Documente Academic
Documente Profesional
Documente Cultură
Paper No. 45
Introducing GS-ESS:
A New Framework for Assessing Fair Value
in Emerging Markets Hard-Currency Debt
Spreads in hard currency sovereign debt experienced significant volatility in the past
3 years, often deviating from a countrys true ability to pay. This situation will likely
persist in the future, creating attractive investment opportunities.
GS-ESS can help identify those opportunities by modeling a sovereigns fair value
spread as a function of solvency and liquidity variables.
In London
Linda Britten, E.D. & Global Economics Mgr, Support & Systems
In Toronto
In Hong Kong
Sun Bae Kim, M.D. & Director of Asia Pacific Economic Research
In Paris
In Frankfurt
In Singapore
Goldman Sachs Research personnel may be contacted by electronic mail through the Internet at firstname.lastname@gs.com
Main Points
We introduce Goldman Sachs Equilibrium Sovereign Spread (GSESS), our preferred model of long-run equilibrium sovereign spreads
in emerging market economies. In the same spirit as GSDEEMER
and GS-DERBY, GS-ESS develops a systematic guide to valuing
emerging market sovereign spreads over the long run.
Our results indicate that out of the 15 countries considered, 2 are now
trading rich to fair value, 12 are undervalued, and one country close
enough to equilibrium to be considered fairly valued. Of the liquid
credits, Venezuela is the most undervalued and Indonesia is the most
overvalued.
Introduction
The number and value of bonds issued by emerging markets have surged
dramatically over the course of the previous decade. So much so, it is
now one of the fastest growing sources of external development financing
being only second to FDI. In addition, their terms have improved due to
increased investor participation in emerging-market securities, as well as
better long-term economic prospects in a number of countries. However,
emerging-market sovereign spreads also have undergone marked
turbulence, in light of the plethora of financial crises experienced by the
emerging countries since the Mexican crisis in 1994-95. These large
movements have created attractive opportunities over time, both for
buyers as well as for sellers of emerging-market bonds, where price
movements do not often reflect changes in the countries ability to pay.
In this paper we introduce Goldman Sachs Equilibrium Sovereign Spread
(GS-ESS), which represents our preferred model for long-run equilibrium
sovereign spreads in emerging market economies. In the same spirit as
GSDEEMER for the case of real exchange rates and GS-DERBY in the
case of local currency bonds, GS-ESS develops a systematic guide to
valuing emerging market sovereign spreads over the long run.
Empirical studies in the area of emerging spreads, in contrast to studies on
financial crises, have been limited both in scope and in country coverage.
Most studies have focused on identifying the key variables that influence
spreads and justifying their role in a theoretical context. Even those
studies that have offered some empirical support for their theoretical
postulates have not tried to use the model as a means of valuation or
prediction. In light of this, our effort contributes in two innovative ways
(i) first, we perform an empirical analysis of the determination of
sovereign spreads in emerging-market economies; and (ii) second, we
construct a model that provides a valuation tool for the same spreads.
We also tested GS-ESS against other models commonly used to forecast
spreads. We find that from a statistical point of view, GS-ESS performs
better. We believe that this is due to the appropriate combination of
theoretical and state-of-the-art econometric approaches. We also examine
the forecasting ability of GS-ESS and show how it performs within an
asset allocation framework. Here again we find the returns delivered by
GS-ESS to be superior to those that a neutral investor (i.e., one holding
the market) would obtain. Specifically, we find that an investor using
GS-ESS to guide investment decisions would have outperformed the
market by an average annualized return of 386 basis points over the
period December 1997 to April 2000.
Theoretical Determinants of
Sovereign Spreads
logSPREADt = + i X i + t
where = log(1 + i*) and i* is the risk-free world interest rate, the Xis are
the fundamental determinants of the probability of default, the s are the
GS-ESS: Equilibrium Sovereign Spreads
Start of
Sample
Jan-96
Nov-96
Jan-96
Aug-96
Feb-96
Oct-96
Jan-96
Nov-96
May-96
Dec-96
Jun-97
Jan-96
Nov-96
Oct-96
Jan-96
Total
Observations
50
40
50
43
49
41
50
40
46
39
33
50
40
41
50
822
Mean
Spread
535
633
965
147
401
557
336
289
422
477
443
268
2928
400
658
551
Spread
Standard
Deviation
119
226
390
65
198
427
175
257
102
133
141
180
1553
192
274
432
Notes: 1) All spreads are monthly averages. 2) Unless otherwise noted, all table sources are Goldman Sachs & Co.
Estimates. In some cases, we extrapolated the yields historically using similar bonds.
The DFE estimator, on the other hand, estimates a panel regression but
imposes commonality across all elasticities and variances of each of the
groups, allowing cross-sectional variation only through the intercept term.
The DFE is also sensitive to panels with small groups and seems overly
restrictive.
The PMG estimator, due to Pesaran, Smith and Shin (1999), is an
alternative estimator that involves the same preliminary regressions of the
MG estimator but differs in one important respect. While allowing shortrun specifications and parameters to vary across individual groups, the
PMG estimator restricts the long-run elasticities to be identical across
countries. This seems to be a sensible assumption due to the imposition of
restrictions from economic theory, arbitrage conditions, etc., and
reinforces our view that the long-run impact of fundamentals on spreads
should be the same over countries, in stark contrast to the short-run. Due
to these reasons and the performance of PMG estimates in panel studies
cited in the literature so far, we adopt the PMG estimator to derive our
GS-ESS long-run elasticities.
The pooled mean group estimator of Pesaran, Smith and Shin (1999) is
applicable to panels where we have cross sectional variation in the shortrun dynamics but long-run commonality in the long-run equilibrium
relationship.
In estimating GS-ESS, it is important to note that a number of the
explanatory variables that we use (such as economic growth) tend to
exhibit relatively high volatility. This is because there is a gap between
the actual and permanent or potential value of these variables. For
valuation purposes, it appears more appropriate to use only the
permanent, or trend component of the fundamental variables. Hence,
while we use the actual data when we run the regressions, we only use the
permanent component of the data at the fitting stage. To do this, we use
an econometric procedure known as the Hodrick-Prescott (HP) filter to
effectively extract the permanent or trend component from the relevant
variables. The variable we apply the HP filter to is GROWTH.
For each country, the complete data set used for the empirical
specification consists of ten variables (the dependent variable, seven
economic fundamentals, global liquidity conditions and the debt
restructuring dummy variable). These data are pooled for all countries in
the sample, and panel PMG estimation is performed on this 15-country
data set.
Asymptotic
Variable
Coefficient
t-statistic
(in basis points)
Intercept
-2.7
-439.3
Long-Run Real GDP Growth
-5.1
-7
-691.3
Total Amort/Reserves Ratio
162.1
8.3
2
Total External Debt/GDP Ratio
7.5
10.1
7
Nominal Budget Balance
-34.2
-2.0
-34
Total NFGS Exports/GDP Ratio
-2.57
-5.8
-3
FX Real Misalignment
210.4
2.2
2
Long-Run LIBOR
45.3
1.7
45
Debt Restructure Dummy
165.0
5.0
165.1 (if country restructures)
Significant
Conclusion
R-bar Squared
0.55
F-Statistic
58.019
1%
Explanatory variables jointly significant
Breusch-Pagan LM
2.21
No
Residuals are not cross-correlated
Notes: Asymptotic t-statistics are based on robust standard errors. Long-run real GDP growth refers to trend growth. NFGS
refers to exports of non-factor goods and services. This ratio is corrected by transforming exports values by the long-run real
exchange rate as measured by GSDEEMER. The Breusch-Pagan LM test is a test for cross-sectional correlation of residuals,
distributed chi-square with 1 degree of freedom. Source: GS estimates.
GS-ESS Results
Table 2 above presents the results from the panel regression and the PMG
long-run elasticities. The estimated coefficients correspond to the longrun panel model. As indicated, all coefficients are associated with
theoretically expected signs, suggesting that the explanatory variables all
work to influence EM spreads over the long-run as we think they should.
In addition, all variables are statistically significant at the 5% level or
higher, with the exception of LIBOR which is significant at the 7% level.
Even though this variable is not significant at conventional levels, we
refrain from excluding it on the grounds of economic theory.
In the last column of this table, we provide an estimate of how spreads
would change given a 1-percent increase in each of the explanatory
variables. For example, if long-run real GDP growth were to increase by
1%, assuming all else is constant, this would reduce the long-run
equilibrium spread by about 7 basis points. On the other hand, the
marginal impact of a 100 basis points increase in LIBOR is an increase in
1
spreads of about 45 basis points. If the benchmark bond used is one that
corresponds to restructured debt, the spread will be, on average, 165 basis
points wider compared to those that have not been restructured.
In the table, we report additional summary statistics resulting from the
estimation, such as the adjusted coefficient of determination (or R2-bar) of
0.551. This means that over 55% of the variation of sovereign spreads is
explained by the explanatory variables. When interpreting this statistic, it
is important to note that the model is designed to provide the predicted
values for long-run equilibrium spreads. Consequently, we should not
necessarily expect a very high degree of fit because this implies, by
definition, that current spread levels are misaligned to some degree from
their long-run level. We also report an F-test that suggests that the
explanatory variables taken as a group all contribute significantly to
explaining long-run sovereign spreads. Finally, we test to see whether
1
Name of
Bond
Country
May-00
SPREAD
May-00
May-00
Under (-) or
GS-ESS Over (+) Valuation
Argentina
Republic 17
665
455
Brazil
Republic 27
720
473
Bulgaria
IAB11
728
542
China
Republic 08
158
153
Colombia
Republic 07
836
316
Indonesia
Republic 06
578
743
Korea
Republic 08
214
165
Malaysia
09
221
190
Mexico
UMS 16
374
322
Peru
PEPDI 17
575
442
Philippines
Republic 19
600
361
Poland
PLPDI 14
197
302
Russia
28
1150
932
South Africa
Republic 09
388
271
Venezuela
Republic 27
893
363
Notes: All spreads are monthly averages. Source: GS-ESS estimates.
-210
-247
-186
-5
-520
165
-49
-31
-52
-133
-239
105
-218
-117
-530
May-00
Valuation
Cheap
Cheap
Cheap
Fair
Cheap
Expensive
Cheap
Cheap
Cheap
Cheap
Cheap
Expensive
Cheap
Cheap
Cheap
10
11
With the introduction of GS-ESS, we now have two valuation models for
bonds in emerging markets: one for local currency bonds (GSDERBY)
and one for hard currency bonds (GS-ESS). Assuming integrated capital
markets, real interest rate parity requires that, in equilibrium, similarmaturity instruments of similar risk should carry the same real yield.
Differences, may arise due to transaction costs, capital account
restrictions, and other barriers to free capital flows. However, we would
expect our models to deliver results that are broadly consistent with each
other. We turn below to a comparison of GSDERBY and GS-ESS.
12
10
ARGENTINA
MEXICO
12
10
7
6
4
3
GS-ESS
GS-DERBY
0
Sep-97
GS-ESS
GS-DERBY
0
Feb-98
Jul-98
Dec-98
May-99
Oct-99
Mar-00
Sep-97
Feb-98
Jul-98
Dec-98
May-99
Oct-99
Mar-00
16
SOUTH AFRICA
POLAND
14
GS-ESS
GS-DERBY
12
10
5
8
4
6
GS-ESS
2
GS-DERBY
1
0
Sep-97
0
Feb-98
Jul-98
Dec-98
May-99
Oct-99
Mar-00
Sep-97
Feb-98
Jul-98
Dec-98
May-99
Oct-99
Mar-00
13
implies that even if there are temporary deviations between the real yields
during some periods over the sample, they still establish an equilibrium
relationship over the long run. This also confirms our preceding test
based on a single month in the sample, and reassures us that deviations
are not only transitory, but can be explained readily if we consider factors
not explicitly accounted by the two models.
Does GS-ESS Predict Better than
Alternative Models?
This section is devoted to a formal analysis to assess the ability of GSESS to forecast sovereign spreads on a country by country basis using
statistical methods. In the next section, we address the same question
from a market-oriented perspective, examining the performance of an
investor following the advice of GS-ESS.
To assess the models statistical ability to predict spreads, we use a
measure known as the Diebold-Mariano (1995) statistic. This statistic has
been shown in the literature to be the best measure for accomplishing
such an objective. Briefly, this statistic is designed to compare the
performance of a model against alternative models in terms of forecast
errors. Naturally, the smaller the forecast errors, the better the forecasting
ability. The Diebold-Mariano test can determine whether the forecasting
ability of a model is, in a statistical sense, significantly better, not
significantly different, or significantly worse than alternative models.
We present the results of applying this test to compare GS-ESS to two
alternative models, described as:
1. Model A ("Drifted Random Walk"): assumes a random walk with
a drift term, i.e. the best predictor of next months sovereign spread
will be this months sovereign spread adjusted for a trend.
2. Model B ("Mean-Reversion"): assumes that sovereign spreads for
each country, over the long run, will revert to their own means,
respectively. This is a hypothetical model in our exercise but one that
mimics well technical rules used for trading in many developed and
developing bond markets.
We compare the predictive performance of GS-ESS over the period
2
September 1998 to March 2000. The first two columns in Table 5 (p. 15)
show the Diebold-Mariano test statistics resulting from the statistical
comparison of the forecasting performance of GS-ESS against Models A
and B. A * denotes significance at the 10 per cent level, a ** at the 5 per
cent, and a *** at the 1 per cent. If the statistic is significant and
negative, it implies that GS-ESS is a better predictor of spreads than the
alternative model. If the statistic is not significant, it implies that GS-ESS
is not a statistically inferior predictor of spreads than the alternative
model. If the statistic is significant and positive, it means that the
alternative model is a better predictor of spreads than GS-ESS.
In comparison to Model A, GS-ESS is a significantly superior predictor in
9 countries, and not significantly worse in any. When compared to Model
B, GS-ESS provides a similar display of forecast performance, with the
rival model not significantly better than GS-ESS for any country. These
results are truly exemplary given that GS-ESS uses a panel framework.
Traditionally, panel models are not used for forecasting and models based
on just the time series methods are often the preferred forecasting models
in the literature.
2
14
GS-ESS vs.
Mean-Reversion
Diebold-Mariano
PesaranTimmerman
Market
Timing
Country
Using RMSE Criterion
Using RMSE Criterion
Out-of-Sample
Sign (%)
Argentina
-1.39*
-2.47**
2.55***
55.1
Brazil
-2.15**
-2.25**
2.64***
51.5
Bulgaria
-2.75*
-2.88***
3.15***
44.8
China
-1.75*
-1.55*
3.55***
69.2
Colombia
-2.12**
-0.05
2.81***
58.7
Indonesia
-2.11**
-1.64
2.64***
62.7
Korea
-1.23
-3.11***
3.23***
59.8
Malaysia
-1.12
-2.25**
3.15***
51.4
Mexico
0.56
0.36
3.29***
45.8
Peru
-0.66
-0.98
2.00***
67.7
Philippines
-2.15*
-0.15
1.75*
56.7
0.55
-3.54***
2.19**
59.4
Poland
Russia
-2.24**
-2.01
1.96*
53.44
South Africa
-0.12
-2.36**
3.22***
51.3
Venezuela
-1.59*
-1.88**
1.81*
71.5
Notes: *, ** and *** denote significance at the 10, 5 and 1 per cent levels, respectively. If the figures are not significant
this implies that we cannot reject, at conventional levels of significance, the hypothesis of equal expected meansquare error - i.e. GS-ESS is not a statistically significantly inferior predictor of the spread than is the alternative
model. If significantly negative, this favors GS-ESS outright against the alternative models. If significantly positive,
this favors the alternative model over GS-ESS. We use GS-ESS as the benchmark model. The DM test is associated
with the null hypothesis of equal forecasting accuracy in forecasting performance. For further details, see F.X.
Diebold and R.S. Mariano (1995). The PT test explicitly tests for independence between forecast and actual values.
Market timing refers to the percentage of times GS-ESS predicted correct turning points out of sample. For details of
the PT test see Pesaran, M.H. and A. Timmerman (1992). Full references are provided in separate reference list.
Source: GS estimates.
15
Using the expected returns that we obtain and the variance-covariance matrix as
inputs, we calculated two sets of efficient portfolios. An efficient portfolio is a
combination of investments that maximizes the expected rate of return on that
portfolio for a given variance of return. To derive such portfolios, we used the
mean-variance model of asset choice. The model has been extensively used in
finance since its inception, more than four decades ago, by Harry Markowitz. In
our calculations, we specifically followed this approach as developed in Huang
and Litzenberger (1988, Chapter 3). The combination of all such portfolios
defines the portfolio frontier, with each point on the curve representing a
particular combination of weights for each of the bonds.
16
Expected
ns in im u m
3 .8 4%3M returm
-0.99%
va ria n c e p ortfo lio
3 .8 3%
-1.00%
3 .8 2%-1.01%
-1.02%
3 .8 1%
-1.03%
3 .8 0%
-1.04%
3 .7 9%
-1.05%
minimum
variance portf olio
3 .7 8%
-1.06%
-1.07%
3 .7 7%
-1.08%
3 .7 6%
0.175%
0.175%
0 .78 5 %
0 .7 8 5 %
0.175%
0 .78 5 %
0.175%
0 .7 8 5 %
0.175%
0 .7 85 %
0.175%
0 .7 8 5 %
0.175%
0.175%
0 .78 5 %
0 .7 8 5 % volatility
vo la lit y
minimum
variance portfolio
-1.03%
-1.04%
-1.05%
-1.06%
-1.07%
-1.08%
0.175%
0.175%
0.175%
0.175%
0.175%
0.175%
0.175%
0.175%
volatility
17
110
100
EMBI Global
90
80
70
60
Dec-97
Mar-98
Jun-98
Sep-98
Dec-98
Mar-99
Jun-99
LONG ONLY
Sep-99
Dec-99
Mar-00
EMBI Global
Avg. Returns
Std. Dev
Skewness
11.3%
26.79%
(1.11)
7.3%
56.56%
(3.30)
7.5%
76.06%
(1.65)
Sharpe Ratio
42.3%
12.9%
9.8%
18
Appendix 1
19
20
Maximum
Eigenvalue
Statistic
Max Eigenvalue
90%
Critical Value
Trace
Statistic
Trace
90%
Critical Value
Significance of
LR Relationship
Argentina
17.81
12.98
24.43
23.08
High
Korea
22.93
12.98
28.48
23.08
High
Mexico
18.61
12.98
26.85
23.08
High
Poland
24.06
12.98
31.27
23.08
High
South Africa
18.75
12.98
23.41
23.08
Moderate
Philippines
20.22
12.98
25.96
23.08
High
Notes: Maximum eigenvalue and trace are statistics testing the null hypothesis that there is no long-run
relationship between the two variables. A statistic higher than the 90% critical values suggests evidence in
favor of a long-run relationship between GS-DERBY and GS-ESS real yields. The final column summarizes
the strength of the statistical significance with high and moderate indicating significance at the 95% and
90% critical values. These tests are based on the procedure prescribed by Soren Johansen (1991),
"Estimation and hypothesis testing of cointegrating vectors in Gaussian vector autoregressive models",
Econometrica, 59, 1551-1580.
21
References:
Barr, D.G. and B. Pesaran (1997), An assessment of the relative importance of real interest rates, inflation, and the
term premium in determining the prices of real and nominal U.K. bonds, Review of Economics and Statistics, 79,
362-366.
Breusch, T.S. and A.R. Pagan (1980), The Lagrange multiplier test and its applications to model specification in
econometrics, Review of Economic Studies, 47, 239-254.
Calvo, G.A., L. Leiderman, and C.M. Reinhart (1993), Capital inflows and real exchange rate appreciation in Latin
America: The role of external factors, IMF Staff Papers, 40, 108-151.
Choi, I. (1992), Durbin-Hausman tests for a unit root, Oxford Bulletin of Economics and Statistics, 54, 289-304.
Choi, I. and B.C. Ahn (1999), Testing the null hypothesis of stationarity for multiple time series, Journal of
Econometrics, 88, 41-77.
Cline, W. (1983), Interest and Debt: Systematic Risk and Policy Response, Institute for International Economics,
MIT Press, Washington, D.C..
Dickey, D.A. and W.A. Fuller (1979), Distribution of the estimators for autoregressive time series with a unit root,
Journal of the American Statistical Association, 74, 427-431.
Diebold, F.X. and R.S. Mariano (1995), Comparing predictive accuracy, Journal of Business and Economic
Statistics, 13, 253-263.
Dooley, M.P., E. Fernandez-Arias, and K.M. Kletzer (1996), Is the debt-crisis history? Recent private capital
inflows to developing countries, World Bank Economic Review, 10, 27-50.
Eichengreen, B. and A. Mody (1998), What explains changing spreads on emerging market debt: Fundamentals or
market sentiment?, Working Papers Series, National Bureau of Economic Research, W6408.
Edwards, S. (1984), LDC foreign borrowing and default risk: An empirical investigation 1976-1980, American
Economic Review, 74, 726-734.
Edwards, S. (1986), The pricing of bonds and bank loans in international markets: An empirical analysis of
developing countries foreign borrowing, European Economic Review, 30, 565-589.
Henriksson, R.D. and R.C. Merton (1981), On market timing and investment performance: II. Statistical procedures
for evaluating forecasting skills, Journal of Business, 54, 513-533.
Huang, C.F. and R.H. Litzenberger (1988), Foundations for Financial Economics (Prentice-Hall: New Jersey).
Im, K.S., M.H. Pesaran and Y. Shin (1998), Testing for unit roots in heterogeneous panels, University of Cambridge
DAE Working Paper.
Johansen, S. (1991), Estimation and hypothesis testing of coitegrating vectors in Gaussian vector autoregressive
Kao, C. (1999), Spurious regression and residual-based tests for cointegration in panel data, Journal of
Econometrics, 90, 1-44.
Levin, A. and C.-F. Lin (1993), Unit root tests in panel data: New results, Department of Economics UC-San Diego
Discussion Papers.
Markowitz, H., (1959), Portfolio Selection: Efficient Diversification of Investments (John Wiley: New York).
22
References Continued:
Pesaran, M.H. and A. Timmerman (1994), A generalization of the non-parametric Henriksson-Merton test of
market timing, Economics Letters, 44, 1-7.
Sarno, L. and M.P. Taylor (1998), Real exchange rates under the recent float: Unequivocal evidence of meanreversion, Economics Letters, 60, 131-137.
Stock, J.H. and M.W. Watson (1993), A simple estimator of cointegrating vectors in higher order integrated
systems, Econometrica, 61, 783-820.
23
Frankfurt
London
Tokyo
Paris
Singapore
Hong Kong
Korea