Journal Issue
Share
Article

Emerging Market Spread Compression

Author(s):
Laura Kodres, Kristian Hartelius, and Kenichiro Kashiwase
Published Date:
January 2008
Share
  • ShareShare
Show Summary Details

I. Introduction

The determinants of the difference between the yield on a country’s emerging market debt securities and the U.S. 10-year Treasury note, denoted the emerging market spread, has been subject of academic inquiry for some time. The dramatic decline in spreads to record-low levels has renewed attention in the subject. This decline occurred alongside a dramatic decline in policy interest rates in industrial countries and a marked increase in money growth as these countries loosened monetary policy in an effort to revive their economies following the bursting of the equity bubble in 2000, leading some commentators to claim that the fall in spreads was based on “excess global liquidity.” And despite a consistent tightening by major central banks up until recently, and the market turmoil of summer 2007, spreads have continued to remain quite low. The spread compression has also been accompanied by a seeming improvement in the “real” fundamentals of many emerging market countries—debt ratios declined in some countries, current account deficits improved, and other structural policies were adopted. Reflecting these improvements, the average credit rating of the countries included in the EMBI Global index improved from BB- to BB+ between early 2000 and 2007.1

As interest rates have continued their upward path, attempting to discern how much of the improvement in emerging market spreads is due to improvements in country fundamentals and how much is due to excess liquidity has important ramifications. If much of the compression in spreads is attributable to solid reforms undertaken in recent years, then tighter monetary conditions in industrial countries should have only a small effect on spreads. However, if much of the spread compression was due to liquidity, then tighter monetary policy and a drying up of liquidity could mean reversals in spreads, which could be especially pronounced if excessive liquidity had also led to leveraged positions.

This paper attempts to distinguish between the two factors—fundamentals and liquidity—by constructing a model that takes into account several features of emerging market spreads and how they adjust to domestic fundamentals and interest rates.2 Compared to previous examinations of spreads, we innovate in several ways. First, we use data that capture the more forward-looking character of market prices reflected in spreads, namely, by adding credit rating agency outlooks, in addition to the ratings themselves, for country fundamentals and using 3-month Fed Funds futures rates rather than the Fed Funds rate itself, as a measure of liquidity. Second, we develop a more accurate relationship between credit ratings and their likely effect on spreads by modeling the observed non-linear relationship between the two.

Many studies use credit ratings because it encapsulates a host of economic variables.3 While a handy and efficient measure, the measure is relatively “coarse;” that is, there are a fixed number of categories (e.g. AAA, AA+, A, … C-, and SD, referring to default), and alterations among them are not associated with a fixed (or linear) response in spreads. The model below attempts to enrich the informational content of ratings by: (i) using the indications for future up- or down-grades represented by the rating outlook; and (ii) scaling the ratings variable using logarithms to account for their non-linear relation with spreads.

One of the motivations for adding the credit rating outlook to the series of actual ratings is an observation that markets react first and foremost to hints of future ratings changes rather than the actual event when it occurs. Sy (2002) provides formal evidence of the early reaction by market participants. He observes that when a country’s spreads are “excessively high,” a rating downgrade frequently follows; similarly, “excessively low” spreads are often followed by upgrades, suggesting markets anticipate future ratings changes (or alternatively that credit rating agencies are late to alter ratings).4 Below we look more carefully at spreads and changes in the outlook and show a somewhat stronger and more timely response to the outlooks than the actual rating changes.

Previous studies that have examined the influence of global interest rates on emerging market spreads have variously used a short-term interest rate, such as a 3-month eurodollar deposit rate or the Fed Funds rate, a long-term interest rate, such as the U.S. 10-year Treasury note or some weighted average of 10-year government securities from industry countries, or the slope of the yield curve, as measured by, say the difference between the U.S. 10-year note and 3-month Treasury bill. The use of this latter variable is prompted by the use of “carry trades” in which participants borrow funds at the Fed Funds rate or similar rate and then invest in longer-term higher yielding securities.

In many cases, the results have been less than satisfactory, with some authors finding support for the global liquidity story in a positive relation between the given interest rate and spreads and other finding either no relation or a negative relation.5 We chose to use a forward-looking short-term interest rate—the Fed Funds 3-month ahead futures rate to better capture the view that financial markets are forward-looking and therefore incorporate expectations of future interest rates into their current trading decisions. We also use the volatility of the Fed Funds futures rates to proxy for the uncertainty surrounding changes in liquidity conditions—another way to capture forward-looking trading behavior.

In the next section we describe the data in more detail and then move to the results. We then examine the proportion of spread dynamics explained by fundamentals and external factors. Finally, we sum up our results and suggest some areas of further research.

II. Data

A. Variables

We collected a number of daily and monthly series during the period of January 1991 and February 2007 for 33 major emerging market economies as well as financial data from the U.S. economy.

Emerging Market Bond Spreads

To examine the factors that determine emerging market spreads, we collected daily observations of the Emerging Market Bond Index (EMBI), and the EMBI Global for individual countries, as well as the composite index. Due to specific criteria for country inclusion and liquidity, the two bond indices consist of a varied number of member countries and financial instruments. Hence, starting points of the series differ across countries and the indices (Table 1).6 There are some instances where a country’s index has missing observations in the middle of its series.7 For this reason, when necessary, the spreads based on these indices are spliced to obtain a longer time series in the sample period of January 1991 and February 2007. It is important to note that a sovereign’s market weight changes over time as inclusions or exclusions of a country or various debt issuances occur.8 Despite such complexities, the EMBI and the EMBI Global have become market benchmarks and provide us with the most comprehensive, readily available, and easy-to-analyze data.9

Table 1.Availability of EMBI and EMBI Global
EMBI1/EMBI Global
Aggregate Index91M1~02M697M12~07M2
Algerian.a.99M3~03M21/
Argentina93M4~02M693M12~07M2
Brazil91M12~02M694M4~07M2
Bulgaria94M11~02M694M7~07M2
Chilen.a.99M5~07M2
Chinan.a.94M3~07M2
Colombian.a.97M2~07M2
Cote d’Ivoiren.a.98M4~07M2
Croatian.a.96M8~04M61/
Dominican Republicn.a.01M11~07M2
Ecuador95M6~02M695M2~07M2
Egyptn.a.01M7~07M2
El Salvadorn.a.02M4~07M2
Hungaryn.a.99M1~07M2
Lebanonn.a.98M4~07M2
Malaysian.a.96M10~07M2
Mexico91M12~02M693M12~07M2
Moroccon.a.97M12~06M51/
Nigeria92M1~02M62/93M12~07M2
Pakistann.a.01M6~07M22/
Panama97M2~02M696M7~07M2
Peru97M5~02M697M3~07M2
Philippines91M12~97M11/97M12~07M2
Poland94M11~02M694M10~07M2
Russia98M12~02M697M12~07M2
South African.a.94M12~07M2
South Korean.a.93M12~04M41/
Thailandn.a.97M5~06M41/
Tunisian.a.02M5~07M2
Turkeyn.a.96M6~07M2
Ukrainen.a.00M5~07M2
Uruguayn.a.01M5~07M2
Venezuela91M12~02M693M12~07M2

These series are terminated.

There are some missing observations during the period. For Nigeria, the EMBI series does not have observations during

April 1998 through April 1999. For Pakistan, the EMBI Global series has missing observations during March 2003 through March 2004.Source: J.P. Morgan Chase.

These series are terminated.

There are some missing observations during the period. For Nigeria, the EMBI series does not have observations during

April 1998 through April 1999. For Pakistan, the EMBI Global series has missing observations during March 2003 through March 2004.Source: J.P. Morgan Chase.

Graphically, the spreads appear to trend downward beginning in 1999, a year in which several countries are added to the sample. We test for non-stationarity and find that a large number of countries’ spreads exhibit unit roots. We discuss this issue in the context of the panel data sets below.

Credit Ratings and Outlooks

Monthly sovereign credit ratings and their outlooks are collected from Ratings Direct, provided by the Standard & Poor’s. We used long-term credit ratings for assessing the sovereign’s foreign currency default risk since debt instruments included in the EMBI family are mostly long-term and denominated in U.S. dollars. In general, most rating agencies issue both long- and short-term credit ratings in domestic and foreign currencies. In addition, they assign credit outlooks for each country in both currencies, though the rating agencies note that their outlook does not guarantee any future changes in the ratings themselves. As shown in Figure 1, variation in the credit outlook appears to add information for examining country risk. There have been 234 and 300 monthly observations over the sample period that recorded changes in long-term credit ratings and outlooks, respectively.10 Moreover, there have been 169 monthly observations for which changes in outlooks occurred when long-term credit ratings remained the same. Thus, we can enrich our variable on countries’ risk by using the outlook.11 Short-term ratings, however, do not appear to add very much fundamental information since the timing of their changes generally coincides with that of long-term ratings changes. There were only six monthly observations for which there was a change in short-term ratings but no simultaneous change in long-term credit ratings or outlooks.

Figure 1.Changes in Sovereign Credit Ratings and Outlook: January 1991 ~ February 2007 1/

(Number of months)

1/ Based on 33 emerging market economies included in J.P. Morgan Emerging Markets Bond Index (EMBI) and EMBI Global.

Source: Standard & Poor’s, Moody’s Investors Service, and Authors’ calculations.

Fed Funds Futures

The Fed uses a target rate for Fed Funds to transmit of its monetary policy objectives and this rate has become a market-wide benchmark for various financial activities. Leveraged carry traders who borrow at the short-end of the yield curve to invest in emerging market bonds keenly watch short-term benchmark rates, such as the Fed Funds rate. For this reason, we look at the implied yield on the 3-month Fed Funds futures and evaluate how market expectations of future U.S. monetary policy affect the emerging market bonds. The Fed Funds futures rate has the advantage that it influences interest rates all along the U.S. yield curve.12 It is noteworthy that the 3-month Fed Funds futures rate has a 0.99 correlation with 3-month LIBOR and U.S. Treasury bill rate, and 0.69 with 10-year U.S. Treasury notes at a daily frequency over the entire sample period. Despite the use of an expected Fed Funds rate, this variable exhibits a unit root using an Augmented Dickey-Fuller test (with a constant).

Volatility in the Fed Funds Futures Market

Uncertainty of future U.S. monetary policy is perceived to have a large impact on the financial markets, making decisions about financial risk allocation more difficult. To measure this uncertainty, we used the difference between the implied yield of three-month Fed Funds futures contract and the target rate at a daily frequency. In a rolling 90-calendar-day window, we calculated the standard deviation of the difference. The daily series of standard deviation was then averaged over each month. This measure of uncertainty of future U.S. monetary policy, expressed in standard deviation, was closely linked to the widening of bond spreads during the tightening cycle of 1994–1995 (Figure 2). Heightened anxiety at the beginning of the most recent tightening cycle (beginning mid-2004) concerning the speed and magnitude of the tightening to come was also initially reflected in rising spreads.

Figure 2.Volatility of Fed Funds Futures Market and Emerging Market Bond Spread

1/ This monthly volatility measure is based on a 90-day rolling standard deviation.

Sources: Bloomberg L.P.; J.P. Morgan Chase; and Authors’ calculation.

Volatility Index of S&P 500 (VIX)

The Chicago Board Options Exchange (CBOE) Volatility Index, denoted “VIX,” is based on the S&P 500 options prices. The VIX is often used as a proxy for investor’s attitude toward risk and appears to explain movements of the emerging market bond spread in recent years. The spread compression seems to coincide with the reduction of the VIX, which is generally interpreted as increased investor risk appetite.13 However, during the Fed tightening cycle of 1994–1995, the relation between the fluctuations of the VIX and the emerging market bond spread appears to have broken down (Figure 3). In recent years, the VIX has mirrored bond spreads quite well.

Figure 3.VIX and Emerging Market Bond Spread

1/ The SPX Volatility Index, measuring the implied volatility in the prices of a basket of options on the S&P 500..

Sources: Bloomberg L.P.; J.P. Morgan Chase; and Authors’ calculation.

B. Total Credit Rating-Outlook Index (CROI)

One of the main contributions of this paper is to introduce a Total Credit Rating-Outlook Index (CROI) which encompasses both the information of sovereign long-term credit ratings and outlooks to better proxy for underlying fundamentals that influence emerging market bond spreads.

Log Linearity Between the Spreads and Ratings

Although many studies take a log of the spreads and look at the relationship between the spreads and an index of cardinal numbers assigned to sovereign long-term credit ratings, there is little supporting evidence provided for this specification. Often the index is constructed such that its value increases by some increment as the rating deteriorates, or vice versa. In this paper, we begin our exercise by observing a linear relationship in log between the spreads and an index of ratings increasing from 1 to 22 in log scale by an increment of 1 as the ratings deteriorate from AAA all the way down to SD. With this logarithmic scaling, a one-notch decline in the ratings, (i.e., AA+ to AA), is represented by an increase of 100 percent in the ratings index level. Figure 4 indicates that the log relationship is linear, implying that there is a nonlinear relationship in level between the spreads and the rating index.

Figure 4.Average vis-à-vis Estimated Bond Spreads on Long-Term Sovereign Credit Ratings 1/

(log scale: basis points)

1/ Based on the monthly observations during December 1991 through May 2005 for the countries included in the EMBI, EMBI Plus, and EMBI Global.

2/ Estimated spreads are from the univariate regression of the spreads in log on a constant term and a dummy variable of long-term sovereign credit ratings in log. The dummy variable increases by an increment of 1 in log, representing a one-notch decline in the rating.

Sources: Bloomberg L.P.; J.P. Morgan Chase; and Authors’ calculation.

Construction of the Total Credit Rating-Outlook Index (CROI)

To define the Total Credit Rating-Outlook Index (CROI), we use information on both the long-term credit rating and outlook for a sovereign’s foreign currency denominated instruments. We use regression results to estimate the individual response to both the rating and the outlook.14 The index value of the CROI in logs is found on Table 2. At each time t for each country i CROI takes on one of the values in the table depending on the configuration of the current rating and the outlook.

Table 2.Total Credit Rating-Outlook Index (CROI)(Index value in log)
Category:Sovereign Long-term

Credit Ratings (Ri)
Credit Outlook (O)
Stable (STB)Positive (POS)Negative (NEG)
Investment Grade (G1)
AAA1.00.02.7
AA+2.01.03.7
AA3.02.04.7
AA-4.03.05.7
A+5.04.06.7
A6.05.07.7
A-7.06.08.7
BBB+8.07.09.7
BBB9.08.010.7
BBB-10.09.011.7
Noninvestment Grade: Tier 1 (G2)
BB+11.010.112.7
BB12.011.113.7
BB-13.012.114.7
B+14.013.115.7
B15.014.116.7
B-16.015.117.7
CCC+17.016.118.7
Noninvestment Grade: Tier 2 (G3)
CCC18.018.018.0
CCC-19.019.019.0
CC20.020.020.0
C21.021.021.0
SD22.022.022.0
Source: Authors’ calculations.
Source: Authors’ calculations.

It is important to recognize that the CROI is an index that, by construction, attempts to calibrate how ratings and outlooks relate to spreads—that is, it uses (past) information about how spreads react to changes in ratings and outlooks to make this calibration. Table 2 indicates that there are four unique properties embedded into the CROI that reflect various non-linear responses of spreads to changes in ratings or outlooks. For instance, given log linearity, a spread of a country with a credit rating above the first tier of non-investment grade category and an outlook of “positive” is lower than that of a country with a one-notch higher credit rating and an outlook of “negative.” The CROI does not distinguish between “positive” and “stable,” nor “negative” and “stable” outlooks for the countries in the second tier of noninvestment grade category. The index values depend entirely on the credit ratings since the addition of outlooks were statistically insignificant in this category. As well, when a country’s outlook changes from “stable” to “positive” with its rating remaining the same, this improvement is reflected in a larger reduction of the CROI for countries in the investment grade category (-0.97) relative to those in the first tier of noninvestment grade category (-0.90). On the other hand, a deterioration in the outlook increases the value of the CROI by a larger degree for countries in the first tier of noninvestment grade category than those in the investment grade category. And, given a credit rating above the first tier of noninvestment grade category, an increase in the CROI responding to a “negative” outlook is higher in an absolute magnitude than the decline in the index responding to a “positive” outlook.

Using market capitalization of EMBI Global for each member, we aggregated the newly constructed CROI and the Long-term Credit Rating (LTCR). Figure 5 compares the CROI and the LTCR in log level plotted against the EMBI Global in basis points. One of the most striking findings is that CROI appears to be more closely linked to spreads. In fact, during the period of January 1994 through February 2007 the correlation of spreads with the CROI is 89 percent compared to only a 74 percent with the LTCR.

Figure 5.Aggregate “Fundamentals”: Total Credit Rating-Outlook Index (CROI) vis-a-vis Long-term Credit-Rating Index (LTCR)

(EMBI Global: left scale in log, basis points and CROI and LTCR: right scale, index 1/)

1/The higher the index, the worse the “fundamentals”. Aggregate “fundamentals” were calculated based on the countries’ market capitalizations in EMBI Global.

Sources: Bloomberg L.P.; J.P. Morgan Chase; Standard & Poor’s, Ratings Direct; and Authors’ calculations.

III. Results

We analyze both effects of economic fundamentals and liquidity on the bond spreads. The liquidity effects result from the examination of the following variables; (i) the anticipated level of U.S. interest rates, (ii) the volatility of the Fed Funds futures markets, and (iii) the level of risk appetite as proxied by the VIX index. In this analysis, we make comparisons of the results from regressions based on the newly constructed Total Credit Rating-Outlook Index (CROI) and the standard Long-term Credit Rating (LTCR).

As a starting point, we examine the time series properties of our underlying variables. While there is little theoretical reason to believe that any of the variables are non-stationary in the long run, the trends from 1998 onward suggest non-stationary tests are warranted, particularly because many countries entered the sample in this latter period. According to several panel tests of non-stationarity, the EMBIG spreads and the credit ratings variables have unit roots over the sample. The Fed Fund’s futures rate is also found to be non-stationary, while the volatility of Fed Funds and VIX are not. Since both independent and dependent variables in our model are non-stationary, we further test for cointegration in our preferred model below.

A. Basic Model

We first estimate a fixed effects panel regression model with the log of bond spreads (SPRD) as the dependent variable and only two explanatory variables; fundamentals as captured by the CROI (or LTCR), and the 3-month ahead U.S. Fed Funds futures’ rate (FF3M).15Table 3 indicates the regression results based on this basic model. Although we collected data for the spreads for 33 countries, due to limited data availability on the credit ratings and outlooks, 30 countries are included in the estimation.16

Table 3.Basic Model Results: CROI vs. LTCR, December 1991 ~ February 2007 1/
Dependent variable: Log of bond spreads (SPRD)
Explanatory

variables
CoefficientStandard

error
p-valueExplanatory

variables
CoefficientStandard

error
p-value
CROI0.2460.0230.000LTCR0.2420.0290.000
FF3M0.0220.0170.222FF3M0.0130.0190.503
Constant2.7100.2890.000Constant2.8230.3560.000
R2R2
within0.414within0.327
between0.726between0.724
overall0.594overall0.554

The results are based on fixed effects regressions using 3038 monthly observations. Standard errors corrected for heteroskedasticity and autocorrelation within groups.

Sources: Bloomberg, L.P.; J.P. Morgan Chase; The PRS Group, International Country Risk Guide; Standard & Poor’s, Ratings Direct.

The results are based on fixed effects regressions using 3038 monthly observations. Standard errors corrected for heteroskedasticity and autocorrelation within groups.

Sources: Bloomberg, L.P.; J.P. Morgan Chase; The PRS Group, International Country Risk Guide; Standard & Poor’s, Ratings Direct.

At an aggregate level, the correlation between the log of SPRD and log of CROI was significantly higher than that between the log of SPRD and the log of LTCR. However, the regression using CROI only explains 4 percent more of the variation in bond spreads (overall R2 is 0.594 as opposed to 0.554). Although the coefficient values and R2 statistics are similar, the improvement in R2-within reflects the closer time-series association obtainable using the log of CROI where a country’s outlook changes more frequently.

B. Extended Model with Volatility

Next, we estimate the following fixed-effect panel regression model by OLS.17

where eit is a random error. The explanatory variables included in this regression are: 3-month Fed Funds futures rate (FF3M); the volatility of the Fed Funds futures market (V_FF) represented by the 90-day rolling standard deviation of the difference in FF3M and FFTR; the Volatility Index (VIX) for the S&P 500; and the fundamentals. For the measure of the fundamentals, we again apply both the log of CROI and the log of LTCR.

The use of the CROI improves the overall fit, yielding a higher R2 of 0.704 relative to 0.682 under the model with LTCR. And again the embedded information in the outlooks better explains the variation of the spreads over time, but does not add much value to the explanation of spreads across countries.

Table 4.Extended Model Results: CROI vs. LTCR, January 1991 ~ February 2007 1/
Dependent variable: Log of bond spreads (SPRD)
Explanatory

variables
CoefficientStandard

error
t-

statistics
p-valueExplanatory

variables
CoefficientStandard

error
t-

statistics
p-value
FundamentalsFundamentals
CROI0.1930.01711.50.000LTCR0.1920.0228.80.000
Other explanatory variables:Other explanatory variables:
FF3M0.0500.0163.20.003FF3M0.0450.0162.90.008
V_FF1.9030.2029.40.000V_FF2.0390.2418.50.000
VIX0.0410.00410.20.000VIX0.0450.00411.60.000
Constant2.1760.20910.40.000Constant2.1300.2528.50.000
R2R2
Within0.635Within0.598
Between0.772Between0.776
Overall0.704Overall0.682

The results are based on fixed effects regressions using 3038 monthly observations. Standard errors corrected for heteroskedasticity and autocorrelation within groups.

Sources: Bloomberg, L.P.; J.P. Morgan Chase; The PRS Group, International Country Risk Guide; Standard & Poor’s, Ratings Direct; and authors’ calculation.

The results are based on fixed effects regressions using 3038 monthly observations. Standard errors corrected for heteroskedasticity and autocorrelation within groups.

Sources: Bloomberg, L.P.; J.P. Morgan Chase; The PRS Group, International Country Risk Guide; Standard & Poor’s, Ratings Direct; and authors’ calculation.

The p-values of explanatory variables suggest that they all be included in the models. The estimated coefficient for the CROI signifies that the spreads will increase by 0.193 in logs when the index value of the CROI increases by 1 unit in logs. Given a country with a rating of “B” and outlook of “stable,” when its rating falls by a one-notch to “B-,” ceteris paribus, the country’s spread will increase by 21 percent from its initial level.

The 3-month Fed Funds futures rate (FF3M) has estimated coefficients of 0.050 and 0.045 in the regressions using the CROI and the LTCR for fundamentals, respectively, and they are both statistically significant at 1 percent level. When participants of the Fed Funds futures market expect there will be a one percent (100 bps) increase in the U.S. policy rate over the next three month window, the spreads today will increase by about 5 percent from their initial levels, ceteris paribus. Relative to the “fundamentals,” the level in the expected interest rate contains less explanatory power for spreads.

Uncertainty about future U.S. monetary policy can have a significant effect on emerging market bond spreads. In fact, we find that the volatility of the Fed Funds futures market influences spreads to a greater degree than does the level of the interest rate. The estimated coefficients of V_FF are 1.903 and 2.039 for the regression of CROI and that of LTCR, respectively. They are both statistically significant at the 1 percent level. If the volatility of the Fed Funds futures market increases by a standard deviation, ceteris paribus, spreads widen by about 13.5 percent in the model using CROI.18

Lastly, the CBOE Volatility Index (VIX) captures volatility on the S&P 500 options prices and is viewed as a proxy for investor’s attitude toward risk. Over the entire sample period, regardless of the choice of fundamental variables, the estimated coefficient of the VIX displays statistical significance at 1 percent level. An increase of the VIX by one standard deviation will increase spreads by about 30 percent in the model with CROI.19 If the VIX surges to the level observed in September 1998, the regression with CROI estimates that spreads will leap by 238 percent if other factors remain constant.20

In this section, we have presented the results from fixed-effect panel regressions run by OLS. Since using several different tests we find spreads and the credit outlook variables to be non-stationary in the panel, we proceed to test for panel-based cointegration to assure that the coefficients are robust and spurious trending relationships are not responsible for our results. Using residual-based tests devised by Kao (1999) and Pedroni (1995), we reassuringly find that the variables in above specification cointegrate and thus the estimated coefficients are robust to possible time trends.21

C. Graphical Interpretation of the Models

To evaluate how well the model is able to perform, we aggregate the estimated EMBI global spreads, weighted by market capitalizations of the countries in the sample, and compare them against the actual spread. The aggregation is based on 30 countries, and the actual spread is based on splicing EMBI and EMBI Global.22

Panel (A) in Figure 6 shows that the in-sample fit of the extended model with CROI mirrors the data fairly well. The model does less well in the early part of the sample in part because the sample is sparse and the volatility of actual spreads is quite high. Panel (B) shows that the out-of-sample forecasting ability of the model has increasingly deviated from actual spreads—with the estimated spreads at the end of our sample some 180 basis points above actual spreads. This may reflect an additional “search for yield” that is not captured by the VIX index and the low level of interest rate volatility. Deviations could also be explained by structural shifts in the parameters or a faster decline in issuance of external debt than the previous period, which we do not control for. Panel (C) shows how the model performs for a specific country, Brazil. The panel coefficients are used to produce in-sample and out-of-sample forecasts of Brazilian EMBIG spreads. The model captures the dramatic events surrounding 2002 quite well. The current level of the spread is well-approximated (in-sample) by the model, whereas the out-of-sample forecast suggests somewhat higher spreads.

Figure 6.Actual vs. Estimated Spreads Extended Model with CROI as Fundamentals

(scale in basis points)

Sources: Bloomberg L.P.; J.P. Morgan Chase; PRS Group, International Country Risk Guide; Standard & Poor’s, Ratings Direct; and Authors’ calculations.

D. Contributions to EMBI Spreads

By decomposing the model dynamics over a certain time period, we can get an insight into how important changes in fundamentals have been, relative to changes in external financial variables, for the dynamics of the EMBI spread over that period.23

Table 5 shows that from December 2002 to the end of our sample, the variable that represented the country-specific fundamentals, the CROI, explained 197 basis points (46 percent) of the model-predicted 431 basis point decline in spreads. The other variables, representing the external factors, VIX, the fed funds futures rates, and the volatility of fed funds futures, explained a total of 234 basis points of decline (54 percent). Of the actual decline in spreads, some 455 basis points, the CROI accounts for 43 percent and the other three external variables for 51 percent of the decline. Just examining the two variables representing U.S. monetary policy (the Fed Funds futures rate and its volatility), the explanatory power amounts to less—around 10 percent—while the rise in risk appetite, proxied by the VIX, is an important external factor—accounting for nearly 44 percent—of the recent movement in the EMBIG. Overall, then the recent decline spreads is more related to risk appetite and uncertainty regarding U.S. monetary policy than it is due to country fundamentals as these variables account for over half of the decline in spreads. Rising forecasted U.S. interest rates, by contrast, have taken a back seat.

Table 5.Determinants of Change in the EMBIG Spread, December 2002–February 2007

IV. Conclusions

While it is difficult to parse out a pure “liquidity effect” from the model, the estimations show that US interest rate variables clearly have an effect on emerging market debt spreads. This implies that the Fed can play a role in reducing the risk of any disruptions in the emerging bonds market. A clear communication strategy by the Fed that helps guide market expectations can promote financial stability by keeping the volatility of the expected U.S. monetary policy low, thus contributing to a more modest widening in emerging market spreads when fundamentals start to worsen.

While the Fed plays an important role, emerging market economies also have a role. To avoid abrupt increases in spreads they must put policies in place during “good times” to help insure that their overall fundamentals will not deteriorate. Even when the U.S. interest rate increases, the model shows that they can still offset any negative impact by continuing to improve good economic policies that contribute to better credit ratings.

The model attempts to gauge the role of U.S. interest rate effects, risk appetite, and emerging market fundamentals by using a more refined variable for fundamentals. The new variable utilizes not only rating changes (a variable used in previous studies to proxy for country economic fundamentals) but incorporates the outlooks for ratings as well. This adds some, but not a great deal, of explanatory power. The model is explicitly designed as a descriptive model for the determinants of emerging market bond spreads and does not account for the supply-side of the sovereign emerging market bond market. Future research could attempt to model both the demand and supply side of the market to better hone the effects of U.S. interest rates on emerging market bond spreads.

Appendix 1.A: A Procedure of Constructing the CROI

There are four steps required to construct the Total Credit Rating-Outlook Index (CROI). The CROI, by its construction, includes a factor that differentiates prices of bonds by investment/noninvestment grade categories as well as by the negative, stable, and positive outlooks. Since the CROI integrates these unique properties, the first step is to explain how we embed such features in the index.

Step 1: Instead of applying a unified log-linear relationship between the spreads and the ratings (Ri)24 across the entire rating spectrum, we conducted regression analyses and assessed if there are more than one log-linear relationship across the rating spectrum. We divided the spectrum into three categories: (i) investment grades, denoted G1; (ii) first tier of noninvestment grades, denoted G2; and (iii) second tier of noninvestment grades, denoted G3. By creating a set of intercept and slope dummy variables including ones for outlooks, we ran a variance-constrained OLS regression based on a pooled sample to evaluate levels of statistical significance for these three grade categories as well as for the rating and outlook variables.25

When defining investment grades as ratings of AAA through BBB−, first tier of noninvestment grades as BB+ through CCC+, and the second tier as CCC through SD, F-statistics of joint hypothesis suggest that separately distinguishing both the investment and the first tier of noninvestment grade categories is important as both are statistically significant at the 1 percent level (Appendix Table 1.A.1).26 However, the second tier of noninvestment grade category was not. The adjusted R2 reported by this regression was 0.715, compared to 0.676 reported under the univariate regression.27 There are varying levels of individual and joint significance for the intercept and slope dummies for “negative and “positive” outlooks across the various grade categories.

Appendix Table 1.A.1:Selecting an Underlying Model for the Total Credit Rating- Outlook Index(CROI)
(1) Break between G1 and G2 :(1) BBB-/BB+
(2) Break between G2 and G3 :(2) CCC+/CCC
Feasible GLS
Variance-Constrained OLS based on the pooling sampleWith U.S. Policy Rate

(Heteroskedasticity corrected)
Without U.S. Policy RateWith U.S. Policy Rate
Coef.Std. Err.p-valueCoef.Std. Err.p-valueCoef.Std. Err.p-value
FFTR0.0820.0050.0000.0830.0040.000
(1) Investment grades: AAA ~ BBB - (G1)
Constant2.7940.1090.0002.5320.1050.0002.7830.0810.000
R0.2560.0130.0000.2560.0120.0000.2330.0090.000
G1* POS0.7160.2250.0010.6920.2140.0010.4390.1640.007
R* G1* POS-0.1030.0260.000-0.1070.0250.000-0.0790.0190.000
G1*NEG0.1350.2710.6180.0720.2580.7810.0900.1920.639
R* G1 * NEG0.0430.0310.1680.0390.0290.1890.0350.0230.117
(2) 1st tier of noninvestment grades: BB+ ~ CCC+(G2)
G2*STB1.7400.1610.0001.4990.1540.0001.0170.1190.000
R* G2 * STB-0.1280.0150.000-0.1110.0150.000-0.0680.0120.000
G2*POS0.4990.2510.0470.0580.2400.810-0.2970.2070.152
R* G2* POS-0.0410.0210.051-0.0150.0200.4750.0210.0170.230
G2*NEG0.7030.2330.0030.5610.2220.0110.4230.1880.024
R* G2 * NEG-0.0290.0200.146-0.0200.0190.2950.0030.0160.839
(3) 2nd tier of noninvestment grades: CCC ~ SDG(G3)
G3*STBn.a. 2/n.a. 2/n.a. 2/
R* G3* STB-0.0700.0240.004-0.0620.0230.007-0.0510.0140.000
G3*POSn.a. 2/n.a. 2/n.a. 2/
R* G3 * POSn.a. 2/n.a. 2/n.a. 2/
G3*NEG0.4170.9440.6580.7860.8980.3820.4650.8300.575
R* G3 * NEG-0.0350.0460.448-0.0500.0430.246-0.0220.0400.589
R2(Adj.)/Log likelihood0.7150.742-1601.2
RMSE0.5040.480n.a. 2/
F-stat/Chi^2-stat 1/
Overall
“Positive”11.0(***)15.5(***)83.9(***)
“Negative”38.4(***)32.8(***)295.0(***)
Group 1 (G1)
“Positive” and “Negative”39.4(***)40.2(***)179.7(***)
“Positive”18.1(***)30.5(***)71.7(***)
“Negative”46.7(***)33.1(***)70.5(***)
Group 2 & Group 338.1(***)36.7(***)468.7(***)
Group 2 (G2)
“Stable”, “Positive”, and “Negative”53.8(***)52.3(***)454.5(***)
“Stable”91.8(***)73.1(***)157.9(***)
“Positive”1.982.2(*)2.83
“Negative”24.1(***)20.9(***)119.4(***)
Group 3 (G3)
“Stable”, “Positive”, and “Negative”3.1(**)2.9(**)0.32
“Stable”8.3(***)7.3(***)n.a. 2/
“Positive”n.a. 2/n.a. 2/n.a. 2/
“Negative”1.811.80.32
Chi^2-stat (H0: constant variance) 3/1.0625.1(***)n.a. 2/

The number of asterisks indicates the level of significance: *** at 1% level; ** at 5% level, and * at 10% level.

Estimated coefficient is not available.

Breusch- Pagan/Cook -Weisberg test for heteroskedasticity. The null hypothesis is based on a constant variance for the fitted value of the spreads. The level of significance is expressed by the number of asterisks where (***) indi cates 1%, (**) 5%, and (*) 10%.

Source: Authors’ calculations.

The number of asterisks indicates the level of significance: *** at 1% level; ** at 5% level, and * at 10% level.

Estimated coefficient is not available.

Breusch- Pagan/Cook -Weisberg test for heteroskedasticity. The null hypothesis is based on a constant variance for the fitted value of the spreads. The level of significance is expressed by the number of asterisks where (***) indi cates 1%, (**) 5%, and (*) 10%.

Source: Authors’ calculations.

Examining heteroskedasticity indicated that acceptance of the null hypothesis of no heteroskedasticity.28

Under the hypothesis that liquidity influences the spreads, to avoid any omitted variable biases on the estimated coefficients, we re-ran the regression while controlling for Federal Funds target rate (FFTR). The result from Table 1.A.1 indicates that the regression with FFTR gives a better fit with an adjusted R2 of 0.75. In addition, FFTR is statistically significant at 1 percent level. Some of the individual intercept and the slope dummies for “positive” and “negative” outlooks altered somewhat with the inclusion of the FFTR and now the test of no heteroskedasticity is rejected.

By running a Feasible Generalized Least Squares (FGLS), we correct for heteroskedasticity. Table 1.A.1 indicates that some further changes in the statistical significance of various parameters. After excluding the variables that are not statistically significant to explain the spreads, we ran another FGLS and obtained the following estimated result for the next step.

Step 2: Given the coefficients from the regression result in Step 1, we estimated the spreads (SPRD) in log based on the combination of all possible credit ratings (R) and outlook (O) on average. Based on the estimated spreads, we calculated unweighted difference of the spreads (DSPRDU) between “positive” (POS) and “stable” (STB) outlook and that between “negative” (NEG) and “stable” (STB) outlook for each credit rating as follow.

Step 3: We then calculated weighted differences of the spreads (DSPRDGs,OW) for each grade category (Gs), where G1 is an investment grade category, G2 is the first tier, and G3 is the second tier of noninvestment grade categories. Our weight (WGTRi,OGs) is based on the number of observations (OBS) sorted by credit ratings (Ri) and outlook (O) under the category of Gs. For example, for the investment grade category,

where

The weighted differences of the spreads (DSPRDGs,OW) are calculated as:

For the investment grade (G1) category, we obtained the following estimates in log.

For the first tier of noninvestment grade (G2) category, the estimates are:

For the second tier of noninvestment grade (G3) category, the estimates are:

These estimates signify discount/penalty factors based on a given outlook under the three grade categories. A discount of 0.226 given to countries with a “positive” outlook in the investment grade (G1) category implies that their spreads are lower by 20.6 percent relative to other countries with a “stable” outlook in the same grade category. Similarly, a penalty of 0.346 given to countries with a “negative” outlook in the first tier of noninvestment grade (G2) category implies that their spreads are higher by 34.6 percent relative to other countries with a “stable” outlook in the same grade category. In contrast, a “positive” outlook in the second tier of noninvestment grade category does not bear a discount factor, and neither does a “negative” outlook a penalty factor. In addition, the discount/penalty factors account for asymmetric responses by the investors. The widening of the spreads in response to a “negative” outlook is larger in an absolute magnitude than the spread tightening in response to a “positive” outlook. Investors, instead of handsomely rewarding countries with “positive” outlook, penalize them harshly with a “negative” outlook.

Step 4: These discount/penalty factors feed into a calculation of adjusted spreads for each rating (Ri) with outlooks (O) of both “positive” and “negative”. While we use the estimated spreads (SPRDRiSTB) for a “stable” outlook from Step 1, the adjusted spreads (SPRDRi,OAdj) are defined as follow.

Using these adjusted spreads and the estimated coefficients of the regression from Step 1, we calculate the values of ratings (Ri). These values are, in essence, the index values of the Total Credit Rating-Outlook Index (CROI). Thus, the CROI takes the following form when an outlook is “stable.”

When an outlook (O) is either “positive” or “negative,” the CROI takes the following form.

Table 2 shows values of the CROI based on the formulae above. The initial CROI is altered to incorporate the four unique properties that are attributed to the discount/penalty factors calculated earlier. These properties are discussed in the section B.1.

Appendix Figure 1.Actual and Estimated Spreads: Extended Model with CROI as Fundamentals

(log scale; basis points)
(log scale; basis points)
(log scale; basis points)
(log scale; basis points)
(log scale; basis points)
(log scale; basis points)
(log scale; basis points)
(log scale; basis points)
References

    ChiangM-H and C.Kao(2002) “Nonstationary Panel Time Series Using NPT 1.3 –A User GuideCenter for Policy Research, Syracuse University.

    • Search Google Scholar
    • Export Citation

    EichengreenB. and A.Mody1998What Explains Changing Spreads on Emerging Debt: Fundamentals or Market Sentiment?NBER Working Paper 6408 (Cambridge: National Bureau of Economic Research).

    • Search Google Scholar
    • Export Citation

    ErbC.HarveyC.R.ViskantaT.2000Understanding Emerging Market BondsEmerging Markets Quarterly. Spring2000 pp. 723.

    • Search Google Scholar
    • Export Citation

    IMF2004Appendix 1: Determinants of the Rally in Emerging Market Debt—Liquidity and FundamentalsGlobal Financial Stability Report; World Economic and Financial Surveys (Washington: International Monetary FundApril) pp. 6070.

    • Search Google Scholar
    • Export Citation

    KaminS.B. and K.von Kliest1999The Evolution and Determinants of Emerging Market Credit Spreads in the 1990sBIS Working Paper 68 (Basel: Bank for International Settlements).

    • Search Google Scholar
    • Export Citation

    KaminskyG.C.Reinhart and C.Vegh2003The Unholy Trinity of Financial ContagionJournal of Economic Perspectives. Fall2003Vol. 17No. 4 pp. 5174.

    • Crossref
    • Search Google Scholar
    • Export Citation

    KaminskyG.C.Reinhart and C.Vegh2004When It Rains, It Pours: Procyclical Capital Flows and Macroeconomic PoliciesNBER Working Paper 10780 (Cambridge: National Bureau of Economic Research).

    • Search Google Scholar
    • Export Citation

    KaoC.(1999) “Spurious regression and residual-based tests for cointegration in panel dataJournal of Econometrics90 pp.1-44.

    KuttnerKennethN.2001. “Monetary Policy Surprises and Interest Rates: Evidence from the Fed Funds Futures Market.” Journal of Monetary Economics47 (2001) 523544.

    • Crossref
    • Search Google Scholar
    • Export Citation

    McGuireP. and M.Schrijvers2003Common Factors in Emerging Market SpreadsBIS Quarterly Review (Basel: Bank of International Settlements).

    • Search Google Scholar
    • Export Citation

    ReisenH.2002Ratings since the Asian CrisisWorld Institute for Development Economics Research (WIDER), United Nations UniversityJanuary2002Discussion Paper No. 2002/2.

    • Search Google Scholar
    • Export Citation

    PedroniP.1995A Residual-Based Test for the Null of Cointegration in Panel DataEconometric Reviews17 pp. 57-84.

    SlokT. and M.Kennedy2004Factors Driving Risk PremiaOECD Working Paper No. 385 (Paris: Organization for Economic Co-operation and Development).

    • Search Google Scholar
    • Export Citation

    Standard & Poor’s2004Sovereign Credit Rating: A PrimerRatingsDirect, The McGraw-Hill CompaniesMarch152004New York.

    Standard & Poor’s2002Standard & Poor’s Ratings DefinitionsRatingsDirect, The McGraw-Hill CompaniesNew York.

    SyA.2002Emerging Market Bond Spreads and Sovereign Credit Ratings: Reconciling Market Views with Economic FundamentalsEmerging Market ReviewVol. 3Issue 4 pp 380408.

    • Crossref
    • Search Google Scholar
    • Export Citation

    The PRS Group Inc.2004The International Country Risk Guide DatabaseNew York.

    WilsonPeterVenezuela Debt Rating Cut to Selective Default by S&PCaracas, VenezuelaJanuary182005.

1Improvement in the ratings is based on the average of Moody’s and Standard & Poor’s ratings for each country aggregated using EMBIG weights.
2The model represents a type of forecasting model for spreads and, as such, does not distinguish between supply and demand factors for debt securities and their influence on spreads.
3See for example, the Global Financial Stability Report (April 2004) and Sy (2002).
4Rating agencies, however, may in turn argue that higher debt spreads increase borrowing costs and implying a greater risk of default. Thus, changes to ratings are not “behind the curve” but accurately reflect the increased risk of default as spreads widen.
5Eichengreen and Mody (1998), Kamin and von Kleist (1999), Sløk and Kennedy (2003), and McGuire and Schrijvers (2003) all find a negative or inconclusive relationship.
6To be included in the EMBI index, bonds had to meet strict liquidity criteria. For this reason, there were only five countries in the original EMBI. Their series start at the end of 1991. The EMBI Global, introduced in January 1998, uses more relaxed liquidity criteria. Countries’ admission requirements under the EMBI Global are different from those under the EMBI. Countries to be included in the EMBI must be rated BBB+ or lower by Standard & Poor’s. On the other hand, countries under the EMBI Global only need to satisfy one of the following criteria: (i) classified as having low or middle per capita income by the World Bank; (ii) has restructured external or local debt in past 10 years; or (iii) currently has restructured external or local debt outstanding.
7Nigeria, for example, has missing observations during the period between April 1998 and April 1999. Likewise, Pakistan has missing observations during the period of February 2003 through March 2004. In case of the Philippines, its EMBI series terminated in January 1997 (Table 1).
8Similarly, when a sovereign retires a debt instrument or amortizes principal of the debt, and if the issue’s current outstanding face value falls short of a required level under the criteria, the issue is removed from the country’s index. Market weight data are made available by J.P. Morgan Chase. This monthly series starts in December 1993.
9Sy (2002) uses EMBI Plus sovereign spreads, noting the additional advantage that they control for floating coupons, principal collateral, and rolling interest guarantees.
10There have been seven monthly observations which country’s long-term credit rating and/or outlook changed twice during the same month, though they are counted as one monthly change.
11Kaminsky et al. (2003) and Sy (2002) also refer to the importance of outlooks in their analysis of the spreads.
12Using Fed Funds futures rates, Kuttner (2001) disentangled expected from unexpected policy actions, and concluded that the impact of unexpected rate decisions on the interest rates of both short- and long-term maturities were significantly positive.
13See Global Financial Stability Report (2004).
14A detailed discussion of how to construct the CROI is found in Appendix 1.A.
15A Hausman test indicates that we cannot reject the null hypothesis of no systematic difference between the estimates of the random effect and those of the fixed-effect models. However, because we suspect that a country specific factor is not completely independent of the CROI or the LTCR, we apply the fixed-effect model of panel regression.
16Algeria and Cote d’Ivoire are excluded due to lack of the data on sovereign credit ratings and outlook. A total of 3,038 monthly observations are included for estimation. Initially, the basic model included Argentina, but because the crisis values for its spreads in 2001–2002 represented extreme outliers relative to any other historical period, Argentina is excluded from the remainder of the empirical work.
17As with the basic model, a Hausman test indicates that we cannot reject the null hypothesis of no systematic difference between the estimates of the random effect and those of the fixed-effect models. However, because we suspect that a country specific factor is not completely independent of the CROI or the LTCR, we apply the fixed-effect model of panel regression.
18The sample mean of V_FF is 12.2 percent, with a standard deviation of 6.6 percent.
19One standard deviation of the VIX is roughly 6.4 percent during the sample period of 1991 through 2007.
20At the dawn of the LTCM crisis on August 31, 1998, S&P 500 Index fell 6.8 percent from the previous day, and the VIX surged to 44.3. The VIX was jittery through the end of October 1998. It reached a historical high of 45.7 October 8, 1998.
21The tests were performed using the code provided by Chiang and Kao (2002). These tests can only be performed on balanced panels. Thus, we use two balanced samples, one including only June 1995 through February 2007 for Brazil, Mexico, Poland, South Africa, and Venezuela, and one shorter but broader panel for the period from January 1998 to February 2007 which additionally includes Colombia, Lebanon, Malaysia, Panama, Peru, Philippines, Russia, and Turkey. The results are robust across both panels.
22Aggregate EMBI series through November 1997, and aggregate EMBI Global series from January 1998 through February 2005 are spliced as to create the combined series.
23The decomposition is done for each country by calculating the share of the change in the log of the EMBIG spread that the model attributes to each variable for the chosen time period. These shares are then multiplied by the change in the model spread over the same period.
24The rating variable (Ri) takes a cardinal number in log such that AAA=1, AA+=2, …, and SD=22.
25Where STB is a dummy variable for a “stable” outlook, POS for a “positive” outlook, NEG for a “negative” outlook, R is for log-term credit ratings in log, and SPRD is log of spreads in basis points. We suppressed (n) for indicating the n-th observation of the sample.
26We divided the rating spectrum in a number of ways to see what combinations of rating groups provide the best fit and plausible significance levels for estimated coefficients.
27This univariate regression is based on the long-term credit ratings (R) only, without the three categories.
28Based on Breusch-Pagan/Cook-Weisberg test for heteroskedasticity, the null hypothesis is defined as a constant variance for the fitted value of SPRD.

Other Resources Citing This Publication