The Barra Global Equity Model (GEM2) : Research Notes
The Barra Global Equity Model (GEM2) : Research Notes
The Barra Global Equity Model (GEM2) : Research Notes
com
September 2008
Jose Menchero
Andrei Morozov
Peter Shepard
Table of Contents
1. Introduction …..………………………………….………………………….. 3
3. Factor Exposures
4. Factor Returns
6. Specific Risk
8. Conclusion ………………………..……………………………….....……. 38
1. Introduction
This document describes the new Barra Global Equity Model, GEM2, and provides an in-depth
comparison with the model it replaces, the Global Equity Model (GEM). With GEM, originally
released in 1989, Barra pioneered the use of factor models for the forecasting of global equity
portfolio risk. In the nearly 20 years since the launch of GEM, changes have occurred that
warrant the introduction of a completely new model. These changes fall into two categories.
First, additional data sources are available. These new data permit risk to be captured and
quantified along dimensions that were previously unattainable. Second, there have been
significant methodological advances in global equity risk modeling. GEM2 exploits both of these
changes to provide users with more accurate risk forecasts within a highly intuitive model
structure.
Two key elements are required to construct an accurate equity risk model. First, the model must
employ a set of factors that explain as fully as possible the cross section of stock returns. The
factors should capture all sources of return commonality, and thereby cleanly separate the
systematic component of return from the idiosyncratic, or stock-specific, component. The second
element pertains to the temporal resolution of the factor structure. In other words, the factor
returns must be observed at a sufficiently high frequency to accurately estimate their volatilities
and correlations.
a revised and updated industry factor structure based on the Global Industry Classification
Standard (GICS®), better reflecting the global economy’s current industrial structure
a broader estimation universe based on the MSCI Global Investable Market Indices (GIMI) for
greater precision in factor return estimates
an improved and expanded set of style factors that captures additional sources of style risk
the introduction of a World factor, giving a cleaner separation of country and industry effects
a structural specific risk model that generates more accurate specific risk estimates
greater stability in the proportion of risk coming from equity factors, currencies, and stock-
specific sources
GEM2 is being offered in short-term (GEM2S) and long-term (GEM2L) versions. Both versions
share the same factor structure, and have identical factor returns. They differ in their
responsiveness. GEM2S is designed to provide users with the most accurate and responsive risk
forecasts. GEM2L, by contrast, is intended for longer-term investors who value more stable risk
forecasts.
GEM also comes in two versions, one using MSCI industries, and the other based on FTSE
classifications. Both versions of GEM employ the same methodology and produce very similar
risk forecasts. This paper, however, considers only the MSCI version, which is the more widely
used of the two.
Global investors derive return from two basic sources: price appreciation in the local currency of
the asset, and repatriation of the asset value back to the base currency (numeraire) of the
investor.
For concreteness, consider an investment from the perspective of a US portfolio manager. Let
C (n ),$
C (n) denote the local currency of stock n , and let rFX be the return of currency C (n) due to
exchange-rate fluctuations against the dollar. The return of stock n in US dollars, denoted Rˆ n ,
$
Rˆ n$ = rˆn + rFX
C ( n ),$
+ rˆn rFX
C ( n ),$
, (2.1)
C ( n ),$
where r̂n is the local return of the asset. The cross term, rˆn rFX , is typically minute, and can be
safely ignored for risk purposes.1
Our primary interest is to forecast the volatility of excess returns (i.e., above the risk-free rate). If
rf$ denotes the risk-free rate of US dollars, then Rn$ = Rˆ n$ − rf$ is the excess return of stock n
measured in US dollars.
1
A notable exception to this is during times of extreme inflation, when the cross term can become significant.
Suppressing the cross term, and utilizing Equation 2.1, we decompose the excess return into an
$
equity component rn and a currency component qn :
qn$ = r fC ( n ) + rFX
C ( n ),$
− r f$ . (2.4)
This term represents the excess return in US dollars due to holding cash denominated in the local
currency of the stock.
We adopt a multi-factor framework to explain the local excess returns. This approach yields
valuable insight into the underlying sources of portfolio return by separating systematic effects
from the purely stock-specific component that can be diversified away. More specifically, we
posit that the local excess returns are driven by a relatively small number, K E , of global equity
factors, plus an idiosyncratic component unique to the particular stock,
KE
rn = ∑ X nk f k + un . (2.5)
k =1
Here, X nk (k ≤ K E ) is the exposure of stock n to equity factor k , f k is the factor return, and
un is the specific return of the stock. The specific returns un are assumed to be uncorrelated
with the factor returns. The factor exposures are known at the start of each period, and the factor
returns are estimated via cross-sectional regression.
Suppose that there are K C currencies in the model. Ordering the currencies after the equity
factors, we can express the excess currency returns as
K E + KC
q =
$
n ∑X
k = K E +1
nk f k$ , (2.6)
where X nk (k > K E ) is the exposure of stock n to currency k , and f k is the excess return of
$
the currency with respect to US dollars (which we calculate from risk-free rates and exchange
rates, as in Equation 2.4). We take the currency exposures X nk to be equal to 1 if k
corresponds to the local currency of stock n , and 0 otherwise.
K
Rn$ = ∑ X nk f k$ + un , (2.7)
k =1
where K = K E + K C is the total number of combined equity and currency factors. The elements
of X nk define the N × K factor exposure matrix, where N is the total number of stocks. Note
that the factor exposure matrix is independent of numeraire.
Our treatment thus far has considered only a single asset. Investors, however, are more
concerned with portfolio risk. The portfolio excess return (in US dollars) is given by
N
RP$ = ∑ hnP Rn$ , (2.8)
n =1
P
where hn is the portfolio weight of asset n . Note that, in general, the assets include both stocks
and cash. The portfolio exposure to factor k is given by the weighted average of the asset
exposures,
N
X kP = ∑ hnP X nk . (2.9)
n =1
In order to estimate portfolio risk, we also require the factor covariance matrix and the specific
risk forecasts. The elements of the factor covariance matrix are
(
Fkl$ = cov f k$ , f l $ , ) (2.10)
where the dollar superscript indicates that at least some of the matrix elements depend on the
base currency of the investor. Note, however, that the K E × K E block of the factor covariance
matrix corresponding to equity factors is independent of numeraire.
Δ mn = cov(um , un ) . (2.11)
For most stocks, we assume that the specific returns are uncorrelated, so that the off-diagonal
elements of Δ mn are zero. However, Equation 2.11 represents the generalized case in which the
specific returns of some securities are linked.2
The portfolio risk in US dollars can now be obtained as the square root of the variance,
1/ 2
⎡ ⎤
σ ( R ) = ⎢∑ X kP Fkl$ X lP + ∑ hmP Δ mn hnP ⎥
$
P . (2.12)
⎣ kl mn ⎦
2
We relax the assumption of uncorrelated specific returns for different share classes of the same stock.
Although Equation 2.12 is written for portfolio risk, it is equally valid for tracking error by replacing
portfolio weights and exposures with their active counterparts.
Many global investors, of course, are interested in risk forecasts from different numeraire
perspectives. The only term on the right-hand-side of Equation 2.12 that depends on base
$
currency is the factor covariance element, Fkl . These elements, however, can be transformed to
γ
some other numeraire, γ . Substituting these transformed factor covariances Fkl into Equation
2.12 yields risk forecasts with respect to the new base currency γ .
3. Factor Exposures
3.1. Estimation Universe
The coverage universe is the set of all securities for which the model provides risk forecasts. The
estimation universe, by contrast, is the subset of stocks that is used to estimate the model.
Judicious selection of the estimation universe is a critical component to building a sound risk
model. The estimation universe must be sufficiently broad to accurately represent the investment
opportunity set of global investors, without being so broad as to include illiquid stocks that may
introduce spurious return relationships into the model. Furthermore, the estimation universe
should be reasonably stable to ensure that factor exposures are well behaved across time.
Representation, liquidity and stability, therefore, represent the three primary goals that must be
attained when selecting a risk model estimation universe.
A well-constructed equity index must address and overcome these very issues, and therefore
serves as an excellent foundation for the estimation universe. The GEM2 estimation universe
utilizes the MSCI All Country World Investable Market Index (ACWI IMI), part of the MSCI Global
Investable Market Indices family which represents the latest in MSCI index-construction
methodology. MSCI ACWI IMI aims to reflect the full breadth of global investment opportunities
by targeting 99 percent of the float-adjusted market capitalization in 48 developed and emerging
markets. The index-construction methodology applies innovative rules designed to achieve index
stability, while reflecting the evolving equity markets in a timely fashion. Moreover, liquidity
screening rules are applied to ensure that only investable stocks with reliable pricing are included
for index membership.
The official MSCI ACWI IMI history begins in June 1994. To obtain a deeper model history,
however, we backward-extend the estimation universe for two years by including select securities
with existing GICS codes and clean daily returns. More specifically, for the period June 1992 to
June 1994, we construct the estimation universe by including the constituents of the MSCI All
Country World Index (ACWI), as well as all US and Japanese small-cap stocks with MSCI ACWI
IMI membership as of June 1994.
The MSCI ACWI IMI aims to capture the full opportunity set open to global equity investors. As a
result, if a country is deemed excessively restrictive to foreign investment, it is excluded.
Nevertheless, many of these frontier markets represent important opportunities for certain global
investors. For instance, since 2002, Qualified Foreign Institutional Investors (QFIIs) are permitted
to invest directly in the Chinese domestic A-share market, which is excluded from MSCI ACWI
IMI. Similarly, the six Gulf Cooperation Council (GCC) markets (Saudi Arabia, Oman, Qatar,
UAE, Kuwait, and Bahrain) are important to many global investors, although they are also
excluded from MSCI ACWI IMI. We therefore supplement the estimation universe with stocks
from China Domestic and the six GCC countries, represented by the MSCI China A Index and the
MSCI GCC Countries Index. We down-weight these markets in the regression in order to not
distort the factor relationships within the investable universe. In the future, as frontier markets
become more liquid and data becomes available, we plan to extend GEM2 coverage to many of
these markets.
In Figure 3.1, we show the number of securities in MSCI ACWI IMI over time. Several trends are
discernable. First, the number of assets rises significantly in the early history (1994-1997); this is
due to generally rising markets and an increase in the number of companies meeting the
minimum capitalization requirements for index membership. The second clear trend is the
gradual decline in the number of constituents that occurs from 1997 to 2002. This is a
consequence of a series of financial shocks, beginning with the Asia and Russia crises of
1997/1998, and ending with the collapse of the internet bubble in 2000 and the ensuing bear
market. Since June 2002, with the general recovery of the financial markets, we again observe a
secular increase in the number of MSCI ACWI IMI constituents.
The most prominent event in Figure 3.1 occurs in June 2002, which corresponds to the transition
to float-adjusted market-cap weighting and the application of liquidity screening rules for MSCI
ACWI IMI. In this event, the number of constituents drops sharply from nearly 12,000 to under
7,000. Although most of the excluded assets were small-cap stocks, this event must be treated
carefully in the estimation universe to avoid potentially spurious jumps in factor exposures. We
eliminate the discontinuity in the estimation universe by backward-excluding all stocks that were
dropped from MSCI ACWI IMI in June 2002. We allow exceptions to this rule in cases where
additional stocks are needed to populate countries with few assets.
In Figure 3.1, we also plot the number of securities in the GEM2 estimation universe over time.
We report only assets that receive full regression weight (i.e., we exclude China domestic and
GCC countries). While the large discontinuity in MSCI ACWI IMI (June 2002) has been removed
from the estimation universe by our backward-exclusion policy, we observe another discontinuity
in June 1994, coinciding with the start of MSCI ACWI IMI history. Although this would appear to
violate our goal of estimation universe stability, this event does not present model difficulties for
two reasons. First, the transition occurs well before the first available factor covariance matrix, so
that it cannot lead to spurious jumps in risk forecasts. Second, the impact is mitigated by the fact
that the affected assets are small-cap stocks so that, on a regression-weighted basis, the factor
exposures are quite stable.
• June 1994 to June 2002: MSCI ACWI IMI assets less those dropped in June 2002
(assets dropped are backward-excluded)
• June 1992 to June 1994: MSCI ACWI assets plus US and Japanese assets in MSCI
ACWI IMI as of June 1994 (MSCI ACWI IMI members backward-included)
• MSCI China A Index and MSCI GCC Countries Index, with down-weighting in regression
The GEM estimation universe is based on the MSCI World Index, which includes developed
markets. The number of constituents in this index is shown in Figure 3.1. Clearly, the MSCI
World Index is far narrower than the GEM2 estimation universe. Since estimation error increases
as the sample size is reduced, the broader estimation universe of GEM2 results in more efficient
estimates of the factor returns.
rn = f w + ∑ X nc f c + ∑ X ni f i + ∑ X ns f s + un . (3.1)
c i s
Mathematically, the World factor represents the intercept term in the cross-sectional regression.
Economically, it describes the aggregate up-and-down movement of the global equity market.
Typically, the World factor is the dominant source of total risk for a diversified long-only portfolio.
For most institutional investors, however, the primary concern is the risk of active long/short
portfolios. If both the portfolio and benchmark are fully invested – as is typically the case – then
the active exposure to the World factor is zero. Similarly, if the long and short positions of a
long/short portfolio are of equal absolute value, the net exposure to the World factor is zero.
Thus, we must look beyond the World factor to other sources of risk.
Country factors play a critical role in global equity risk modeling. One reason is that they are
powerful indicator variables for explaining the cross section of global equity returns. A second,
related, reason is that the country allocation decision is central to many global investment
strategies, and portfolio managers often must carefully monitor their exposures to these factors.
We therefore include explicit country factors for all markets covered.
In Table 3.1, we present a list of the 55 countries covered by GEM2, together with their
corresponding currencies. The country exposures X nc in GEM2 are set equal to 1 if stock n is
in country c , and set equal to 0 otherwise. We assign country exposures based on country
membership within the MSCI ACWI IMI, MSCI China A Index and MSCI GCC Countries Index.
Note that depository receipts and cross-listed assets are assigned factor exposures for the
underlying or primary asset, as defined by the MSCI Equity Indices.
We also report the average and ending weights over the period January 1997 to January 2008. It
is interesting to note the relative decline of the US and Japanese markets, and the rise of markets
such as Brazil, Russia, India and China.
Industries are also important variables in explaining the sources of global equity return co-
movement. One of the major strengths of GEM2 is to employ the Global Industry Classification
Standard (GICS®) for the industry factor structure. The GICS scheme is hierarchical, with 10 top-
level sectors, which are then divided into 24 industry groups, 68 industries, and 154 sub-
industries. GICS applies a consistent global methodology to classify stocks based on careful
evaluation of the firm’s business model and economic operating environment. The GICS
structure is reviewed annually by MSCI Barra and Standard & Poor’s to ensure it remains timely
and accurate.
Identifying which industry factors to include in the model involves a combination of judgment and
empirical analysis. At one extreme, we could use the 10 GICS sectors as industry factors. Such
broad groupings, however, would certainly fail to capture much of the cross-sectional variation in
stock returns. At the other extreme, we could use all 154 sub-industries as the factor structure.
Besides the obvious difficulties associated with the unwieldy numbers of factors (e.g., risk
reporting, thin industries), such an approach would present a more serious problem for risk
2
forecasting: although adding more factors always increases the in-sample R of the cross-
sectional regressions, many of the factor returns would not be statistically significant. Allowing
noise-dominated “factors” into the model defeats the very purpose of a factor risk model.
In GEM2, selection of the industry factor structure begins at the second level of the GICS
hierarchy, with each of the 24 industry groups automatically qualifying as a factor. This provides
a reasonable level of granularity, without introducing an excessive number of factors. We then
analyze each industry group, carefully examining the industries and sub-industries contained
therein to determine if a more granular factor structure is warranted. The basic criteria we use to
guide industry factor selection are: (a) the groupings of industries into factors must be
economically intuitive, (b) the industry factors should have a strong degree of statistical
significance, (c) incorporating an additional industry factor should significantly increase the
explanatory power of the model, and (d) thin industries (those with few assets) should be
avoided.
The result of this process is the set of 34 GEM2 industry factors, presented in Table 3.2.
Industries that qualify as factors tend to exhibit volatile returns and have significant weight. We
2
find that this relatively parsimonious set of factors captures most of the in-sample R explained
by the 154 sub-industries, but with a much higher degree of statistical significance. Also reported
in Table 3.2 are the average and end-of-period industry weights from January 1997 to January
2008. The weights were computed using the entire GEM2 estimation universe (i.e., including
China Domestic and GCC countries). Only five industries have end-of-period weights less than
100 bps, and these tend to be highly volatile, thus making them useful risk factors.
In Table 3.3, we report the underlying GICS codes that map to each of the GEM2 industry factors.
In five of the ten GICS sectors, there is a direct mapping between industry groups and factors. In
the other sectors, however, the factors are composed of collections of industries or sub-
industries. For instance, in the energy sector, the Oil & Gas Exploration & Production sub-
industry forms a separate factor, with the other four sub-industries constituting another factor.
Extracting this particular sub-industry is warranted due to its high statistical significance and
reasonable weight within the estimation universe. Other notable examples of industry factor
refinement include (a) carving out the Airlines industry from Transportation, (b) including
Biotechnology as a separate industry within the Health Care sector, and (c) using an Internet
factor within the Information Technology sector. In all of these cases, the resulting factor
groupings are financially intuitive and statistically significant.
Investment style represents another major source of systematic risk. Style factors, also known as
risk indices, are designed to capture these sources of risk. They are constructed from financially
intuitive stock attributes called descriptors, which serve as effective predictors of equity return
covariance. Since the descriptors within a particular style factor are meant to capture the same
underlying driver of returns, these descriptors tend to be significantly collinear. For instance,
price-to-book ratio, dividend yield, and earnings yield are all attributes used to identify value
stocks, and they tend to exhibit significant cross-sectional correlation. Although these descriptors
have significant explanatory power on their own, naively including them as separate factors in the
model may lead to serious multi-collinearity problems. Combining these descriptors into a single
style factor overcomes this difficulty, and also leads to a more parsimonious factor structure.
Unlike country and industry factors, which are assigned exposures of either 0 or 1, style factor
exposures are continuously distributed. To facilitate comparison across style factors, they are
Raw
standardized to have a mean of 0 and a standard deviation of 1. In other words, if d nl is the
raw value of stock n for descriptor l , then the standardized descriptor value is given by
d nlRaw − μl
d nl = , (3.2)
σl
where μl is the cap-weighted mean of the descriptor (within the estimation universe), and σl is
the equal-weighted standard deviation. We adopt the convention of standardizing using the cap-
weighted mean so that a well-diversified cap-weighted global portfolio, such as MSCI ACWI IMI,
has approximately zero exposure to all style factors. For the standard deviation, however, we
use the equal-weighted mean to prevent large-cap stocks from having an undue influence on the
overall scale of the exposures.
Some of the style factors are standardized on a global-relative basis, others on a country-relative
basis. In the former case, the mean and standard deviation in Equation 3.2 are computed using
the entire global cross section. In the latter case, the factors have mean 0 and standard deviation
1 within each country. When deciding which standardization convention to adopt, we consider
both the intuitive meaning of the factor and its explanatory power.
X nk = ∑ wl d nl , (3.3)
l ∈k
where wl is the descriptor weight, and the sum takes place over all descriptors within a particular
risk index. Descriptor weights are determined using an optimization algorithm to maximize the
explanatory power of the model.
One of the major advances of GEM2 is its refined and expanded set of style factors. GEM2 uses
eight style factors, compared to four in its predecessor. Below, we provide a qualitative
description of each of the style factors:
The Volatility factor is typically the most important style factor. In essence, it captures market
risk that cannot be explained by the World factor. The most significant descriptor within the
Volatility index is historical beta relative to the World portfolio (as proxied by the estimation
universe). To better understand this factor, consider a fully invested long-only portfolio that is
strongly tilted toward high-beta stocks. Intuitively, this portfolio has greater market risk than a
portfolio with beta equal to one. This additional market risk is captured through positive
exposure to the Volatility factor. Note that the time-series correlation between the World
factor and the Volatility factor is typically very high, so that these two sources of risk reinforce
one another in this example. If, by contrast, the portfolio is invested in low-beta stocks, then
the risk from the Volatility and the World factors is partially cancelled, as intuitively expected.
We standardize the Volatility factor on a global-relative basis. As a result, the mean
exposure to Volatility within a country can deviate significantly from zero. This
standardization convention is a natural one for a global model, as most investors regard
stocks in highly volatile markets as having more exposure to the factor than those in low-
volatility markets. This view is reflected in the data, as we find that the explanatory power of
the factor is greater using the global-relative standardization.
The Momentum factor often ranks second in importance after Volatility. Momentum
differentiates stocks based on recent relative performance. Descriptors within Momentum
include historical alpha from a 104-week regression and relative strength (over trailing six and
12 months) with a one-month lag. Similarly to Volatility, Momentum is standardized on a
global-relative basis. This is also an intuitive convention for a global model. From the
perspective of a global investor, a stock that strongly outperforms the World portfolio is likely
to be considered a positive momentum stock, even if it slightly underperforms its country
peers. The empirical results support this view, as the Momentum factor standardized globally
has greater explanatory power than one standardized on a country-relative basis.
The Size factor represents another well-known source of return covariance. It captures the
effect of large-cap stocks moving differently from small-cap stocks. We measure Size by a
single descriptor: log of market capitalization. The explanatory power of the model is quite
similar whether Size is standardized globally or on a country-by-country basis. We adopt the
country-relative standardization, however, since it is more intuitive and consistent with
investors’ perception of the markets. For instance, major global equity indices, such as the
MSCI Global Investable Market Indices, segment each country according to size, with the
largest stocks inside each country always being classified as large-cap stocks. Moreover,
standardizing the Size factor on a global-relative basis would serve as an unintended proxy
for developed markets versus emerging markets, and increases collinearity with the country
factors.
The Value factor describes a major investment style which seeks to identify stocks that are
priced low relative to fundamentals. We standardize Value on a country-relative basis. This
again is consistent with the way major indices segment each market, with each country
divided roughly equally into value and growth sub-indices. This convention also circumvents
the difficulty of comparing fundamental data across countries with different accounting
standards. GEM2 utilizes official MSCI data items for Value factor descriptors, as described
in the MSCI Barra Fundamental Data Methodology handbook.3
Growth differentiates stocks based on their prospects for sales or earnings growth. It is
standardized on a country-relative basis, consistent with the construction of the MSCI Value
and Growth Indices. Therefore, each country has approximately half the weight in stocks
with positive Growth exposure, and half with negative exposure. The GEM2 Growth
descriptors also utilize official MSCI data items, as described in the MSCI Barra Fundamental
Data Methodology handbook.
The Non-Linear Size (NLS) factor captures non-linearities in the payoff to the size factor
across the market-cap spectrum. NLS is based on a single raw descriptor: the cube of the
log of market capitalization. Since this raw descriptor is highly collinear with the Size factor,
we orthogonalize it to the Size factor. This procedure does not affect the fit of the model, but
does mitigate the confounding effects of collinearity, and thus preserves an intuitive meaning
3
http://www.msci.com/methodology/meth_docs/MSCI_Sep08_Fundamental_Data.pdf
for the Size factor. The NLS factor is represented by a portfolio that goes long mid-cap
stocks, and shorts large-cap and small-cap stocks.
The Liquidity factor describes return patterns to stocks based on relative trading activity.
Stocks with high turnover have positive exposure to Liquidity, whereas low-turnover stocks
have negative exposure. Liquidity is standardized on a country-relative basis.
Leverage captures the return difference between high-leverage and low-leverage stocks.
The descriptors within Leverage include market leverage, book leverage, and debt-to-assets
ratio. This factor is standardized on a country-relative basis.
In Appendix A, we provide additional detail on the individual descriptors comprising each style
factor.
The overlap of country factors for GEM and GEM2 is very large. Both models cover the 48
developed and emerging markets contained within MSCI ACWI IMI. GEM also contains the
following eight non-MSCI ACWI IMI countries: Sri Lanka, Venezuela, Nigeria, Slovakia,
Zimbabwe, Saudi Arabia, Oman, and Bahrain. GEM2, on the other hand, supplements the 48
MSCI ACWI IMI markets with China Domestic (A-shares traded in mainland China) and the six
GCC countries. Given the growing importance of the GCC region and the China domestic
market, we included these in GEM2. By contrast, due to tight capital controls, illiquidity, or
extreme economic instability (e.g., hyperinflation), Sri Lanka, Slovakia, Venezuela, Nigeria, and
Zimbabwe are not presently included in GEM2.
A more fundamental difference between the models is that, whereas GEM2 uses country
exposures of 0 or 1, GEM determines country exposure as historical beta estimated by a time-
series regression. More specifically, the GEM country exposures are found by regressing 60
months of local excess stock returns against the local excess country returns within MSCI ACWI.
While this has some intuitive appeal, it also presents difficulties. For instance, the country betas
can only be estimated with the usual sampling error. Furthermore, since the regression takes
place over the trailing 60 months, the estimates can be stale. As a result, the GEM country
exposures – one of the key drivers of global equity risk – are sometimes unintuitive. This can
lead to unexpected active exposures to GEM country factors, even when the active country
weight is zero.
GEM uses 38 industry factors, with exposures given by 0 or 1, as in GEM2. The GEM structure
was a good representation of the global industrial structure at the time of the model’s release but
is now significantly out of date. In Table 3.4, we report the mapping of GEM2 industries to GEM
industries. In some sectors, such as Energy, we see that more than one GEM2 factor maps to a
single GEM factor. In other sectors, such as Information Technology, we find the reverse (i.e.,
single GEM2 industries mapping to multiple GEM industries). In yet other sectors, such as
Financials, there is a near a one-to-one mapping between GEM2 and GEM industries.
In Table 3.5, we report the mapping of GEM industries to GEM2 industries. As of January 2008,
13 of the 38 GEM industries have weights less than 100 bps. Many of these thinly populated
GEM industries are not particularly volatile, such as Textiles & Apparel, or Forestry and Paper
Products, so that they no longer constitute ideal risk factors.
GEM contains four style factors. Some are very similar to their GEM2 counterparts, while others
are significantly different. Size is an example of a style factor that is essentially the same in GEM
and GEM2. Both models define the Size factor by a single descriptor: log of total market
capitalization.
The GEM Value factor is similar to the GEM2 Value factor. Both models use forward and trailing
earnings-to-price ratios, dividend yield, and book-to-price ratios as descriptors. GEM2 also
includes a fifth descriptor, cash earning-to-price ratio, in the Value factor. Descriptor weights, of
course, are different between the two models.
The Success factor in GEM roughly corresponds to the GEM2 Momentum factor. Whereas
GEM2 uses three descriptors in the factor, GEM uses only 12-month relative strength. Another
difference is that the GEM2 factor applies a one-month lag to relative strength, whereas GEM
does not. The GEM2 treatment is more consistent with the way most practitioners define
momentum.
The biggest difference is with the GEM Variability in Markets (VIM) factor, which roughly
corresponds to the GEM2 Volatility factor. The GEM factor is a single-descriptor risk index which
uses the residual volatility (historical sigma) from a CAPM country regression. Since this
descriptor does not directly capture covariance, however, it is weaker than the Volatility factor of
GEM2.
A final comment is in order regarding country factors and volatility factors. GEM estimates
country exposure by historical beta, and includes VIM as a separate factor. GEM2, by contrast,
uses indicator variables for country exposure and includes Volatility (with historical beta as a
descriptor) as a separate factor. The most meaningful direct comparison between the GEM and
GEM2 country factors, therefore, should include not only the country factors, but also the Volatility
and VIM factors as well.
4. Factor Returns
4.1. Data Quality and Outlier Treatment
The most difficult and time-consuming part of constructing a global equity risk model lies in
preparation of the input data. If the data inputs are garbage, the risk forecasts will likewise be
garbage, no matter how sophisticated the model. Assuring a high degree of data quality,
therefore, is a vital part of building a reliable risk model.
In order to obtain the highest quality inputs, GEM2 leverages the same data infrastructure used
for the construction of MSCI Global Investable Market Indices. Data items such as raw
descriptors, GICS codes, country classifications, and clean daily stock returns are obtained from
in-house sources that have already undergone extensive quality control.
No matter how stringent the data quality assurance process, we can never exclude the possibility
of extreme outliers entering the data set. These extreme outliers may represent legitimate values
or outright data errors. Either way, such observations must be carefully handled to prevent a few
data points from having an undue impact on model estimation.
In GEM2, we employ a multi-step algorithm to identify and treat outliers. The algorithm assigns
each observation into one of three groups. The first group represents values so extreme that
they are treated as potential data errors and removed from the estimation process. The second
group represents values that are regarded as legitimate, but nonetheless so large that their
impact on the model must be limited. We trim these observations to three standard deviations
from the mean. The third group of observations, which forms the bulk of the distribution, consists
of values less than three standard deviations from the mean. These observations are left
unadjusted.
The thorough data quality assurance process, coupled with the robust outlier algorithm, ensures
that the input data are clean and reliable. There is still the issue, however, of dealing with
missing data. Data could be missing either due to lack of availability or due to removal by the
outlier algorithm. The data items most likely to be missing include descriptors for the Value,
Growth, Leverage, and Liquidity style factors. More rarely, stocks may be missing exposure to
Volatility and Momentum factors; this occurs for securities which lack return history, such as
recent spin-offs or IPOs. Other data items, such as Size, industry, and country exposures, are
never missing, since these items are required for coverage universe membership.
If the underlying data required to compute factor exposures are missing, then the factor
exposures are generated using a data-replacement algorithm. A simple approach would be to
assign zero exposure to all missing values. A better approach, however, is to exploit known
relationships between factor exposures. For instance, if we know that a stock is in the
Semiconductor industry and has a positive exposure to Volatility, then it is likely that the stock
also has a positive exposure to Growth.
We apply the data-replacement algorithm to the Value, Growth, Leverage, and Liquidity factors.
The algorithm works by regressing these four factors (using only non-missing exposures) against
Size, Volatility and industries. The slope coefficients of the regression are then used to estimate
the factor exposures for the stocks with missing data. Countries are not used as explanatory
variables in the regression since the four style factors (Value, Growth, Leverage and Liquidity) are
standardized to be mean zero with respect to countries. The algorithm is not applied to the other
four style factors (Size, Non-Linear Size, Volatility and Momentum) since these factor exposures
are rarely, if ever, missing. Note also that the algorithm is applied at the factor level as opposed
to the descriptor level. In other words, if a stock has data for some descriptors within a risk
index, but not others, then the non-missing data will be used to compute the factor exposures.
Only if all descriptors are absent does the replacement algorithm become active.
KE
rn = ∑ X nk f k + un . (4.1)
k =1
GEM2 uses weighted least squares, assuming that the variance of specific returns is inversely
proportional to the square root of total market capitalization.
As described in Section 3, the GEM2 equity factors include the World factor, countries, industries,
and styles. Every stock in GEM2 has unit exposure to the World factor, and indicator variable
exposures of 0 or 1 to countries and industries. As a result, the sum of all country factors equals
the World factor, and similarly for industries, i.e.,
∑Xc
nc = 1, and ∑X i
ni = 1, (4.2)
for all stocks n . In other words, the sum of all country columns in the factor exposure matrix
gives a column with 1 in every entry, which corresponds to the World factor. The same holds for
industry factors. The GEM2 factor structure, therefore, exhibits exact two-fold collinearity.
Constraints must be applied to obtain a unique solution.
In GEM2 we adopt an intuitive set of constraints that require the cap-weighted country and
industry factor returns to sum to zero,
∑w
c
c c f = 0, and ∑w f
i
i i = 0, (4.3)
where wc is the weight of the estimation universe in country c , and wi is the corresponding
weight in industry i . These constraints remove the exact collinearities from the factor exposure
matrix, without reducing the explanatory power of the model.
We can now give a more precise interpretation to the factors. Consider the cap-weighted
E
estimation universe, with holdings h . The return of this portfolio RE can be attributed using the
n
GEM2 factors,
RE = f w + ∑ wc f c + ∑ wi f i + ∑ X sE f s + ∑ hnE un . (4.4)
c i s n
The constraints imply that the first two sums in Equation 4.4 are equal to zero. The third sum is
also zero since the style factors are standardized to be cap-weighted mean zero; i.e., X s = 0 ,
E
for all styles s . The final sum in Equation 4.4 corresponds to the specific return of a broadly
diversified portfolio, and is approximately zero (note it would be exactly zero if we used
regression weights instead of capitalization weights). Thus, to an excellent approximation,
Equation 4.4 reduces to
RE ≈ f w . (4.5)
In other words, the return of the World factor is essentially the cap-weighted return of the
estimation universe.
To better understand the meaning of the pure factor portfolios, we report in Table 4.1 the long
and short weights (January 2008) of the World portfolio and several pure factor portfolios in
various market segments. The World portfolio is represented by the cap-weighted GEM2
estimation universe. The pure World factor is 100 percent net long and the net weights closely
match those of the World portfolio in each segment. The other pure factors all have net weight of
zero, and therefore represent long/short portfolios.
As a first approximation, the pure country factors can be regarded as going long 100 percent the
particular country, and going short 100 percent the World portfolio. For instance, going long 100
percent Japan and short 100 percent the World results in a portfolio with roughly 91 percent
weight in Japan, and -91 percent in all other countries. The pure country factors, however, have
zero exposure to industry factors. This is accomplished by taking appropriate long/short
combinations in other countries. For instance, the Japanese market is over-represented in the
segment corresponding to the Automobile factor. To partially hedge this exposure, the pure
Japan factor takes a net short position of -1.08 percent in the US Automobile segment. A similar
short position would be found in the German Automobile segment.
The pure Automobile factor can be thought, as a first approximation, to be formed by going 100
percent long the Automobile industry and 100 percent short the World portfolio. A more refined
view of the factor takes into account that the net weight in each country is zero. The pure
Automobile factor naturally takes a large long position in Japanese automobiles, but hedges the
Japan exposure by taking short positions in other Japanese segments.
The pure Volatility factor is perhaps the easiest to understand, as it takes offsetting long and short
positions within all segments corresponding to GEM2 factors (e.g., Japan, US, and Automobiles).
Note that the weights are not equal to zero for segments that do not correspond to GEM2 factors,
such as Japanese automobiles.
In this section, we present and discuss some of the quantitative characteristics of the GEM2
factors. In particular, we investigate the degree of collinearity among the factor exposures, and
report on the statistical significance, performance and volatility of the factor returns.
A feature of the multi-factor framework is that it disentangles the effects of many variables acting
simultaneously. Multi-collinearity among factors, however, can confound this clean separation of
effects.
One measure of collinearity is the pair-wise cross-sectional correlation between factor exposures.
In Table 4.2, we report regression-weighted correlations among style factors and industries,
averaged over the period January 1997 to June 2008. In general, the correlations are intuitive in
sign. For instance, Volatility is positively correlated with Growth, Liquidity, and Semiconductors,
and negatively with Value and Utilities. Value, as expected, is positively correlated with Banks
and Utilities, and negatively with Growth and Biotechnology. Although none of the correlations
are particularly large, the correlations between pairs of style factors are typically larger, on
average, than those between styles and industries.
The statistical significance of factor returns plays a key role in the construction of a risk model.
Let f k be the factor return for a particular period, and se( k ) be the standard error of this
estimate. The t-statistic is given by
fk
tk = . (4.6)
se(k )
Generally, a t-statistic with absolute value greater than 2 is considered significant at the 95
percent confidence level. Ideally, a risk factor should have high t-statistics that are persistent
across time. Useful measures of this are the average squared t-statistic over time, and the
percentage of observations with significant t-statistics.
In Table 4.3, we report the statistical significance of the GEM2 World and country factor returns
over the sample period January 1997 to June 2008. Not surprisingly, the World factor scores the
highest average squared t-statistic. Countries that have highly significant factor returns are
Japan, US, and China Domestic. Generally, countries with large weights and volatile returns tend
to be highly significant. Also note that most countries have a high proportion of statistically
significant factor returns, far higher than the five percent one would expect if the factors were
driven by pure noise.
In Table 4.3, we also report the sample kurtosis for weekly factor returns, average factor returns,
volatilities, and correlations with the estimation universe over the January 1997 to June 2008
sample period. All of the country factor returns have kurtosis in excess of 3 (which is the level of
kurtosis of a normal distribution), indicating that the returns are not normally distributed. Four
countries – Hong Kong, Malaysia, Russia, and Singapore – have kurtosis greater than 10.
Since the country factors are net of the World factor, positive factor returns indicate that the
country portfolio has outperformed the World portfolio. We see that the GCC markets have
performed particularly well over the sample period, while several Asian markets have
underperformed. We also report the annualized factor volatility. Not surprisingly, the US factor
has the lowest volatility, followed by Canada and the developed European markets. This is a
reflection of the fact that these countries tend to track the World portfolio relatively closely. The
Sharpe Ratio is the annualized factor return divided by the factor volatility. Again, the GCC
markets have performed well on a risk-adjusted basis. The final column in Table 4.3 reports the
time-series correlation with the estimation universe return. It is interesting to observe that this
correlation for the World factor is 0.994, confirming that it represents the cap-weighted estimation
universe.
In Table 4.4, we report statistical characteristics for the GEM2 industry factors. Clearly, the
GEM2 industry factors are strongly significant, with roughly two-thirds of them exhibiting
significant t-statistics over 50 percent of the time. Table 4.4 provides strong empirical support for
our industry factor structure: in all instances where we added a refined industry structure (e.g.,
carving out Internet, Biotechnology or Airlines), the resulting factors are highly volatile and
statistically significant. Note that industry factors, like country factors, all have kurtosis levels
exceeding 3.
There are several other interesting observations in Table 4.4. First, the worst performing industry
factor over this period was Airlines (-8.54%). This contrasts sharply with the Transportation Non-
Airline factor, which is derived from the same GICS industry group and had a positive return over
the sample period. The top-performing industry factor, both on an absolute basis and a risk-
adjusted basis, was Biotechnology. Note that defensive sectors, such as Consumer Staples or
Utilities, were negatively correlated with the estimation universe, whereas aggressive sectors,
such as Information Technology, were positively correlated.
In Table 4.5, we report the summary statistics for the GEM2 style factors. In this case, we divide
the sample period into two sub-periods. In the first sample period (January 1997 to June 2002),
the style factors were generally more volatile and had a greater degree of statistical significance
than in the second sub-period (July 2002 to June 2008). This is not too surprising, since the first
sub-period contains the rise and fall of the internet bubble. Note that the kurtosis of the Value
factor in the second sample period is unusually high. This is caused primarily by two six-sigma
events in August 2007.
We can discern from Table 4.5 a clear risk hierarchy in the style factors. Alone at the top is the
Volatility factor, which is far more volatile and significant than any of the other factors. In the
second tier are Momentum and Size, which are clearly set apart from the third group, containing
the other five styles. In this third group, the Value factor tends to be the most significant.
Performance, however, tells a different story. Volatility performed very poorly during the first sub-
period, and was essentially flat during the second. Value and Momentum, on the other hand,
were strongly positive over both sample periods.
There is substantial correlation between some of the style factors with the estimation universe.
For instance, from Table 4.5, we see that Volatility is strongly correlated (about 80 percent) with
the estimation universe. This makes sense since, during up-markets, high-beta stocks tend to
outperform low-beta stocks, so that the return to Volatility – which includes beta – is also positive.
A similar argument holds for down-markets. Liquidity and Size also tend to be positively
correlated with the estimation universe over both sample periods.
While the GEM regression is formally equivalent to that in GEM2, there are nonetheless
significant differences between the two regressions. The first major difference is that the GEM
regression imposes constraints on all of the country factor returns,
~
f k = RkW , (4.8)
W
where Rk is the cap-weighted return of country k in MSCI ACWI. In other words, GEM does
not estimate the country factor returns by regression; rather, it takes them from the country
indices. The effect of these multiple constraints is to reduce the explanatory power of the model.
Another difference is that GEM does not include a World factor. Reflecting the global economy of
its time, GEM embeds the market effect in the country factor, giving them greater importance than
industry factors, which are net of the market. In GEM2, by contrast, the market is captured by the
World factor, and both industries and countries are treated on an equal footing net of the market.
R2 = 1 −
∑wu n n
2
n
, (4.9)
∑wr n n n
2
where wn is the regression weight of stock n , un is the specific return, and rn is the local
2
excess return. The R can be interpreted as the ratio of the average explained squared return to
the average total squared return.
To investigate the explanatory power of GEM2 relative to GEM, we carry out monthly cross-
2
sectional regressions for both models. Since differences in R can be quite small, it is important
to control all external variables that can impact the R values.4 Therefore, in our regressions, we
2
use (a) identical set of stocks (namely, the GEM2 estimation universe excluding China domestic
and GCC), (b) identical input stock returns, and (c) identical regression weighting (i.e., square
root of market cap). The only quantity that was varied was the factor exposure matrix. For the
GEM regressions, we also impose the constraint on country factor returns (that they be equal to
the country index returns), given by Equation 4.8.
2
In Figure 4.1, we plot the trailing 12-month average R for GEM2 and GEM. The GEM2 factors
consistently dominate their GEM counterparts over the entire sample period. On average, GEM2
2
has an R advantage of about 350 basis points (bps) over GEM, although this difference rises to
well over 500 bps during the period 2000-2002. For equity factor models, where even a 50 bp
2
boost in R is considered a significant achievement, an increase of 350 bps is unusually large.
4
When comparing the explanatory power of two sets of factors, it is usually important to use the adjusted R-squared, which applies a slight
correction based on the number of explanatory variables. This adjustment is unnecessary in this case, since GEM and GEM2 have the
same number of factors.
In Table 4.6, we investigate the sources of this increase in explanatory power. We examine
seven special cases, carefully selecting combinations of factors and running side-by-side
regressions to isolate the desired effect. The sample period is from January 1997 to June 2008.
Regressions labeled “A” use only GEM2 factors, whereas those labeled “B” employ either GEM
factors or a mix of GEM2 and GEM factors.
In Case 1 of Table 4.6, we compare GEM2 country factors against the GEM country factors.
Recall that GEM country exposures are given by the country betas, whereas GEM2 uses simple
indicator variables. One expects that the GEM country factors, in isolation, would explain more
than the GEM2 countries. Surprisingly, we see that GEM2 country factors outperform GEM’s by
94 bps.
This puzzling result can be explained in terms of the GEM constraints (i.e., Equation 4.8). In
Case 2, we run identical regressions as in Case 1, except this time removing the GEM
constraints. As expected, the GEM country factors now explain more than the GEM2 factors.
As discussed in Section 3.3, because the GEM country factors incorporate covariance with the
market, a fair comparison between the GEM2 and GEM country factors must include the Volatility
factor (for GEM2) and VIM (for GEM). In Case 3, we run this comparison and see that GEM2
country factors now outperform the GEM factors by 137 bps. This also reflects the strength of the
2
GEM2 volatility factor, which boosts the R by 361 bps. Case 3 therefore demonstrates that
GEM2 country factors, in proper combination with Volatility, outperform the GEM country factors.
In Case 4 of Table 4.6, we compare the explanatory power of GEM2 style factors versus GEM’s.
In Regression A, we include the full set of GEM2 factors. In Regression B, we swap the GEM2
style factors with the GEM style factors. This comparison, therefore, isolates the effect of style
factors. We see that the GEM2 style factors outperform their GEM counterparts by 71 bps over
the sample period.
In Case 5 of Table 4.6, we compare the explanatory power of GEM2 industry factors versus
GEM. Again, this was accomplished by pulling out the GEM2 industry factors and replacing them
with the GEM industry factors in Regression B. The result is that GEM2 industries explain, on
average, 109 bps per month more than the GEM industries.
In Case 6, we compare the explanatory power of the complete factor sets for both models, but
without imposing the GEM country constraints. GEM2 outperforms GEM by 142 bps per month.
In the final case, we run the actual regressions for GEM, which includes the country constraints.
Over this sample period, GEM2 explains, on average, 355 bps more each month than GEM. This
represents an impressive increase in explanatory power.
2
While Table 4.6 provides insight into the sources of the difference in R between GEM2 and
2
GEM, it is also illuminating to consider the sources of absolute R in GEM2. In Figure 4.2, we
2
plot the trailing 12-month R of GEM2 factor groups over time. The first case we examine is the
World factor in isolation. This is the single most important factor and, on average, explains about
2
10 percent in R . Its explanatory power, however, fluctuates over time. For instance, in 2000
the World factor explained only a few percentage points, whereas this figure rose to nearly 20
percent by 2003.
From Figure 4.2, it is evident that most of the explanatory power of the model comes from the
combined effect of country and industry factors. Interestingly, the relative explanatory power of
these factor groups changes significantly over time. In the early period, industries explained
relatively little compared to countries. Then, some time in 1999, the explanatory power of
industries suddenly exploded. From 2000-2003, industries actually explained a greater
2
proportion of trailing 12-month R than countries. Since 2003, the situation has partially
reversed, with country factors reasserting their dominance although industry factors remain very
important. This example illustrates some the insights obtained by the GEM2 factor structure,
which places countries and industries on an equal footing.
2
Figure 4.2 also demonstrates that style factors contribute about 250 bps in R beyond that
explained by countries and industries alone. This amount, however, also varies considerably
over time. In the early period, style factors added only about 100 bps, whereas this widened to
about 500 bps during the 2000-2004 period. More recently, style factors have added about 200
bps.
2
The black dots in Figure 4.2 represent the monthly R values for the complete set of GEM2
factors. This quantity varies dramatically from month to month, and tends to be largest during big
market moves. In extreme months, commonalities in returns swamp asset-specific returns. For
2
example, the largest single R observed during the sample period was 0.75. This occurred in
August 1998, when the market dropped by more than 15 percent.
We use exponentially weighted moving averages (EWMA) to estimate the factor covariance
matrix for both equity and currency blocks. This approach gives more weight to recent
observations and is a simple, yet robust, method for dealing with data non-stationarity. An
alternative approach would be to use generalized auto-regressive conditional heteroskedasticity,
or GARCH. We find that EWMA estimators are typically more robust to changing market
dynamics and produce more accurate risk forecasts than their GARCH counterparts.
In GEM2, we must also account for the possibility of serial correlation in factor returns, which can
affect risk forecasts over a longer horizon. Suppose, for instance, that high-frequency returns are
negatively correlated. In this case, long-horizon risk forecasts estimated on high-frequency data
will be lower than that implied using simple square-root-of-time scaling, since returns one period
tend to be partially offset by opposing returns the following period.
The prediction horizon in Barra risk models is typically one month. Models that are estimated on
daily or weekly returns, therefore, must adjust for serial correlation. Note that models estimated
on monthly observations, such as GEM, need not adjust for serial correlation, since the
observation frequency coincides with the prediction horizon.
Full treatment of serial correlation must account for not only the correlation of one factor with itself
across time, but also for the correlation of two factors with each other across different time
periods. In GEM2, we model serial correlation using the Newey-West methodology5 with two
lags. This assumes that the return of any factor may be correlated with the return of any other
factor up to two weeks prior.
It is useful to think of the factor covariance matrix as being composed of a correlation matrix,
which is scaled by the factor volatilities. Volatilities and correlations can then be estimated
separately using EWMA with different half-life parameters.
5
Newey, W., and K. West, 1987. “A Simple, Positive Semi-Definite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix.”
Econometrica, Vol. 55, No. 3: 703-708.
The volatility half-life determines the overall responsiveness of the model. Selecting the proper
volatility half-life involves analyzing the trade-off between accuracy and responsiveness on the
one hand, and stability on the other. If the volatility half-life is too long, the model gives undue
weight to distant observations that have little to do with current market conditions. This leads to
stable risk forecasts, but at the cost of reduced accuracy. By contrast, a short volatility half-life
makes the model more responsive, generally improving the bias statistics, but at the cost of
jumpier risk forecasts. Of course, the volatility half-life cannot be lowered without bound; if the
half-life is too short, then sampling error can become so large that the risk forecasts are not only
less stable, but also less accurate.
We measure the accuracy of the risk forecasts using bias statistics, which essentially represent
the ratio of realized risk to forecast risk. A bias statistic of 1 is considered ideal, but sampling
error ensures that realized bias statistics will deviate from 1, even for perfect risk forecasts. In
Appendix B, we provide a review of bias statistics.
When carrying out bias statistic testing, one must consider the number of months to include in the
observation window. Sampling error suggests that one should use the largest interval possible,
since that would minimize sampling error. This would be a compelling argument for stationary
data. In reality, however, financial markets are non-stationary. It is possible to over-predict risk
for several years and then under-predict for others, while obtaining a bias statistic close to 1.
Getting the right average risk forecast over 10 or 20 years is small consolation to a portfolio
manager who can be wiped out due to poor risk forecasts in any single year.
Rolling 12-month intervals provide a more relevant framework for evaluating the accuracy of risk
forecasts. Although sampling error is larger, this approach penalizes the model for poor risk
forecasts during any individual year. For a given portfolio, we compute the rolling absolute
deviation (RAD) of the bias statistic as follows. First, we calculate the bias statistic over the first
year of the observation window, using 12 monthly observations. We then take the absolute value
of the deviation of the bias statistic from 1. We then roll the 12-month window forward one month
at a time, computing absolute deviations each month, until we reach the end of the observation
window. For instance, a 138-month observation window will accommodate 127 separate 12-
month windows. The RAD for the portfolio is the average of the absolute deviations over these
127 sub-periods. The RAD measure captures in a single number the accuracy of risk forecasts
over rolling 12-month intervals by penalizing all observations throughout the sample history that
deviate from the ideal bias statistic of 1.
As shown in Appendix B, for perfect risk forecasts and normally distributed returns (kurtosis 3),
the critical value of the RAD over a 138-month observation window is 0.22. In other words, if the
observed RAD is larger than 0.22, then with 95-percent confidence we can reject the null
hypothesis that the risk forecast was accurate. Returns, of course, are not normally distributed.
For a kurtosis of 5, which corresponds to slightly fat tails, the critical value of RAD for a 138-
month observation window is 0.29. For even fatter tails, say a kurtosis of 10, the critical value
increases to 0.34.
To study the question of the stability of risk forecasts, we compute the monthly absolute change
in volatility for each factor k ,
σ k (t + 1) − σ k (t )
vk (t ) = . (5.1)
σ k (t )
If the volatility forecast changes over one month from 6.0 percent to 5.7 percent, for instance,
then the monthly absolute change is 5 percent.
The mean variability provides a measure of the stability of the risk forecasts, and is given by the
average of the monthly absolute changes,
1
v=
KT
∑ v (t ) ,
k ,t
k (5.2)
In Figure 5.1, we show the mean RAD for all GEM2 equity factors, plotted as a function of
volatility half-life. The bias statistics were computed over rolling 12-month windows using a 138-
month observation window from January 1997 to June 2008. We also plot the mean monthly
variability v of the equity factors. The volatility half-lives of the GEM2S and GEM2L models are
indicated by vertical lines.
The solid lines in Figure 5.1 are the actual GEM2 results, which incorporate a two-lag
autocorrelation adjustment. For comparison, we also show by dashed lines the corresponding
results without serial correlation adjustments. Including autocorrelation adjustments makes the
risk forecasts more accurate, but also slightly increases the monthly variability in the forecasts.
From Figure 5.1, we see that the volatility half-life of GEM2S (18 weeks) minimizes the mean
RAD, and therefore optimizes the accuracy of the risk forecasts. The mean RAD for GEM2S is
about 0.24, versus 0.27 for GEM2L, which has a volatility half-life of 52 weeks. While GEM2S is
therefore more accurate than GEM2L, the price one pays is jumpier risk forecasts. For GEM2S,
the monthly variability is about 6.7 percent, versus 2.5 percent for GEM2L.
Figure 5.1 also clearly illustrates the effect of sampling error. Namely, as the half-life is reduced
below about 15 weeks, the RAD increases dramatically while the variability of the risk forecast
also increases.
In Figure 5.2 we show the mean RAD and variability for the GEM2 currency factors. The results
were weighted by market capitalization within the GEM2 estimation universe. Qualitatively and
quantitatively, the results for GEM2 currency factors are very similar to those for equity factors.
For GEM2S, the 18-week volatility half-life minimizes the mean RAD at about 0.26, versus about
0.30 for GEM2L. The GEM2L forecasts, however, are more stable, with mean monthly variability
of 2.6 percent versus 6.6 percent for GEM2S. Figure 5.2 also demonstrates the increased
accuracy of risk forecasts that comes with treatment of autocorrelation.
Special care must be exercised when selecting the correlation half-life. To ensure a well-
conditioned correlation matrix, the number of observations T must be significantly greater than
the number of factors K . In the extreme case that T < K , the correlation matrix will have zero
eigenvalues. This can present problems for portfolio optimization, since it would be possible to
find spurious active portfolios with seemingly zero factor risk.
For the short model, we use a correlation half-life of 104 weeks. A useful rule of thumb is that the
number of effective observations is roughly three times the half-life. Therefore, a correlation half-
life of 104 weeks corresponds to roughly 300 effective observations. For the long model, we use
a correlation half-life of 156 weeks. The volatility and correlation half-lives used in GEM2 are the
same as those used in other Barra single-country models, such as USE3S and USE3L.
Another complication in estimating the factor covariance matrix arises for the case of missing
factor returns. One possible reason for missing data is holidays. For instance, stocks do not
trade in China during golden week, so it is not possible to directly estimate a return to the China
Domestic factor. More commonly, however, missing factor returns arise from using time series of
differing lengths. For instance, the GCC stocks appear in the estimation universe as of July
2000, so the factor returns prior to this date are missing.
We use the EM algorithm6 to estimate the factor covariance matrix in the presence of missing
factor returns. This method employs an iterative procedure to estimate the covariance matrix
6
Dempster, A.P., Laird, N.M., and D.B. Rubin, 1977. “Maximum Likelihood from Incomplete Data via the EM Algorithm.” Journal of the
Royal Statistical Society, Vol. 39, No. 1: 1-38.
conditional on all available data. The EM algorithm assures that the factor covariance matrix is
positive semi-definite.
The next step is to scale the correlation matrix with the factor volatilities to obtain the first-pass
covariance matrix. Here, different methods are used to estimate the volatilities of different
factors. For the US, Japan, and Sweden, the country factor volatilities are taken from a GARCH
model. The volatilities of the other equity factors are estimated using EWMA with half-lives
ranging from 36 months to 90 months. The currency volatilities, on other hand, are taken from a
separate fixed income risk model that uses highly responsive GARCH estimators or EWMA
based on daily data with a 24-day half-life.
At this stage of the GEM covariance matrix construction process, the currency volatilities match
those of the fixed income risk model, but the correlations do not. The fixed income risk model
uses an EWMA with a 17-week half-life to estimate the currency correlations. To ensure that the
forecasts of pure currency portfolios are the same in GEM and the fixed income risk model, a
transformation is applied to the factor covariance matrix.
The average kurtosis of monthly currency returns over the sample period – above 10 – is quite
large compared to that of typical equity factors. The range of kurtosis values is also very large.
On the one hand, many developed European countries have kurtosis values of 3 or less. By
contrast, the kurtosis of emerging-market currencies is typically much higher. The extreme case
is the Russian ruble, which exhibits a sample kurtosis of almost 60 over the sample period.
From Table 5.1, we see that on an equal-weighted basis, GEM2S had the most accurate
currency risk forecasts, with a mean RAD of 0.36. GEM2L had a mean RAD of 0.40, followed by
GEM, with a mean RAD of 0.43. On a cap-weighted basis, which is more relevant to most
international investors, GEM2S produces a mean RAD of 0.26, versus 0.30 for GEM2L and 0.32
for GEM.
When comparing these values, the reader should keep in mind the mean RAD for perfect risk
forecasts and normally distributed returns (kurtosis of 3) is about 0.17. If the returns have fat tails
(kurtosis greater than 3), then the mean RAD will be greater than 0.17. For instance, our
simulations indicate that for a kurtosis of 10, the mean RAD for perfect risk forecasts is about
0.23. Relative to the theoretical lower bound of perfect risk forecasts and stationary returns, a
mean RAD of 0.26 is quite impressive.
In Table 5.1 we also report the mean variability of the currency risk forecasts for the three
models. For GEM, which uses a highly responsive currency model, the mean monthly variability
of currency risk forecasts is 11.2 percent on a cap-weighted basis. GEM2S and GEM2L provide
more stable currency risk forecasts. The mean monthly variability for GEM2S is 6.6 percent and
for GEM2L it is 2.6 percent.
6. Specific Risk
6.1. GEM2 Specific Risk Model
GEM2 uses a structural model to estimate specific risk. This methodology is similar to that used
in other Barra risk models, but employs weekly rather than monthly observations. These higher-
frequency observations allow us to make the forecasts more responsive while also reducing
estimation error.
Mathematically, the specific risk forecast of a stock is given as the product of three components
( )
σ n = Sˆ 1 + Vˆn K M , (6.1)
where Ŝ is the forecast mean absolute specific return of stocks in the estimation universe, Vˆn is
the forecast relative absolute specific return, and K M is a multiplicative scalar which depends on
the market-cap segment M of the stock. The overall responsiveness of the model is dominated
by Ŝ , whereas Vˆn determines the cross-sectional variation in specific risk. The role of K M is to
convert from weekly absolute specific returns to monthly standard deviation. The motivation for
modeling specific risk in terms of absolute specific returns is that the estimation process is less
prone to outliers.
The realized mean absolute specific return for a given period t is computed as the cap-weighted
average,
where unt is the realized specific return for stock n . We use the cap-weighted average because
large-cap assets are less prone to extreme outliers. The forecast mean absolute specific return is
given by the exponentially weighted average of the realized values,
Sˆ = ∑ γ t St . (6.3)
t
For GEM2S, we use a half-life parameter of 9 weeks for γt , versus 24 weeks for GEM2L. To
determine these half-life parameters, we constructed a group of active portfolios, and then
computed their average forecast variability. We then selected the half-life parameters γt so that
the specific risk forecast variability of these portfolios roughly matched the variability in forecasts
of the factors, while also taking into account model performance. This results in highly accurate
specific risk forecasts and greater stability in the proportion of risk coming from factors and stock-
specific components.
In Figure 6.1, we plot the realized mean weekly absolute specific return over the period June
1992 to June 2008. In the mid-1990s, the absolute specific returns were quite low (about two
percent), but then began to rise, hitting a peak of nearly six percent in early 2000. After the
collapse of the internet bubble, absolute specific return levels fell sharply over the next five years.
Since 2007, however, specific risk levels have again begun to rise.
In Figure 6.1, we also plot the forecast values from GEM2S and GEM2L. Overall, both models do
an excellent job forecasting the absolute specific return levels. In the early history, up to 1997,
the forecasts are virtually identical. GEM2S, however, catches the upward trend more quickly
than GEM2L, and likewise responds more quickly when the volatility levels decrease following the
collapse of the internet bubble.
We use a factor model to forecast the relative absolute specific return, whose realized values are
given by
unt − St
ε nt = . (6.4)
St
GEM2 factors already provide a sound basis for explaining ε nt , since the level of relative
absolute specific risk depends on the country, industry, and style exposures of the individual
stocks. For instance, large-cap utility stocks tend to have lower specific returns than, say, small-
cap internet stocks.
While the model factors are well suited to forecast specific risk, the most powerful explanatory
variable is simply the trailing realized specific volatility. We therefore augment the GEM2 factor
exposure matrix with this additional factor for the purpose of forecasting the relative absolute
specific returns.
To estimate the relationship between factor exposures and relative absolute specific returns, we
perform a 104-week pooled cross-sectional regression,
~
ε nt = ∑ X nkt g k + λnt , (6.5)
k
~t
where X nk represents the augmented factor exposure matrix element. We use capitalization
weights in the regression to reduce the impact of high-kurtosis small-cap stocks, and apply
exponential weights of 26 weeks for the short model and 52 weeks for the long model. The
forecast relative absolute specific return is given by
~
Vˆn = ∑ X nk g k , (6.6)
k
~
where X nk is the most recent factor exposure, and g k is the slope coefficient estimated over the
trailing 104 weeks.
The product Sˆ (1 + Vˆn ) is the weekly forecast absolute specific return for the stock. We need to
convert this to a monthly standard deviation. If the specific returns were normally distributed
(temporally), then the time-series standard deviation of the ratio
u nt
(6.7)
S t (1 + Vˆnt )
would be π / 2 , where Vˆnt is the forecast relative absolute specific return at time t . However,
since the specific returns exhibit kurtosis, the standard deviation is usually slightly greater than
this theoretical value. For this reason, the multiplicative scalar K M is sometimes called the
kurtosis correction.
There is one final adjustment that must be applied to the kurtosis correction. The prediction
horizon of the risk model is one month. We observe, on average, a small but persistent negative
serial correlation in weekly specific returns. As a result, the monthly specific volatility will be
slightly less than that suggested by simple square root of time scaling. If we rank stocks by
degree of serial correlation, however, we find no relationship between the rank of stocks over
non-overlapping periods. We therefore define the standardized return with two-lag
autocorrelation,
u n ,t + u n ,t −1 + u n ,t −2
bnt = . (6.8)
3S (1 + Vˆ )
t nt
K M = σ (bnt ) , (6.9)
where the standard deviation is computed over all observations within a particular market-cap
segment M over the trailing 104 weeks.
If there are fewer than 24 months of returns, then the specific risk is estimated by regressing
historical sigma values against country, industry, and size factors. The slope coefficients produce
fits given the country, industry and size exposures of the stock.
In Panel A of Table 6.1, we consider the 66-month sample period from January 1997 to June
2002. This was a period of generally rising specific risk levels. Both GEM2S and GEM2L
perform extremely well over this period. The mean bias statistics are close to 1, indicating that
both models adapt quickly to rising volatility levels. For the cap-weighted case, over 85 percent
of observations in GEM2S fall within the confidence interval, versus about 82 percent for GEM2L.
On an equal-weighted basis, which over-represents small-cap stocks, the percentage falling
within the confidence interval is slightly lower, but still in the neighborhood of 80 percent.
GEM does not fare so well over the first sample period. The mean bias statistic is 1.21 on a cap-
weighted basis, and 1.32 on an equal-weighted basis, indicating that risk was significantly under-
predicted. The mean RAD is also significantly higher, e.g., 0.37 for GEM versus 0.24 for GEM2S
on a cap-weighted basis.
In Panel B of Table 6.1, we consider the 72-month sample period from July 2002 to June 2008.
This was a time of generally falling specific risk levels. The GEM2 bias statistics are still very
close to 1, indicating that GEM2 did an excellent job adapting to new specific risk levels. GEM,
by contrast, exhibits a mean bias statistic of 0.80 for the cap-weighted case, indicating that
specific risk was over-predicted during this period. The GEM2 models also have a higher
percentage of observations falling within the confidence interval than GEM.
In Panel C of Table 6.1, we examine the entire 138-month sample period. The results are
consistent with Panels A and B. Namely, GEM2S and GEM2L have mean bias statistics very
close to 1, have small RAD compared to GEM, and have a greater percentage of observations
falling within the confidence interval.
In Figure 6.2, we plot the frequency distribution of the rolling 12-month bias statistics for the
GEM2 and GEM models. The results are cap-weighted, and the analysis period covers 138
months from January 1997 to June 2008. The confidence interval is indicated by the two vertical
lines. The performance of the GEM2S and GEM2L specific risk models is evidently very similar.
The GEM2S distribution is slightly more peaked in the center, and has slightly smaller tails than
the GEM2L distribution.
7. Model Performance
In this section, we evaluate the accuracy of risk forecasts for a combination of country, industry
and style portfolios. The country and industry portfolios were constructed by carving out the
constituents of the GEM2 estimation universe from the appropriate segment, and then cap-
weighting the stocks. The style portfolios were constructed by ranking all stocks according to
their exposures to the particular style factor, and then cap-weighting the stocks that ranked in the
top and bottom 20 percent. For each of the long-only country, industry, and style portfolios thus
formed, we also construct long/short active portfolios by selecting MSCI ACWI IMI as the
benchmark. The sample period in all cases was from January 1997 to June 2008.
In Table 7.1, we report the 12-month RAD and sample kurtosis for long-only country portfolios
and active portfolios. The average kurtosis was about 4.5 for both sets of portfolios. Overall, the
GEM2S model produces the most accurate risk forecasts. The 12-month mean RAD for the
country portfolios was 0.23, versus 0.25 for GEM2L, and 0.29 for GEM. For the active portfolios,
the differences were even more pronounced. The mean RAD for GEM2S was 0.23, versus 0.26
GEM2L and 0.33 for GEM.
The GEM mean RAD is more than 40 percent higher than the mean RAD of GEM2S. This
difference is quite significant, especially when one considers the theoretical lower bound.
Assuming normally distributed returns, the mean RAD is about 0.17 for perfect risk forecasts. For
a kurtosis of 4.5, however, the mean RAD is about 0.20 for perfect risk forecasts. By this
measure, the GEM2S model has a mean RAD only 0.03 above the theoretical lower bound,
versus 0.13 for GEM.
In Table 7.2, we report the 12-month RAD and sample kurtosis values for the industry portfolios.
Industry factor kurtosis levels are slightly below those of country factors. Similar to the case for
the country portfolios, the GEM2S model produces the most accurate risk forecasts, followed by
GEM2L, and then GEM. For instance, for active industry portfolios, the mean RAD for GEM2S is
0.24, versus 0.27 for GEM2L, and 0.38 for GEM. It is interesting to note that GEM fares
particularly poorly in the Biotechnology and Internet industries, suggesting that the older GEM
industry structure is not well suited to capture the risk of these newer industries.
In Table 7.3, we report the 12-month RAD and sample kurtosis values for the style portfolios.
Note that the kurtosis levels for the active portfolios are considerably higher than those for the
long-only portfolios. Once again, the GEM2S model produces the most accurate risk forecasts,
followed by GEM2L and GEM. For example, the mean RAD of the active style portfolios is 0.28
for GEM2S, 0.31 for GEM2L, and 0.44 for GEM.
In Table 7.4, we report summary results of 12-month rolling bias statistics for the complete set of
country, industry, and style portfolios from Tables 7.1, 7.2, and 7.3. The sample period is divided
into two sub-periods. Sub-period A is from January 1997 to June 2002, and is a time of generally
rising volatility. All risk models slightly underpredict risk during sub-period A, although GEM2S
underpredicts by a smaller margin than either GEM2L or GEM. GEM2S and GEM2L have
smaller mean RAD than GEM, and a higher percentage of observations falling within the
confidence interval.
Sub-period B runs from July 2002 to June 2008, and generally corresponds to a falling volatility
environment. From Table 7.4, we see that the mean bias statistics for GEM2S are very close to
1, indicating that the model is able to quickly adjust to the lower volatility levels. GEM2L also
performed reasonably well, with mean bias statistics of 0.89 for both long-only and active
portfolios. The mean bias statistics for GEM during this period were 0.78 for long-only portfolios
and 0.70 for active portfolios, indicating that the relatively unresponsive GEM is slow to adjust to
the reduced volatility environment.
The entire sample period, running from January 1997 to June 2008, is shown in Panel C.
GEM2S has the greatest proportion of observations (above 86 percent) falling within the
confidence interval. The GEM2L model also performs well, with about 81 percent of the
observations fall within the confidence interval. For GEM, only about 66 percent of the
observations for active portfolios fall within the confidence interval. Most of the cases falling
outside correspond to over-forecasting of risk during sub-period B, and are the result of attaching
too much weight to the highly volatile events surrounding the internet bubble period.
In Figure 7.1, we plot the frequency distribution of the 12-month rolling bias statistics for the
complete set of country, industry, and style active portfolios. The sample period was January
1997 to June 2008. The vertical lines denote the approximate confidence interval assuming
normally distributed returns. This graph clearly illustrates that the GEM2S bias statistic
distribution is more centered about the ideal value of 1, and has thinner tails. The GEM2L
distribution is shifted somewhat to the left, and has slightly broader tails. The GEM distribution is
shifted significantly more to the left and has fatter tails on both sides of the distribution.
8. Conclusion
GEM was the pioneering global equity factor model. It has helped managers measure the risk of
their portfolios, understand the sources of the risk and construct better portfolios. GEM2 applies
the methodological advances and greater data availability to produce a new standard in global
equity modeling.
GEM2 contains many improvements over GEM. In this document, we have presented many of
the methodological innovations and enhancements in the new GEM2 model. The highlights are:
• Use of weekly factor returns, allowing us to build a more responsive and accurate model
• A cleaner separation of country and industry effects through the introduction of the World
factor
These enhancements, combined with painstaking data and analytical quality checks, make GEM2
a great advance in global equity risk modeling.
Figure 3.1
Number of stocks versus time, for MSCI ACWI IMI, GEM2 ESTU, and MSCI World.
16000
MSCI
14000 ACWI
IMI
12000
Number of Stocks
10000
GEM2
8000 ESTU
6000
4000
MSCI
2000 World
0
1992 1994 1996 1998 2000 2002 2004 2006 2008
Year
Table 3.1
GEM2 country factors and currencies. Weights are computed within the GEM2 estimation
universe using total market capitalization. Average is taken over the period from January 1997 to
January 2008.
Table 3.2
GEM2 Industry Factors. Weights are computed within the GEM2 estimation universe using total
market capitalization. Average is taken over the period from January 1997 to January 2008.
Table 3.3
Mapping of GEM2 industry factors to GICS codes.
Table 3.4
Mapping from GEM2 Industries to GEM Industries, as of January 2008. Mapping is based on
capitalization weights using common assets between the GEM2 estimation universe and the
GEM coverage universe. Mapped weights below 10-percent threshold are not shown.
GEM2 GEM Mapped
Code GEM2 Industry Factor Name Code GEM Industry Factor Name Weight
1 Energy Equipment & Services 14 Energy Equipment & Services 97
2 Oil, Gas & Consumable Fuels 1 Energy Sources 98
3 Oil & Gas Exploration & Production 1 Energy Sources 100
4 Chemicals 4 Chemicals 98
5 Construction, Containers, Paper 3 Building Materials & Components 54
5 5 Forestry and Paper Products 35
6 Aluminum, Diversified Metals 6 Metals - Non-Ferrous 55
6 7 Metals – Steel 16
6 38 Gold Mines 12
6 1 Energy Sources 10
7 Gold, Precious Metals 38 Gold Mines 63
7 6 Metals - Non-Ferrous 37
8 Steel 7 Metals – Steel 98
9 Capital Goods 37 Multi-Industry 24
9 16 Machinery & Engineering 24
9 9 Aerospace & Military Technology 13
9 10 Construction & Housing 13
9 12 Electrical & Electronics 10
10 Commercial & Professional Services 25 Business & Public Services 99
11 Transportation Non-Airline 30 Transportation – Road & Rail 50
11 25 Business & Public Services 24
11 31 Transportation – Shipping 23
12 Airlines 29 Transportation - Airlines 100
13 Automobiles & Components 18 Automobiles 69
13 15 Industrial Components 27
14 Consumer Durables & Apparel 17 Appliances & Household Durables 27
14 22 Recreation & Other Consumer Goods 24
14 23 Textiles & Apparel 22
14 10 Construction & Housing 13
15 Consumer Services 26 Leisure & Tourism 90
15 25 Business & Public Services 10
16 Media 24 Broadcasting & Publishing 91
17 Retailing 27 Merchandizing 84
18 Food & Staples Retailing 27 Merchandizing 100
19 Food, Beverage & Tobacco 19 Beverages & Tobacco 56
19 20 Food & Household Products 44
20 Household & Personal Products 20 Food & Household Products 70
20 17 Appliances & Household Durables 18
20 21 Health and Personal Care 10
21 Health Care Equipment & Services 21 Health and Personal Care 73
21 25 Business & Public Services 27
22 Biotechnology 21 Health and Personal Care 99
23 Pharmaceuticals, Life Sciences 21 Health and Personal Care 100
24 Banks 33 Banking 100
25 Diversified Financials 34 Financial Services 98
26 Insurance 35 Insurance 100
27 Real Estate 36 Real Estate 95
28 Internet Software & Services 25 Business & Public Services 95
29 IT Services, Software 25 Business & Public Services 99
30 Communications Equipment 28 Telecommunication 98
31 Computers, Electronics 13 Electrical Components & Instruments 58
31 11 Data Processing & Reproduction 27
31 25 Business & Public Services 11
32 Semiconductors 13 Electronic Components & Instruments 100
33 Telecommunication Services 28 Telecommunication 100
34 Utilities 2 Utilities-Electrical & Gas 94
Table 3.5
Mapping from GEM Industries to GEM2 Industries, as of January 2008. Mapping is based on
capitalization weights using common assets between the GEM2 estimation universe and the
GEM coverage universe. Mapped weights below 10-percent threshold are not shown.
GEM ESTU GEM2 Mapped
Code GEM Industry Name Weight Code GEM2 Industry Name Weight
1 Energy Sources 10.60 2 Oil, Gas & Consumable Fuels 82
1 3 Oil & Gas Exploration & Production 16
2 Utilities - Electrical & Gas 5.07 34 Utilities 99
3 Building Materials & Components 0.96 5 Construction, Containers, Paper 64
3 9 Capital Goods 34
4 Chemicals 2.78 4 Chemicals 98
5 Forestry and Paper Products 0.35 5 Construction, Containers, Paper 99
6 Metals - Non Ferrous 1.53 6 Aluminum, Diversified Metals 84
6 7 Gold, Precious Metals 13
7 Metals - Steel 2.02 8 Steel 81
7 6 Aluminum, Diversified Metals 18
8 Misc. Materials & Commodities 0.13 6 Aluminum, Diversified Metals 39
8 5 Construction, Containers, Paper 35
9 Aerospace & Military Technology 1.26 9 Capital Goods 100
10 Construction & Housing 1.28 9 Capital Goods 79
10 14 Consumer Durables & Apparel 17
11 Data Processing & Reproduction 0.99 31 Computers, Electronics 97
12 Electrical & Electronics 0.91 9 Capital Goods 91
13 Electrical Components & Instruments 3.31 31 Computers, Electronics 48
13 32 Semiconductors 45
14 Energy Equipment & Services 1.35 1 Energy Equipment & Services 91
15 Industrial Components 0.53 13 Automobiles & Components 88
16 Machinery & Engineering 1.80 9 Capital Goods 99
17 Appliances & Household Durables 0.76 14 Consumer Durables & Apparel 62
17 20 Household & Personal Products 27
18 Automobiles 1.89 13 Automobiles & Components 95
19 Beverages & Tobacco 2.66 19 Food, Beverage and Tobacco 100
20 Food & Household Products 2.94 19 Food, Beverage and Tobacco 69
20 20 Household & Personal Products 30
21 Health & Personal Care 6.30 23 Pharmaceuticals & Life Sciences 63
21 21 Health Care Equipment & Services 23
21 22 Biotechnology 11
22 Recreation & Other Consumer Goods 0.56 14 Consumer Durables & Apparel 86
22 13 Automobiles & Components 12
23 Textiles & Apparel 0.29 14 Consumer Durables & Apparel 90
24 Broadcasting & Publishing 2.08 16 Media 99
25 Business & Public Services 6.01 29 IT Services, Software 42
25 10 Commercial & Professional Services 13
25 28 Internet Software & Services 12
26 Leisure & Tourism 1.16 15 Consumer Services 98
27 Merchandising 3.67 18 Food & Staples Retailing 49
27 17 Retailing 49
28 Telecommunications 8.45 33 Telecommunication Services 80
28 30 Communications Equipment 20
29 Transportation - Airlines 0.36 12 Airlines 91
30 Transportation - Road & Rail 1.01 11 Transportation Non-Airline 99
31 Transportation - Shipping 0.40 11 Transportation Non-Airline 89
32 Wholesale & International Trade 0.62 9 Capital Goods 78
32 17 Retailing 19
33 Banking 11.19 24 Banks 100
34 Financial Services 5.23 25 Diversified Financials 100
35 Insurance 3.95 26 Insurance 100
36 Real Estate 2.40 27 Real Estate 99
37 Multi-Industry 2.46 9 Capital Goods 95
38 Gold Mines 0.75 7 Gold, Precious Metals 56
38 6 Aluminum, Diversified Metals 44
Table 4.1
Segment weights, as of January 2008, for the World portfolio and several pure factor portfolios.
The World portfolio is represented by the cap-weighted GEM2 estimation universe.
Table 4.2
Regression-weighted cross-sectional correlation of style and industry factor exposures. Results
are averages over the period January 1997 to June 2008. Correlations above 0.10 in absolute
value are shaded in gray.
Table 4.3
Country factor summary statistics based on weekly factor returns. Sample period is from January
1997 to June 2008. Averages computed excluding World, China Domestic, and GCC factors.
Table 4.4
Industry factor summary statistics based on weekly factor returns. Sample period is from January
1997 to June 2008.
Table 4.5
Style factor summary statistics based on weekly factor returns. The sample period, January 1997
to June 2008, is divided into two sub-periods, A and B.
Table 4.6
2
Explanatory power of GEM2 factors versus GEM. Results are averages of monthly R values
over the sample period January 1997 to June 2008. Notation: G2 indicates GEM2 factors, G
indicates GEM factors with constraints (Equation 4.8 of text), and GU indicates GEM factors
unconstrained. Furthermore, W, C, I, and S denote World, country, industry, and style factors
respectively; Vol denotes the GEM2 Volatility factor, and VIM denotes the GEM Variability in
Markets factor.
Case 1 compares the explanatory power of GEM2 country factors versus GEM with constraints.
2
Case 2 shows the increase in R explained by countries when the GEM constraints are
removed. Case 3 compares the explanatory power of GEM2 and GEM country factors in
conjunction with the Volatility and VIM factors. Cases 4 and 5 demonstrate the superiority of the
GEM2 style factors and industry factors, respectively. Case 6 demonstrates the increased
explanatory power of GEM2 versus an unconstrained GEM model. The final case shows the
added explanatory power of the actual GEM2 model versus the actual GEM model.
Figure 4.1
2
Trailing 12-month R for GEM2 and GEM models.
0.5
0.4 GEM2
2
12-month Trailing R
0.3
0.2
GEM
0.1
0.0
1994 1996 1998 2000 2002 2004 2006 2008
Year
Figure 4.2
2
Explanatory power of GEM2 factor groups, expressed as trailing 12-month R . Here, W
indicates the World factor in isolation, W/C denotes World factor plus countries, and W/C/I
2
represents World, country, and industry factors. The black dots indicate the actual monthly R
values.
0.8
All
2
0.6
12-month Trailing R
Factors
W/C/I
0.4
0.2
W/C
W
0.0
1994 1996 1998 2000 2002 2004 2006 2008
Year
Figure 5.1
Mean 12-month rolling absolute deviation (RAD) and mean Variability (v ) of GEM2 equity
factors, plotted as a function of volatility half-life. The sample period was 138 months from
January 1997 to June 2008. The solid lines indicate results with the two-lag autocorrelation that
is used GEM2, and the dashed lines are for no autocorrelation adjustment. The vertical lines
indicate the actual half-life parameters for GEM2S and GEM2L.
0.5
GEM2S GEM2L
0.4
RAD, Variability
0.3
Mean
0.2 2 Lags
RAD
0 Lags
0.1
v
0.0
0 10 20 30 40 50 60 70 80 90 100
Half-Life (weeks)
Figure 5.2
Mean 12-month rolling absolute deviation (RAD) and mean Variability (v ) of GEM2 currency
factors, plotted as a function of volatility half-life. Results are cap-weighted averages as of June
2008. Sample period was 138 months from January 1997 to June 2008. The solid lines indicate
results with the two-lag autocorrelation used in GEM2, and the dashed lines are for no
autocorrelation adjustment. The vertical lines indicate the actual half-life parameters for GEM2S
and GEM2L.
0.5
GEM2S GEM2L
0.4
RAD, Variability
0.3
Mean
0.2 RAD
2 Lags
0 Lags
0.1
v
0.0
0 10 20 30 40 50 60 70 80 90 100
Half-Life (weeks)
Table 5.1
Currency 12-month RAD and forecast variability for GEM2S, GEM2L, and GEM models. The
sample period contains 138 months from January 1997 to June 2008. Also reported is the
sample monthly kurtosis. Monthly standardized returns were trimmed at 10σ , and monthly
variability was trimmed at 2. Trimming primarily affected only the GEM reported values. Results
are quoted from a US Dollar numeraire perspective.
Figure 6.1
Plot of weekly realized and forecast mean absolute specific returns for the GEM2 estimation
universe.
7
Mean Absolute Specific Returns (%)
6 Realized
GEM2S
5 GEM2L
0
1992 1994 1996 1998 2000 2002 2004 2006 2008
Year
Table 6.1
Specific risk bias statistic results, based on 12-month rolling windows. Results were computed on
the GEM2 estimation universe, and aggregated on cap-weighted and equal-weighted bases. The
full sample period, January 1997 to June 2008, is divided into two sub-periods, A and B. The 12-
month confidence interval is taken as [0.59, 1.41].
Figure 6.2
Frequency distribution of 12-month rolling bias statistics for specific risk forecasts. Sample period
is January 1997 to June 2008. Results are cap-weighted within the GEM2 estimation universe.
Vertical lines indicate confidence interval [0.59, 1.41].
0.10
GEM2L
0.08 GEM2S
Probability Distribution
0.06 GEM
0.04
0.02
0.00
0.0 0.5 1.0 1.5 2.0 2.5 3.0
Rolling 12m Bias Statistic
Table 7.1
Country 12-month RAD and monthly kurtosis values for sample period January 1997 to June
2008. Portfolios are cap-weighted carve-outs of GEM2 estimation universe. Active portfolios are
obtained by running portfolios against MSCI ACWI IMI as the benchmark.
Kurtosis GEM2S GEM2L GEM
Portfolio Name Portfolio Active Portfolio Active Portfolio Active Portfolio Active
Argentina 4.55 4.77 0.32 0.31 0.30 0.30 0.38 0.38
Australia 3.04 3.48 0.16 0.19 0.20 0.22 0.20 0.31
Austria 2.96 3.39 0.20 0.23 0.22 0.26 0.24 0.31
Bahrain 3.78 2.58 0.16 0.13 0.19 0.14 0.12 0.16
Belgium 4.42 3.56 0.18 0.24 0.23 0.28 0.29 0.34
Brazil 4.26 4.60 0.38 0.33 0.35 0.32 0.39 0.48
Canada 4.61 2.85 0.24 0.22 0.26 0.21 0.29 0.24
Chile 4.73 3.71 0.22 0.21 0.25 0.25 0.25 0.30
China International 5.61 6.71 0.24 0.27 0.32 0.35 0.51 0.59
Colombia 3.18 3.04 0.23 0.25 0.23 0.25 0.27 0.27
Czech Republic 3.77 3.84 0.19 0.19 0.20 0.24 0.29 0.32
Denmark 3.07 2.59 0.16 0.19 0.21 0.23 0.20 0.25
Egypt 4.66 5.14 0.30 0.27 0.27 0.24 0.29 0.26
Finland 4.22 4.60 0.19 0.20 0.22 0.21 0.28 0.37
France 3.72 3.86 0.21 0.21 0.27 0.24 0.26 0.32
Germany 5.49 6.54 0.23 0.24 0.29 0.26 0.28 0.38
Greece 3.95 4.64 0.22 0.23 0.29 0.34 0.29 0.41
Hong Kong 5.67 6.94 0.18 0.26 0.27 0.31 0.37 0.45
Hungary 5.40 4.82 0.21 0.21 0.21 0.21 0.27 0.33
India 2.62 2.94 0.20 0.17 0.17 0.17 0.19 0.26
Indonesia 4.58 4.36 0.20 0.22 0.28 0.29 0.33 0.33
Ireland 3.80 3.03 0.18 0.19 0.21 0.21 0.23 0.30
Israel 3.78 4.98 0.15 0.23 0.19 0.27 0.30 0.32
Italy 4.14 5.34 0.18 0.25 0.22 0.29 0.25 0.42
Japan 2.81 3.72 0.22 0.23 0.23 0.25 0.23 0.23
Jordan 5.99 4.53 0.26 0.24 0.25 0.24 0.29 0.22
Korea 6.62 10.28 0.21 0.27 0.31 0.39 0.41 0.49
Kuwait 3.34 3.46 0.13 0.23 0.15 0.23
Malaysia 8.15 6.98 0.21 0.18 0.29 0.30 0.47 0.46
Mexico 6.16 4.53 0.24 0.26 0.23 0.27 0.30 0.34
Morocco 4.23 4.62 0.27 0.27 0.26 0.26 0.27 0.26
Netherland 3.91 4.47 0.22 0.30 0.30 0.32 0.36 0.36
New Zealand 3.64 3.25 0.19 0.21 0.22 0.23 0.24 0.23
Norway 4.42 2.93 0.26 0.29 0.26 0.30 0.24 0.28
Oman 3.62 3.41 0.28 0.34 0.31 0.35 0.24 0.36
Pakistan 4.54 4.96 0.23 0.25 0.23 0.27 0.27 0.30
Peru 4.77 3.44 0.33 0.23 0.30 0.21 0.38 0.40
Philippines 7.25 6.11 0.19 0.17 0.26 0.20 0.32 0.26
Poland 4.36 3.47 0.20 0.21 0.14 0.15 0.16 0.21
Portugal 3.25 2.98 0.21 0.18 0.27 0.24 0.31 0.35
Qatar 7.09 6.22 0.32 0.34 0.34 0.40
Russia 7.95 7.08 0.22 0.21 0.22 0.20 0.42 0.45
Saudi Arabia 3.02 3.04 0.25 0.19 0.27 0.21 0.31 0.27
Singapore 5.31 6.07 0.26 0.21 0.32 0.31 0.36 0.41
South Africa 4.80 4.55 0.26 0.24 0.25 0.26 0.22 0.30
Spain 3.48 2.81 0.17 0.16 0.23 0.20 0.25 0.27
Sweden 3.61 4.21 0.20 0.21 0.24 0.21 0.23 0.26
Switzerland 4.28 3.41 0.23 0.21 0.27 0.26 0.29 0.34
Taiwan 3.80 4.59 0.20 0.22 0.26 0.26 0.24 0.29
Thailand 4.26 4.51 0.24 0.26 0.24 0.25 0.31 0.34
Turkey 5.09 5.29 0.28 0.23 0.25 0.22 0.24 0.27
UAE 5.88 6.50 0.38 0.31 0.39 0.37
UK 2.96 3.10 0.17 0.26 0.20 0.28 0.23 0.35
US 3.48 3.43 0.21 0.25 0.28 0.27 0.28 0.26
Average (Equal Wt) 4.48 4.45 0.23 0.23 0.25 0.26 0.29 0.33
Table 7.2
Industry 12-month RAD and monthly kurtosis values for sample period January 1997 to June
2008. Portfolios are cap-weighted carve-outs of GEM2 estimation universe. Active portfolios are
obtained by running portfolios against MSCI ACWI IMI as the benchmark.
Table 7.3
Style 12-month RAD and monthly kurtosis values for sample period January 1997 to June 2008.
Portfolios are obtained by cap-weighting the 20 percent of stocks with maximum and minimum
exposure to the factor. Active portfolios are obtained by running portfolios against MSCI ACWI
IMI as the benchmark.
Table 7.4
Summary 12-month rolling bias statistics results for country, industry and style portfolios
contained in Tables 7.1, 7.2, and 7.3. The full sample period, January 1997 to June 2008, is
divided into two sub-periods, A and B.
Figure 7.1
Frequency distribution of 12-month rolling bias statistics for country, industry and style active
portfolios (long/short) contained in Tables 7.1, 7.2, and 7.3. Sample period is January 1997 to
June 2008. Vertical lines indicate the confidence interval [0.59, 1.41].
0.10
GEM2L
GEM2S
0.08
Probability Distribution
0.06 GEM
0.04
0.02
0.00
0.0 0.5 1.0 1.5 2.0 2.5 3.0
Rolling 12m Bias Statistic
T T
Z (T ) = ∑ ln(1 + rnt ) − ∑ ln(1 + rft ) . (A3)
t =1 t =1
where rnt is the local return of stock n for month t , and r ft is the risk-free
return of the local currency. The cumulative range is given by
6 6
RSTR 6 = ∑ ln(1 + rnt ) − ∑ ln(1 + r ft ) , (A5)
t =1 t =1
with a one-month lag. Here, rnt is the local return of stock n for month t , and
r ft is the risk-free return of the local currency.
12 12
RSTR12 = ∑ ln(1 + rnt ) − ∑ ln(1 + r ft ) , (A6)
t =1 t =1
⎛V ⎞
STOM = log⎜⎜ t ⎟⎟ , (A7)
⎝ St ⎠
where Vt is the trading volume of the asset for the month, and St is the
corresponding number of shares outstanding.
⎛1 V ⎞
STOQ = log⎜⎜ ∑ t ⎟⎟ , (A8)
⎝ 3 t St ⎠
where Vt is the trading volume of the asset for month t , and St is the number of
shares outstanding. The sum runs over the past 3 months.
⎛1 V ⎞
STOA = log⎜⎜ ∑ t ⎟⎟ , (A9)
⎝ 12 t S t ⎠
where Vt is the trading volume of the asset for month t , and St is the number of
shares outstanding. The sum runs over the past 12 months.
MCAP + PREF + LD
MLEV = , (A10)
MCAP
where MCAP is the market value of common equity at previous month-end,
PREF is the most recent book value of preferred equity, and LD is the most
recent book value of long-term debt. For details, refer to Section 5 of MSCI
Barra Fundamental Data Methodology.
BV + PREF + LD
BLEV = , (A11)
BV
where BV is the most recent book value of common equity, PREF is the most
recent book value of preferred equity, and LD is the most recent book value of
long-term debt. For details, refer to Section 5 of MSCI Barra Fundamental Data
Methodology.
DTOA Debt-to-assets
Computed as
TD
DTOA = , (A12)
TA
where TD is the book value of total debt (long-term debt, LD , and current
liabilities, CL ), and TA is most recent book value of total assets. For details,
refer to Section 5 of MSCI Barra Fundamental Data Methodology.
A commonly used measure to assess a risk model’s accuracy is the bias statistic. The bias
statistic, conceptually, represents the ratio of realized risk to forecast risk.
Let rnt be the return to portfolio n over period t , and let σ nt be the beginning-of-period volatility
forecast. Assuming perfect forecasts, the standardized return,
rnt
bnt = , (B1)
σ nt
has expected standard deviation 1. The bias statistic for portfolio n is the realized standard
deviation of standardized returns,
Bn =
1 T
∑ (bnt − bn )2 , (B2)
T − 1 t =1
where T is the number of periods in the observation window.
Assuming normally distributed returns and perfect risk forecasts, for sufficiently large T the bias
statistic Bn is approximately normally distributed about 1, and roughly 95 percent of the
observations fall within the confidence interval,
[
Bn ∈ 1 − 2 / T , 1 + 2 / T . ] (B3)
If Bn falls outside this interval, then we reject the null hypothesis that the risk forecast was
accurate.
If returns are not normally distributed, however, then fewer than 95 percent of the observations
will fall within the confidence interval, even for perfect risk forecasts. In Figure B1, we show
simulated results for the percentage of observations actually falling within this interval, plotted
versus observation-window length T , for several values of kurtosis k .
For the normal case (kurtosis k = 3 ), except for the smallest values of T , the confidence interval
indeed captures about 95 percent of the observations. As the kurtosis increases, however, the
percentage falling within the interval drops significantly. For instance, even for a fairly modest
kurtosis level of 5, only 86 percent of bias statistics fall inside the confidence interval for an
observation window of 120 periods.
A more relevant question is to study how accurate the risk forecasts were during any 12-month
period. For this purpose, we define the rolling 12-month bias statistic for portfolio n ,
1 τ +11
Bnτ = ∑ (bnt − bn ) 2 ,
11 t =τ
(B4)
where τ denotes the first month of the 12-month window. The 12-month windows are rolled
forward one month at a time until reaching the end of the observation window. If T is the
number of periods in the observation window, then each portfolio will have T − 11 (overlapping)
12-month windows. It is often informative to consider the frequency distribution of rolling 12-
month bias statistics. These histograms quickly indicate whether there were any 12-month
periods for which portfolio risk was significantly either over-forecast or under-forecast. Figure 7.1
is an example of such a distribution. If there are N portfolios, then the histogram will contain
N (T − 11) data points, resulting in fairly smooth distributions.
It is also useful to consider the mean of the rolling 12-month bias statistics, given by
~ 1
B= ∑
N (T − 11) n ,τ
Bnτ . (B5)
This number is reported in Table 6.1 and Table 7.4, and indicates whether, on average, risk was
over-forecast or under-forecast during the observation window.
Again, it is possible to over-forecast the risk of some portfolios within certain time periods, and to
under-forecast the risk of other portfolios during different time periods, while producing a mean
~
rolling bias statistic B close to 1. What is needed is a way to penalize every deviation away from
the ideal bias statistic of 1. The 12-month rolling absolute deviation (RAD) for portfolio n ,
defined as
1 T −11 τ
RAD n = ∑ Bn − 1 ,
T − 11 τ =1
(B6)
captures this effect. Smaller RAD numbers, of course, are preferable to larger ones. To assess
whether an observed RAD is statistically significant or not, we must consider the properties of the
RAD distribution.
In Figure B2, we plot the probability distributions for 12-month RAD over observation windows of
12, 36, and 150 months. Results were obtained via numerical simulation using normally
distributed returns. The plot labeled by T = 12 corresponds to a single period. The most likely
RAD occurs near zero in this case, and the mean of the distribution is approximately
1 / 12π = 0.163 . A more precise computation, which accounts for small sample sizes, gives a
mean of 0.17.
Increasing the observation window to 150 months leads to more than 12 completely independent
observations. In this case, the distribution becomes narrower, and the peak shifts further to the
right. As T increases further still, the central limit theorem assures that the distribution becomes
more sharply peaked about the mean value 0.17, with an approximately normal distribution.
~C
In Figure B2, we indicate by vertical lines the critical mean RAD values, denoted QT , for rolling
12-month windows. These represent the 95th percentile of the cumulative RAD distributions. In
other words, assuming perfect risk forecasts and normally distributed returns, over a 36-month
observation window we would expect to find an RAD above 0.30 only five percent of the time. If
we observe an RAD above this value, we reject the null hypothesis that the risk forecasts were
accurate.
In reality, of course, returns are not normally distributed. In Figure B3, we plot the critical RAD
versus size of observation window T , for various levels of kurtosis. Not surprisingly, the normal
distribution ( k = 3) has the smallest critical values. One useful case to consider is the T → ∞
limit, which converges to a critical value of 0.17 for kurtosis k = 3 . As kurtosis increases,
however, the critical values also increase. For instance, using the same 36-month observation
window as before but now assuming a kurtosis k = 5 , RAD values in excess of 0.38 would occur
about five percent of the time.
Often we are interested in the mean RAD of a collection of portfolios. This is defined as
where wn is the weight given to observation n . Usually we report equal-weighted mean RAD,
as in Figure 5.1 or Table 7.4, although sometimes cap-weighted mean RAD are also reported
(e.g., Figure 5.2 or Table 6.1).
Figure B1
Percent of observations falling within confidence interval [1 − 2 / T , 1 + 2 / T ] , where T is the
number of periods in the observation window. Results were simulated using a normal distribution
(k = 3) , and using a t-distribution with kurtosis values k = 5 and k = 10 . The standard
deviations were equal to 1 in all cases.
For the normal distribution, the percentage of observations inside the confidence interval quickly
approaches 95 percent. As kurtosis is increased, however, the proportion within the confidence
interval declines considerably.
Simulated Results
100
Percent in Confidence Interval
k=3 (Normal)
95
90
k=5
85
80 k=10
75
0 20 40 60 80 100 120 140
Number of Periods, T
Figure B2
Probability distribution of 12-month rolling absolute deviation (RAD). Results were simulated
using normally distributed returns, for observation windows of 12, 36, and 150 months.
Since a 12-month rolling window within a 12-month observation window corresponds to a single
fixed window, the curve labeled T = 12 is peaked near zero. A 36-month observation window,
however, corresponds to roughly three independent observations, and the peak of the distribution
shifts to the right. As the number of periods T becomes large, the central limit theorem assures
that the distribution converges to a normal distribution centered at approximately 0.17.
~C
Critical values QT are indicated by the vertical lines. They indicate the 95th percentile of the
cumulative distribution.
Simulated Results
0.14
T=150
~C
0.12 Q150
Probability Distribution
0.10
T=36 ~ ~
0.08 Q36C Q12C
0.06
T=12
0.04
0.02
0.00
0.0 0.1 0.2 0.3 0.4 0.5 0.6
Rolling Absolute Deviation (12m windows)
Figure B3
Critical values of rolling absolute deviation (RAD), for rolling 12-month windows, versus the
number of periods in the observation window. For perfect risk forecasts and normally distributed
returns, over a 60-period observation window we expect to observe an RAD above 0.26 only five
percent of the time. For a kurtosis of 10, however, the critical value at 60 periods rises to about
0.40.
Simulated Results
0.5
0.4
~ k=10
QTC
0.3
k=5
0.1
0 20 40 60 80 100 120 140
Number of Periods, T
Contact Information
[email protected]
Americas
Amsterdam + 31.20.462.1382
Cape Town + 27.21.683.3245
Frankfurt + 49.69.2166.5325
Geneva + 41.22.817.9800
London + 44.20.7618.2222
Madrid + 34.91.700.7275
Milan + 39.027.633.5429
Paris 0800.91.59.17 (toll free)
Zurich + 41.44.220.9300
Asia Pacific
www.mscibarra.com
MSCI Barra is a leading provider of investment decision support tools to investment institutions worldwide. MSCI Barra
products include indices and portfolio risk and performance analytics for use in managing equity, fixed income and multi-
asset class portfolios.
The company’s flagship products are the MSCI International Equity Indices, which are estimated to have over USD 3
trillion benchmarked to them, and the Barra risk models and portfolio analytics, which cover 56 equity and 46 fixed income
markets. MSCI Barra is headquartered in New York, with research and commercial offices around the world. Morgan
Stanley, a global financial services firm, is the controlling shareholder of MSCI Barra.