How Efficient Is Your Frontier
How Efficient Is Your Frontier
How Efficient Is Your Frontier
Executive Summary
An appropriate asset allocation is a key component to achieving an institutions investment
objectives (and statistically the largest contributor to portfolio performance)
Traditional mean-variance optimization techniques are flawed: they are limited by assumptions
that past performance is predictive of the future and that the distribution of returns is normal
and stationary
Monte Carlo simulation-based models represent a more powerful analytical technique:
they generate a range of potential returns and the volatility of those returns, allowing for more
dynamic statistical analysis such as distribution of returns and probabilities of achieving
investment objectives
The Commonfund Allocation Planning Model is a proprietary model based on the term
structure of interest rates that uses Monte Carlo simulation to project future economic
scenarios for up to 20 years
Introduction
In the world of investment management, asset allocation is one of the most discussed and perhaps
most important decisions facing nonprofit institutions. Indeed, it is well known that Brinson,
Hood, and Beebower concluded that more than 90 percent of the variability of investment returns
is explained by asset allocation.1 We know that many factors such as spending, manager
performance, and disciplined rebalancing contribute to the investment success of a nonprofit
institution, and we recognize that the trustees of such institutions are charged with the challenging
task of asset allocation. Whether trustees are reengineering their institutions entire asset
allocation, determining appropriate bands or ranges around a policy portfolio, or simply making
minor adjustments to an already effective asset allocation, they must tackle asset allocation
discussions and decisions as part of their fiduciary role. An appropriate asset allocation is a key
component to achieving an institutions investment objectives and can have long term
consequences, both positive or negative, to the financial health of the institution. The challenge is to
determine an appropriate asset allocation that will best help an institution achieve its stated goals
by providing the highest return given a tolerable level of risk. But how do we achieve the highest
return, how do we consider risk, and what tools are available to assist with these challenges?
There are several models that are widely accepted as aides to developing an asset allocation. We will
focus on two of the most commonly used in portfolio management: Mean Variance Optimization
and Monte Carlo simulation, considering their attributes as well as limitations and ultimately
presenting our reasons for regarding our simulation-based approach to be the superior methodology.
2003 COMMONFUND
Brinson, Gary P., L. Randolph Hood and Gilbert L. Beebower, Determinants of Portfolio Performance,
(Financial Analysts Journal, July-August 1986) 39-44.
2003 COMMONFUND 3
Optimization Exercise
Return
18%
16%
14%
12%
10%
8%
6%
4%
2%
0%
0%
2%
4%
6%
8%
10%
Risk
At the bottom left corner of our example efficient frontier (lowest return, lowest risk) is a
portfolio comprised of 96.6 percent cash, 2.3 percent absolute return, 0.7 percent global bonds,
and 0.4 percent emerging markets. This portfolio has an expected annualized return of 5.9
percent with a standard deviation of 0.6 percent. At the other end of the spectrum lies a
portfolio that results in the highest expected return of all of the portfolios: 100 percent hedge
funds, which has an expected annualized return of 17.1 percent and a standard deviation of 9.3
percent. These two portfolios represent the corners of all optimal portfolios given the index
data input along the efficient frontier. Infinite combinations of portfolios create the line that is
the efficient frontier. As with any optimization model, we can now focus on an expected return
that we would like to achieve or the risk level we are willing to tolerate and the model will find
the point on the efficient frontier and the resultant asset allocation. For example if we require a
return of nine percent, which would result in a standard deviation of 1.5 percent, the model tells
us our asset allocation should be: 36.6 percent absolute return, 34.0 percent real estate, 20.4
percent cash, 5.2 percent global bonds, 2.6 percent core bonds, 1.0 percent commodities, and
0.2 percent hedge funds.
Due to liquidity or other institution-specific factors, many institutions cannot allocate 70
percent of their portfolio to absolute return strategies and real estate, and so are forced to put
limits around various asset classes. As we begin constraining asset classes, such as real estate and
2
2003 COMMONFUND
absolute return to 10 percent each, the efficient frontier continues to move down in a sub
optimal direction (that is down the x-axis, return, and out the y-axis, risk). The efficient frontier,
with these constraints, is notably less optimal. In order to construct a more reasonable
allocation, we constrained a group of asset classes: at least 20 percent in large cap U.S. stocks (in
the unconstrained model all domestic equity asset classes were conspicuously missing from the
efficient frontier) and no more than 10 percent each in private capital, equity hedge, emerging
markets, real estate, commodities, global bonds, absolute return or distressed debt. The result is
an efficient frontier in the chart below (the gray line) that is demonstratively less optimal than
the original frontier without constraints (represented by the blue line).
Return
18%
16%
14%
12%
10%
8%
6%
4%
2%
0%
0%
Risk
2%
4%
Unconstrained
6%
8%
10%
Constrained
Returning to our nine percent targeted return, we see on the constrained efficient frontier a
higher standard deviation of 4.6 percent made up of a portfolio comprising 39.6 percent cash,
20.0 percent large cap U.S. stocks, 10 percent in each of absolute return, real estate and global
bonds (the maximum end of the constraint), 4.2 percent in private capital, 3.2 percent in
commodities, and three percent in core bonds. Essentially we are adding cash to our constrained
asset classes (defined above) and generating a portfolio with an expected return of nine percent
and a standard deviation of 4.6 percent. We must remember how sensitive the optimization is to
these inputs, and how dependent the outcome is on history. Cash, for example, has a historical
return of 5.7 percent (see Appendix I), with very low volatility. Although one may exist, we
doubt there are many nonprofit institutions that are implementing this asset allocation and we
are even more skeptical if they are that they will be able to achieve a nine percent targeted return,
particularly given the current three month Treasury bill rate of 0.94 percent (at 9/30/03).
While we have several concerns as to the validity of the inputs of a mean variance optimization
model, we also question the meaningfulness of the output. Would we as investors prefer a
portfolio with an expected return of nine percent and a standard deviation of 9.1 percent or a
portfolio with an expected return of 9.2 percent and a standard deviation of 9.4 percent? It is
difficult to understand the ramifications of asset allocation decisions in the context of only these
two numbers. Additionally, MVO models assume that data used (usually historical) as inputs
2003 COMMONFUND 5
have normal or symmetrical returns. MVO assumes the capital markets are in equilibrium and
that the correlation matrix and standard deviations of the asset classes do not change. Since the
future is unknowable, how can we account for events in an investment world that is decidedly
NOT normal? Roger Lowenstein explores this concept in his book about the failure of LongTerm Capital Management, When Genius Failed. In the text, he discusses the statistical
probability of the stock market crash of Black Monday, when the Dow Jones Industrial Average
fell 22 percent on October 19, 1987. Economists later figured that, on the basis of the markets
historic volatility, had the market been open every day since the creation of the Universe, the
odds would still have been against its falling that much on any single day. In fact, had the life of
the Universe been repeated one billion times, such a crash would still have been theoretically
unlikely.3 Lowenstein goes on to point out that historical volatilities do not prepare investors
for future shocks of inexplicably volatile markets.
Theoretically, MVO is a logical and useful concept. For practical implementation of working
with asset allocations, however, we question whether this method, given its reliance on history
and human assumptions, can provide the most valid statistical picture of the future or specific
upcoming periods. A model is, by definition, nothing more than a tool that can help human
decision; a model cannot determine without doubt what the future holds. Without complete
confidence in the assumptions that are input into an optimization model, how can we use the
output in our decision making process? Finally, using standard mean variance work and
efficient frontiers, it is difficult to understand what the total range of possible outcomes may
be with a specific asset allocation of a nonprofits portfolio. While an analysis using mean
variance method does provide an estimate, it assumes a normal distribution of returns. A more
robust method, in our opinion, to examine future expected returns is to use a Monte Carlo
simulation based approach.
2003 COMMONFUND
Roger Lowenstein, When Genius Failed: The Rise and Fall of Long-Term Capital Management. (Random House, 2000) 72.
distribution of expected returns resulting from the optimization. In this case the values that are
generated are investment returns and the variables are the mean and standard deviation from the
MVO and are confined to the distribution associated with that mean and standard deviation.
One of the challenges with the mean variance technique, even when a Monte Carlo simulation is
used to enhance the output, is that it assumes a normal distribution of outcomes. However, the
distribution of U.S. stock market returns, for example, is far from normal and is in fact log
normally distributed; that is, the distribution of U.S. stock market returns demonstrates
skewness or exhibits relationships where the median can be less than or greater than the mean.
Monte Carlo simulation is therefore constrained by the type of model in which the simulation is
run. Used in other ways, such as a multi variable investment model, Monte Carlo simulation
analysis does not assume normality and generates random returns and their distributions.
Instead of using a single value for each variable, the simulation uses many different values,
running the model over and over again, each time using a different value for each of the variables
in the model. Therefore, for each asset class there is a range of potential returns for a given
investment period. The selection of the value for the variable, such as interest rates, for each trial
is random. But, the permitted values of the variable are constrained to account for reasonable
randomness, such as mean reversion, and are carefully constructed with some assumptions about
the future and how the variables behave.
Commonfunds Allocation Planning Model (APM), which we discuss in detail in the following
section, is a proprietary model based on the term structure of interest rates that uses Monte
Carlo simulation to project future investment returns for up to 20 years. For each of 1,000
scenarios, the APM using Monte Carlo simulation projects new yield curves by randomly
changing factors such as inflation, interest rates, and corporate spreads. Using regression analysis,
the model creates a variance/covariance matrix that represents the asset class relationships to
these different yield curves and to each other and allows us to ultimately paint forward 20 years
of projected returns for each asset class and entire portfolios. The model, with 20,000 data
points for each asset class (1,000 scenarios by 20 years), takes into account many different
investment environments such as a high inflation/low return environment, high inflation/high
return environment, low inflation/high return environment, etc. While we do not input any
manual constraints, as in an optimization, the APM is constrained by the term structure model.
In terms of the output, the APM generates an entire distribution of returns for a given asset
allocation (as opposed to the optimization model which generates the optimal portfolio) which
is generally not normally distributed. From this distribution, we can calculate medians, standard
deviations, and probabilities of achieving goals such as intergenerational equity.
Many of our nonprofit institutions look at the world, in terms of expected returns from their
endowments, in a linear fashion. As example an institution might calculate five percent
spending, plus three percent inflation, plus one percent management cost, all implying a
required return of nine percent. The MVO model tells us the allocation we need to achieve
this return, but it is unable to show us how many times we will or will not achieve our goal. The
MVO will also be unable to capture the effect of long term compounding of excess returns or
losses of the portfolio. Commonfunds APM can tell us in the context of 1,000 different future
scenarios per year for twenty years the probabilities of reaching the nine percent required
return. Because we are able to examine a full distribution of potential outcomes, we are also able
to evaluate tail risk, which although statistically unlikely to occur, is nonetheless present in the
investment world in which we live.
2003 COMMONFUND 7
To demonstrate the output of the APM, we decided to run the allocation that the constrained
optimization produced in the section above on a $100 million portfolio. Recall that with the
goal of a nine percent return, the optimization model suggested an asset allocation of 39.6
percent cash, 20 percent large cap U.S. stocks, 10 percent in each of absolute return, real estate
and global bonds, 4.2 percent in private capital, 3.2 percent in commodities, and three percent
in core bonds. The APM generates for this allocation a median return of seven percent with a
standard deviation of 4.6 percent. This distribution of returns is presented in the following chart.
Chart Three Distribution of 20 Year Annualized Returns
Frequency
300
Median= 7.0%
250
150
100
50
15
.0
15
.0
to
16
.0
16
.0
to
17
.0
>1
7.0
.0
to
14
14
.0
13
13
.0
t
12
12
.0
t
o
.0
t
11
.0
.0
.0
11
.0
10
.0
to
10
9.0
o
to
9.0
8.0
8.0
t
to
7.0
7.0
6.0
o
6.0
t
5.0
to
5.0
t
4.0
4.0
3.0
to
3.0
t
2.0
<
2.0
Return
If we required a return of nine percent, we may or may not be comfortable with a median return of
seven percent. Since a median is different than a mean and represents the mid-point of a series of datapoints, in this case 1,000, half of the 1,000 future scenarios will result in returns less than seven percent.
Moreover, when we consider issues such as intergenerational equity we realize that this allocation
might be undesirable given a five percent spending rate.4 The probability that this allocation will
achieve a net market value of zero (net after inflation and spending) is 23 percent, excluding the
impact of contributions. That means that more than 75 percent of the time, an institution that
has this asset allocation and a five percent, three year rolling average spending policy will not
achieve intergenerational equity or equilibrium. We define intergenerational equity as when the
net market value after spending is equal to zero, meaning that the real value of the portfolio has
been maintained after spending and inflation. The next chart portrays the distribution of net
market values for this asset allocation utilizing a five percent spending rate. In this case, the
median net market value of all scenarios is a negative $45.1 million. It is staggering to consider
4
2003 COMMONFUND
For more information on intergenerational equity please refer to Why Do We Feel So Poor, a Commonfund
white paper authored by Commonfund President and CEO Verne Sedlacek and Managing Director Sarah Clark.
that in half of the scenarios this portfolio will lose almost 50 percent of its value, in real terms,
over a twenty year period. The distribution graph not only demonstrates a negative median net
market value, but also reveals to the investor a negative skewness with a mean that is less than
the median (-$53.0 million for the mean) as well as the risks in the tails of the distribution.
Chart Four Distribution of Net Market Values
Frequency
300
250
Zero
200
150
100
50
0t 9
o
-4
0t 9
o
4
50 9
to
10 99
0t
o
15 149
0t
o
20 199
0t
o
25 249
0t
o
30 299
0t
o
35 349
0t
o
40 399
0t
o
45 449
0t
o
50 499
0t
o
55 549
0t
o
59
>6 9
00
49
-9
-5
00
0t
99
-1
to
49
-1
to
-1
99
-2
to
50
00
-1
49
-2
to
50
-2
99
-3
to
00
-2
49
-3
to
50
-3
99
-4
to
00
50
-4
-4
-3
49
-4
to
-5
to
-5
00
to
<-
-5
50
60
-5
99
Ultimately, the power of a model that incorporates Monte Carlo simulation, unlike MVO, lies
in the ability to produce a range of outcomes and generate meaningful statistical analysis from
the distribution. Mean variance optimization is in many respects akin to the analogy of the cart
leading the horse. With historical-based inputs and/or user inputs, a MVO model can produce
an efficient frontier along which reside optimal portfolios for a given expected return and
standard deviation. Monte Carlo simulation based models, on the other hand, consider asset
allocations from the users perspective and then generate expected returns, standard deviations,
distributions, and probabilities associated with that asset allocation. With this type of analysis,
the user is able to understand the likelihood of achieving goals rather than merely focusing on a
mean and standard deviation of an optimal portfolio produced by MVO.
From the fiduciarys perspective, the utilization of a model that incorporates Monte Carlo
simulation can be more effective in understanding a particular mix of asset classes and styles.
Since there is no one perfect allocation of assets for all institutions, difficult decisions must be
made in creating the most effective mix for any given institution. More robust models naturally
provide more comprehensive tools with which fiduciaries can consider these difficult decisions.
The following section will describe, in detail, how Monte Carlo simulation is incorporated into
Commonfunds APM and enhances its utility in considering asset allocation decisions.
2003 COMMONFUND 9
The APM is, at the core, a term structure model. That is, it is based on the term structure of
interest rates. We believe that the investment returns of the asset classes included in the model
have been and will continue to be a function of the economic environment and, in particular,
changes in the yield curve. Our model takes todays yield curve, uses Monte Carlo simulation to
project 1,000 different yield curves for next year by changing economic factors that affect the
curve, and projects returns for each of 19 asset classes in each of the new yield curve
environments. The projected returns are based on the regression of the historical relationship
between these asset classes and the yield curve. The model then takes each of the 1,000 new
yield curves as the next starting point and repeats the process, building another 1,000 yield
curves, and projecting returns in those environments. In order to have the ability to focus on the
long term, the model runs these simulations for twenty years into the future.
Fundamentally there are two core processes at work in the APM: defining the asset classes in
terms of their historical relationship to the yield curve and projecting the returns of those asset
classes in 1,000 different future economic scenarios for each year. The first process, defining the
asset classes, is a complex regression analysis that uses historical monthly index data going back
as far as 1970 to define each asset class returns in the context of four factors:
inflation,
risk-free interest rates,
corporate spreads,
and a world market premium.
U.S. Government bonds represent the risk-free rate and any other investment should earn above
that rate to be compensated adequately for any additional risk. Corporate spreads represent the
default risk which is the risk of holding non-government secured securities and requires additional
compensation. Finally the total world premium is a weighted index inclusive of world equity and
fixed income returns which is regressed against the first three factors (inflation, risk-free rates, and
corporate spreads) and represents the portion of return that is attributed to holding the market.
10
2003 COMMONFUND
In addition, for each asset class there is an asset class-specific or unique component of return
that is either above or below the return attributable to the original four factors. An example
might serve to clarify. If an investor buys large cap U.S. stocks, that investor is expecting that the
investment will outpace inflation and the risk-free rate as well as be compensated for taking
default and market risk. Presumably the investor is hoping this asset class selection will
outperform all four factors (or else the investor would buy Treasuries, or the basket of the world
market). The degree to which large cap U.S. stocks do, or do not outperform, is the
uniqueness of that asset class relative to the four market factors and is ultimately defined by the
regression analysis. It should be noted that not all asset classes are exposed to all of these risks.
Treasury Inflation Protected Securities (TIPS) for example are an investment option that is not
exposed to the risk of inflation or the market.
Along with generating the uniqueness of each asset class return, the regression analysis
generates a variance/covariance matrix for each asset class, further defining them against the four
factors as well as to each other. Essentially this matrix determines how the returns fit together.
The covariance part of the matrix defines how asset class returns move relative to each other and
the variance is the dispersion of the returns, or how far they vary relative to each other.
The second fundamental process in the APM, which occurs in the engine of the model, is
generating projected returns of the asset classes in 1,000 different scenarios. As described earlier,
the model uses Monte Carlo simulation to randomly adjust different economic scenarios that
affect the yield curve. Starting with todays yield curve, the Monte Carlo simulation alters one or
any combination of four factors:
inflation,
short term interest rates,
long term interest rates,
and corporate spreads, thus creating a new yield curve.
For the purpose of this text, we will call this yield curve Scenario 1, Year 1. Using the regressions
and variance/covariance matrix for the asset classes, the model is then able to project how each asset
class will perform in that environment (Scenario 1, Year 1). From this yield curve, another
simulation is run to alter again one or any combination of inflation, short term rates, long term
rates, and corporate spreads creating another yield curve which is Scenario 1, Year 2. Again, the
model projects the performance of the asset classes, using the regressions and variance/covariance
matrix, in the Year 2 yield curve environment. This process is repeated for twenty years and is done
for 1,000 different scenarios effectively generating 20,000 data points (returns) for each asset class.
The Monte Carlo simulation that is used in the APM generates random economic conditions that
affect and therefore change the yield curve. Within the context of the term structure model, the
random changes are, however, reasonable. While these alterations can be aggressive and incorporate
literally thousands of scenarios of low inflation/high returns, high inflation/high returns, low
inflation/low returns, high inflation/low returns, etc., the evolution of the yield curve in each scenario
is tested against the term structure model and will not generate in one year drastic or unreasonable
changes such as a change in one year from negative inflation (deflation) to hyperinflation.
2003 COMMONFUND 11
Ultimately, the power of a model that incorporates Monte Carlo simulation, unlike MVO, lies in
the ability to produce a range of returns and generate meaningful statistical analysis from the
distribution. With historical-based inputs and/or user inputs, a MVO model can produce an
efficient frontier along which reside optimal portfolios for a given expected return and standard
deviation. The APM, in contrast, considers asset allocations from the users perspective and then
generates projected returns, standard deviations, distributions, and probabilities associated with
that asset allocation. With this type of analysis, the user is able to understand the likelihood of
achieving goals rather than merely focusing on a mean and standard deviation of an optimal
portfolio produced by a MVO model.
12
2003 COMMONFUND
The APM has many advantages over MVO. In addition to generating a distribution of
potential outcomes and different economic scenarios as described above (which cannot
be accomplished with MVO), the APMs term structure model has advanced features
that distinguish it from most other forecasting models that use Monte Carlo simulation.
The model consistently forecasts the term structure of interest rates at every point in
simulation time, which provides a more realistic set of the expectations that drive interest
rates and a better formulation of the documented dynamic properties of inflation and
interest rates. The APM forecasts four term structure components whereas other models
known to incorporate term structure models forecast only one or two. Finally, the open
design architecture of the APM makes it relatively easy to update and further develop.
The APM has been designed to be a state-of-the-art investment-planning tool. Although
no analytical model can completely replace informed professional judgment, the APM
can provide a better foundation on which to base professional judgment.
What are the limitations?
No model or simulation can predict the future or account for the infinite number of
possible outcomes. The projections generated by Commonfunds APM are based on
assumptions about performance and risk characteristics of investments in various asset
classes. Those assumptions are based on historical data that are believed to be accurate
and on which the APM relies. The utility of the APM depends greatly on the accuracy of
that historical data and its meaningfulness in forecasting future events. Commonfund
cannot guarantee the accuracy of the data nor does it represent that the data will
necessarily represent market conditions in the future.
The model simulates the range of probable outcomes over a twenty-year time horizon of
varying combinations of asset allocations, inflation expectations, spending policies,
capital gifts and rebalancing rules. The reasonableness of the input assumptions made by
the user will affect the reasonableness of the simulations. In all cases, the statistical
confidence in the predictions falls as the forecast period gets shorter.
Because the model uses asset class returns, it should not be used to evaluate or simulate
the results of any specific investment program.
No allocation planning model simulation can replicate the exact experience of an institution.
As such, the results of the Commonfund APM should only be used as a general guide. In no
way should the APM be a substitute for the important policy choices that an institution must
make in developing its investment program.
2003 COMMONFUND 13
Appendix
I. Data and Indexes used in Mean Variance Optimization
MODEL GENERATED
Asset Class
Large Cap
Mid Cap
Small Cap
International
Emerging Markets
Private Capital
Hedge Funds
Absolute Return
Distressed Debt
Commodities
Private Real Estate
Core Bonds
Cash
Global Bonds
Index
S&P 500 Index
S&P MidCap 400 Index
S&P SmallCap 600 Index
MSCI EAFE Index
MSCI Emerging Markets
Venture Economics
HFR Equity Hedge
HFRI Fixed Income
Altman NYU Distressed Debt
Goldman Sachs Commodities
NCREIF
Lehman Aggregate
3-Month T-bill
Citigroup World Govt. Bond
Mean
Return
12.6
15.9
11.5
11.2
5.3
15.7
19.7
10.6
3.7
12.7
9.1
9.9
5.7
11.1
Standard
Deviation
16.0
17.0
18.8
17.0
23.4
10.9
9.4
3.6
14.3
18.5
3.4
5.0
0.7
7.0
MODEL GENERATED
20 year projections*
Asset Class
U.S. Large Cap
U.S. All Cap
U.S. Small Cap
International
Emerging Markets
Private Capital
Hedge Funds
Absolute Return
Distressed Debt
Energy
Timber
Private Real Estate
Core Bonds
Cash
Global Bonds
*Based on 06/30/03 Model
14
2003 COMMONFUND
Median
Return
8.0
8.6
9.2
8.6
11.8
13.4
12.4
3.6
12.9
13.3
9.9
13.2
4.2
3.7
5.6
Standard
Deviation
17.5
18.7
23.5
18.9
21.3
17.0
11.5
4.1
8.1
21.6
6.8
5.0
7.1
3.2
7.0
Start
Date
Jan-74
Mar-81
Jan-94
Jan-70
Jun-92
Oct-73
Jan-90
Jan-90
Jan-87
Jan-70
Jan-78
Jul-82
Jul-82
Jan-94
End
Date
Jul-03
Jul-03
Jul-03
Jul-03
Jul-03
Sep-03
Jun-03
Jun-03
Jun-03
Jul-03
Dec-03
Jul-03
Jul-03
Jul-03
2003 COMMONFUND 15
Tel 203-563-5000
Tel 888-823-6246
www.commonfund.org