentiles and around the CTE measures; they are
quite useful. It takes a surprisingly large number of scenarios to reduce the
confidence intervals.
Avoid the creation of black boxes. There's nothing worse than having a model that
no one understands. If you do create a model that is on a common platform that
people around the company use, that's a good thing, because at least there's a
common understanding as to what the model is and what it does. One problem with
some software languages is open code. Anyone can go in and change things. It can
be hard when you're running Excel VBA code to lock it down.
There are other issues to wrestle with. Some models generate more volatility than
others. Look at the various approaches when it comes to financial reporting. For
example, if you're using a regime switching lognormal model with two regimes,
you'll get much more volatility in the results over time than if you're using a simple
log-normal process. The trouble is, when you're using a log-normal process, you're
not getting thick enough tails, whereas the regime switching manages to satisfy the
weaknesses of a simpler model. In short, there are always trade-offs to deal with.
However, with a regime switching model, it may be more difficult to interpret the
results.
For example, in a regime switching model with two regimes, you end up with six
parameters: two means, two volatilities and two transition probabilities. You
calibrate to your data using maximum likelihood estimation. You add in one more
monthly return and your revised parameter estimates appear to be dramatically
different. You ask yourself how you rationalize one parameter set from the other.

Given those six sets of parameters, you can calculate the expected return and the
expected volatility. Now, you've distilled it down to a mean and a volatility, and it's
much simpler to understand and compare.
There are other things to keep in mind when you're looking at calibration. I've
observed when calibrating, especially in a regime switching lognormal model, if you
add three months of data into a time series that's even thirty years long, and those
data points happen to be very extreme movements, then the calibrated expected
mean can significantly overshoot or undershoot the historical mean. This really
came to light when we were performing some earnings-at-risk analysis. We added
three monthly data points that were extreme events to our data. When we re-ran
the model, the quantified risk level declined. People were asking how the risk could
decline given that we have added three months of highly-volatile data and we only
changed the calibration or the scenario generator. It turned out that the model in
this case happened to undershoot the mean and volatility, and in the previous case
it overshot the mean or volatility. So the results can be sensitive to the dynamics of
the calibration procedure.
We also want to make sure that the fixed income returns are integrated with the
equity returns in the model, especially if you're modeling embedded options that
have joint interest rate and equity risks. The question is how to capture the
correlations between the markets. Another question is how many scenarios to use.
You never get a good answer on that, and I'm not going to give you a good answer.
One of the common answers is that you just keep running more sets of 1,000 or
5,000 and so on until you see that you don't get a material difference in the end
results between the sets of scenarios. Some people just give up after 10,000, based
upon systems imitations.
Policyholder behavior is a critical element when assessing the risk embedded within
GMIBs and GMDBs, but how do we model it? We don't have a material amount of
data on policyholder behavior. Whatever data we do have can be highly dependent
upon the company and the product design, so how to model policyholder behavior
is a tough issue to wrestle with.
MR. HENRY M. MCMILLAN: (Pacific Life) I'm curious about your calibration. It
appears that you're normally working with monthly data for stock prices and
interest rates. Is that correct?
MR. HARASYM: For the example I presented, that is correct.
MR. MCMILLAN: How far back in time do you typically go to get the information
that you're looking at? How many months?
MR. HARASYM: It depends on what we're doing. For example, if we're doing
Canadian annual statement reserves for segregated fund products, we'll have to go
back to the 1950s because the regulations tell us to do so. If we're performing a

pricing exercise, we may go back only 10, 15 or 20 years. If we're doing an
earnings-at-risk, we'll probably look at various combinations of 10, 15, 20 and 30
years and try to get a process that we're comfortable with. We try to settle on a
calibration process so we don't keep changing our process each month or quarter.
I guess the answer to your question would depend upon the product. For example,
if we were doing hedging of a segregated fund or a GMDB, if we were trying to do
dynamic hedging and fitting data to it, then we've gone to daily information. That's
where we found there were a lot of data problems with fund information that was
being provided.
MR. MCMILLAN: So the answer is that it depends. But if you were using monthly
data, you would probably always look for something in the area of 300 to 500 data
observations when you're trying to estimate the parameters that you have here. If
that's daily data, then you have 500 daily observations, et cetera?
MR. HARASYM: Yes, you certainly want to have a reasonable time series. There is
one thing to keep in mind though if, for example, you're looking at interest rates.
We often talk about stock prices, but if you're looking at calibrating for interest
rates and there's a different interest rate regime that you're in or a different macro-
economic environment, then it may make sense to truncate it somewhere.
MR. MCMILLAN: That's absolutely right. I was wondering how you possibly dealt
with that when you were getting to the interest rate process because that's
obviously changed. I would expect that with you doing international activity as well,
you might be interested in foreign exchange movements, and then you have
foreign exchange regimes that you have to deal with.
MR. HARASYM: Absolutely correct; that's a good point. A lot of it is gut feel. The
interesting aspect about foreign exchange is that we do have to model it for
earnings-at-risk purposes because we have operations around the world. A number
of academic papers out there say a basic geometric Brownian motion with a few
minor adjustments is as good as anything. Much to my surprise, when talking to
people in the banking community, that's all they were doing too.
MR. WILSON: I think it's safe to say if you're looking at foreign exchange, it
depends on the time frame over which you were intending to look. If you're looking
at earnings-at-risk, which probably covers a one-year period and is totally random,
a Brownian motion probably works rather well. But if you're looking at doing some
sort of projection from a strategy motivation, which looks over a 10- or 15-year
period, then there probably is a relationship between interest rates and currency
changes. But over a short period of time there isn't one.
For example, we have a fairly large operation in the Philippines. The Philippines
currency against either the U.S. currency or the Canadian currency, over the last
fifty years, has basically done nothing except decline. That shows in the interest

rates as well. The interest rates in the Philippines have tended to be about 500 or
600 basis points above U.S. rates at the short end and maybe 10 percent points
above U.S. rates at the long end. But their currency has actually strengthened over
the last year against the U.S. dollar, not driven by the fundamentals of the
Philippines economy, but driven by the fact that the U.S. dollar has sunk.
MR. HARASYM: There is one point that I'd like to make about the calibration
procedures. It's easy to think that you have achieved a global solution by setting up
the model using Solver in Excel. However, in some cases if you change your
starting conditions you'll get a different answer. You have to look at your maximum
likelihood function, whether you're minimizing it or maximizing it and whether you
have a better answer or not. You can end up with a local solution without realizing
it. I've seen cases where the Solver routine is placed into VBA code and the process
is repeated in order to increase the likelihood that they have found a "better"
solution.
MR. WILSON: I have another example, which I'm stealing from Jeff Hancock, who
presented it at the Appointed Actuaries Seminar in Toronto a couple of years ago.
When the CIA did its study that was included as part of the American Academy of
Actuaries' recommendation on risk-based capital for calibrating the regime
switching log-normal model, Jeff looked at monthly data from the end of 1956
through the end of 1999. It was monthly, but the results differed depending on
whether you picked the first business day of the month, the second business day of
the month, third, et cetera. So if you picked the first of the month versus the last
trading day of the month for every month, you got a different answer, and it was
significantly different in terms of the calibration of the model. That's one reason you
want to use something like a CTE.
Remember the fact that these are models. There's no model out there that says
"airplane hits World Trade Center." Those just become data points that the model
tries to say are stochastic and, of course, are not stochastic in the normal sense but
are in a chaotic sense. On the interest rates, the oil price shock in the 1970s, which
helped drive interest rates up and stock markets down, is certainly not something
that just flows out of the stochastic model. The model is devoid of any economic
and sociopolitical ramifications.
You can get significant regime change. Andrew Smith, one of the theoreticians at
B.W. Deloitte over in the U.K., came up with a model that had a provision for a
random shock in it. So you have this nice stochastic model, and then suddenly it
would go haywire for a period, which is probably more realistic. For years, we've
been doing dynamic capital adequacy testing as required under Canadian
regulations. When I became an appointed actuary at the company in 1998, we
started putting in what we referred to as the "Japan scenario" for financial condition
reporting. Of course, many people thought this was a totally ridiculous scenario that
should never be included; it's a waste of our time to do it because it will never
happen here; we're not Japan. The scenario had the interest rates going down to

200 basis points and the stock market dropping 50 percent. Last time I looked,
NASDAQ was down 80 percent and interest rates are below 2 percent at the short
end of the spectrum.
I once asked some people what embedded options they had in their products. They
were selling regular and ordinary everyday life insurance, and they said "None." I
said, "Oh? What do you call a 4 percent minimum interest rate?" On Wall Street
that's called an interest rate floor. What do you call a cash value? On Wall Street
they call that a put option. But they hadn't considered pricing it because in their
minds that's not an option. "We've always had these things; we've had them for 50
years." They're options.
Danico Life is the largest life insurance company in Denmark. Its dividend formula
has a provision that it has stochastically determined as to what the option cost is by
interest rate guarantee. For its business that has 4 percent guarantees, the
company has a 40-odd basis point charge in the dividend formula. I think it was
ING or Aegon, one of those two companies, that had a factor in its European
dividend formulas for even a 2.5 percent guarantee that was about 10 basis points,
and then something like 50 basis points for a 4 percent guarantee. You can study
these things stochastically, but you have to take all the answers with a grain of
salt.
MR. MICHEL HEBERT: (Swiss Re) You mentioned that the random number
generators are not all random. Do you have a comment about the random number
generator in Excel?
MR. HARASYM: That is an interesting question because it seems to depend upon
what version of Excel you're running. If you went back to Excel 95, some people
had found that with a certain seed number, you could get the random sequence of
numbers to repeat after 300 or so random numbers. From this perspective, the
random number generator is unsatisfactory. On the other hand, we've done some
testing on the random number generator in Excel 2000 and the results were much
more favorable. The formula that Excel uses can be found on Microsoft's Web site.
What we do in our in-house random number generator is somewhat of a random
pick. We generate a series of random numbers and then pick one randomly. We
throw out the rest and then start again. On that basis, we perform testing on the
random number sequences. There are two considerations: does the same number
repeat or does the same number come back in the same sequence? Of course,
other statistical tests are relevant. I would look on the Internet. There are a
number of algorithms to test random number generators that are in the public
domain, as well as random number generators themselves. People have gone
through and tested them, so a lot of the leg work is already done. But you should
still do the tests yourself to confirm them.
The random number generator is a very important thing and very easily

overlooked. It's not going to get you on the front page of your Web site for
generating a good random number generator. On the other hand, you might end up
in the news for a bad product if you mispriced it because you had a bad random
number generator. There is not a lot of glory in the random numbers, but they're
very important.
MR. JAMES A. DEMOPOLOS: (ING) We're talking about situations where the
amount of capital requirements will be dependent on whether or not stochastic
modeling is used. Could you comment on the extent to which stochastically
modeled results are auditable?
MR. HARASYM: Do you mean internally or externally?
MR. DEMOPOLOS: I was thinking in terms of external, but any comments would
be welcome.
MR. HARASYM: The unusual aspect of auditing stochastic modeling is that it's
often done in black box form. But good practices would be that the models are
documented. Were you looking for some of the thoughts I'd have on what to search
for when auditing?
MR. DEMOPOLOS: I was thinking more in terms of the fact that, because some of
the interactions between variables are highly subjective, you could come up with
two qualified actuaries giving different answers for stochastic measurements of
liabilities. Would that create a problem in an audit situation?
MR. WILSON: Stochastic approaches for doing actuarial liabilities in Canada have
been there for the last four years, so that our reserves on segregated fund
guarantees, GMIBs and GMDBs are indeed calculated with a stochastic model. The
regulator has come up with a long paper on what you have to do to be able to use
your own models for capital purposes. We use the same model for doing our
reserves. You can audit it, but you're right, you're going to end up with two
actuaries who might come up with different answers. But two actuaries might come
up with a different assumption for mortality. The assumptions under U.S. GAAP, as
well as under Canadian, are the property of management. They are not the
property of the auditor per se. The auditor can scream and yell if he or she doesn't
like your expense assumption, your lapse assumption or your mortality assumption,
but they're management's assumptions. As long as you can show the auditor that
you have gone through a process that makes sense, you should be okay.
For the purposes of getting capital credit in Canada, our models have had to be
certified by the regulator, so the regulator actually has come in. The regulator
doesn't typically have people with this expertise, so they hire experts to work with
their people to see whether or not the model makes sense. Every time we change
the model we have to go back and have it audited by the regulator.

That comes from the banking world, because the Basil agreement on capital
requirements for banks has moved to use of own models for credit risk in banks.
They have to go through the same process of having all of their models certified by
the regulator. That would apply to American banks as well if they wanted to use
own models. There generally is an ability to use a factor approach, but it' generally
produces larger capital requirements. Use of own models generally will give you a
better capital position than if you just use factors. That's the direction that Europe
is going as well, in the insurance industry. The insurance industry is going to have
to have models to set its capital requirements because the world is being
recognized as a very stochastic place.

Chart 1
11
Is There Really A Starting and Ending Point? … No!Is There Really A Starting and Ending Point? … No!
Output
Historical
Economic Data
Historical
Policyholder
Data
RandomNumber
Generator
Economic
Scenario
Generator (ESG)
StochasticESG
Parameters &
Assumptions
Policyholder
Input Data
Economic
Scenarios
Data Validation
&
ESG Calibration
Random
Numbers
Stochastic
Asset / Liability
Models
Liability Data
Validation
Deterministic &
StochasticLiability
Assumptions
Deterministic &
Stochastic Asset
Assumptions
Result Tabulation,
Validation, & Review
Reported
Financial Results,
RiskManagement
Measures
Chart 2
12
Where does one Start? Key Steps Are ...Where does one Start? Key Steps Are ...
• Identify the key objectives and potential roadblocks before
considering ways of solving the problem.
• Identify key issues and potential road blocks.
• Describe the process/model in general terms before proceeding to
the specific.
• Develop the model: assumptions, input parameters, data, output.
• Fit the model: gather and analyze data, estimate input
parameters
• Implement the model.
• Analyze and test sensitivity of the model results.
• Communicate the results.

Chart 3
25
Present Value vs. Average Interest Rate per Scenario Scatter Plot
Stochastic Base Case: Target Equity Return = 8%, Target Interest Rate = 6%
2%
4%
6%
8%
10%
12%
-$300 -$250 -$200 -$150 -$100 -$50 $0 $50 $100
AverageInterestRateoverProjectionHorizon
2%
4%
6%
8%
10%
12%
Chart 4
26
Present Value vs. Average Equity Return per Scenario Scatter Plot
Stochastic Base Case: Target Equity Return = 8%, Target Interest Rate = 6%
-5%
0%
5%
10%
15%
20%
25%
-$300 -$250 -$200 -$150 -$100 -$50 $0 $50 $100
AverageEquityReturnoverProjectionHorizon

Chart 5
23
Stochastic Simulation Results:
• Recall
• GMIB Guaranteed Account Value of $1.4B
• Market Account Value of $1.0B
CTE GMIB ($millions)
95% $204.3
90% $177.2
80% $145.8
75% $133.9
70% $123.8
65% $114.9
60% $106.9
0% $43.4
Chart 6
24
(Negative) PV of GMIB Cash Flow by CTE
$0
$50
$100
$150
$200
$250
$300
$350
$400
$450
0% 20% 40% 60% 80% 100%
Conditional Tail Expectation ( % )
Base Case
Equity Return = 6%
Lapse Rate x0.5
PV of GMIB Cash Flow by Percentile
-$300
-$250
-$200
-$150
-$100
-$50
$0
$50
0% 20% 40% 60% 80% 100%
Percentile (%)
Base Case
Equity Return = 6%
Lapse Rate x0.5

Chart 7
29
GMIB CTE Measures: Investment Assumption Sensitivity Testing
$0
$50
$100
$150
$200
$250
$300
BaseCase
EquityReturn=10%
EquityReturn=9%
EquityReturn=7%
EquityReturn=6%
LTYield=8%
LTYield=7%
LTYield=5%
LTYield=4%
CTE(95%)
CTE(90%)
CTE(80%)
CTE(70%)
CTE(60%)
CTE(0%)
Base
Case
Chart 8
28
GMIB CTE Measures: Liability Assumption Sensitivity Testing
$0
$50
$100
$150
$200
$250
$300
BaseCase
RiderCharge-10bps
CurrentPricing
Spread-10bps
Pre-AnnMortDecr
10%
Post-AnnMortDecr
10%
LapseRatex2
LapseRatex0.5
AnnuitizationRatex2
AnnuitizaionRate
x0.5
CTE(95%)
CTE(90%)
CTE(80%)
CTE(70%)
CTE(60%)
CTE(0%)
Base
Case

Stochastic Modeling - Financial Reporting - Record

More Related Content

Similar to Stochastic Modeling - Financial Reporting - Record (20)

Stochastic Modeling - Financial Reporting - Record