Volatility Spillovers and Dynamic Correlations among Foreign Exchange Rates and Bond Markets of Emerging Economies

Acknowledgement: Preliminary empirical results were presented at conferences including ASSA Meetings (MEEA) in Boston (January 2015), World Finance Conference in Buenos Aires (July 2015), 4th Economics and Finance Con-ference in London (August 2015). We benefited from comments and suggestions from Shawkat Ham-moudeh, Mehmet Balcilar and other participants of these conferences. We thank them all. We also thank the editor of the journal and two anonymous referees for valuable suggestions. This work was supported by the Scientific and Technological Research Council of Turkey (TUBITAK) (Grant No: 114K519). Volatility Spillovers and Dynamic Correlations among Foreign Exchange Rates and Bond Markets of Emerging Economies

The 2007-2008 global financial crisis reached its peak when Lehman Brothers declared its bankruptcy on September 15, 2008 (the Lehman shock in the paper) which resonated throughout the world's financial markets including those of developing countries. To avoid the risk of a financial collapse, the U.S. Federal Reserve (the Fed) has taken steps to launch several quantitative easing programs. As a result, billions of US dollars poured into the equity and bond markets of emerging countries. Sizable capital inflows significantly increased asset prices in emerging markets. However, after mid-2013 this trend tended to be reverse mainly due to the tapering decision of the Fed. The Fed's first exit signal from the quantitative easing programs was given by Chairman Bernanke on May 22, 2013 during a Congress hearing (the Tapering 1 shock). Eventually, the Fed announced its tapering decision (the Tapering 2 shock) on December 18, 2013 which amounted to reducing asset purchases beginning in January 2014.
These historical shocks are well known to have had a huge impact on foreign exchange, money and credit markets, especially in emerging countries such as Brazil,

Literature Review
Due to increasing interdependency and contagion between countries under the highly globalized financial markets, understanding the interaction dynamics of the financial indicators has been very crucial for decision makers recently. In particular, through magnifying volatility, financial shocks have great effects on the financial variables and markets across the world economy. There is a large empirical literature on volatility transmission models across financial markets and financial indicators. However, to the best of our knowledge, the literature on volatility transmission between exchange rates and interest rates for the countries of the Fragile Five, except individual country studies, hardly exists. We first briefly review the literature of the methodologies for volatility transmission and then the literature on the empirical studies for volatility transmission across financial indicators by focusing on the forex and bond markets.
The multivariate GARCH (MV-GARCH) models have been widely used for modelling volatility transmission recently. These models allow one to analyze the interrelations between financial markets. In a multivariate framework they permit one to model not only the mean of the variables but also the variance of the variables in both the short-and the long-run. For example, Tim Bollerslev (1986) transforms the ARCH model of Robert F. Engle (1982) into the generalized autoregressive conditional heteroskedasticity (GARCH) model by allowing the conditional variance to be an ARMA process in ARCH models. Since then, the ARCH-family models have developed rapidly and have been used to characterize financial markets extensively (Bollerslev, Ray Y. Chou, and Kenneth F. Kroner 1992;Bollerslev 2009). Luc Bauwens, Sébastien Laurent, and Jeroen V. K. Rombouts (2006) distinguish three approaches for constructing multivariate GARCH models. The first approach provides direct generalizations of the univariate GARCH model of Bollerslev (1986) such as the VEC, BEKK and

Author(s) Year Method Period Findings
John Wei-Shan Hu et al. 1997Causality, GARCH 1992-1996 daily Differentiated interrelationships among the stock markets.
Charles M. Jones, Owen Lamont, and Robin L. Lumsdaine 1998ARCH-M 1979-1995 Announcement-day volatility does not persist at all, consistent with the immediate incorporation of information into prices.
Edward S. Lim, John G. Gallo, and Peggy E. Swanson 1998VECM 1988-1993 Bidirectional causality exists between stock market returns and bond market returns.
George A. Christodoulakis and Stephen E. Satchell 2002Corr-GARCH 1985-1996 Fitting well the CorrARCH/CorGARCH to the data set from the G7 stock markets and stylized facts in the literature, such as time variation, persistence, volatility and correlation comovement.
Vasiliki D. Skintzi  Using varying methodologies to analyse volatility spillovers, the authors investigate different samples of countries and periods. Therefore, the results vary across studies. One of the main findings is that significant volatility spillovers are largely market-specific and different. The authors also find significant time-varying correlations between financial assets and volatilities of these assets which usually substantially increase during crises including the recent global financial crisis. Earlier research using linear GARCH-DCC models suffers from asymmetry and long memory problems. However, recent papers take into account these issues and provide more robust findings. As the above review suggests, most papers in the literature focus on the linkage between stock markets among developed or emerging markets. Moreover, although studies using VIRF analysis and FIAPARCH-DCC models separately looking at different assets are available in the literature, none of them specifically focus on foreign exchange and bond markets in the Fragile Five countries. We contribute to the literature to fill this gap by examining the dynamic interactions between exchange rates and bond rates of the Fragile Five in the context of the global financial crisis and the taper tantrum.

Model and Data
We obtained daily data for the ten-year bond rates and foreign exchange rates of Turkey, Brazil, India, Indonesia and South Africa from Bloomberg. In the exchange rate analysis, the sample ranges from 02 January 2007 to 04 November 2016 with 2569 observations, whereas 1683 observations of the ten-year bond interest rates of these five countries from 27 January 2010 to 03 November 2016 are used in the ten-year bond market investigation. Note that the ten-year bond data for Turkey prior to January 2010 is not available. The continuously compounded return of a market at time t is calculated as follows: where P it and P it-1 are the exchange rates of country i against the US-dollar at time t and t-1, respectively. The ten-year bond rates (i.e., yields) provided by Bloomberg are measured in local currency. The data for weekends and holidays are excluded from the sample. It is important to note that the levels of bond rates and the returns of exchange rates are stationary. But the levels of exchange rates are nonstationary. Throughout the paper, whenever we talk about exchange rates, we refer to their compounded returns unless otherwise stated. However, by the bond rates we simply mean their levels rather than their compounded returns. Table 2 presents the summary statistics and the Augmented Dickey-Fuller (ADF) unit root test statistics for the daily returns of the exchange rates and the level of ten-year bond rates. The exchange rate returns and the bond rates demonstrate common characteristics observed in the financial time series. All the exchange rate returns except Brazil and Indonesia are positively skewed and highly leptokurtic. Those returns are not normally distributed as shown by Jarque-Bera statistics. Regarding bond rates, we see that the Brazil bond rate is positively skewed, whereas all the others exhibit negative skewness. All the bond rates are also highly leptokurtic. They are not normally distributed indicated by Jarque-Bera statistics as expected as all of them are significant at 1% level. Daily values of the exchange rates and bond rates in the sample are illustrated in Figures A1 and A2 in the Appendix, respectively. We observe tremendous increases in volatility after the Lehman and Tapering shocks. The ADF test statistics associated with exchange rate returns and levels of ten-year bond rates are all significant at 1% level indicating that they are stationary.

Volatility Impulse Responses
In this study the volatility impulse responses are obtained using the diagonal BEKK model developed by Engle and Kroner (1995). The main advantage of BEKK specification is that it provides positive variances and the number of parameters to be estimated is reduced. The BEKK (1, 1) model we consider is as follows 2 : where matrix H is symmetric and contains conditional variances and covariances, C is a lower triangular matrix, and A and G are the parameter matrices. Matrix A captures how the conditional variance is correlated with past squared residuals, and hence measures the impact of shocks on volatility. Matrix G indicates the degree to which the conditional variances and covariances are linked to conditional variances and covariances of the previous period. Thus, it measures volatility spillovers. The residuals (u t ) are obtained from a VAR model and following Engle and Kroner (1995): where = 0, = and K is the number of variables. In all cases, we accommodate the residuals to follow a student t distribution. This is justified since series have excess kurtosis and skewness as shown in Table 1.
The student t-distributed residuals also accommodate the volatility impulse responses advanced by Hafner and Herwartz (2006). In fact, the volatility impulse response function delineates the impact of an "independent" shock on volatility of the variables. The independence of shocks from each other permits shocks to be obtained from historical data to construct volatility impulse response functions. However, in a multivariate framework it is hard to assume that shocks are independent if they occur at the same time. Cholesky decomposition is widely used for the orthogonalization of residuals. As an alternative to the Cholesky decomposition method, Hafner and Herwartz (2006) use the Jordan decomposition to obtain independent shocks. Consequently, the impulse responses that are obtained turn out to be free from issues involving ordering of the variables. Moreover, to trace out the effects of a historical shock related to a specific event, we estimate the model for the whole sample rather than using a sub-period corresponding to that event.
Using the Jordan decomposition Hafner and Herwartz (2006) decompose H t in order for identical and independent shocks to be obtained from Equation (1) as follows: After the decomposition, the independent shocks can be given as: = . (5) The volatility impulse responses are defined as the difference between the expectation of volatility conditional on an initial shock and the history where F is the coefficient matrix obtained from vech representation of Equation (1) and and represent the Moore-Penrose inverse and duplication matrix, respectively. For t ≥ 2, the volatility impulse responses are obtained as follows: for ≥ 2. (7)

Dynamic Correlations
To analyze dynamic correlations between exchange rates and bond rates we employ a multivariate AR(1)-FIAPARCH-DCC model. The FIAPARCH specification allows for an asymmetric response of volatility to positive and negative shocks and also long memory volatility dependence. The FIAPARCH combines the FIGARCH specification of Baillie, Bollerslev, and Mikkelsen (1996) and the APARCH model of Zhuanxin Ding, Clive W. J. Granger, and Engle (1993). In this paper, we use the FIAPARCH (1, d, 1) specification as follows: where ω > 0, δ > 0, β < 1 and λ < 1. The innovations which follow student-t distributions are derived from an AR(1) model. In Equation (8),  stands for the leverage parameter and  represents the power term and d (0 ≤ ≤ 1) is the long memory parameter, which captures persistence and long memory in volatility. It should be noted that for the conditional variances to be positive the sufficient condition (Christian Conrad and Berthold R. Haag 2006 and Conrad 2010) is that: Moreover, the FIAPARCH reduces to the APARCH when d = 0 and to GARCH when  = 2 and  = 0.
After the FIAPARCH specification is fitted for all series in the first step, in the second step to estimate the DCC parameters the series are transformed using the estimated standard deviations obtained from the FIAPARCH models. We apply the DCC approach developed by Engle (2002) to obtain the time varying correlations. Following Engle (2002), suppose * , R t and D t denote the × conditional covariance matrix, the conditional correlation matrix, and a diagonal matrix with time-varying standard deviations on the diagonal, respectively. Then we can write * as follows: * = , where the terms ℎ , * are conditional variances obtained from AR(1)-FIAPARCH(1, d, 1) model.
where = , is nxn symmetric positive definite matrix and is the unconditional correlation matrix of the standardized residuals * ( * = , , * ). The parameters θ and α are positive and satisfy the condition that θ + α < 1. The innovations , are assumed to be conditionally t-distributed and are obtained from an AR (1) specification to each series: The dynamic correlations can be obtained as follows: 3. Estimation Results -Volatility Impulse Responses

Volatility Impact from Past Shocks: Foreign Exchange Markets
In this section, we illustrate the volatility impulse response function analysis for the foreign exchange markets. We consider three historical shocks: Lehman Bankruptcy, Tapering 1 (i.e., Bernanke's first exit signal from the quantitative easing programs) and Tapering 2 (the Fed's tapering decision), which are within the sample period. Figure 1 shows the time pattern of the impulse response of volatilities after each historical event. Regarding the Lehman shock, it demonstrates that there are significant volatility increases stemming from this shock on four conditional variances (except India) after the Lehman bankruptcy. This suggests that the Lehman bankruptcy event has a significant but different impact on each currency in the sample. We find the largest positive effect for the post-Lehman era for the returns of Brazil's currency. Its onestep ahead expected conditional variance is increased by around 8.5%. The post-bankruptcy positive impact in expected conditional variance of exchange rate returns of South Africa, Turkey, Indonesia and India are about 7%, 4%, 2% and 0.4%, respectively for the one-step ahead expected conditional variance. The impact is not statistically significant for India however. Hence, in terms of impact (the maximum value of the VIRF, which is to say the peak value), we have the following ranking for the countries in the sample (from highest to lowest): Brazil > South Africa > Turkey > Indonesia > India.

The Lehman Bankruptcy
Source: Authors' own calculations. Except for Indonesia, the impact is instantaneous for all other countries. The largest response of conditional variance happens after 25 days for Indonesia. The results reported above indicate that all four countries' exchange rates except India exhibit a responsiveness to shocks and that their magnitudes are higher than that of India's exchange rate. These findings may reflect the fact that India is more insulated and Brazil is more fragile to shocks relative to other emerging markets. The trade characteristics of Brazil indicate that its export revenues are highly sensitive to commodity prices as commodity products account for most of its exports. Besides, commodity prices are expressed in US dollars and the two (i.e., commodity prices and the US dollar) have been negatively correlated between 2003 and mid-2016. Therefore, it is unsurprising to see that volatility in Brazil's currency strongly depends on the US economy and movements in the USD (Riadh Aloui, Mohamed Safouane Ben Aïssa, and Nguyen 2011). On the other hand, unlike Brazil, Indian exports rely mostly on manufactured products and we observe a relatively low degree of openness in trade for India. This factor in turn insulates the Indian rupee against global financial shocks such as the Lehman bankruptcy (Dimitriou, Kenourgios, and Simos 2013). Furthermore, the effects of the shock reveal that the durations of the impacts are relatively long. In other words, the impact of the shock seems to be highly persistent for all countries.

Figure 1 Volatility Impulse Responses Functions for Exchange Rate Returns with 99% Confidence Intervals
Overall, the findings associated with the Lehman shock suggest that the sensitivity of foreign exchange markets from the impact of shocks are largely market-specific and different, although the foreign exchange markets concurrently soak up shocks. While Brazil seems to be more volatile and responsive than other markets regarding the volatility impulse response analysis associated with the Lehman collapse, India is the least responsive to the shock both in terms of size and persistency.

The Tapering Events
Tapering talks have inundated the global financial markets especially after Chairman Bernanke's speech during a U.S. Congress hearing on May 22, 2013 when he gave the first exit signal from the Fed's quantitative easing programs. Eventually, the Fed announced its tapering decision on December 18, 2013 signaling that it would reduce its purchases of bonds and other financial assets from banks beginning in January 2014. Since tapering basically indicates the end of easy money in financial markets, emerging economies especially those of the Fragile Five have been adversely affected by this decision. In this study, in addition to the Lehman shock, we investigate the impact of these two events, the exit signal by Bernanke on May 22, 2013 (Tapering 1) and the Fed's announcement of tapering decision on December 18, 2013 (Tapering 2) on foreign exchange rates volatilities as well. Figure 1 shows that there are statistically significant volatility increases resulting from impact of the Tapering 1 shock on conditional variances of exchange rate returns of Turkey, Indonesia and India after Bernanke's exit signal. However, the size of the impacts is much lower than that of the Lehman shock. The positive impact for the post-Tapering 1 on expected conditional variance of exchange rate returns of India, Indonesia and Turkey are about 0.6%, 0.35% and 0.1 % respectively for the one-step ahead expected conditional variance. Unlike the Lehman shock, while India seems to be more volatile and more responsive than other markets regarding the volatility impulse response analysis associated with Tapering 1, Brazil is the least sensitive to the shock both in terms of size and persistency. Indeed, the impact of the shock is not statistically significant for Brazil and South Africa. To sum up, both the size and persistency of the impact of the shock on exchange markets are much less than the Lehman shock.
Tapering 2 seems to have a more significant impact than Tapering 1 on the volatilities of exchange rate returns. Specifically, the positive impact for the post-Tapering 2 on the expected conditional variance of exchange rate return of Turkey, Indonesia, Brazil, India and South Africa are about 0.6%, 0.45%, 0.45%, 0.4% and 0.35%, respectively, for the one-step ahead expected conditional variance. Hence, in terms of impact (peak), we have the following ranking for the countries (from highest to lowest): Turkey > Brazil > Indonesia > India > South Africa. We also find that the impact tends to gradually disappear within significantly longer period than Tapering 1, but in much shorter period than the Lehman impact does.
Overall, our findings indicate that in terms of both size and persistency the Lehman shock has a more significant impact on the exchange rate volatilities in the Fragile Five than both tapering events. During the taper tantrum, capital flowed out of emerging economies which caused their currencies to depreciate. To avoid sudden or large depreciation, authorities intervene in foreign exchange markets. Guillermo A. Calvo and Carmen M. Reinhart (2002) document that some emerging market economies with floating exchange rates are actually unwilling to allow their currencies fluctuate in response to shocks, which they call the "fear of floating". Indeed, the Fragile Five countries which were susceptible to the fear of floating took measures such as currency intervention and policy rates increases to maintain stability in their currency markets during the taper tantrum. Our findings related to the tapering shocks are consistent with the earlier literature and these measures (Yildirim 2016).

Volatility Impact from the Past Shocks-Ten-Year Bond Markets
In this section, we present the volatility impulse response functions analysis for the ten-year bond markets. The ten-year bond data for Turkey is not available prior to March 12, 2010. Hence, we consider two historical shocks (Tapering 1 and Tapering 2) that are within the sample period. Figure 2 shows the time pattern of the impulse response of volatilities after each historical event. It demonstrates that there are large statistically significant volatility increases stemming from this shock on all five conditional variances after Bernanke's exit signal. This indicates the tapering signal has a significant but different impact on each bond market in the sample. We find that the largest positive impact for this postsignal can be discerned for Brazil's ten-year bond rates. Its one-step ahead expected conditional variance is increased by around 140%. The positive impact for the postsignal on expected conditional variance of ten-year bond rates of Indonesia, Turkey, India and South Africa are about 50%, 37%, 29% and 16%, respectively for the onestep ahead expected conditional variance. Therefore, in terms of the impact (peak), we The impact is instantaneous for all countries except Indonesia for which the largest response of conditional variance occurs after 30 days. The results reported above indicate that all five countries' bond rates especially Brazil exhibit a high responsiveness to the shock. These findings may reflect the fact that Brazil's ten-year bond market is much less insulated from shocks relative to other emerging markets. Furthermore, the impacts are relatively perpetuated, although less so than is the case in foreign exchange markets. In other words, the impact of the shock seems to be persistent for all countries as they do not die out after 200 days on average. The reason for this has to do with several eigenvalues of the matrix AA + GG being very close to unity.

Tapering 1
As a conclusion, the findings suggest that the sensitivity of bond markets to the impact of shocks is largely market-specific despite the fact that the shocks are absorbed by bond markets concurrently. Brazil seems to be more volatile and responsive than the other markets regarding the volatility impulse response analysis associated with Tapering 1. We also find that South Africa is the least sensitive to the shock in terms of size. Figure 2 shows that there are statistically significant volatility increases stemming from the Tapering 2 shock on only two conditional variances after the Fed's tapering announcement (i.e., Brazil and India). The size of the impacts is much lower than the Tapering 1 shock. The positive impact for the post-Tapering 2 on expected conditional variance of bond rate return of Brazil and India are about 12% and 9%, respectively for the one-step ahead expected conditional variance. The Tapering 2 shock has significantly much less impact on bond markets than the Tapering 1 shock in terms of size. We also find that the impact steadily vanishes within significantly shorter period than Tapering 1. Again, similar to Tapering 1, while Brazil seems to be more volatile and responsive than the other markets regarding volatility impulse response analysis associated with Tapering 1, Indonesia is the least responsive to the shock in terms of size. To sum up, both the size and persistency of the impact of the shock is much less than for Tapering 1.

Tapering 2
Overall, regarding the volatility impulse response analysis in the foreign exchange and bond markets in the Fragile Five, we have the following results. First, our findings suggest that Brazil is the most sensitive country among the Fragile Five to the shocks under investigation both in the exchange rate and the bond markets. This underscores the fragility of Brazil's financial markets to such shocks. Second, in terms of size, we observe that shocks specifically Tapering 1 and Tapering 2 seem to have a more impact on the bond markets than on the exchange rate markets. As a result, we can infer that the Fed's tapering decision affects bond markets more than exchange rate markets. This result might be attributable to the fact that during the taper tantrum, foreign investors have gradually increased their share in bond markets of emerging economies. Surges in foreign ownership in turn make bond rates more responsive to global shocks. These results are consistent with David Bowman, Juan M. Londono, and Horacio Sapriza (2015) and Yildirim (2016).  Table 3 presents the estimation results of the AR (1) -FIAPARCH (1, d, 1) -DCC model for exchange rates. We find that most of the coefficients are significant and except for the asymmetry term (), all coefficients have expected sign. Positive and insignificant coefficients () of asymmetry terms indicate the absence of leverage effects.
Furthermore, the condition that conditional variances are positive is satisfied for all models. The long memory coefficients (d) are also significant at 5 % and 1 % level of significance suggesting that exchange rate volatilities are persistent. The power term coefficients () are all positive and significantly different from 2 confirming the appropriateness of the FIAPARCH model in our analysis. As seen from the DCC estimates in Table 3, correlations between exchange rates are positive and statistically significant at 5% or 1% level of significance. The highest average correlation is between the Turkish Lira and the Rand while the lowest one is between the Real and Rupiah. Furthermore θ and α parameters are significant for the DCC model. This implies that the DCC model provides a good fit. On the other hand, the sum of the θ and α parameters is less than unity indicating that conditional correlations are mean reverting.
Diagnostic test results for the AR (1)-FIAPARCH-DCC model are presented at the bottom of Table 3 with probability values of the Ljung-Box Q 2 (50), Hosking squared (50) and Li-McLeo squared (50) statistics show that there is no remaining ARCH effect in the residuals. Figure 3 plots the estimates of dynamic conditional correlations (DCC), which are obtained from AR (1)-FIAPARCH-DCC model, between the exchange rate returns of different country combinations during the period January 2007-November 2016. We have total 10 combinations of the DCC plots since there are five countries in the sample for the exchange rate analysis. We observe that the DCC plots exhibit relatively high and positive correlations for most cases and low but positive correlations in others.
We first consider the estimates of the dynamic conditional correlations (DCC) between the exchange rate returns of Turkey (TR) and those of Brazil (BR), India (II), Indonesia (IO) and South Africa (SA) during the period January 2007-November 2016. The results suggest that the DCCs increase sharply after both the Lehman bankruptcy and the Tapering shocks. This result is in line with earlier studies (see Table 4 in Itir Ozer-Imer and Ibrahim Ozkan 2014 for a list of studies). The result may be attributable to several factors. Due to their shared macroeconomic characteristics (i.e., after all, they are categorized as emerging markets), large investors in forex markets including international banks might treat them similarly which in turn could cause positive and high correlations between their currencies. During turmoil times participants in FX markets might prefer safer assets and therefore flight to quality behavior tend to increase. In other words, capital flows from relatively risky emerging market assets to safer ones like the US treasuries. Source: Authors' own calculations.

Figure 3 The DCCs for the Exchange Rates
We also find that DCCs for South Africa are systematically larger than those for Brazil, which in turn are larger than those for India. And DCCs for Indonesia exhibit the lowest correlations. Regarding other country combinations, we see similar DCC patterns in the sense that the DCCs increase after the shocks, consistent with the literature. Overall, we also observe relatively high and positive correlations among the exchange rate returns of Turkey, Brazil and South Africa and, to some extent, India. However, the DCCs of Indonesia and those of others are relatively lower. In other words, we can group the sample into two subgroups: (a) Turkey, Brazil, South Africa and India; (b) Indonesia. This might be partly due to the rising influence of China on the Indonesian economy as it serves as Indonesia's the largest export and import market, but also due to specific monetary (policy rate hikes, currency intervention) and fiscal (reduction of energy subsidies etc.) policies of Indonesia (M. Chatib Basri 2017). Table 4 shows the estimation results of the AR (1)-DCC-FIAPARCH(1, d, 1) specification for ten year bond rates. The results are similar to those obtained with exchange rates. Most of the coefficients are significant and except for the leverage coefficient all coefficients have the expected signs. The fractional integration coefficient (d) is significant for all bond markets. Moreover, the power term coefficients () are significantly different from 2 justifying the use of a FIAPARCH specification. The FIAPARCH coefficients satisfy the restrictions shown in Equation (9) indicating that conditional variances are positive. The DCC estimates in Table 4 suggest that the correlation coefficients obtained from the DCC model are significant except for the one between Turkey and India. We also find that the coefficients θ and α are significant and that their sum is less than unity indicating that the conditional coefficients are mean reverting. The insignificant diagnostic test statistics presented at the bottom of the table confirms the appropriateness of the AR (1)-FIAPARCH-DCC model used in the study. Figure 4 plots the estimates of dynamic conditional correlations (DCC) between the ten-year bond rates of different country combinations during the period January 2010-November 2016. Note that ten-year bond rate data is not available prior to January 2010 for Turkey. We observe that DCC plots exhibit both positive and negative correlations during the sample period.

Ten-Year Bond Markets
Our results indicate that the correlations between bond markets are lower than those between foreign exchange markets. For example, while the average of DCCs between ten-year bond rates of Turkey and those of India is 0.02, the average of exchange rate return correlations turns out to be 0.45 (see summary statistics of the DCC series in Tables A1 and A2 in the Appendix). We observe the same differences with other country combinations. This result, which is one of the main findings of our study, suggests that bond markets, at least ten-year bond markets, provide a better diversification opportunity for investors than foreign exchange markets in the emerging markets under investigation. When we consider the dynamic correlations of Turkey with others, we observe that the DCCs between Turkey and South Africa are relatively higher than those with Indonesia, which are in turn higher. Than those with Brazil. The DCCs between Turkey and India exhibit lower correlations. Indeed, the DCCs of India with others demonstrate the lowest correlations overall. In other words, we can infer that the Indian ten-year bond market moves independently from others during the sample period.
To compare the dynamic conditional correlations between the ten-year bond markets and the foreign exchange markets, we conduct mean-and variance-tests. The average correlations in the foreign exchange markets are significantly larger than those in the bond markets. Several factors may explain these patterns. Central banks' coordinating actions concerning exchange rates help increase correlations. Moreover, it might have to do with the fact that the forex market is the largest and the most liquid asset market in the world. Indeed, according to the BIS triennial report of 2016, the average daily trading volume in the foreign exchange market was around 5.1 trillion USD in April 2016. However, average daily trading in global bond markets was around 700 billion USD. And the US Dollar, being the most critical currency, accounts for 85% of forex trading volume (Bank for International Settlements -BIS 2016). Since we measure each currency in our sample against the USD and movements in the value of the USD would automatically lead to appreciation or depreciation in other currencies which in turn cause higher correlations. Unlike the currencies in our sample, bond rates are measured in local currencies rather than USD. Therefore, idiosyncratic domestic factors are more instrumental in determining the sensitivity of bond rates than forex markets to global financial shocks (Fernanda Nechio 2014). Regarding the volatilities (captured by the variance) of the DCC series, the variance-tests suggest that the correlations among the ten-year bond markets exhibit less volatility than the ones among the exchange rate markets. This result may also be attributed to the fact that global exchange rate markets are much larger and more liquid than global bond markets.

Conclusion
Different methods have been utilized in the literature to analyze various financial markets focusing on different episodes and country samples including advanced and emerging markets. Results might differ from each other depending on the method and sample used in the analysis. Nonetheless, most of the literature documents significant volatility spillover effects among financial markets. As far as we know, this is the first paper using the FIAPARCH-DCC model and volatility impulse response analysis a la Hafner and Herwartz (2006) which examines the volatility spillover effects among the foreign exchange and ten-year bond markets of the emerging economies of the socalled the Fragile Five. Through VIRF, we measure not only the size but also the persistence of volatility spillovers of the above-mentioned financial markets. For this purpose, we consider three important historical shocks, the Lehman bankruptcy and two shocks associated with tapering decisions (Bernanke's first tapering signal and the Fed's eventual tapering decision) that have had a tremendous impact on the global financial markets.
Regarding the VIRF analysis of the foreign exchange markets, our results suggest that in terms of both size and persistency the Lehman shock had a more significant impact on the exchange rate volatilities for each country in the Fragile Five than both of the tapering events. Between Tapering 1 and Tapering 2, we find that the latter seems to affect exchange markets more than the former. For ten-year bond markets, due to lack of data we exclude some observations prior to January 2010 from the analysis. Focusing on the impact of only the tapering shocks, we find that in contrast with the foreign exchange markets, both the size and persistency of the impact of the Tapering 2 shock is much less than the Tapering 1 shock. We also find that the Fed's tapering decision affects bond markets more than it affects foreign exchange markets in the Fragile Five. a significant increase in foreign ownership in bonds of emerging economies during the taper tantrum might explain this result, which is consistent with earlier literature (Bowman, Londono, andSapriza 2015 andYildirim 2016).
We also investigate separately the dynamic conditional correlations (DCCs) among the exchange rate and the ten-year bond markets of the Fragile Five by employing an AR (1)-FIAPARCH-DCC model. We find that the DCCs of exchange rate returns increase sharply after both the Lehman bankruptcy and the two Tapering shocks. Similarly, those of bond rates increase only slightly or decrease after the two Tapering shocks. Our results also suggest that the DCC series of bond rates exhibit much lower correlations than those associated with exchange rate returns. This might be due to the fact that global exchange rate markets are much larger and more liquid than global bond markets (Nechio 2014). This result provides useful information for investors as it indicates that ten-year bond markets provide a better diversification opportunity than foreign exchange markets in the Fragile Five.
The VIRF and DCC results also yield important lessons for central banks and other regulators as to how to devise appropriate measures regarding financial stability. Bond markets overall seem to be more responsive than forex markets to financial shocks. Central banks have several monetary policy instruments to influence the economy. The most commonly used one is targeting of short term interest rates (policy rates) which indirectly determine the long-term interest rates in the economy.
According to the impossible trinity in international finance, it is impossible to have all three of the following simultaneously: (a) free capital flows; (b) a fixed foreign exchange rate; (c) an independent monetary policy (Maurice . In other words, in the absence of capital controls, there is a trade-off between a fixed exchange rate and independently set interest rates. Our results imply that central banks should be more cautious about interest rates rather than exchange rates when confronted with the trade-off as foreign exchange rates are found to be relatively more stable than long-term interest rates.
Our results also provide a ranking of the Fragile Five countries in terms of their resilience to financial shocks. According to the ranking, Brazil seems to be the most fragile country among them. Therefore, Brazilian authorities should be more proactive in curbing instability in its currency and bond yields. The other countries could also assess their own resilience to shocks by comparing their position with others so that they can devise more effective policy measures. This paper could be extended by analyzing how other historical and hypothetical shocks that may occur in the future would affect the financial markets of the Fragile Five. It would also be interesting to examine the determinants of the DCCs among these countries to see the rationale behind the observed correlation patterns because the insights would be very helpful for policy makers and portfolio managers. We would like to address these issues as future research topics as well.