Portfolio Construction
Portfolio Construction
Portfolio Construction
1. Introduction
1
We begin with two basic examples. The first is a single-period case, for a 130/30
portfolio, based on the Russell 2000 Value Universe. This simple example illustrates the
basic effects that cost modeling may have on the optimization process, including
increased diversification and the minimization of overreaction to strong alpha signals,
consistent with Coppejans and Madhavan [2007]. The second example extends the
analysis to a multi-period setting, incorporating a 20-month back-test with monthly
rebalancing, for a market and dollar neutral portfolio. Net return performance is shown
to be sharply higher in the case in which transaction costs are explicitly incorporated
into the optimization problem. Although the position count tends to be higher than in
the case where costs are ignored, the dollar turnover falls, reinforcing the single-period
lesson with respect to reaction to alpha signals. Simply put, taking costs into account ex
ante increases net returns and diversification, while diminishing the turnover effect on
the cost of trading.
We then examine the impact of transaction costs on portfolio composition in a
series of global mean-variance portfolio problems. The stock universe shifts from the
U.S. market to a global setting, covering North America, Western Europe, Latin
America, and East Asia. Results are provided for periods predating and then spanning
the credit crisis of 2007-2008. We first adopt the perspective of a U.S. investor, who
views the riskless asset as a U.S. Treasury bill, and construct optimal portfolios with
and without inclusion of transaction costs in the stock selection process. Trading costs
by country are related to shifts in regional investment allocations, in which portfolio
weights shift, relative to those calculated without inclusion of transaction costs, by
double-digit percentages. These results are strongly affected by the interplay of
expected returns and expected costs within and across regions. We therefore turn to an
examination of cost-adjusted minimum variance portfolios, which have the advantage
of not requiring estimates of expected returns, and provide even sharper results.
Moving from a paper portfolio to a cost-aware minimum variance problem results in
regional weight changes ranging from a negative 68% to over 160%.
2
2. Two Examples
3
The factor covariance matrix is scaled using an option-implied adjustment coefficient to
exploit the options market information with respect to the future levels of risk. The
loadings are estimated in a time-series framework on a per-stock basis, unlike some
cross-sectional paradigms.xiii
The optimization problem maximizes risk-adjusted expected returns and
minimizes short borrowing costs, taking the basic form:
− 0 TC ( − 0 , K )
max − − − BC
(1)
100
subject to the constraints
( − ) ( − )
B B
2
,
= + − − ,
where
N is the number of stocks in the Russell Value 2000 universe,
is the vector of stock-specific expected returns (in excess of the risk-free rate,
in percentage units),
is the covariance matrix of the stock-specific returns,
TC ( ) is the vector of stock-specific expected transaction costs per share (in bp),
which takes into account both implementation shortfall and commission costs and
consequently is a function of the traded shares (or equivalently the weight changes
relative to the initial portfolio value times the portfolio value),
K is the initial portfolio value,
4
is the tracking error bound.
The last three constraints ensure that the total portfolio exposure is indeed 160%. It is
achieved by introducing two new portfolio weight variables ( + for positive and − for
negative weights) as well as a binary variable . We divide the TC component in
Equation (1) by a factor of 100 since transaction costs are computed in basis points while
returns and borrowing costs are reported in percent.
The initial value of the portfolio is set to $200 million for the purpose of this
example, and we assume that we start from a cash position, i.e. 0,i = 0 for i = 1,..., N .
We also set = 1 , which means that both alpha maximization and cost minimization are
given equal weight in the objective function. The benchmark weights B are obtained
from Russell Investment Company, and the tracking error bound is set to be 5%.
We adopt the following terminology. We refer to the solution of the problem
including transaction costs ex ante as being cost-aware, while the non-cost-aware portfolio
is optimized, setting costs to zero and then subtracting expected costs from returns ex
post. Virtually by construction, the gross expected return from the cost-aware portfolio
is lower than the gross expected return without the inclusion of transaction costs. In our
example, the gross expected return for the non-cost-aware portfolio is 2.24%, while it is
only 1.68% for cost-aware portfolio.
However, the expected cost incurred in acquiring the non-cost-aware portfolio is
$7.4 million, reducing the expected net return from a positive 2.24% to a negative 1.46%.
In contrast, transaction costs associated with the cost-aware problem total $2.2 million,
keeping the portfolio net return positive at 0.57%.
The improvement clearly does not come through a reduction in turnover, since
we are considering a single period. However, the cost-aware portfolio incurs costs over
a larger set of stocks, illustrated in Exhibit 1, in which the horizontal axis indexes the
names, by number, in the portfolio. The number of positions held in that portfolio is
366, compared to 211 in the non-cost-aware problemxiv. In the absence of cost
considerations, expected return is the dominant factor in determining final portfolio
5
weights, subject to the risk constraint. Since costs increase with trade size, at some point
the cost to trade will negate the alpha in a name. This type of result is the fundamental
driver of fund capacity analysis, but it appears here in the form of increased
diversification stemming from expected costs of trading. The diversification effect is
apparent by looking at the expected total portfolio risk: it goes down from 15.24% in the
non-cost-aware portfolio to 13.99% in the cost-aware portfolio. The tracking error
relative to benchmark, however, is kept at 5% for both portfolios, as specified in the
problem formulation.
A corollary to the diversification argument is illustrated in Exhibit 2: the trade
sizes relative to ADV are uniformly lower across the entire range of portfolio weights.
We define the weights 0-0.3% of total portfolio as small; 0.3%-1% as medium, and
exceeding 1% as large.
When costs are considered, multiple securities with similar alpha characteristics
can be given more weight than a single security with a higher alpha estimate. The cost
to trade a smaller amount of many securities is less than the cost to trade the same
dollar value of a single security, an effect emphasized by typical diversification effects.
An alternative interpretation is that overreaction to strong alpha signals is minimized
when expected costs are taken into account, consistent with the theoretical prediction of
Coppejans and Madhavan [2007].
Exhibit 3 illustrates how the allocation of securities changes with respect to the
stocks’ liquidity, proxied here by the 21-day average daily share volume (ADV)
numbers. While the non-cost-aware optimizer allocates larger weights (over 1%) to less-
liquid stocks, the cost-aware optimizer allocates larger weights to more-liquid
securities. These changes in allocation have a direct impact on the size and composition
of the transaction cost components, as presented in Exhibit 4. We decompose
transaction costs into four parts: bid-ask spread (which remains largely independent of
trade size), commission cost (which we assume is $0.02 for every security), market
impact or implementation shortfall cost (proportional to trade size), and borrowing cost
(50bp for every stock). Of course, borrowing costs appear on the short side only.
6
Besides the obvious effect of reducing the overall level of transaction costs, the
key lesson is in the change in the relative weights of the cost components. In particular,
the short side of the portfolio benefits the most relative to the non-cost-aware case: the
overall level of costs falls from nearly 750bp to 143bp, with the market impact
component decreasing from 613bp to 65bp, almost a tenfold drop. This dramatic
reduction in costs is, probably, explained by the distribution of expected alphas in our
sample. Since the stock market was trending down during the 2007-2008 period, the
expected alphas fed into the optimizer are negative for three quarters of all securities in
the Russell 2000 Value index. Moreover, the distribution of expected alphas has a
relatively heavy left tail (negative alphas) and a relatively light right tail (positive
alphas). This results in the non-cost-aware portfolio allocation being concentrated in a
few stocks on the short side. Once transaction costs are incorporated in the problem,
idiosyncratic risk and stocks’ liquidity become important, and the optimizer is able to
decrease the costs by moving out of extreme weights on the short side. The optimizer
simply picks more short-side names, from roughly 30 in the non-cost-aware portfolio to
almost 100 in the cost-aware portfolio. For comparison, the number of stocks picked for
the long side grows from around 175 to 270, which is a much smaller relative increase.
At 100 names picked on the short side of the cost-aware portfolio, the optimizer is
nowhere near exhausting the supply of the candidates for short positions; however, 270
names on the long side constitute around 80% of the names with positive expected
alpha in the Russell 2000 Value index. Spreading the trading across more names results
in a dramatic reduction in the average relative trade size, especially on the short side, as
illustrated in Exhibit 2, and transaction costs decline. This shifting of securities and
shares in the cost-aware portfolio results in borrowing costs becoming as important as
the market impact costs. In addition, the relative importance of average spread and
commission costs increases for both short and long positions.
7
The single period comparison illustrates the basic effects of cost modeling on the
portfolio optimization process. We now demonstrate that such desirable results are
achievable over time. In this second example, we implement a 20-month back-test, over
the period from December 2006 to August 2008, with monthly rebalancing. The period-
by-period optimization is mathematically similar to the single-period case, including
the use of the Russell 2000 Value Universe. Instead of modeling a 130/30 portfolio with
a tracking error constraint, a market neutral/dollar neutral portfolio with an absolute
risk bound is used. More precisely, we consider the mathematical problem
− 0 TC ( − 0 , K )
max − − − BC (2)
100
subject to the constraints
2 ,
Market = 0 ,
= + − − ,
+ = 1, − = 1 and = 1,1,...,1 ,
where the notation is the same as before and we use the same covariance and
transaction cost estimates as in the single-period optimization case.
We begin with a $50 million budgetxv. We construct a market and dollar neutral
portfolio at each monthly step, with an absolute risk upper bound of 10% and a
transaction cost aversion factor of = 20 , reflecting the fact that we expect to buy and
sell the typical security once every two months. The total rebalancing period is 20
monthsxvi. Expected returns are based on a normally distributed random scheme, here
exploiting an information-coefficient formulation,
i ,t = IC iobserved
,t + i 1 − IC 2 i , i = 1,..., N , (3)
in which,
IC is the information coefficient,
8
iobserved
,t is the realized monthly return observed in month t for stock i;
i ~ N ( 0,1) .
By definition, the correlation of forecasted and realized monthly returns is equal to IC.
In what follows, we assume that the information coefficient takes the value of 10%
(which reflects above average forecasting ability). We draw for each month, and for
both the non-cost and the cost-aware case (i.e. = 0 and = 20 ), 20 random numbers
from the standard normal distribution and obtain 20 draws of i ,t using Equation (3).
The results for both optimization problems are summarized by the means and the 95%
empirical confidence intervals of net performance, total transaction costs, turnover and
number of monthly positions held, demonstrating in all cases statistically significant
differences in the outcomes.
The 20-month period chosen for our analysis was quite challenging for the
markets. The return on the Russell 2000 Value index, used as a benchmark, from
December 2006 to August 2008 was a negative 18.7%. However, due to our high
information coefficient, the paper return of our optimal rebalancing strategies remains
quite high: 20.9% for the non-cost-aware and 12.8% for the cost-aware multi-period
optimization problem. Once the net returns are calculated, the destructive effect of
transaction costs on the performance of our rebalancing strategies becomes apparent. To
achieve net results, the transaction costs of each rebalance are subtracted from the
portfolio value at each rebalancing step during the optimization.
9
about -1.10. In contrast, the cost-aware portfolio performed much better, yielding a
modest but positive Sharpe ratio of 0.07.
Exhibit 5 also helps to identify the sources of the outperformance. First and
foremost, the average dollar costs of implementing the cost-aware rebalancing strategy
are two to four times lower than the costs of implementing the non-cost-aware strategy.
As indicated by the 95% confidence bands, the difference is statistically significant in all
20 months. Second, the costs of implementing the cost-aware rebalancing strategy
remain lower despite the fact that its portfolio size starts to exceed the size of the non-
cost-aware portfolio by as much as 50% in the last several months of the period
considered.
10
sense that diversification differences across portfolios diminish. For example, the sharp
increase in diversification starting in the summer of 2007, as a response to sharply
increased risk levels, is not limited to the cost-aware rebalancing strategy.
As an important corollary, overreaction to certain alpha signals is lessened when
costs are considered at the portfolio construction stage. This also leads to portfolio risk
reduction, which we observed (albeit only in-sample) in the previous section. Exhibit 8
confirms the portfolio risk reduction resulting from higher position counts, decreased
turnover, and lower sensitivity to extreme alpha signals. The chart shows the monthly
realized portfolio volatilities for both the non-cost-aware and cost-aware portfolio,
computed as the mean absolute portfolio return in a given month. The cost-aware
portfolio exhibits lower risk than the non-cost-aware portfolio in all months but one.
The result illustrates that traditional portfolio statistics (such as risk) can be significantly
influenced not only by the stock selection and hedging decisions, but also by a viable
cost-control process.
One way to gauge the impact of transaction costs is to measure changes in the
composition of a portfolio as a result of including costs in the portfolio construction
process. We have already seen the potential for such shifts in the context of the single-
period Russell 2000 Value example. In this section, the stock universe changes to high-
turnover securities in North America, Western Europe, East Asia, and Latin America,
aggregated to these regional levels in terms of portfolio composition. xviii For simplicity,
we focus on two general solutions to the portfolio problem. The first is the tangency
portfolio, adopting the perspective of a U.S. investor, who views the riskless asset as a
U.S. Treasury bill. The second is a minimum variance formulation, which has the
advantage of not requiring expected return estimates for its construction. We produce
results for three different periods: December 2006, December 2007, and August 2008.
The first period arguably predates the recent credit crisis, while the second captures
11
some of the initial reaction to that crisis; the last is the most recent data available at the
time of this computation, just before the collapse of Lehman.
The rule for selecting securities in each of the six cases studied is based on stock
selection criteria for relatively liquid securities in the countries consider. For each
country, we rank securities based on the average volume traded during the 21 days
prior to each date when the optimization is performed. We then add securities to the
universe, starting from the top and going down the list until the total volume traded in
selected securities reaches 70% of total volume in that market. ETFs and ADRs are
excluded in order to keep the analysis as clean as possible. xix We limit the number of
securities from each country to 20, which yields roughly 165 securities in any given
sample. Results are computed for a $500 million portfolio.xx
Exhibit 9 contains the actual dollar-weighted average implementation shortfall
costs of the securities for the trimester prior to the optimization dates considered in our
analysis. The actual costs are retrieved from ITG’s Peer Group Database. xxi Average
costs are the lowest in Western Europe and the highest in East Asia. In addition,
average implementation costs are typically the lowest in 2006 and the highest in 2007. In
what follows, we will use these empirical costs to explain some of our optimization
results.
subject to = 1, = 1,1,...,1 ,
12
The notation is the same as in Section 2. The tangent portfolio is calculated by rescaling
the weights in risky assets obtained from Equation (4) so that they sum to one. The
coefficient of risk aversion is calibrated so as to ensure that the total risk of the
optimal portfolio roughly matches the total risk of MSCI EAFE over the sample
period.xxii The coefficient of cost aversion, , is set to zero, in the case that costs are
ignored at the portfolio construction stage, or to successively higher values, which
proxy for roundtrip turnover levels corresponding to annual ( = 2 ), semiannual
( = 4 ), and quarterly ( = 8 ) rebalancing frequencies. Expected returns are computed
similarly to Section 1, i.e., via a shrinkage estimator as a rolling window three-year
average return in excess of the risk-free rate.xxiii Expected implementation shortfall costs
are generated from the same source as discussed in the previous sections,xxiv while the
necessary covariances are taken from ITG’s Global Risk Model and global commission
costs have been retrieved from ITG’s Peer Group Database.xxv
Exhibit 10 reports the weight of securities from each country, aggregated to a
regional level, in the tangent portfolio under four cost scenarios: the non-cost-aware
case, and for the cost-aware cases associated with turnover levels of 100%, 200%, and
400% per year. Our discussion is limited to the 100% annualized turnover statistics;
examination of the table shows that cost effects are sharply exacerbated in the higher
turnover cases, but no new qualitative lessons are available there.
Weights on the North American and Western European regions increase by 1.9%
and 4.2%, respectively, once transaction costs are taken into account for 2006. This is a
relative increase of 13.5% and 10.7%, respectively. The relative decline in East Asian
investment is led by China and Japan. To put this result in perspective, transaction costs
in Japan are approximately 240% higher than in the United States, for example. East
Asia except Japan exhibits costs that are 67% higher than in North America over the
period. In contrast, Western Europe’s trading costs are running at about 31 basis points
during this period, which is a bit lower than in the United States. This accounts for the
increase in Europe’s portfolio weight relative to other regions.
13
U.S. trading costs rose in the fourth quarter of 2007, relative to 2006, reflecting
increased market friction due to market volatility that began in August of 2007. Western
Europe’s costs also rose, but any portfolio effects due to that increase were offset by a
jump in Asian costs by 38%. Korea’s weight in the portfolio dropped by 28%, for
example. This led to a 9.5% decrease in the Asian weight, relative to the paper portfolio,
while investment in North America and Europe grew by small amounts, once costs
were factored into the calculation.
In August 2008, average costs in Western Europe except the U.K. remained at the
same high levels as in 2007, with transaction costs being 80% higher than in December
2006 (mostly attributed to France and Italy). However, expected returns fell for that
region in 2008, leading to a relatively low weight in the paper portfolio, further reduced
by 26% once costs were taken into account. In contrast, U.S. costs fell by about 13%. In
the face of a sharply lower weight for Europe and a decrease in costs for U.S. and
Canadian securities, the North American weight rose over 17% relative to the paper
portfolio. Interestingly, the Asian weight remained basically unchanged, due to
country-specific effects in the region. In particular, Japan’s costs remained
approximately the same as in December 2006, increasing its relative share in the
portfolio, while Korea’s and especially China’s costs increased relative to December
2006.
On a global basis, the story is basically a decrease of liquidity in Western Europe
except the U.K., which has the effect of sharply lowering European investment
exposure relative to a paper portfolio and other regions. Of course, this type of result is
strongly affected by the interplay between costs and expected returns within and across
regions. For example, the results for Western Europe are influenced by a decline in
expected returns by approximately half in August 2008, relative to the previous eight
months. We now turn to a set of scenarios in which the effect of costs is examined
without consideration of return expectations.
14
We define the cost-aware minimum variance portfolio as the solution to the
problem,
− 0 TC ( − 0 , K )
min + (5)
100
subject to = 1, = 1,1,...,1 ,
in which we use the same notation as before. Equation (5) is a special case of Equation
(4) where we have a strong prior that = 0 (i.e., the expected returns for all securities
are set to be zero). The model has the advantage of not requiring estimates of expected
returns, unlike all previous examples in which the effects of alpha and transaction costs
compete with each other. The optimization problem is solved with = 0 and = 2 ,
representing the paper or non-cost-aware portfolio and a cost-aware portfolio with
annual turnover.xxvi Results are presented in Exhibit 11 for a $500 million budget.
Overall portfolio risk increases as one moves from the paper portfolio to the cost-
aware formulation, by construction. The increase is 22% for 2006 and 2007, falling to
16% in 2008. Risk is growing, year over year, but this is true regardless of transaction
costs, given changes in the global economic environment.
The cost-adjusted minimum variance formulation drives the results in a stark
fashion, which was partially masked by the previous inclusion of expected returns.
China is a good example in this context. Costs are sufficiently great so as to reduce
China’s weight in the portfolios by an average 51% over the three periods. These
adjustments are not from a similar base, however, in terms of the paper portfolio.
China’s weight, without regard for transaction costs, decreases by 48% in 2007 relative
to 2006, and then by over 80% in 2008 relative to 2007. Simply put, China loses its
diversification benefit over this period, and its high cost of trading exacerbates this
effect. The correlation between the Shanghai SSE composite and indices such as the SPY
and EAFE increases sharply between the end of 2006 and August of 2008. xxvii Once one
15
factors in the high trading costs in Japan and Korea noted earlier, East Asia weights in
the cost-aware portfolios fall sharply, by between 44% and 58%, depending on the year.
The East Asian effect on the minimum variance portfolio is strong enough to
reverse some results relative to the optimal tangent portfolio. Western Europe’s cost-
aware weight increases in 2008, for example, relative to the paper portfolio, by 42%.
Although changes in the North American weights are small in percentage terms,
moving from a paper formulation to a cost-aware minimum variance problem results in
substantial weight changes ranging from -68% for Latin America in 2006 to 168% for
Western Europe in 2007.
4. Conclusion
16
some cases, and can increase by 168% in others, depending on region and time period.
These changes imply substantial shifts in investment strategy.
Originally driven by regulatory concerns, the interest in transaction costs and
their importance for institutional returns has been renewed with the advent of the
current financial crisis and the corresponding spike in equity market volatility. There is
much interesting work to be done in the future. Proper consideration of transaction
costs can be used in establishing the viability of certain investment strategies. One
example is fund capacity analysis and the optimal timing of fund closures. Another
possibility is illustrated by Hou and Moskowitz [2005], with respect to the construction
of portfolio strategies that exploit the relationship between the cost of trading a stock
and its expected returns.
We offer a basic and very practical message. Cost reduction is not only the
responsibility of trading desks. Mitigation of frictional costs, leading to improvement in
realized returns and better alignment of return with risk, begins with the inclusion of
transaction costs in the stock selection and portfolio construction process.
Endnotes
i See, for example, Domowitz and Krowas [2004] for an early discussion of the extension of the concept of
efficient frontier to trading applications, building on the work by Almgren and Chriss [2000].
ii “Imperfect Knowledge: International Perspectives on Transaction Cost Analysis,” the Tabb Group,
March 2008.
iii “An Extension of the Markowitz Portfolio Selection Model to Include Variable Transactions’ Cost, Short
Sales, Leverage Policies and Taxes,” Gerry A. Pogue, Journal of Finance, 25 [1970], pp.1005-1027.
17
iv
For instance, in one of the more recent papers Lobo, Fazel, and Boyd [2006] study the non-convex
mean-variance problem setting that arises in the presence of fixed costs and suggest expressing and
solving it as a series of convex optimization problems.
V “Is Execution Risk the Same as Investment Risk?” Robert Engle and Robert Ferstenberg, Journal of
Portfolio Management, Winter 2007, and “The Value of Transaction Cost Forecasts: Another Source of
Alpha,” Mark Coppejans and Ananth Madhavan, Journal of Investment Management, Vol. 5, No. 1
(2007).
ViAn exception is an early study on global transaction costs, in which it is suggested that proper
accounting of transaction costs in the portfolio construction process may lead to sharp changes in the
weights given to regional investment objectives in global portfolios; see “Liqudidity, Volatility, and Equity
Trading Costs Across Countries and Over Time,” Ian Domowitz, Jack Glen, and Ananth Madhavan,
International Finance, 4 (2001).
Vii Fabozzi, Focardi, and Kolm [2006], pp. 74-79, consider a rebalancing example using 18 countries in
the MSCI World Index over the period January 1980 to May 2004. For simplicity, roundtrip transaction
costs are assumed to be a constant 50 basis points for all countries and trade sizes.
viii See, for example, Amihud [2002] and Spiegel and Wang [2005] and references therein.
ix This issue is explored in Domowitz, Glen, and Madhavan [2001]. Theoretically, the trading cost variance
can be decoupled from the portfolio problem if the time allowed for trading is short relative to the holding
period of the portfolio; see Engle and Ferstenberg [2007].
x A complete description of the model, and results comparing expected to actual costs, are contained in
are similar to the single-day VWAP strategy. We choose the single-day VWAP strategy to reflect the
typical situations faced by investors who track the Russell Index during the rebalancing time. The
pressure to maintain a low tracking error to the index forces many portfolio managers to accomplish
rebalancing trades in a single day, sometimes pushing up the trading costs relative to the ones obtained
with a more serene pace of trading (corresponding, for example, to a 10% volume participation strategy).
xii Borrowing costs are hard to estimate since available information is scarce . However, the borrowing
cost estimates do not have too much impact on our empirical results. As a robustness check, we use the
borrowing costs from D’Avolio [2002], one of the few published papers that discuss borrowing costs. We
partition our sample to match the size and liquidity groups reported in Table 5 of D’Avolio [2002], assign
the corresponding group-average borrowing costs to each stock, and run the same optimization with
these new estimates. The results obtained are qualitatively similar and are available upon request.
xiii A complete description, including performance testing results, can be found in ITG’s Risk Models,
could be a different security from stock #20 in the bottom chart. Note that 44 securities of the non-cost-
aware portfolio have not been selected in the cost-aware portfolio.
xv We choose a different initial cash position and different optimization constraints than in the single-
period example in order to illustrate that the inclusion of transaction costs in the portfolio construction
process significantly improves net return results for a variety of optimization problems. Results for larger
initial cash positions look qualitatively very similar. In fact, the effects are even more pronounced.
xvi The turnover results of our cost-aware optimization problem are consistent with this statement. The
average turnover is between 150% and 200% for all periods (compared to the maximum possible
turnover of 400% per period). See Exhibit 6 for more details.
xvii In fact, the long-short portfolio turnover can occasionally exceed 400% when, for example, the short
side outperforms the long side in the month following the optimization date, thus reducing the net portfolio
basis.
xviii More specifically, we use the following country universe: North America includes the United States and
Canada; East Asia includes Japan, Korea, and China; Western Europe is comprised of the U.K.,
Germany, France, and Italy; and Latin America incorporates Mexico and Brazil.
18
xix We wanted to ensure that we rely as much as possible on country- or exchange-specific costs to
determine the country/region weights in our global portfolios. For instance, excluding ADRs and GDRs
will force a U.S. investor to buy/sell Sony in Japan.
xx Results are computed for three portfolio sizes, $50 million, $100 million, and $500 million, but the
results are qualitatively the same, and even quantitatively similar, so we limit the reporting to the largest
portfolio size.
xxi ITG’s Peer Group Database is ITG’s proprietary client execution database and comprises client
execution data from more than 120 different large buy-side institutions.
xxii A total risk of roughly 13%, for example, translates into = 2 .
xxiii A 5% trimmed average is computed by stock over the 36 months preceding the optimization date.
Since mean-variance procedures have a tendency to pick extreme returns, average returns for each
stock are shrunk towards the grand average, using shrinkage intensity proportional to the volatility of the
security.
xxiv ACE costs are calculated for a 10% Volume Participation strategy. Results obtained with that strategy
are qualitatively similar to those obtained with optimal implementation shortfall or VWAP 1-day strategies.
We used 10% Volume Participation strategy since we believe it reflects the typical trading style for liquid
securities.
xxv The average commission costs for liquid securities during the period studied vary across countries
between approximately 10bp for the United States and 33bp for China.
xxvi A higher turnover assumption is typically not necessary to maintain a minimum-variance portfolio and
for example.
References
Almgren, R. and N. Chriss. “Optimal Execution of Portfolio Transactions.” Journal of Risk, 3 (Winter
2000/2001), pp. 5-39.
19
Amihud, Y. “Illiquidity and Stock Returns: Cross-Section annd Time Series Effects.” Journal of Financial
Markets, 5 (2002), pp. 31-56.
Coppejans, M., and A. Madhavan. “The Value of Transaction Cost Forecasts: Another Source of
Alpha.” Journal of Investment Management, Vol. 5, No.1 (2007), pp. 65-78.
D’Avolio, G. “The Market for Borrowing Stock.” Available at SSRN: http://ssrn.com/abstract=305479,
2002.
Domowitz, I., J. Glen, and A. Madhavan. “Liquidity, Volatility, and Equity Trading Costs Across
Countries and Over Time.” International Finance, 4 (2001), pp. 221-255.
Domowitz, I., and J. Krowas. “Where risk control meets cost control.” Journal of Investing, Vol. 13, No.3
(2004), pp. 82-87.
Engle, R., and R. Ferstenberg. “Execution Risk: It’s the Same as Investment Risk.” Journal of Portfolio
Management, Vol. 33, No. 2 (2007).
Fabozzi, F., S. Focardi, and P. Kolm. Financial Modeling of the Equity Market: From CAPM to
Cointegration. John Wiley and Sons, 2006.
Hou, K., and T. Moskowitz. “Market Frictions, Price Delay, and the Cross-Section of Expected Returns.”
Review of Financial Studies, Vol. 18, No. 3 (2005), pp. 981-1020.
Lobo, M., M. Fazel, and S. Boyd. “Portfolio Optimization with Linear and Fixed Transaction Costs.”
Annals of Operations Research, Vol. 152, No. 1 (2007).
Pogue, G. “An Extension of the Markowitz Portfolio Selection Model to Include Variable Transactions'
Costs, Short Sales, Leverage Policies and Taxes.” Journal of Finance, 25 (1970).
Spiegel, M., and X. Wang. “Cross-Sectional Variation in Stock Returns: Liquidity and Idiosyncratic
Risk.” Yale ICF Working Paper No. 05-13; EFA 2005 Moscow Meetings Paper, 2005. Available
at SSRN: http://ssrn.com/abstract=709781.
Tabb Group. “Imperfect Knowledge: International Perspectives on Transaction Cost Analysis,” March
2008.
© 2009 Investment Technology Group, Inc. All rights reserved. Not to be reproduced or redistributed without
permission. Approval #030209-41117
Exhibits
Exhibit 1
20
Single Period Allocation Comparison
non-cost aware
4
Weight in Portfolio
2
0
(in Pct.)
2
0
(in Pct.)
-4
-6
-8
Exhibit 2 Exhibit 3
Trade Sizes (rel. to ADV) Liquidity of Stocks
800000
250
700000
Average Daily Share
200 600000
(in Pct. to rel. ADV)
500000
Volume
Trade Size
150
400000
300000
100
200000
50 100000
0
0 short long, long, long,
short long, long, long, small medium large
small medium large
Exhibit 4
Cost Decomposition Relative to Portfolio Weights
21
Non-cost aware Portfolio Cost aware Portfolio
800 160
700 140
600 120
Cost (in b.p.)
spread impact commission borrow ing spread impact commission borrow ing
Exhibit 5
40 40
30 30
20 20
10 10
Sharpe ratio = -1.10 Sharpe ratio = 0.07
0 0
22
Transaction Costs Transaction Costs
(Non-Cost Aware) (Cost Aware)
4 4
2 2
1 1
0 0
Exhibit 6
500% 500%
400% 400%
Turnover (%)
Turnover (%)
300% 300%
200% 200%
100% 100%
0% 0%
Exhibit 7
350 350
300 300
Position Count
Position Count
250 250
200 200
150 150
100 100
50 50
0 0
23
Exhibit 8
Monthly Realized Volatility Comparison
9
8
7
Volatility (in %)
6
5
4
3
2
1
0
Exhibit 9
Global Transaction Cost Comparison
24
Dollar Weighted Averages Dollar Weighted Averages
Per Region Per Country
120 200
180
100 160
140
80
Cost (in b.p.)
20 40
20
0 0
2006 Q4 2007 Q4 2008 Q2
LatinAmerica NorthAmerica
Western Europe Eastern Asia 2006 Q4 2007 Q4 2008 Q2
Exhibit 10
Regional Weights in the Tangent Portfolio
North
America 14.1% 16.0% 16.6% 17.1% 8.6% 8.8% 6.9% 7.7% 39.2% 46.0% 48.7% 54.4%
Latin
America 27.0% 24.1% 21.4% 16.4% 5.5% 6.7% 7.4% 6.4% 3.4% 7.2% 9.6% 13.8%
Western
Europe 39.4% 43.6% 49.0% 55.0% 49.1% 51.2% 56.5% 59.8% 41.8% 31.0% 26.4% 20.1%
East
Asia 19.5% 16.2% 13.0% 11.5% 36.8% 33.3% 29.3% 26.1% 15.6% 15.8% 15.3% 11.7%
Exhibit 11
25
Regional Weights in the Minimum Variance Portfolio
December 2006 December 2007 August 2008
paper 100% paper 100% paper 100% turnover
portfolio turnover portfolio turnover portfolio
26