Você está na página 1de 51

Predictability in Emerging Sovereign Debt Markets

Gergana Jostova∗

George Washington University


Department of Finance
Lisner Hall 540J, 2023 G Street, Washington, DC 20052
Tel: (202) 994-7478, Email: jostova@gwu.edu
Webpage: home.gwu.edu/˜jostova

First draft: July 12, 2000


This revision: October 28, 2003


I thank seminar participants at Boston College, the European Finance Association meeting 2001, the
Midwest Finance Association meeting 2002, the Eastern Finance Association meeting 2003, the Washington
Area Finance Association meeting 2002, JPMorgan, and Deutsche Bank. The study has benefited from the
comments of Alexander Philipov, Pierluigi Balduzzi, Alan Marcus, Wayne Ferson, Doron Avramov, Cesare
Robotti, Robert Savickas, Evan Gatev, Eric Jacquier, Edith Hotchkiss, Edward Kane, and an anonymous
referee.
Predictability in Emerging Sovereign Debt Markets

Abstract

This paper finds strong evidence of predictability in Brady bonds, the most liquid
emerging debt market, by implementing a new model for credit spreads. Predictability
is economically and statistically significant and robust to various considerations. Active
management provides US investors in emerging markets with double the buy-and-hold
returns at lower risk and the equivalent of free options on Brady bonds. Our analysis
suggests that predictability is primarily driven by credit spread deviations from funda-
mentals, rather than time-varying risk or risk premia. We believe this inefficiency is the
result of the restrictions of a non-transparent, institutionally dominated, dealer market
and the lack of a well developed derivatives market for emerging country credit risk.
1. Introduction
Brady bonds1 are the primary financing vehicle of emerging market countries, currently rep-
resenting 80% of their total government external debt2 . Brady debt now totals over $100
billion. Brady bonds are also the primary choice of diversification into emerging markets for
US mutual, pension, and endowment funds. These bonds are the single most traded emerging
market debt instrument, with transaction volume of $1 trillion in 1993 and $2.7 trillion in
1996 (see Hassan (2001)). The sharp increase in Brady bond turnover, relative to the mod-
est increase in their amount outstanding, suggests a large increase in their liquidity. The
most liquid and popular Brady bonds are issued by the four largest emerging market debtors,
Argentina, Brazil, Mexico, and Venezuela, which account for 75% of the market. The high
liquidity of these countries’ Brady bonds is underscored by a typical bid-ask spread of $0.25
(0.4%), although large trades can get even better terms (see Cumby and Pastine (2001) and
Claessens and Pennacchi (1996)). As a result, Brady debt provides an easy and inexpensive
way for US investors to diversify into emerging debt markets.
Since their creation in 1990, Brady bonds have generated an average annual return
of 43% by 2001. However, emerging countries have also been ridden by frequent financial
crises - the Peso (December 1994), the Asian (October 1997), the Russian (August 1998),
and Brazilian (January 1999) crises - leading to sudden declines in emerging debt portfolios.
Understanding credit risk and the ability to time changes in credit fundamentals is essential in
emerging debt markets. The combination of high returns, high liquidity, and frequent turmoil
makes active management all the more attractive.
In this respect, the contribution of this study is twofold. First, the paper proposes a
general two-stage model for credit risk, and develops a formulation of the model that cap-
tures the dynamics of credit spreads in emerging debt markets. Second, the study presents
strong evidence of predictability in the Brady bond market and supports it with a variety
of out-of-sample tests. Predictability in emerging sovereign debt markets has not yet been
investigated. So far, predictable variation in equity and bond returns has been studied in
1
Named after the Nicholas Brady plan of 1989 to restructure the troubled Mexican loans. Earlier loans
were swapped with Brady bonds with lower coupons and shorter maturity by collateralizing the face value of
the bond with US treasuries. The swap transformed Mexico’s government loans into tradable and liquid debt
instruments. Argentina, Brazil, and Venezuela followed suit and all issued Brady bonds in the early 1990s.
Currently 17 emerging market countries have Brady debt.
2
Source: JPMorgan: ”Introducing the Emerging Market Bond Index Plus”,
http://www2.jpmorgan.com/MarketDataInd/EMBI/embi.html.

1
developed markets3 , as well as in emerging market equity returns4 . Studies of emerging debt
markets have focused on the pricing and issuance of Brady debt (see Claessens and Pennac-
chi (1996), Cumby and Pastine (2001), Duffie, Pedersen, and Singleton (2003), Eichengreen
and Mody (1998)), or on identifying fundamental determinants of debt prices (Boehmer and
Megginson (1990)).
We study the predictable component in the changes of the credit spread index, called
the Emerging Market Bond Index 5 (EMBI) spread6 , of the four biggest issuers: Argentina,
Brazil, Mexico, and Venezuela. Brady bonds trade on dealer markets, in which dealers trade
on spreads rather than prices. JPMorgan, one major dealer in the Brady market, derives
the credit spread implied in the price of each Brady bond, of each country’s EMBI index
(a total return index of the country’s most liquid Brady bonds), and of the World EMBI
index. Predictability of credit spread changes translates into predictability of Brady bond
excess returns over US treasuries7 . In this study, the model’s formulation is based on credit
spreads, but its predictive power is evaluated out-of-sample based on both realized credit
spread changes as well as actual holding-period returns to a US investor after accounting for
transaction costs.
The proposed model for credit spreads has two stages. The first stage describes the long-
term equilibrium relation between a country’s credit spread level and local macroeconomic
factors. The financial intuition behind the long-term equilibrium is that the spread implied
in market prices is a premium for holding defaultable sovereign instruments. This premium
depends on the intrinsic credit risk of the government, which is a function of the economic
conditions in the emerging country. Under the assumption of market rationality, there should
be a stable relation between the level of spreads observed on the market and the intrinsic
credit risk of the government, as proxied by the local macroeconomic factors, despite the well-
known instability of emerging market conditions. The stability of this relationship provides
important information in predicting the direction of future credit spread changes.
The second stage relates the short-term dynamics of spread changes to global instru-
ments as well as to the deviation of the spread level from its long-term equilibrium (derived
3
See Avramov and Chordia (2003), Avramov (2002), Ang and Bekaert (2001), Bekaert and Hodrick (1992),
Ferson (1989), Ferson and Harvey (1993, 1999), Harvey (1991), Ilmanen (1995), and Lewellen (1999).
4
See Bekaert (1995), Bekaert and Harvey (1995), and Harvey (1993, 1995).
5
The EMBI spread of a country is the spread above the US treasury spot curve that sets the total market
value of all Brady bonds equal to their discounted payments. (More details in the Data section.)
6
Credit spreads are the direct counterpart of equity excess returns and are a standard focus in the literature
on defaultable bonds.
7
A change in the spread causes a change in the relative discount factors that translate the debt’s promised
cash flows back to the present, and hence in a change in the bond price. As a result, predicting the direction
of the spread change is equivalent to forecasting the sign of the Brady bond excess return over US treasuries.

2
in stage one). The results show that this deviation from fundamentals is the most important
instrument in predicting subsequent spread changes. The inclusion of this deviation can be
viewed as an innovation to traditional single-equation predictive models. The importance
of this innovation is assessed out-of-sample using a broad set of tools for testing both the
statistical and economic significance of the observed predictability.
We take special effort to show that the documented predictability is not spurious or
unexploitable. Predictability is evaluated out-of-sample using three independent tests. The
first test is based on realized holding period returns. An active Brady-bond strategy based on
the model’s predictions allows US investors to double the returns from a buy-and-hold strategy,
while taking less risk and accounting for transaction costs. Previous studies (e.g. Ilmanen
(1995) and Harvey (1991, 1993, 1995)) document only slight predictability in developed and
emerging equity markets. We also show the superior performance of the active strategy is
robust to the timing of the initial investment. Second, we apply Merton’s (1981) equilibrium
test for market-timing value, which is based on the number of correct out-of-sample directional
forecasts of credit spread changes. Merton’s test shows that the model adds significant value
to US investors and provides them with the equivalent of free options on Brady bond indexes.
Finally, to check the robustness of the results to the relatively small size of the out-of-sample
window, we perform Henriksson and Merton’s (1981) nonparametric small-sample market-
timing test, which specifies a sufficient number of correct predictions necessary to reject the
null of no predictability for a particular sample size (the smaller the sample size, the harder
it is to reject no predictability). The null hypothesis of no predictability is rejected at the
1% significance level in Brady markets. The last two out-of-sample tests, both based on the
number of correct predictions, serve to eliminate concerns that the results are driven by a few
’lucky’ periods. All three tests agree that the spread’s deviation from fundamentals is the
instrument that drives all predictability.
Predictability is robust to various considerations. Concerns about spurious predictive
regression and biased t-statistics (see Nelson and Kim (1993)) or biased slope coefficients (see
Stambaugh (1999)) are confined to in-sample results. Under the null hypothesis of no pre-
dictability, spurious regressions would not bias out-of-sample results and would not produce
superior returns out-of-sample (see Ferson, Sarkissian, and Simin (2003) and Lo and MacKin-
lay (1996)). A robust estimator is used to correct for data non-stationarity and cross-country
correlations (contagion) and simulations are conducted to assess possible small sample biases.
Transaction costs are accounted for by buying at the ask and selling at the bid price when
rebalancing. Liquidity is not an issue as Brady bonds are the most liquid emerging debt
market and the active strategy involves only monthly rebalancing.

3
The asset pricing literature has documented time-series return predictability (see e.g.
Fama and French (1988,1989) and Keim and Stambaugh (1986)). This predictability has been
attributed to time-varying risk, market price of risk, and market inefficiency (see Avramov and
Chordia (2003) and Ferson and Harvey (1991, 1999)). The predictability documented in this
paper is primarily due to informational inefficiency as it is driven by the spread’s deviation
from its fundamental value (90% of which is determined by country-specific factors). We show
that global equity and bond instruments do not provide any out-of-sample predictability in
the Brady market.
We believe this informational inefficiency is the result of the characteristics of the Brady
market. First, Brady markets are dominated by large institutional investors following con-
strained investment policies, which slows down the process of price adjustment. There is
an absence of arbitrageurs and unrestricted investors8 due to the scarcity of derivatives and
the large transaction lots. Second, unlike Treasury markets which are also dominated by
large institutional investors, Brady markets lack the completeness provided by fully devel-
oped derivatives markets which would allow the separate pricing of credit risk. Third, Brady
bonds trade in non-transparent dealer markets which are generally associated with lower in-
formational efficiency. Section 6 discusses these sources of predictability in detail.
We focus on Brady bonds rather than other emerging market debt instruments, because
(1) they are by far the most liquid and largest emerging debt market, and (2) they have
significantly longer history than other emerging market bond indexes. While investors in
emerging debt markets have three options: domestic bonds, Eurobonds, and Brady bonds,
the Brady bond market is by far the most liquid and largest market of all (Solnik (2000, p.
369)) as the issue size of Brady bonds is quite large9 . If one is to make a case for active
management, it is more reasonable to go with more liquid instruments. The typical Brady
bond bid-ask spread of $0.25 is very low relative to that of Eurobonds and more so than
that of domestic bonds issued by emerging countries. Second, the history of the Brady bond
indexes is longer than that of Eurobonds, most of which were issued after 1995, allowing us to
study the statistical property of these indexes over a longer period of time. In addition, the
correlation between the excess returns of Brady bonds and other emerging market instruments
is almost perfect to make the analysis applicable to other instruments or to allow investors to
take advantage of their relative mispricing10 .
8
Unrestricted investors here refer to ’speculators’ who can take unrestricted positions in the asset they
consider mispriced.
9
Unlike Brady bonds, whose issue size is very large, the typical Eurobond is $100 million or less. The set
of actively traded Eurobonds is very limited. Source: JPMorgan: ”Introducing the Emerging Market Bond
Index Plus”, http://www2.jpmorgan.com/MarketDataInd/EMBI/embi.html.
10
The analysis of emerging market domestic, Eurobonds, and Brady bond indexes (provided by JPMorgan)

4
The remaining of the paper is organized as follows. Section 2 motivates the two-stage
model for credit spreads. Section 3 describes the data. Section 4 suggests a robust estimation
methodology. Section 5 presents the results from economic and statistical tests of out-of-
sample predictability. Section 6 discusses the causes for the existence of the documented
predictability and section 7 concludes the paper. Appendix A provides methodological details
on the SUCCR11 estimator used in the estimation of the first stage of the model and Appendix
B presents simulation results on the SUCCR small-sample bias reduction.

2. The model
Credit spreads reflect the market estimate of the credit quality of risky bonds, an unobservable
intrinsic characteristic. The intrinsic credit quality of bonds is driven by the debtor’s true
capacity to service its debt. In the case of sovereign debt, this capacity depends on the
economic, fiscal, and financial conditions in the emerging country, which can be proxied by
a set of properly motivated macroeconomic and financial indicators. Under the assumption
of market rationality, the market-determined credit spread should not deviate significantly
from the intrinsic credit risk of the local government. Hence, market rationality dictates that
credit spreads will converge to their long-term equilibrium levels commanded by the true credit
quality of the debtor. In the short-term, however, spreads may deviate from their fundamental
value due to investors sentiment, market momentum, or institutional factors.
When a long-term equilibrium exists, future credit spread changes depend on the cur-
rent deviation of the spread from its long-term equilibrium level with respect to the local
macroeconomic factors. Statistically, long-term equilibrium is represented by cointegration 12 .
Engle and Granger (1987) prove that omitting deviations from long-term equilibrium when
predicting changes of cointegrated variables results in model misspecification, because current
deviations from equilibrium contain information about future changes beyond that provided
by the levels or differences in the variables. Following Engle and Granger’s analysis, standard
predictive models, conditioning returns (or changes in prices) on a set of exogenous instru-
ments, will be misspecified if there exists a long-term equilibrium between the level of prices
and fundamental factors.
The existence of a long-term equilibrium between credit spreads and fundamental fac-
reveals that the correlation between their monthly spread changes range from 0.8 to 0.9, implying the excess
returns for the three bond types should be similar before transaction costs.
11
Seemingly Unrelated Canonically Cointegrating Regressions, Park and Ogaki (1991).
12
Cointegration implies that although individual series may experience permanent shocks, these shocks will
affect all variables in a way that preserves their long-term equilibrium. (See Data section.)

5
tors in Brady bond markets (this assumption is tested in the Data section) motivates the
formulation of a two-stage model for credit spreads:

Spreadi,t = Bi Fi,t + εi,t (1)


∆Spreadi,t+1 = βE,i ∆SpreadEM BI,t + βW,i ∆M SCIt + αi ε!i,t + ei,t+1 (2)

where
E(εi,t ) = 0; E(ε2i,t ) = σi2 ; E(εi,t εj,t ) = σij
Et−1 (εi,t ) = f (εi,t−1 , εi,t−2 , ...) (if a long-term equilibrium exists)
E(ei,t ) = 0; E(e2i,t ) = γi2 ; E(ei,t ej,t ) = 0

and the variables are defined as:


Spreadi,t = country i ’s credit spread above the US treasury spot curve
Fi,t = vector of country i’s local macroeconomic factors
Bi = vector of country i’s spread sensitivities to local factors
εi,t = spread deviation from long-term equilibrium in country i
∆SpreadEM BI,t = change in the World EMBI spread index
∆M SCIt = change in the Morgan Stanley Capital International world equity index
ε!i,t = estimate of Spreadi,t ’s deviation from its long-term equilibrium
βE,i , βW,i = sensitivity to the respective instrument
αi = speed of adjustment to long-term equilibrium
σi2 = variance of the deviation from long-term equilibrium
σij = covariance between spread deviations in countries i and j
γi2 = unexplained variance of country i’s spread changes

The proposed two-stage model has a strong link to existing credit risk models. The first
stage relates to credit scoring models and links the level of spread observed on the market to
the equilibrium spread level commanded by the intrinsic credit risk of the Brady bond. The
second stage relates to predictive models with two important adjustments: (1) expected bond
excess returns are represented by expected changes in credit spreads. This representation
is motivated by the fact that bond excess returns are driven by changes in credit quality or
liquidity13 . In the case of the highly liquid Brady bonds, changes in credit quality is the major
factor driving excess returns. However, it is the change in spreads, rather than returns per
se, that are directly related to the first stage of the model. (2) The deviation from long-term
equilibrium from the first stage is a new instrument included in the predictive stage.
13
Since we are dealing with excess returns, the effect of interest rate changes cancels out.

6
2.1. Determinants of long-term equilibrium credit spread levels

Our choice of factors (summarized in Table 1), used to proxy for the intrinsic credit risk in
emerging countries, is motivated by whether the factors reflect on the expected default proba-
bility of the government and/or the recovery rate in case of default. The factors considered for
each country are the Local Equity Index, Consumer Price Index (CPI), Real Exchange Rate
Index, Short-term Interest Rates, Money Supply, Unemployment, and Gross Domestic Product
(GDP). Our selection draws on economic theory, and previous studies on credit spreads or
asset returns.
The Local Equity Index is used as a proxy for the wealth of the economy. This variable
also reflects on capital gains, tax revenues, and the ability of the government to service its
debt. The inclusion is supported by Barnhill et al. (2000), Bookstaber and Jacob (1986),
Ramaswami (1991), and Shane (1994) who document the co-movement between high-yield
bonds14 and equity indexes. As in Ferson (1989), the CPI is included to control for the
inflationary component of the stock index. By the purchasing power parity, the CPI should
also have an affect on exchange rates, imports, exports, and the balance of payments of a
country, which in turn affect the country’s funds available to service its debt. Real Exchange
Rates have an impact on the country’s terms of trade and current account and appreciating
exchange rates have been the precursor of financial crises in emerging markets, including the
recent crisis in Argentina. Real exchange rates have been found to be a significant factor
in pricing international assets in studies by Adler and Dumas (1983), Dumas and Solnik
(1995), and Sercu (1980). Local Short-Term Interest Rates reflect local intertemporal rates of
substitution. High interest rates often indicate a large local debt as the government increases
its demand for loanable funds. Domestic debt represents an extra burden on the government
in servicing its external debt. High interest rates further lead to underinvestments which
hamper the future growth of the economy and reduce the government’s debt service capacity.
Fama and Schwert (1977) and Ferson (1989) find evidence that short-term interest rates are
an important factor in pricing US long-term bonds and equities. The Money Supply variable
reflects on the government’s monetary policy and discipline. A growing money supply is often
the result of a large fiscal deficit (’monetizing the deficit’), which leads to hyperinflation and
ultimately to real economic crises. Fama (1981), Geske and Roll (1983), and Patelis (1997)
find money supply to be a significant variable in explaining asset returns. Unemployment
reflects the real productive capacity of the economy and Shiller (1984) finds unemployment to
be a significant business cycle indicator. GDP measures the government’s ability to generate
cash flows and service its debt. Since prices are discounted expected future cash flows, Fama
14
A BBB+/Baa1 rating or below defines Emerging Markets in the context of external debt markets.

7
(1990) argues that such a production measure should be significant in explaining yields and
returns.

2.2. Predictors of short-term credit spread changes

In the predictive second stage of the model, spread changes are conditioned on lagged global
instruments, as well as on the spread’s deviation from long-term equilibrium, ε!i,t−1 , derived
in the first stage of the model.
The inclusion of this deviation is an innovation of this study, prompted by the presence
of a long-term equilibrium between credit spreads and underlying macroeconomic fundamen-
tals. Any deviation from the long-term equilibrium expressed in eq. (1) is temporary by
definition. The equilibrium error, εit , should, therefore, be stationary and mean-reverting,
even if the spreads and economic factors experience persistent shocks. Ignoring deviations
from long-term equilibrium biases the expectation of future spread changes, because, while
the expected equilibrium error (E(εi,t )) is zero unconditionally, its conditional expectation
(Et−1 (εi,t )) is different from zero and affects subsequent changes in spreads. The deviation
from long-term equilibrium levels carries information beyond that contained in the differenced
factors or spreads (see Engle and Granger (1987)). For example, if spreads on the market are
below the equilibrium commanded by the level of the underlying factors, they will correct
partially up as the long-term equilibrium relation prevents the spread and factor levels from
drifting too far apart. The degree and speed of correction is measured by αi (the coefficient
of ε!i,t−1 ), which should be between -1 and 0, reflecting a partial adjustment to long-term
equilibrium. The deviation instrument will be significant in predicting spread changes under
two conditions: (1) the long-term equilibrium relation in the first stage exists and is correctly
identified, and (2) the market fails to react instantaneously to temporary misalignment in
credit spread levels.
The remaining instruments in the predictive stage are global equity and bond instru-
ments, included to test whether time-varying risk or risk premia can capture some of the
predictability in Brady markets. The first instrument is the lagged change in the World
Emerging Market Bond Index (World EMBI) spread, SpreadEM BI,t−1 . One can think of the
first instrument as representing global Brady-bond excess returns, thus capturing the return
of the asset class on which the study focuses. The second instrument is the change in the
Morgan Stanley Capital International (MSCI) equity index, ∆M SCI, which is a global total
return equity index. Its inclusion follows Ilmanen (1995) and Harvey (1995) who find global
instruments to have some predictive power in developed and emerging equity markets.

8
3. Data
The study uses data from the four largest emerging market debtors: Argentina, Brazil, Mexico,
and Venezuela, accounting each for about 25%, 20%, 20%, and 10% of total emerging market
Brady debt, respectively. All series consist of end-of-month observations from April 1993
to February 2001. Although some EMBI indexes date back to December 1990, when the
World EMBI index was constructed, the econometric methodology of the study requires equal
observations for all countries and limits the sample to the shortest history of Argentina’s
Brady debt.

3.1. EMBI spreads and total return indexes

All EMBI spread and total return series are provided by JPMorgan. The company derives
the spread of each Brady bond and computes the EMBI spread and total return indexes for
each country with Brady bonds. These indexes are continuously provided on Bloomberg by
JPMorgan and are used directly by dealers and investors to compare sovereign instruments.
We use the country EMBI spread index in our analysis.
The credit spread of a Brady bond is defined as the spread above the US treasury spot
curve that sets the current market price of the bond equal to its discounted payments:
τ
" "T
Ct Ct F ace
P0 = t
+ t
+ (3)
t=1 (1 + rt ) t=τ +1 (1 + rt + Spread) (1 + rT )T

where P0 is the current price of the Brady bond, T is the maturity, Ct is a cash flow scheduled
for period t, and rt is the US treasury spot rate for delivery at time t. The final payment is
discounted at the US treasury bill rate as the face value of Brady bonds is guaranteed by the
US government through the Nicolas Brady plan of 1989. This final payment does not carry
credit risk and does not provide a default premium. The first τ payments are also discounted
at the US treasury spot rates when the Brady plan provides for a rolling interest-rate guarantee
for the immediate τ payments on the debt. The purpose of the spread is to provide a single
measure of the pure sovereign default risk of Brady instruments.
JPMorgan also computes each country’s spread, Spreadi,t , which is the weighted average
spread of all Brady bonds that meet certain size and liquidity requirements, as well as the
World EMBI spread index, SpreadEM BI,t , of all emerging market countries with Brady debt.
The four countries’ EMBI spreads are highly correlated with the World EMBI and
among themselves. Table 2 presents the sample correlations among the EMBI spreads of
the four countries and the World EMBI spread. This high level of cross-country correlation

9
(ranging from 74% to 90%) suggests that spreads may have common trends and shocks and
should be examined together rather than country by country. Figure 1 shows that spreads
have experienced periods of high volatility especially around the Peso (December 1994), Asian
(November 1997), and Russian (August 1998) crises, as well as during the Brazilian devalu-
ation (January/February 1999). Descriptive statistics of spread levels and spread changes of
the four countries are provided in Tables 3 and 4, respectively.

3.2. Macroeconomic factors

The monthly economic variables used are country-specific indicators, that are publicly avail-
able and can be obtained from the IMF and IIF databases (the sources used are summarized
in Table 1). The financial series - local stock index, local interest rates, and exchange rates -
are available continuously for all countries. The economic series - unemployment, GDP, CPI,
money supply - are available on a monthly basis, but are reported with a two- to three-week
lag. To make them contemporaneous with the EMBI spreads in terms of information arrival,
we lag all economic series by one month. For example, September’s consumer price index for
Argentina is reported in October, while October’s sovereign spread is known the same day in
October. Those two variables are contemporaneous in the sense that they become public in
the same month. It is important to make sure that all predictions are based on information
that is publicly available when the active strategy is implemented. Descriptive statistics for
all macroeconomic indicators are presented in Table 5.

3.3. Diagnostic tests

The EMBI spreads and all macroeconomic variables in the long-term equation are nonsta-
tionary and integrated of order one according to the Augmented-Dickey-Fuller (ADF) test 15 .
This result is not surprising for the level of macroeconomic variables. The intuition behind the
nonstationarity of the EMBI spreads (especially given the short history of the Brady market)
lies in the very nature of ’emerging’ markets. EMBI spreads capture the economic conditions
of these countries, which are evolving, volatile, and often unpredictable. Regressions of non-
stationary variables would produce spurious results unless some equilibrium relation ties the
series together. Such a relation is modeled through cointegration. Intuitively, cointegration
implies that, although the individual economic variables may experience permanent exogenous
shocks, such shocks affect all variables in a way that preserves the equilibrium among them.
15
Results from the ADF tests are available from the authors upon request.

10
In the context of this paper, cointegration in the long-term equilibrium stage is essential
to estimating the dynamics of the predictive stage. Engle and Granger (1987) show that in
the presence of cointegration, the level of the macroeconomic factors contains information
beyond that contained in the first differences of the variables. Deviations from the long-term
equilibrium levels of the variables are useful in predicting subsequent changes in EMBI spreads
(the predictive stage).
Johansen’s (1988) multivariate procedure is used to test for cointegration. We find at
least two cointegrating equations between the spreads and macroeconomic variables in each
country16 . The specific features of our econometric methodology, however, require that there
exist a single cointegrating equation in each country (see Appendix A). Therefore, when-
ever more than one cointegrating relation is found for a country, we reduce the number of
macroeconomic factors until a single long-term equilibrium relation among the spread and the
remaining factors is obtained. There is no significant loss of information because the removed
variables are stationary combination of the remaining variables in the set. The final set of
variables used in each country long-term dynamics stage are presented in Table 6.

4. Estimation Methodology
With the structure of the model and characteristics of the data in mind, this section motivates
the use of a robust and efficient estimation methodology for each stage. The parameters
of the long-term equilibrium stage are estimated using Park and Ogaki’s (1991) Seemingly
Unrelated Canonical Cointegrating Regression (SUCCR). This method makes full use of the
nonstationarity and cross-country correlation in our data and produces efficient and unbiased
estimates. The parameters of the short-term dynamics equation are estimated using standard
OLS methodology.

4.1. Long-term equilibrium stage: the SUCCR methodology

The data used in this study have two dimensions, the parameters of which cannot be appropri-
ately estimated using methods traditionally used in the finance literature. In the time-series
dimension, all series have unit roots and are cointegrated within each country. OLS estimates
would be superconsistent, but their limiting distributions would be biased and inefficient 17 .
16
Cointegration results are available from the authors and are omitted here to preserve space.
17
Park and Ogaki (1991), Park (1992) and Pedroni (1996) show that OLS estimates of the coefficients of
nonstationary regressors are inefficient, and their distributions are asymptotically biased and contain nuisance
parameters. The bias in the asymptotic distribution of the OLS (or GLS) estimates is due to the fact that all

11
Fully modified OLS procedures18 would produce correct hypothesis tests in the presence of unit
roots but would not be suitable for the cross-country correlation in spreads and residuals. Zell-
ner’s (1962) Seemingly Unrelated Regression (SUR) would account for the cross-correlation,
but would not correct for the bias due to nonstationarity in the time-series. One methodology
that accounts for both the time-series and cross-sectional properties of our data is Park and
Ogaki’s (1991) Seemingly Unrelated Canonical Cointegrating Regression (SUCCR).
The SUCCR methodology is an optimal statistical procedure for a system of poten-
tially correlated cointegrating regressions. The econometric details on the SUCCR estimator
are presented in Appendix A. Next, we provide briefly the intuition, advantages, and gen-
eral structure of the SUCCR estimator in the context of traditional least-square estimators.
Consider a general panel structure represented by:

y = Xβ + u (4)

where y = (y1" , ..., yn" )" , yi = (yi1 , ...yiT )" , i = 1...n, is a vector of all dependent variables
stacked, X = block-diagonal(Xi ), Xi = (xi1 , ...xiT )" is a block-diagonal matrix with the re-
gressors of country i in the i" th block, β = (β1" , ..., βn" )" is a vector of stacked cointegrated
vectors, and u = (u"1 , ..., u"n )" , ui = (ui1 , ...uiT )" is a the vector of stacked residuals.
The SUCCR estimator is the modified system GLS estimator using the longrun co-
variance matrix19 , thus adjusting for autocorrelation in the errors, while also canonically
transforming all variables to eliminate the bias due to the presence of unit roots:

β!SU CCR = (X"∗ (Ω


! −1 ⊗ I)−1 X )−1 X" (Ω
∗ ∗ ∗
! −1 ⊗ I)−1 y
∗ ∗ (5)

All notations are defined in Appendix A.


The SUCCR estimator generalizes Park’s (1992) CCR (Canonical Cointegrating Re-
gressions) estimator, β!CCR = (X"∗ X∗ )−1 X"∗ y∗ , by using system information in the same way
as the SUR estimator, β!SU R = (X" (Σ
! −1 ⊗ I)−1 X)−1 X" (Σ
! −1 ⊗ I)−1 y, generalizes the usual OLS

estimator, β!OLS = (X" X)−1 X" y, for stationary panels. However, the SUCCR estimator uses
the adjusted longrun variance of the errors, Ω∗ , rather than the shortrun variance, Σ, used
by SUR. The longrun variance, Ω, accounts for the autocorrelation in residuals, while the
canonically adjusted longrun variance, Ω∗ , further corrects for the presence of unit roots in
the regressors.
sample moments converge to random matrices when the variables are nonstationary and cointegrated rather
than to constant matrices for which traditional techniques are designed.
18
Phillips (1988, 1991), Johansen (1988, 1989), Phillips and Hansen (1990), Park (1992), and Stock and
Watson (1991) provide corrections for the asymptotic bias of OLS. The problems of nuisance parameters are
addressed by Phillips and Durlauf (1986) and Park and Phillips (1988, 1989). These techniques are designed
for the estimation of a single equation, rather than a system of correlated equations as in the present study.
19
See Hamilton (1994), p. 248, for Newey-West adjustment for autocorrelated errors.

12
In summary, the advantages of SUCCR over traditional methods, such as SUR and OLS,
can be represented in the following adjustments. First, the SUCCR procedure corrects for
the asymptotic bias introduced by the unit roots by modifying the regressors and covariance
matrices. It utilizes the presence of unit roots in the entire system, not just equation by
equation. Second, the SUCCR estimator accounts for the cross-correlation across equations
(or countries) by using the system-wide covariance structure of the errors. Therefore, the
SUCCR estimator in (5) improves on OLS, SUR, and CCR by correcting simultaneously for the
problems of nonstationary and cross-correlated panel data. The SUCCR methodology allows
for stationary regressors in addition to the nonstationary ones used here. If all regressors
are stationary, the SUCCR procedure reduces to SUR. Appendix B documents simulation
results of the finite sample improvement (in both bias and mean square errors) of the SUCCR
estimator over OLS, GLS, SUR, and CCR. Our results show that the SUCCR estimator has
the lowest small-sample bias among all methods and this bias falls rapidly as the sample size
increases.

4.2. Short-term dynamics stage

The second stage models the dynamics of credit spread changes. All instruments in this stage
are stationary:

∆Spreadi,t+1 = βE,i ∆SpreadEM BI,t + βW,i ∆M SCIt + αi ε!i,t + ei,t+1 (6)

Since spreads are unit-root, their first differences are stationary. The error correction compo-
nent, ε!t , is stationary due to the existence of a cointegrating relation. All remaining instru-
ments are stationary first differences of unit-root variables. Given stationarity, this stage can
be estimated using OLS or GLS methods (the two methods are later compared).

5. Results
The in-sample results show that the long-term equilibrium relation is strong in all countries
and local fundamentals capture essentially all of the variation in credit spread levels. The
out-of-sample results demonstrate predictability in Brady markets is significant and valuable
to US investors. Predictability is driven by the spread’s deviation from equilibrium.

13
5.1. In-sample results

5.1.1. Long-term dynamics

Results from the SUCCR estimation, presented in Table 6, show that the long-term equilib-
rium between each country’s credit spreads and local macroeconomic indicators is significant
and strong. All estimates of the cointegrating vectors are significant and have the expected
sign. The high adjusted R2 in each country (89% to 93%) show that most of the variation in
credit spreads is captured by local factors.
Table 7 compares alternative estimates of the long-run equilibrium parameters for each
country. The SUR and MSUR estimates have the biggest small sample bias. They are
often quite different from the SUCCR or OLS estimates and sometimes have the wrong sign.
Although, both the OLS and SCCR small sample biases are relatively small in panels with 95
time-series observations, our simulations (Appendix B) indicate that the SUCCR improvement
becomes more significant as the sample size increases to 150 monthly observations, while OLS
and SUR show no reduction in small-sample bias (see Tables 13 and 14).
Figures 2 and 3 illustrate the spread deviations from equilibrium estimated with OLS
and SUCCR, respectively. Both the SUCCR and OLS estimated errors are stationary due to
the cointegration among spreads and macroeconomic factors, but the SUCCR errors are less
auto- and cross-correlated than the OLS series. These deviations are used as instruments in
the predictive stage of the model.

5.1.2. Short-term dynamics

Table 8 shows the in-sample estimates of the second, predictive, stage of the model. The OLS
and GLS coefficients, showing the sensitivity of spread changes to the instruments, have the
same sign, but are different in size. In both cases the coefficient measuring the speed of spread
reversion to fundamental value, α, is the most significant instrument - |t − statistic| = 6.23
(for OLS) and |t − statistic| = 4.83 (for GLS). The negative sign shows that spreads do indeed
revert to their equilibrium level. The estimated speed of reversion, α, is higher when estimated
with OLS (-0.26), suggesting that about one fourth of any credit spread misalignment is
corrected within the next month. The GLS estimate (-0.14) implies a slower mean-reversion.
The coefficients on the remaining two instruments, ’traditional instruments’, measure
the in-sample significance of global factors. The equity instrument is insignificant at the 95%
confidence level, while the bond instrument is significant but not as much as the deviation
from equilibrium.

14
The second stage of the model was also estimated allowing for country-specific sensitiv-
ities to the instruments. We find, however, that the restriction of common coefficients, cannot
be rejected and only those results are reported in the paper. These findings suggest that the
speed of correction of any spread misalignment, and hence informational inefficiency, is the
same across countries.

5.2. Out-of-sample performance

Several out-of-sample tests are performed to assess the robustness of predictability. First,
to evaluate its economic significance, we compare the realized returns of an active strategy
based on the predicted spread changes to a riskier buy-and-hold strategy. Second, we apply
Merton’s (1981) market-timing test to estimate the value-added of Brady bond predictability
to a US investor. This test, which is based on the number of correct directional forecasts,
ensures that results are not driven by the size of the returns in a few ’lucky’ periods. Third,
Henriksson and Merton (1981) nonparametric test evaluates the small sample robustness of
the market-timing results.
The out-of-sample period is 35 months long - from April 1998 to February 2001. We
estimate the model using an extended window methodology. The model parameters are es-
timated based on data from period 1 to period t, and a one-step-ahead forecast is made for
each individual country spread change in t + 1. With 35 months and 4 countries, we make
35 × 4 = 140 individual out-of-sample predictions about the direction of each credit spread
change. This period includes the Russian and Brazilian crises. We allow for the maximum
out-of-sample testing window given the limited time-series observations (95 months in total).
Extending the testing period further leaves insufficient observations to estimate the in-sample
parameters of the model (Park and Ogaki (1991) show that the SUCCR method performs
well in samples larger than 60 time-series observations).

5.2.1. Profitability of predictability

An active strategy is used to evaluate the economic significance of predictability based on


realized holding period returns during the out-of-sample period. For each country, the active
strategy is a 0-1 strategy which switches between Brady bonds and US treasury bills based
on the model’s one-step-ahead forecast of the credit spread change:
 
 100% EMBI if E(∆Spreadi,t+1 |Φt ) ≤ 0  ∀i = 1, ..., 4 and
i,t
Active = for
 100% cash(T-bills) if E(∆Spread 
i,t+1 |Φt ) > 0 ∀t = 0, ..., w − 1.

15
where w is the length of the out-of-sample testing window (35 months in this case) and Φ t is
the information available at time t. For each $100 invested in emerging markets, we assign
$25 to each country. It is 100% of each $25 in this particular country that is actively managed
based on country i’s forecast.
The benchmark relative to which we evaluate the active strategy is a passive strategy
equally weighted in the four Latin-American countries, i.e.:

P assive = 100% EMBIit for ∀i = 1, ..., 4 and ∀t = 0, ..., w − 1.

The logical question is why 100%. The goal is to assure that the passive strategy is at least
as risky as the active one. The active investor never holds more exposure to Brady bonds
than the passive investor. In fact, the risk of the passive strategy is a limiting case of that of
the active20 . Any superior returns from the active strategy are, therefore, risk-adjusted. Note
that the benchmark strategy does not imply that the investor holds only Brady bonds, nor
do we suggest that investors should change their global asset allocation. The issue here is, for
any amount allocated to Brady bonds, whether active management adds value. Transaction
costs are incorporated in the active strategy returns, as we are buying at the ask and selling
at the bid when rebalancing. We use a total return EMBI index, which includes capital gains
and distributions, to calculate the holding period return on the two strategies.
Table 9 summarizes the cumulative performance of the passive and active strategies
over the 35-month period out-of-sample testing window. Alternative sets of instruments are
examined to evaluate their relative predictive value. The equilibrium deviation estimated with
SUCCR, set (2), has the highest predictive power. It alone generates compounded returns
of 19.5% per year and a Sharpe ratio of 0.63, more than double the return of the riskier
passive investment (over 9% per year with Sharpe ratio of 0.18). This instrument shows
the combined predictive power of the two-stage model and the SUCCR estimator. The OLS
estimated correction, set (3), shows the value of the first stage, separating the effect of the
SUCCR procedure (an annual return of 15.3% and a Sharpe ratio of 0.41). Both estimates
confirm that the predictive power of the spread deviation is economically significant. The
global equity and bond instruments do not provide any predictive power, suggesting that
predictability is unlikely to be due to variations in risk premia. Their predictions generate an
annual return of 2.8% and a Sharpe ratio of 0.07, largely underperforming the buy-and-hold
strategy. Moreover, they reduce by 2.4% per year the market-timing returns of the deviation
instrument (compare set (1) to set (2)).
20
The active strategy will be as risky as the benchmark passive strategy only if for all periods and all
countries the model predicts a negative spread change. This is a possible, but unlikely scenario.

16
Figure 4 further illustrates the relative performance of the active strategy with alter-
native instruments. The deviation from equilibrium is clearly the most important instrument
and generates the most abnormal profits. An active investor using the SUCCR errors as in-
struments would have been able to generate almost 70% return on his Brady-bond investment
over a 35-month period, while a buy-and-hold investor would have generated less than 30%.
The evolution of the incremental wealth due to the use of the SUCCR rather than OLS is
illustrated in Figure 5. The pre-May 1999 slope is steeper than the one following, suggesting
that initially active management based on SUCCR generates wealth much faster than OLS
and then slows down in the second half of the sample. While the almost monotonic upward
slope in the graph suggests that SUCCR makes more correct directional forecasts than OLS
in up and down markets, its outperformance is more dramatic in the first half, when spreads
are more volatile (see Figure 1). SUCCR does better in periods of more volatile markets
as it accounts for contagion and it is exactly when active management is most important in
emerging markets.
The biggest gain from the active strategy comes from timing the Russian crisis cor-
rectly. Fama (1998) warns that using cumulative returns could artificially augment the out-
performance of a model as a single large positive or negative return will be compounded in
subsequent periods, even with no additional abnormal performance. As a result, Fama recom-
mends using average, rather than cumulative returns. The second column of Table 9 addresses
Fama’s concern and presents simple average returns (with no monthly compounding) of the
active and passive strategies. The results do not change. The two-stage model, coupled with
the SUCCR methodology, generates the highest average returns of 19.8% per year over the
out-of-sample period.
Figure 6 addresses concerns that the difference between the passive and active strate-
gies is driven by the choice of a particular starting point of investment. The figure illustrates
the relative performance of the strategies for different starting points. Each subsequent point
represents the cumulative wealth at the end of February 2001 assuming the initial $1 invest-
ment is shifted by a month (the X-axis represents the month of initial investment). The active
strategy using SUCCR always dominates regardless of the starting point. However, the gap
between the active and passive strategies narrows due to a shortening holding period and,
more importantly, to the fact that after August 1999 Brady markets have mostly gone up.
Given the way the strategies are defined (p. 15-16), correct market timing in up-markets
makes the active and passive strategies identical, both invested 100% in emerging markets.
Thus, in up-markets a successful active strategy does not ’pull ahead’ as in down markets. We
therefore turn to our next test, which focuses on the model’s ability to consistently generate

17
correct predictions about the direction of market prices.

5.2.2. The value of market-timing: Merton’s test

The ensure that our findings are not driven by a few ’lucky’ periods, we use a second market-
timing test, which is based on the number of correct predictions of up or down markets.
The tests depends only on the number of times the sign (not the size) of returns is correctly
predicted.
Merton (1981) develops an equilibrium theory for the value of market-timing skills and
shows an isomorphic correspondence between successful market-timing and free options on
the market. This correspondence is independent of investors’ preferences and prior probability
distributions and is based only on the manager’s ability to make one of two possible predictions:
a risky asset will either outperform or underperform the risk-free investment. This fits well
the specifics of this study as the active strategy is based on positive or negative spread change
forecasts (i.e., forecasts of negative or positive excess returns over US treasuries).
Merton demonstrates that a necessary and sufficient condition for market timing to be
valuable is that there be a significant number of correct forecasts in down and in up markets 21 ,
i.e. the conditional probabilities, p1 and p2 , to satisfy:

p1 + p 2 > 1 (7)

where p1 (p2 ) is the probability of correctly forecasting down (up) markets 22 . The larger p1 +p2 ,
the more valuable the forecast information is, as (p1 + p2 − 1) represents the percentage of a
free option that a market-timing model provides by correctly predicting market movements.
The results for the 140 (35 months × 4 countries) out-of-sample directional forecasts
of credit spread changes are reported in Table 10. In all four countries, the sum of the two
conditional probabilities exceeds one. By Merton’s argument, this is a necessary and sufficient
condition for our market-timing model to have positive value. This sum is 1.35 on average
for the four countries, implying that the model provides on average 35% of a free option on
the Brady bond market. The results in Table 10 are due to the following three instruments:
∆SpreadEM BI,t , ∆M SCIt , and ε!SU CCR,t .
Table 11 compares the value added of alternative sets of instruments. Using the equi-
librium deviation as an instrument produces the highest market-timing value. Global instru-
21
Merton shows that an unconditional probability of providing a correct forecast of more than 50% of the
times does not prove market-timing ability.
22
In terms of the present model, due to the inverse relation between price and spread, p 1 is the probability
of a correct forecast when actual spreads increase, while p2 is the probability of correctly forecasting a decrease
in spreads.

18
ments (∆SpreadEM BI,t , ∆M SCIt ) do not provide out-of-sample predictability in emerging
debt markets (Merton’s combined probabilities are 1.00 and 1.02, respectively). The use of
Park and Ogaki’s robust SUCCR estimator is partially responsible for the higher conditional
probability of timing the market correctly (Merton’s probability is 1.28 using ε!SU CCR,t alone
vs. 1.22 when using the OLS estimate ε!OLS,t ).

5.2.3. Nonparametric small-sample tests of market-timing

Given the unfortunately limited history of Brady markets, we evaluate the robustness of the
previous results to the sample size of this study. Henriksson and Merton (1981), henceforth
HM, derive small sample nonparametric tests for the significance of market timing. These
tests are independent of the distribution of security returns and provide a critical number
of correct predictions in up and down markets necessary to reject the null hypothesis of no
predictability for a given out-of-sample testing window. The smaller the sample size, the
higher the necessary percentage of correct predictions needed to reject the null.
Given the null hypothesis of no predictability derived in Merton’s (1981) study, i.e.

H0 : p 1 + p 2 = 1 (8)

HM show that the critical number of correct forecasts, x∗ (c), for rejecting the null of no
predictability is the solution to the following equation:
n1
) *) * ) *
" N1 N2 N
/ =1−c (9)
x=x∗ x n−x n
where N1 (N2 ) ≡ number of observations where ∆Spreadsit ≥ 0 (∆Spreadsit < 0), N ≡
N1 + N2 = total number of observations, n1 (n2 ) ≡ number of (un)successful predictions,
given ∆Spreadsit ≥ 0 (∆Spreadsit < 0), n ≡ n1 + n2 = number of times the forecast is
∆Spreadsit ≥ 0, c ≡ chosen confidence level, and n1 ≡ min(N1 , n) ≡ upper bound on correct
predictions of ∆Spreadsit ≥ 0. The null of no predictability will be rejected if the number
of correct forecasts in up markets is higher than the test critical value (n 1 ≥ x∗ (c)) for the
desired confidence level (1 − c) and given sample size (N ).
Table 12 presents the results from HM’s market-timing test. Global instruments do
not provide enough correct predictions to reject the null hypothesis of no predictability (p-
value of 46%). When the SUCCR estimate of the deviation, ε!SU CCR , is used as a predictor,
no predictability can be rejected at the 99% confidence level, while the OLS estimates allow
rejection with 97% confidence. HM’s test shows that our findings are robust to the size of our
out-of-sample window and confirms our previous findings that predictability in emerging debt
markets is genuine and driven by the credit spread’s deviation from its long-term equilibrium.

19
6. What is the source of predictability?
Our model reveals Brady bond spreads do not adjust instantaneously to new information.
We believe this is due to a combination of characteristics of the Brady market. It is a non-
transparent dealer market, dominated by large institutional investors with regulatory and
investment policy restrictions. We believe the lack of unrestricted investors and arbitrageurs,
arising from the large transaction lots and the lack of fully developed derivatives markets, are
key features that differentiate the Brady market from more informationally efficient dealer
markets such as the US treasury market.
Brady bond markets are dominated by large institutional investors, such as mutual,
endowment, and pension funds, due to a minimum transaction size of $2 million23 . Although
these investors actively manage their emerging market investments, they broadly follow a
benchmark, relatively over- or underweighting their exposure to a specific country (generally
with a limit of ±20% of the country’s benchmark weight24 ). Due to active management, credit
spreads experience price pressure when emerging country fundamentals change. Our results
show that over a period of a few months the credit spreads implied in market prices do change
to reflect the changing fundamentals. Yet it takes longer for the market to clear and spreads
to adjust as institutional money managers cannot drastically rebalance their portfolios in and
out of those countries, thus allowing for short-term predictability.
Large players also dominate US Treasury bond markets, where trades are usually in $1
million lots. Yet these markets are functioning efficiently. Inefficiencies in the Treasury bill
market are small, because they can be arbitraged (or quasi-arbitraged) away (see Rendleman
and Carabini (1979)). US interest rate derivatives are highly liquid, exchange traded, and
inexpensive. Moreover, the ability to strip coupon bonds into multiple zero coupon bonds
facilitates the pricing of Treasuries. Therefore no-arbitrage conditions prevent the Treasury
market from being informationally inefficient. Derivatives play a key role in improving the
pricing of risk by providing price discovery for the cash market.
Derivatives on emerging country sovereign credit exist but their market is still illiquid
due to the lack of a secondary derivatives market and to the illiquidity of the repo market
for Brady debt. Credit derivatives25 could offer efficiency gains in the Brady bond market by
enabling the separate pricing of credit risk. While the global credit derivatives market has
23
Source: Emerging Market Trade Association, http://www.emta.org/emarkets/.
24
For example, if the country’s benchmark weight is 10%, the manager’s weight should be between 8% and
12%. This information is based on interviews with institutional money managers.
25
Credit derivatives separate the credit risk from an underlying and enable investors to gain or reduce
exposure to credit risk.

20
grown from under $250 billion in 1997 to over $1.5 trillion in 2001, the emerging credit deriva-
tives market took off in late 1997. Despite their rapid growth, emerging market derivatives
currently account for only 1% of global derivatives26 . More importantly, the absence of a sec-
ondary market (emerging market credit derivatives are issued over-the-counter by banks like
DeutscheBank and JPMorgan), their lack of liquidity, and the need for hedging using the repo
market, are reducing the potential efficiency gains derived from their introduction. Credit
derivatives are constrained by the illiquidity of the Brady repo market because long default
swap positions are hedged by short positions in bonds. As a result, derivative premiums are
still quite expensive due to the hedging risks incurred by protection sellers. In addition, the
default swap27 , the derivative accounting for 85% of notionals, is a default-triggered derivative
whose valuation has been shown by Chen and Sopranzetti (2003) to have little correlation with
changes in credit spreads, thus offering poor hedging potential in the case of no default.
The illiquidity of emerging market derivatives is exacerbated by their limited use by
investors due to regulatory and investment policy restrictions. A 1998 survey by Levich,
Hayt, and Ripston (1999) of derivatives usage among US institutional investors shows that
only 46% of institutions are allowed to use derivatives and only 27% actually use them due
to excessive capital requirement28 or investment policy restrictions. More importantly, while
more than 83% of institutions are allowed to use US interest rate derivatives, only 40% are
allowed to use emerging market bond derivatives and only 20% actually use them. Where
derivatives are used, the positions are small relative to total assets (the mode being 1% of
total assets). Further, the principal reasons for using derivatives are risk reduction (55%) and
asset allocation (26%), rather than market timing (15%).
The lack of pre- and post-trade transparency further reduces the speed of price discov-
ery in the Brady market. Brady bonds trade in over-the-counter markets composed of brokers,
dealers, and investors worldwide, linked informally through a network of broker screens. Deal-
ers’ bids and offers are anonymous. Actual trading is conducted orally through inter-dealer
brokers. The identity of the broker’s counterparties are not revealed even after the trade 29 .
Research shows that while such lack of transparency typically leads to higher liquidity (as
traders are unwilling to reveal their intentions and market makers can more easily dispose
of large inventories), it is generally associated with less informative prices (see Bloomfield
26
Source: International Monetary Fund, ”Selected Topic: The Role of Financial Derivatives in Emerging
Markets, Global Financial Stability Report, Market Developments and Issues”, December 2002.
27
A credit default swap is a financial contract under which the protection buyer pays a periodic fee in return
for a payment by the protection seller contingent on the occurrence of default.
28
Current bank regulations require that banks hedging positions via credit swaps reserve capital against
both the loan and the derivative contract, rather than netting the position. Source: DerivativesStrategy.com.
29
Source: Emerging Market Trade Association, http://www.emta.org/emarkets/.

21
and O’Hara (1999), Gemmill (1996), O’Hara (2003), Porter and Weaver (1998), and Simaan,
Weaver, and Whitcomb (2003)).

7. Conclusion
Using a new two-stage model for credit spreads, we present evidence of significant predictabil-
ity in the largest, most accessible, and liquid emerging debt market. An active strategy based
on this model provides Brady-bond investors with returns twice as large as those of a riskier
buy-and-hold strategy. Merton’s (1981) and Henriksson and Merton’s (1981) statistical tests,
based on the relative number of correct forecasts, confirm that the market-timing profits in
this market are economically and statistically significant. The observed predictability provides
US investors with the equivalent of free options on Brady bond indexes.
The results suggest that the two-stage model captures well the credit risk structure
of emerging sovereign debt markets. The strong long-term equilibrium relation between the
level of credit spreads (default premiums) and local macroeconomic conditions (fundamental
risk) in emerging debt markets suggests market rationality as the information gets fully re-
flected in market prices. Local factors explain 90% of the variation in credit spreads, leaving
little to global factors and residual variance. These findings are consistent with Erb, Harvey,
and Viskanta (2000), Cumby and Pastine (2001), and Claessens and Pennacchi (1996) who
view emerging market volatility as largely idiosyncratic. Yet prices and spreads fail to react
instantaneously to new information, giving rise to the documented predictability.
We find that global instruments do not have genuine out-of-sample predictive power,
suggesting that time-varying risk or time-varying risk-premia are unlikely explanations for
the documented predictability. The predictability can be attributed to the use of the spread’s
deviation from its long-term equilibrium as an instrument, implying market inefficiencies. We
believe that the absence of unrestricted investors and arbitrageurs, due to the large transaction
lots and the lack of fully developed derivatives markets, coupled with the lack of pre- and post-
trade transparency, are key features that differentiate the Brady market from other more
informationally efficient bond markets with large transactions sizes such as the US treasury
market.

22
A. The SUCCR methodology
The appendix presents the details of the SUCCR procedure using Park and Ogaki’s (1991)
original notation.
The system of regressions is given by:

y1t = x"1t β1 + u1t


..
.
yit = x"it βi + uit (A1)
..
.
ynt = x"nt βn + unt

where i = 1, ..., n is the cross-sectional dimension, while t = 1, ..., T is the time-series dimen-
sion. yit is a scalar and the dependent variable in each regression. Each xit is an si × 1 vector
of regressors, and si is the number of regressors in regression i. Each βit is an si × 1 vector
of sensitivities to the regressors. uit is the error term of regression i and is assumed to be
stationary.
The system can also be rewritten in matrix format as:

y = Xβ + u (A2)

where for i = 1, ..., n.

yi = (yi1 , ...yiT )" (A3)


Xi = (xi1 , ...xiT )" (A4)
ui = (ui1 , ...uiT )" (A5)

and

y = (y1" , ..., yn" )" (A6)


β = (β1" , ..., βn" )" (A7)
u = (u"1 , ..., u"n )" (A8)
X = block-diagonal(Xi ) (A9)

All {xit } are integrated processes of order one, which can assume deterministic trends.
Park and Ogaki specify the regressors in the following three different ways:

M(a) : xit = x0it (A10)

23
M(b) : xit = πi pit + x0it (A11)
M(c) : xit = (p"it , q"it ), qit = πi pit + x0it (A12)
M(d) : xit = pit (A13)

where {pit } is a general deterministic trend and {x0it } is a purely stochastic integrated process.
With M(a), each regression represents cointegration represents cointegration in the sense
of Engle and Granger ’87. Since {xit } do not have any deterministic components, neither would
{yit } for the relation in eq. (A1) to hold.
Under M(b), both {xit } and {yit } contain the deterministic trend {pit }. Park and Ogaki
show that the i’th relation in eq. (A1) in this case is stronger than with M(a).
When {xit }’s dynamics are described by M(c), {yit } may or may not have a deterministic
component. The inclusion of {pit } in the regression effectively detrends both series. The
cointegrating relation is between the stochastic components of {yit } and {xit } only, which the
authors call stochastic cointegration.
The SUCCR methodology is general enough to allow stationary regressors as in M(d)
in addition to regressors as in M(a)-M(c). When all {xit } are as in M(d), the procedure is
reduced to the usual SUR. In the present paper, all our regressors are nonstationary, hence
case M(d) is does not apply.
The long-run relations in eq. (A1) with any specification M(a)-M(c) are testable through
tests of cointegration.
The vector of stationary processes driving the system is defined as {wit }, where:

wt = (u"t , ∆x0"
t )
"
(A14)

where ut = (u1t , ..., unt )" and ∆x0t = (∆x0" 0" "
1t , ..., ∆xnt ) . This process has a covariance structure
given by:  
T
) *) T
*"
1 " " Ω11 Ω12
Ω = lim E wt wt =  (A15)
T →∞ T Ω21 Ω22
t=1 t=1

where the partition is made conformable with that of {wt } in (A14). Ω is the longrun variance
of {wt }. The usual, shortrun, variance of {w} is given by:
 
T
1" Σ11 Σ12
Σ = lim E (wt wt" ) =   (A16)
T →∞ T Σ Σ
t=1 21 22

with partition similar to that of Ω. When {wt } is a martingale difference sequence Ω = Σ.


Note, that the author sometimes denote the variance of the errors in eq. (A1), Σ 11 and Ω11 ,
as Σ0 and Ω0 .

24
The longrun variance30 :
Ω = Σ + Γ + Γ" (A17)

where q ) *
" a
Γ= 1− wt−a wt" (A18)
a=1 q+1
where q is the number of lags considered in the autocorrelation.
Park and Ogaki assume Ω22 > 0 which implies neither redundant nor cointegrated
variables in {xt }. The requirement is that there exist a single cointegrating relation among
the unit root regressors and the dependent variable in the model. Park and Ogaki (1991)
argue that having more than one cointegrating equation (say k cointegrated vectors) will not
add information to the system, as some (k − 1) variables are redundant in the sense that they
are stationary combinations of the remaining factors. Having more than one cointegrating
vectors in eq. (A1) prevents the coefficient vector βi from being uniquely determined.
To comply with this SUCCR requirement, we first establish the number of cointegrating
vector per country (say k)and then exclude (k−1) of macroeconomic factors. We keep the ones
that produce the lowest AIC value. One can use some other criterion for exclusion without
significantly changing the estimation results.
Park (1992) provides a way to efficiently estimate an equation with integrated regressors.
The estimator is called Canonically cointegrating regressions, or CCR, represented by:

β!CCR = (X"∗ X∗ )−1 X"∗ y∗ (A19)

This statistical procedure adjust for the asymptotic bias introduced by the unit roots by
modifying the regressors:

yit∗ = yit − ω12


i
Ω−1 0 " −1
22 ∆xt − βi Λi Σ wt (A20)
x∗it = xit − Λi Σ−1 wt
i
where ω12 is the i-th row of Ω12 ,
T
1"
Λi = lim E(xit wt" ) (A21)
T →∞ T
t=1

while the other variables are defined before. A consistent estimate of Λ can be obtained as:

Λ = Γ 2 + Σ2 (A22)

where Γ ≡ Γ1 + Γ2 and Σ ≡ Σ1 + Σ2 where the partition is made conformable with wt =


(u"t , ∆x0" " 0" 0"
t ) . Λi are the columns of Λ corresponding to the regressors ∆xit in ∆xt .
30
See Hamilton (1994) for Newey-West adjustment for autocorrelated errors.

25
The cointegrating relations in each regression in (A1) continue to hold as all transfor-
mation in A20 involve only stationary terms (hence the term canonically cointegrated). The
regressors are transformed in such a way that the usual least squares procedure yields both
efficient estimates and chi-square tests on the coefficients.
The SUCCR system can be rewritten as:


y1t = x"1t β1 + u∗1t
..
.
yit∗ = x∗" ∗
it βi + uit (A23)
..
.

ynt = x∗" ∗
nt βn + unt

where
u∗it = uit − ω12
i
Ω−1 0
22 ∆xt (A24)

such that
T
1"
lim E(x∗t u∗"
t ) = 0 (A25)
T →∞ T
t=1

which implies that the SUCCR errors are asymptotically independent of the regressors. From
(A24), we have
u∗t = ut − Ω12 Ω−1 0
22 ∆xt (A26)

Defining y∗ , X∗ , and u∗ similarly to (A2), we can rewrite the SUCCR model in matrix
format:
y∗ = X∗ β + u∗ (A27)

Park and Ogaki (1991) extend the CCR methodology which was developed for a single
cointegrating regression. The SUCCR estimator is the modified system GLS estimator using
the longrun variance of the SUCCR errors {u∗t }.:

Ω∗ = Ω11 − Ω12 Ω−1


22 Ω21 (A28)

Park and Ogaki (1991) SUCCR estimator is given explicitly as:

β!SU CCR = (X"∗ (Ω


! −1 ⊗ I)−1 X )−1 X" (Ω
∗ ∗ ∗
! −1 ⊗ I)−1 y
∗ ∗ (A29)

The SUCCR estimator generalizes Park’s (1992) CCR estimator by using system infor-
mation in the same was as SUR

β!SU R = (X" (Σ
! −1 ⊗ I)−1 X)−1 X" (Σ
! −1 ⊗ I)−1 y (A30)

26
generalizes the usual OLS estimator

β!OLS = (X" X)−1 X" y (A31)

for stationary panels.


Yet the authors show that the usual GLS estimator can increase the asymptotic bias
of OLS in a general SUCR system as in A27, as the usual GLS (or SUR) uses the shortrun
variance of the errors, Σ0 , not the longrun variance, Ω0 . The shortrun variance does not
correct for autocorrelation in the errors in (A1).
As the CCR estimator as in (A19), the SUCCR methodology in (A29) corrects for biases
and nuisance parameters introduced by using OLS or SUR in a SUCR system as in (A1). It
does so using the same stationary adjustment to the variables in the model as in (A20).
The SUCCR estimator in (A29) improves OLS, SUR, and CCR by correcting all prob-
lems of a nonstationary cross-correlated panel.
Tests can be conducted on a general hypothesis:

H0 = ϕ(β) (A32)

where ϕ is assumed to be continuously differentiable with first derivative Φ evaluated at the


true value of β. With q restrictions under H0 , the test statistic is the same as in the standard
SUR, except that the longrun variance is used:

D
ϕ(β!SU CCR )" (Φ(X"∗ (Ω−1 −1 " −1 ! 2
∗ ⊗ I)X∗ ) Φ ) ϕ(βSU CCR ) −→ χq (A33)

For a simple t-test on the coefficient, ϕ(β!SU CCR ) = diag(β!SU CCR ), hence Φ = I, the
above can be simplified to:

diag(β!SU CCR )" (I" (X"∗ (Ω−1 −1 −1 !


∗ ⊗ I)X∗ ) I) diag(βSU CCR )

= diag(β!SU CCR )" (I" (X"∗ (Ω−1 −1 −1 !


∗ ⊗ I)X∗ ) I) diag(βSU CCR )

= diag(β!SU CCR )" ((X"∗ (Ω−1 −1 −1 !


∗ ⊗ I)X∗ ) ) diag(βSU CCR )
D
= diag(β!SU CCR )" (X"∗ (Ω−1 ! 2
∗ ⊗ I)X∗ )diag(βSU CCR ) −→ χq (A34)

27
B. SUCCR simulation results
A Monte Carlo simulation exercise is performed to assess the relative merit of the SUCCR
estimation methodology in small samples. I simulate a multivariate error-correction process
with parameters similar to the ones estimated for the variables in this study. The cointegrating
vector is then estimated using the five estimation techniques mentioned in the paper: OLS,
SUCCR, SUR, SCCR, and MSUR. The goal of this simulation is to analyze the improvement
in small sample bias of SUCCR over the remaining methods. I simulate series with 95 and
150 time periods to assess the sensitivity of the SUCCR methodology to changes in the length
of the estimation window. The results show that SUCCR produces the smallest small sample
bias relative to the four other methods. The SUR and MSUR estimates exhibit the maximum
bias. The SUCCR improvement becomes even more apparent as the sample size increases
from 95 to 150 time periods.
The simulation results reported in Tables 13 and 14 are based on a system consisting of
four equations (four countries) and four explanatory variables, Xit , each. The cointegrating
equations for each country are all standardized to be βi = (βi1 , βi2 , βi3 , βi4 )" = 1, where
i = 1, 2, 3, 4.
An error-correction process is simulated such that
   
∆yit +Yit
  = A (εt−1 ) +   (B1)
∆Xit +X
it

where yit is a scalar and the dependent variable in our model, Xit is a 4 × 1 vector of ex-
planatory variables in each country, εt−1 = (ε1,t−1 , ε2,t−1 , ε3,t−1 , ε4,t−1 )" , a 4 × 1 vector, rep-
resents the deviation from the equilibrium cointegrating equation for each country i, i.e.
+Yit
εi,t−1 = (yi,t−1 − βi Xi,t−1 ), βi is the 4×1 cointegrating vector of each country i. X ∼ N (0, Σ)
+it
and
   
αY 0 0 0  ρ"0 0 0 0
 4  
   
 0
 α4Y 0 0 



0 ρ"0 0 0 

   R1 

 0 0 α4Y 0 


0 0 ρ"0 0 

   
 0 0 0 Y 
α4   0 0 0 ρ"0 
A=

 

 ,Σ = 
 ρ 0 0 0



(B2)
 α1X 0 0 0   0 R3 R2 R2 R2 
   
   
 0 α1X 0 0   0 ρ0 0 0 R2 R3 R2 R2 
   

 0 0 α1X 0 

 0 0 ρ0 0 R2 R2 R3 R2 

   
0 0 0 α1X 0 0 0 ρ0 R2 R2 R2 R3

where αiY , a scalar, is the error correction coefficient of the dependent variable of country i.

28
In the current simulation exercise, we set α1Y = α2Y = α3Y = α4Y = −0.2, −0.1, or 0.2. We try
different specification to analyze the sensitivity of the SUCCR estimation to changes in the
process parameters. αiX , a 4 × 1 vector, is the error correction coefficient of the dependent
variable of country i. As in Park and Ogaki (1991), we set α1X = α2X = α3X = α4X =
(0.2, 0.2, 0.2, 0.2)" . ρ0 = (ρ0 , ρ0 , ρ0 , ρ0 )" . The blocks of the covariance matrix are as follows:
     
1 ρ 1 ρ1 ρ1 ρ2 ρ2 ρ2 ρ2 1 ρ 3 ρ3 ρ3
     
     
 ρ1 1 ρ 1 ρ1   ρ ρ2 ρ2 ρ2   ρ 1 ρ 3 ρ3 
R1 =   , R2 =  2  , R3 =  3  (B3)
     
 ρ1 ρ1 1 ρ1   ρ2 ρ2 ρ2 ρ2   ρ3 ρ3 1 ρ 3 
     
ρ1 ρ1 ρ1 1 ρ2 ρ2 ρ2 ρ2 ρ3 ρ3 ρ3 1

The covariance matrix assumes the error components are standard normal random
variables and the following correlations among the variables:

ρ0 = the correlation among changes in the dependent and independent variables of country
i, ∆yit and ∆Xit ,

ρ1 = the correlation among changes in the dependent variables ∆yit and ∆yjt for i *= j;

ρ2 = the correlation among changes in the independent variables across countries, ∆Xit and
∆Xjt for i *= j;

ρ3 = the correlation among changes in the independent within a country, ∆Xitk and ∆Xitj ;

The results in Tables XIII and XIV are based on 2,000 iterations.
Note that as the sample size increases from 95 to 150 time periods, the SUCCR small
sample bias becomes noticeably smaller than the remaining methods. SCCR is the next best
method to use for a cointegrated and cross-country correlated multivariate process. The SUR
method which utilizes the system information without correcting for cointegration does worst
of all methods. OLS performs relatively well, especially with 95 series observations, but lags
behind SUCCR in the 150 sample size experiment.
Increasing the sample size from 95 to 150 observations does not significantly improve
the bias problem in the OLS, SUR, and MSUR methods. In some cases the bias even in-
creases. With the SUCCR and SCCR techniques, however, the improvement is significant
and always positive. Park and Ogaki’s (1991) simulation results find even larger improvement
with sample size of 300 time-series observations. These results are promising and confirm the
initial expectation that the value of the SUCCR methodology should increase with time and
should find more applications in other financial areas as well.

29
References
Adler, M. and B. Dumas, 1983, ” International portfolio selection and corporation finance:
a synthesis,” Journal of Finance, 46, 925-984.

Ang, A. and G. Bekaert, 2001, ” Stock return predictability: is it there?” working paper,
Columbia University, New York.

Avramov, D., 2002, ”Stock return predictability and model uncertainty,” Journal of
Financial Economics, forthcoming.

Avramov, D. and T. Chordia, 2003, ”Stock returns are predictable: A firm level analysis,”
working paper, University of Maryland and Emory University.

Barnhill, T., Joutz, F., and W. Maxwell, 2000, ”Factors affecting the yields on nonin-
vestment grade bond indices: a cointegration analysis,” Journal of Empirical Finance, 7,
57-86.

Bekaert, G., 1995, ”The time-variation of expected returns and volatility in foreign ex-
change markets,”, Journal of Business and Economic Statistics, 13, 397-408.

Bekaert, G. and C. Harvey, 1995, ”Time-varying world market integration”, Journal of


Finance, 50, 403-444.

Bekaert, G. and R. Hodrick, 1992, ”Characterizing predictable components in excess


returns on equity and foreign exchange markets,” Journal of Finance, 47, 467-511.

Boehmer, E. and W. Megginson, 1990, ”Determinants of secondary market prices for


developing country syndicated loans,” Journal of Finance, 45, 1517-1540.

Bloomfield, R. and M. O’Hara, 1999, ”Market transparency: Who wins and who loses?”,
Review of Financial Studies, 12 (1), 5-35.

Bookstaber, R. and D. Jacobs, 1986, ”The composite hedge: controlling the credit risk
of high-yield bonds,” Financial Analysts Journal, 42, 25-35.

Chen, R.R. and B.J. Sopranzetti, 2003, ”The valuation of default-triggered credit deriva-
tives”, Journal of Financial and Quantitative Analysis, forthcoming.

Claessens, S. and G. Pennacchi, 1996, ”Estimating the likelihood of Mexican default from
the market prices of Brady bonds,” Journal of Financial and Quantitative Analysis, 31,
109-126.

Cumby, R. E. and T. Pastine, 2001, ”Emerging market debt: measuring credit quality and
examining relative pricing,” Journal of International Money and Finance, 20, 591-609.

Duffie, D., Pedersen, L., and K. Singleton, 2003, ”Modeling sovereign yield spreads: A
case study of Russian debt,” Journal of Finance, forthcoming.

Dumas, B. and B. Solnik, 1995, ”The world price of foreign exchange risk,” Journal of
Finance, 50, 445-479.

30
Eichengreen, B. and A. Mody, 1998, ”What explains changing spreads on emerging mar-
ket debt: fundamentals or market sentiment?” NBER Working Paper Series, working
paper 6408.

Engle, R. F. and C.W.J. Granger, 1987, ”Co-integration and error correction: represen-
tation, estimation, and testing,” Econometrica, 55, 251-276.

Erb, C., Harvey, C., and T. Viskanta, 2000, ”Understanding emerging market bonds,”
Emerging Markets Quarterly, Spring 2000, 1-17.

Fama, E., 1981, ”Stock returns, real activity, inflation, and money,”, The American
Economic Review, 71, 545-565.

Fama, E., 1990, ”Stock returns, expected returns, and real activity,” Journal of Finance,
45, 1089-1108.

Fama, E., 1991, ”Efficient capital markets: II,” Journal of Finance, 46, 1575-1617.

Fama, E., 1998, ”Market efficiency, long-term returns, and behavioral finance,” Journal
of Financial Economics, 49, 283-306.

Fama, E. and K. French, 1988, ”Permanent and temporary components of stock prices,”
Journal of Political Economy, 96, 246-273.

Fama, E. and K. French, 1989, ”Business conditions and expected returns on stocks and
bonds,” Journal of Financial Economics 19, 3-29.

Fama, E. and G.W. Schwert, 1977, ”Asset returns and inflation,” Journal of Financial
Economics, 5, 115-146.

Ferson, W., 1989, ”Changes in expected security returns, risk, and the level of interest
rates,” Journal of Finance, 44, 1191-1217.

Ferson, W. and C. Harvey, 1991, ”The variation in economic risk premiums”, Journal of
Political Economy, 99, 385-415.

Ferson, W. and C. Harvey, 1993, ”The risk and predictability of international equity
returns,” The Review of Financial Studies, 6, 527-566.

Ferson, W. and C. Harvey, 1999, ”Conditioning variables and the cross-section of stock
returns,” Journal of Finance, 54, 1325-1360.

Ferson, W., S. Sarkissian, and T. Simin, 2003, ”Spurious regressions in financial eco-
nomics?”, Journal of Finance, 58, 1393-1414.

Gemmill, G., 1996, ”Transparency and liquidity: A study of block trades on the London
Stock Exchange under different publication rules”, Journal of Finance, 51 (5), 1765-1790.

Geske, R. and R. Roll, 1983, ”The fiscal and monetary linkage between stock returns and
inflation,” Journal of Finance, 38, 1-33.

Hamilton, J., 1994, Time Series Analysis, Princeton University Press, Princeton.

31
Harvey, C., 1991, ”The world price of covariance risk,” Journal of Finance, 46, 111-157.

Harvey, C., 1993, ”Risk exposure to a trade-weighted index of currency investment,”


working paper, Duke University.

Harvey, C., 1995, ”Predictable risk and returns in emerging markets,” The Review of
Financial Studies, 8, 773-816.

Hassan, K. M., 2001, ”Portfolio investment of the OIC countries and their implications
on trade,” working paper, University of New Orleans.

Hayt, G. and R. Levich, 1999, ”Who Uses Derivatives?”, Risk (August).

Henriksson, R. and R. Merton, 1981, ”On market timing and investment performance.
II. Statistical procedures for evaluating forecasting skills,” Journal of Business, 54, 513.

Ilmanen, A., 1995, ”Time-varying expected returns in international bond markets,” Jour-
nal of Finance, 50, 481-506.

International Monetary Fund, ”Selected topic: The role of financial derivatives in emerg-
ing markets, Global financial stability report, Market developments and issues”, Decem-
ber 2002.

Johansen, S., 1988, ”Statistical analysis of cointegrating vectors,” Journal of Economic


Dynamics and Control, 12, 231-254.

Johansen, S., 1989, ”Estimating and hypothesis testing of cointegrating vectors in Gaus-
sian vector autoregressive models,” working paper, University of Copenhagen, Copen-
hagen.

Keim, T. and R. Stambaugh, 1986, ”Predicting returns in the stock and the bond mar-
kets,” Journal of Financial Economics, 17, 357-390.

Levich, R., G. Hayt, and B. A. Ripston, 1999, ”1998 Survey of derivatives and risk man-
agement practices by U.S. institutional investors”, Working Paper, New York University.

Lewellen, J., 1999, ”The time-series relations among expected return, risk, and book-to-
market,” Journal of Financial Economics, 54, 5-43.

Lo, A. and A. C. MacKinlay, 1996, ”Maximizing predictability in the stock and bond
markets,” Working paper LFE-1019-96, MIT Laboratory for Financial Engineering.

Madhavan, A., 2000, ”Market microstructure: A survey”, Working Paper, University of


Southern California.

Merton, R., 1981, ”On market timing and investment performance. I. An equilibrium
theory of value for market forecasts,” Journal of Business, 54, 363-406.

Nelson, C. and M. Kim, 1993, ”Predictable stock returns: The role of small sample bias,”
Journal of Finance, 48, 641-661.

32
O’Hara, M., 2003, ”Presidential address: Liquidity and price discovery”, Journal of Fi-
nance, 58, 1335-1354.

Park, J., 1992, ”Canonical cointegrating regressions,” Econometrica, 60, 119-143.

Park, J. and M. Ogaki, 1991, ”Seemingly unrelated canonical cointegrating regressions,”


working paper, The Rochester Center for Economic Research, Rochester, New York.

Park, J., and P. Phillips, 1988, ”Statistical inference in regressions with integrated pro-
cesses: part 1,” Econometric Theory, 4, 468-498.

Park, J. and P. Phillips, 1989, ”Statistical inference in regressions with integrated pro-
cesses: part 2,” Econometric Theory, 5, 95-132.

Patelis, A., 1997, ”Stock return predictability and the role of monetary policy,” Journal
of Finance, 52, 1951-1972.

Pedroni, P., 1996, ”Fully modified OLS for heterogeneous cointegrated panels and the case
of purchasing power parity,” working paper, Indiana University, Bloomington, Indiana.

Phillips, P.C.B., 1988, ”Spectral regression for cointegrated time series,” working paper,
Cowles Foundation Discussion 872, Yale University, New Haven, Connecticut.

Phillips, P.C.B., 1991, ”Optimal inference in cointegrated systems,” Econometrica, 59,


283-306.

Phillips, P.C.B. and S. Durlauf, 1986, ”Multiple time series with integrated variables,”
Review of Economic Studies, 53, 473-496.

Phillips, P.C.B. and B.E. Hansen, 1990, ”Statistical inference in instrumental variables
regression with I(1) processes,” Review of Economic Studies, 57, 99-126.
more transparent?” Working paper, Marquette University.

Porter, D. and D. Weaver, 1998, ”Post-trade transparency on Nasdaqs national market


system”, Journal of Financial Economics, 50, 231-252.

Ramaswami, M., 1991, ”Hedging the equity risk of high-yield bonds,” Financial Analysts
Journal, 47, 41-50.

Ranciere, R., 2001, ”Credit derivatives in emerging markets”, IMF Policy Discussion
Paper.

Rendleman, R. and C. Carabini, 1979, ”The efficiency of the Treasury bill futures mar-
ket,” Journal of Finance, 34, 895-914.

Sercu, P., 1980, ”A generalization of the international asset pricing model,” Finance,
91-135.

Shane, H., 1994, ”Comovements of low-grade debt and equity returns of highly leveraged
firms,” Journal of Fixed Income, 3, 79-89.

33
Shiller, R., 1984, ”Stock prices and social dynamics,” Brookings Papers on Economic
Activity, 2, 457-498.

Simaan, Y., D. G. Weaver, and D.K. Whitcomb, 2003, ”Market maker quotation behavior
and pretrade transparency”, Journal of Finance, 58, 1247-1267.

Solnik, B., 2000, International Investments, fourth edition, Addison Wesley Longman,
Inc.

Stambaugh, R., 1999 , ”Predictive regressions,” Journal of Financial Economics, 54,


375-421.

Stock, J. and M. Watson, 1991, ”Stochastic trends and economic fluctuations,” The
American Economic Review, 81, 819-840.

Zellner, A., 1962, ”An efficient method of estimating seemingly unrelated regressions and
tests for aggregation bias,” Journal of the American Statistical Association, 57, 348-368.

34
Table 1
Local Macroeconomic Factors
In Long-Term Equilibrium Stage of the Model
The table presents all macroeconomic variables, Fi,t , used in the long-term equilibrium stage
of the credit risk model:

Spreadi,t = Bi Fi,t + εi,t

These variables are extracted for each country - Argentina, Brazil, Mexico, and Venezuela -
from the sources provided below. Notation: IIF = Institute of International Finance, IMF =
International Monetary Fund, MSDW = Morgan Stanley Dean Witter.
Local Macroeconomic Factors Sources

National Stock Market Index Bloomberg (provided by MSDW)


Real Exchange Rate Index IMF, IIF databases
Consumer Price Index [CPI] IMF, IIF databases
Local Short-Term Interest Rates IMF database
Money Supply IMF database
Unemployment Rate IMF database
Real Gross Domestic Product [GDP] IMF, IIF databases

Table 2
Sample Correlations
among Sovereign Spread Levels
The table presents the correlations among the World and Latin Emerging Market Bond Index [EMBI]
Spreads and the EMBI Spreads of Argentina, Brazil, Mexico, and Venezuela. The correlations are
estimated using end-of-month spread observations from April 93 to February 2001.

EMBI World Latin Argentina Brazil Mexico Venezuela


World 1 0.99 0.91 0.93 0.86 0.88
Latin 1 0.94 0.92 0.91 0.89
Argentina 1 0.80 0.83 0.81
Brazil 1 0.75 0.84
Mexico 1 0.74

35
Table 3
Descriptive Statistics of
Emerging Market Bond Index Spreads
Emerging Market Bond Index [EMBI] Spreads of Argentina, Brazil, Mexico, Venezuela, and the
World EMBI index in basis points. The EMBI spread indexes are computed as the spread above the
US treasury spot curve that equates the discounted cash flows on the Brady bond with the current
bond price. The sample period is from April 1993 to February 2001. The sample is limited by the
history of Argentina, which issued Brady debt in April 1993.

Statistic Argentina Brazil Mexico Venezuela World EMBI


Mean 785 821 654 1,093 820
Median 769 780 531 959 800
Standard Deviation 277 278 288 550 281
Kurtosis 0.38 -0.27 0.62 0.65 -0.52
Skewness 0.66 0.46 1.11 0.79 0.36
Range 1,302 1,227 1,371 2,897 1,205
Minimum 329 362 266 264 350
Maximum 1,631 1,589 1,637 3,161 1,555
Count 95 95 95 95 95

Table 4
Descriptive Statistics of
Emerging Market Bond Index Spread Changes
Emerging Market Bond Index [EMBI] spread changes of Argentina, Brazil, Mexico, Venezuela, and
the World EMBI Index in basis points. The EMBI spread indexes are computed as the spread above
the US treasury spot curve that equates the discounted cash flows on the Brady bond with the
current price of the bond. The sample period is from April 1993 to February 2001 (1 observation
is lost due to differencing). The sample is limited by the history of Argentina, which issued Brady
debt in April 1993.

Statistic Argentina Brazil Mexico Venezuela World EMBI


Mean 0.33 -4.00 -0.17 -0.35 0.60
Median -22.52 -28.67 -15.33 -23.13 -24.00
Standard Deviation 159 154 128 333 139
Kurtosis 9.99 9.00 4.63 26.28 14.83
Skewness 2.05 2.08 1.17 2.99 2.71
Range 1,290 1,190 924 3,586 1,120
Minimum -398 -365 -387 -1,279 -259
Maximum 892 825 538 2,307 861
Count 94 94 94 94 94

36
Table 5
Descriptive Statistics of Macroeconomic Factors
in the Long-Term Equilibrium
All variables from Table 1 were considered as possible factors in the equilibrium relationship between
spreads and macroeconomic activity:

Spreadi,t = Bi Fi,t + εi,t

The SUCCR procedure, used to estimate the parameters of the above equation, imposes a restriction
on the number of variables to include in each country. The restriction is that there exist a single
equilibrium (cointegrating equation) between spreads and the chosen factors. As a result some
variables are excluded from a specific country regression to satisfy this restriction (see Appendix
A for more details). Descriptive statistics are provided for those variables that remain in the final
country equation.

Country Factor Mean Median St.Dev Skew Kurt Min Max


Argentina Stock Market Index 157.89 158.15 25.68 -0.61 -0.07 100.00 212.37
Real Exchange Rate 98.36 98.14 6.24 -1.01 -0.28 84.45 107.98
Real GDP 150.28 152.79 10.57 -1.55 -0.20 130.39 165.40
CPI 99.63 100.31 2.34 1.51 -1.49 91.86 102.37
Brazil Stock Market Index 7,872 7,011 5,026 -0.87 0.22 21.04 18,055
Unemployment 6.02 5.61 1.18 -1.27 0.32 4.05 8.10
Money Supply 23,462 24,450 13,986 -1.08 -0.22 42.65 46,459
CPI 1,200 1,393 486.63 1.10 -1.50 10.16 1,701
Mexico Stock Market Index 3,906 3,647 1,611 -0.91 0.45 1,550 7,473
Real Ex. Rate 105.45 106.54 15.24 -0.57 -0.43 66.76 130.95
Money Supply 87.28 79.12 44.25 -0.51 0.74 35.13 207.65
Local Interest Rates 18.34 15.28 10.54 3.07 1.73 7.17 57.45
Venezuela Stock Market Index 4,287 4,452 2,665 -1.04 0.23 1,000 10,489
Real Exchange Rate 115.54 106.84 31.82 -1.44 0.25 66.14 169.26
Real GDP 143.57 142.19 6.30 -0.59 0.53 132.41 158.43
CPI 97.02 90.61 66.94 -1.42 0.22 11.90 209.85

37
Table 6
SUCCR Estimates
of the Long-Term Equilibrium Stage for
The Long-Term Equilibrium Stage (stage 1 of the model) describes the relationship between the
credit spread of a country and local macroeconomic factors:

Spreadi,t = Bi Fi,t + εi,t

The factors are reported with a two-to-three week lag. We lag them by a month to make them
contemporaneous with the credit spreads in terms of information arrival.
This stage is estimated with Park and Ogaki’s (1991) Seemingly Unrelated Canonically Cointegrating
Regressions [SUCCR] methodology. The SUCCR estimator

β!SU CCR = (X"∗ (Ω


! −1 ⊗ I)−1 X∗ )−1 X" (Ω
∗ ∗
! −1 ⊗ I)−1 y∗

and notation are described in details in Appendix A.


All tests of SUCCR coefficients are χ2 tests. All t−statistics are positive as they are the square-roots
of χ2 statistics.
The table reports the estimation results for the entire period April 1993 to February 2001.

Argentina Brazil
Macroeconomic Factor Coefficient t-statistic Macroeconomic Factor Coefficient t-statistic
Stock Market Index -7.67 19.01 Stock Market Index -0.08 6.88
Real Exchange Rate 6.89 10.77 Unemployment 138.94 9.23
Real GDP -5.14 12.05 Money Supply 0.01 2.71
CPI 34.62 54.20 CPI 0.27 3.61
Adjusted R2 93% Adjusted R2 92%

Mexico Venezuela
Macroeconomic Factor Coefficient t-statistic Macroeconomic Factor Coefficient t-statistic
Stock Market Index -0.09 3.08 Stock Market Index -0.21 6.29
Real Exchange Rate 1.93 1.75 Real Exchange Rate 0.72 2.54
Money Supply 4.47 3.38 Real GDP -12.01 11.28
Local ST Interest Rates 22.67 4.29 CPI 3.73 2.56
Adjusted R2 92% Adjusted R2 89%

38
Table 7
Comparison of Alternative Estimates
of the Long-Term Equilibrium
Comparison among OLS, SUCCR, SCCR, SUR, MSUR estimates of the coefficients, B i , of the
equilibrium stage of the model:

Spreadi,t = Bi F actorsi,t + εi,t

The estimators are described in more details in Appendix A.


The table reports the estimation results for the entire period April 1993 to February 2001.

Country Factor SUCCR OLS SCCR SUR MSUR


Argentina Stock Market Index -7.67 -6.78 -6.98 -6.44 -7.64
Real Exchange Rate Index 6.89 7.90 7.23 -14.39 -16.27
Real GDP -5.14 -6.47 -5.90 12.63 8.84
CPI 34.62 36.18 35.01 -14.95 -8.71
Brazil Stock Market Index -0.08 -0.08 -0.08 -0.12 -0.10
Unemployment 138.94 137.27 139.03 53.07 83.93
Money Supply 0.01 0.01 0.01 0.01 0.02
CPI 0.27 0.28 0.28 0.89 0.57
Mexico Stock Market Index -0.09 -0.09 -0.10 -0.04 -0.01
Real Exchange Rate Index 1.93 1.77 1.94 -2.08 -2.99
Money Supply 4.47 4.65 4.62 4.52 4.31
Local Interest Rates 22.67 23.12 22.80 33.01 34.31
Venezuela Stock Market Index -0.21 -0.21 -0.21 -0.26 -0.25
Real Exchange Rate Index 0.72 0.60 0.81 -4.91 -4.18
Real GDP -12.01 -11.95 -12.08 -9.71 -9.77
CPI 3.73 3.84 3.78 2.48 2.89

Table 8
Estimates of the Parameters
of the Predictive Stage of the Model
OLS and GLS estimates of the short-term dynamics or predictive stage (stage 2) of the model:

∆Spreadi,t+1 = βE ∆SpreadEM BI,t + βW ∆M SCIt + αε!i,t + ei,t+1

The significance of the error-correction coefficient reflects the importance of the long-term equilibrium
stage (stage 1) of the model. The size of the coefficient of the equilibrium deviation shows the speed
of adjustment to long-term equilibrium, i.e. the percentage of deviation that gets corrected within a
given month. The OLS coefficient (-0.26) suggests a faster error-correction than the GLS coefficient
(-0.14).

Regressor β!OLS t − statOLS β!GLS t − statGLS


ε!t -0.26 -6.23 -0.14 -4.83
∆SpreadEM BI,t 0.16 2.03 0.24 2.37
∆M SCIt 0.19 1.87 0.07 0.53

39
Table 9
Out-of-Sample Realized Returns of
Active vs. Passive Strategy
The active strategy is based on one-step-ahead monthly forecasts based on the predictive stage of
the model (stage 2):

∆Spreadi,t+1 = βE,i ∆SpreadEM BI,t + βW,i ∆M SCIt + αi ε!i,t + ei,t+1

The active strategy can be summarized as:


1 2
100% EMBIi,t if E(∆Spreadi,t+1 |Φt ) ≤ 0
Active = for ∀i = 1, ..., 4 and ∀t = 0, ..., 34.
100% cash(T-bills) if E(∆Spreadi,t+1 |Φt ) > 0

The benchmark is a passive strategy equally weighted in the four Latin American countries, i.e.:

P assive = 100% EMBIi,t for ∀i = 1, ..., 4 and ∀t = 0, ..., 34.

The passive strategy involves investing 100% in country i’s index to assure it is at least as risky as
the active strategy. Transaction costs are incorporated by buying at the ask and selling at the bid
when rebalancing.
The two strategies are compared over a 35-month-long out-of-sample testing window - March 1998 to
February 2001 - including the period of the Russian crisis (August 1998) and Brazilian devaluation
(January/February 1999).
The table compares both the total compounded return of each strategy and the simple average
monthly return. Both numbers are annualized.

Set Instruments Used Compounded Returns Simple Returns Sharpe Ratio


(annualized) (annualized)
1 ∆EM BI, ∆M SCI, ε!SU CCR 0.171 0.175 0.43
2 ε!SU CCR 0.195 0.198 0.63
3 ε!OLS 0.153 0.157 0.41
4 ∆EM BI, ∆M SCI 0.028 0.045 0.07
5 Passive Investment 0.090 0.125 0.18

40
Table 10
Sample Probabilities of
Significant Market-Timing Value
The table presents for each country the unconditional sample probability of a correct forecast, p, the
conditional probability of a correct forecast in market upturns, p1 , the conditional probability of a
correct forecast in market downturns, p2 , and Merton’s sufficient statistic for positive market-timing
value: p1 + p2 > 1. The predictions are made based on the second stage of the model:

∆Spreadi,t+1 = βE,i ∆SpreadEM BI,t + βW,i ∆M SCIt + αi ε!i,t + ei,t+1

The out-of-sample window covers the period from March 1998 to February 2001.

Country p! p!1 p!2 p!1 + p!2


Argentina 0.69 0.53 0.83 1.36
Brazil 0.69 0.50 0.84 1.34
Mexico 0.63 0.67 0.60 1.27
Venezuela 0.74 0.95 0.47 1.42
Average 0.69 0.66 0.69 1.35

Table 11
Sample Probabilities of
Significant Market-Timing Value
Using Alternative Sets of Instruments
Estimated sample probability values for the out-of-sample period window - from March 1998 to
February 2001. The table presents for each set of instruments the unconditional probability of a
correct forecast, p, the conditional probability of a correct forecast in market upturns, p 1 , the condi-
tional probability of a correct forecast in market downturns, p2 , and Merton’s combined conditional
probability measure, p1 + p2 .

Set Instruments Used p! p!1 p!2 p!1 + p!2


1 ∆SpreadEM BI , ∆M SCI, ε!SU CCR 0.69 0.66 0.69 1.35
2 ε!SU CCR 0.66 0.63 0.65 1.28
3 ε!OLS 0.63 0.58 0.64 1.22
4 ∆SpreadEM BI 0.47 0.66 0.34 1.00
5 ∆SpreadEM BI , ∆M SCI 0.51 0.44 0.58 1.02

41
Table 12
Results from Small Sample Tests
for Significant Market-Timing Value
Using Alternative Sets of Instruments
The small sample tests are based on Henriksson and Merton (1981) [HM]. The null hypothesis is that
the instruments have no predictive value: H0 : p1 + p2 = 1. If p − value < 5%, the null hypothesis
can be rejected at the 95% confidence level based on HM’s critical number, x ∗ (c), of correct forecasts
in up and down markets, where x∗ (c) is the solution to:
n1
) *) * ) *
" N1 N2 N
/ =1−c
x=x∗
x n−x n

where

N1 (N2 ) ≡ number of observations where ∆Spreadsit ≥ 0 (∆Spreadsit < 0),


N ≡ N1 + N2 = total number of observations,
n1 (n2 ) ≡ number of (un)successful predictions, given ∆Spreadsit ≥ 0 (∆Spreadsit < 0),
n ≡ n1 + n2 = number of times the forecast is ∆Spreadsit ≥ 0,
c ≡ chosen confidence level
n1 ≡ min(N1 , n) ≡ upper bound on correct predictions of ∆Spreadsit ≥ 0,

Set Instruments Used HM p − value Hypothesis of No Predictability


1 ∆SpreadEM BI , ∆M SCI, ε!SU CCR 0.01 Rejected
2 ε!SU CCR 0.01 Rejected
3 ε!OLS 0.03 Rejected
4 ∆SpreadEM BI , ∆M SCI 0.46 NOT Rejected

42
Table 13
Simulation Results for Small Sample Bias
in Non-Stationary Systems with 95 Time-Series Observations
Simulation results (T=95) for the small sample bias in the coefficients of the cointegrating vector
estimated with OLS, SUCCR, SUR, SCCR, and MSUR. The figures in the table represent the average
bias of the estimated coefficients across 2,000 iterations and are based on simulated panel data with
4 cross-sectional and 95 time-series dimensions. The bias figures are multiplied by 10,000. As the
true betas of the cointegrating vectors are 1, i.e. βi = (βi1 , βi2 , βi3 , βi4 )" = 1, the numbers in the
table represent bias in terms of hundredths of a percent (i.e. basis points) of the true value.

α ρ1 OLS SUCCR SUR SCCR MSUR


0.3 2.59 2.20 3.93 2.31 3.51
-0.2 0.5 2.67 2.19 3.82 2.36 3.06
0.8 2.61 2.30 4.37 2.39 3.70
0.3 2.57 2.21 4.09 2.29 3.31
-0.1 0.5 2.40 2.00 3.78 2.10 3.20
0.8 2.84 2.48 4.93 2.61 4.06
0.3 3.35 2.82 4.38 3.01 3.77
0.2 0.5 3.75 3.30 5.65 3.39 5.16
0.8 4.12 3.66 6.34 3.82 5.67

Table 14
Simulation Results for Small Sample Bias
in Non-Stationary Systems with 150 Time-Series Observations
Simulation results (T=150) for the small sample bias in the coefficients of the cointegrating vector
estimated with OLS, SUCCR, SUR, SCCR, and MSUR. The figures in the table represent the average
bias of the estimated coefficients across 2,000 iterations and are based on simulated panel data with
4 cross-sectional and 150 time-series dimensions. The bias figures are multiplied by 10,000. As the
true betas of the cointegrating vectors are 1, i.e. βi = (βi1 , βi2 , βi3 , βi4 )" = 1, the numbers in the
table represent bias in terms of hundredths of a percent (i.e. basis points) of the true value.

α ρ1 OLS SUCCR SUR SCCR MSUR


0.3 2.56 1.94 3.80 1.98 3.43
-0.2 0.5 2.68 1.94 3.96 2.06 3.51
0.8 2.36 1.76 3.89 1.77 3.46
0.3 2.39 1.56 3.55 1.71 3.22
-0.1 0.5 2.40 1.82 3.65 1.81 3.26
0.8 2.07 1.28 3.46 1.39 2.94
0.3 3.11 2.16 4.37 2.29 3.93
0.2 0.5 2.64 1.69 3.95 1.86 3.44
0.8 4.11 3.04 6.02 3.18 5.60

43
3500

Argentina
3000 Brazil
Mexico
Venezuela
EMBI Spreads in Basis Points

2500

2000

1500

1000

500

0
3

0
93

94

95

96

97

98

99

00
r9

r9

r9

r9

r9

r9

r9

r0
ct

ct

ct

ct

ct

ct

ct

ct
Ap

Ap

Ap

Ap

Ap

Ap

Ap

Ap
O

Figure 1. History of Emerging Market Bond Index Spreads for Argentina, Brazil,
Mexico, and Venezuela (in basis points.) The sample is limited by the history of Ar-

gentina, which issued Brady debt in April 1993. The period covers the Mexican Peso crisis
(December 1994), the Asian crisis (October 1997), the Russian crisis (August 1998), and the
Brazilian devaluation (January/February 1999).

44
2000
Spread Deviation from Long!Term Equilibrium (bps)

Argentina
Brazil
1500 Mexico
Venezuela

1000

500

!500

!1000
93

93

94

94

95

95

96

96

97

97

98

98

99

99

00

00
n

ec

ec

ec

ec

ec

ec

ec

ec
Ju

Ju

Ju

Ju

Ju

Ju

Ju

Ju
D

Figure 2. Credit spread deviations from their long-term equilibrium for Ar-

gentina, Brazil, Mexico, and Venezuela estimated with OLS (in basis points (bps)).
The figure presents the deviations from the first stage of the model, that are used as instru-

ments in predicting spread changes in the second stage.

45
2500
Spread Deviation from Long!Term Equilibrium (bps)

Argentina
2000 Brazil
Mexico
Venezuela
1500

1000

500

!500

!1000

!1500
93

93

94

94

95

95

96

96

97

97

98

98

99

99

00

00
n

ec

ec

ec

ec

ec

ec

ec

ec
Ju

Ju

Ju

Ju

Ju

Ju

Ju

Ju
D

Figure 3. Credit spread deviations from their long-term equilibrium for Ar-

gentina, Brazil, Mexico, and Venezuela estimated with SUCCR (in basis points
(bps)). The figure presents the deviations from the first stage of the model. These deviations
are used as instruments in predicting spread changes in the second stage of the model.

46
1.7
SUCCR Correction
OLS Correction
1.6
Traditional Instruments
Passive Investment
1.5

1.4
Wealth process

1.3

1.2

1.1

0.9

0.8

0.7
98

99

00
98
98

98

99
99

99

00
00

00
ar

ar

ar
p
n

ec

p
n

ec

p
n

ec
Se

Se

Se
Ju

Ju

Ju
M

M
D

Figure 4. Relative performance of active strategies based on the model pre-


dictions with alternative sets of instruments. All active strategies involve no short-selling
and are based on one-step-ahead forecasts of the model. The passive strategy is a 100%

buy-and-hold investment in Brady bonds, equally weighted among Argentina, Brazil, Mex-
ico, and Venezuela (the passive strategy is riskier than all active ones). Notation: ’SUCCR

Correction’ uses instruments: ∆SpreadEM BI , ∆M SCI, ε!SU CCR , ’OLS Correction’ uses in-
struments: ∆SpreadEM BI , ∆M SCI, ε!OLS , and ’Traditional Instruments’ uses instruments:
∆SpreadEM BI , ∆M SCI.

47
0.2

0.18

0.16

0.14
Wealth(SUCCR)!Wealth(OLS)

0.12

0.1

0.08

0.06

0.04

0.02

0
Mar 98 Oct 98 May 99 Dec 99 Jul 00 Feb 01

Figure 5. Incremental wealth process of the active strategy based on SUCCR es-

timates relative to that based on OLS estimates. The figure illustrates the incremental
wealth due to using SUCCR in estimating the long-term equilibrium stage of the model relative
to the wealth process based on OLS estimates, i.e. W ealthSU CCR − W ealthOLS . Both active

strategies involve no short-selling and are based on one-step-ahead forecasts of the model.
Notation: ’SUCCR’ uses instruments: ∆SpreadEM BI , ∆M SCI, ε!SU CCR , ’OLS Correction’

uses instruments: ∆SpreadEM BI , ∆M SCI, ε!OLS .

48
1.7
SUCCR Correction
OLS Correction
Traditional Instruments
1.6 Passive Investment
Final wealth on February 28,2001

1.5

1.4

1.3

1.2

1.1

1
Mar 98 Jul 98 Nov 98 Mar 99 Jul 99 Nov 99 Mar 00 Jul 00 Nov 00

Month of initial $1 investment

Figure 6. Relative performance of the active and passive strategies when


the timing of the initial investment changes. The figure illustrates the final

wealth on February 28, 2001, resulting from investing $1 in the months shown on the
X-axis under the different investment strategies. All active strategies involve no short-
selling and are based on one-step-ahead forecasts of the model. Notation: ’SUCCR Cor-

rection’ uses instruments: ∆SpreadEM BI , ∆M SCI, ε!SU CCR , ’OLS Correction’ uses in-
struments: ∆SpreadEM BI , ∆M SCI, ε!OLS , and ’Traditional Instruments’ uses instruments:

∆SpreadEM BI , ∆M SCI.

49

Você também pode gostar