Do Local and Global Factors Impact the Emerging Markets’s Sovereign Yield Curves? Evidence from a Data-Rich Environment

Transcription

1. Do Local and Global Factors Impact the Emerging Markets ’s Sovereign Yield Curves? Evidence from a Data-Rich Environment Oğuzhan ÇEPNİ İbrahim Ethem GÜNEY Doruk KÜÇÜKSARAÇ Muhammed Hasan YILMAZ March 2020 Working Paper No: 20/04
2. © Central Bank of the Republic of Turkey 2020 Address: Central Bank of the Republic of Turkey Head Office Structural Economic Research Department Hacı Bayram Mh. İstiklal Caddesi No: 10 Ulus, 06050 Ankara, Turkey Phone: +90 312 507 80 04 Facsimile: +90 312 507 78 96 The views expressed in this working paper are those of the author(s) and do not necessarily represent the official views of the Central Bank of the Republic of Turkey.
3. Do local and global factors impact the emerging markets ’s sovereign yield curves? Evidence from a data-rich environment∗ Oguzhan Cepni1 , I. Ethem Guney1 , Doruk Kucuksarac1 , and M. Hasan Yilmaz1 1 Central Bank of the Republic of Turkey Abstract This paper investigates the relation between yield curve and macroeconomic factors for ten emerging sovereign bond markets using the sample from January 2006 to April 2019. To this end, the diffusion indices obtained under four categories (global variables, inflation, domestic financial variables, and economic activity) are incorporated by estimating dynamic panel data regressions together with the yield curve factors. Besides, in order to capture dynamic interaction between yield curve and macroeconomic/financial factors, a panel VAR analysis based on the system GMM approach is utilized. Empirical results suggest that the level factor responds to shocks originated from inflation, domestic financial variables and global variables. Furthermore, the slope factor is affected by shocks in global variables, and the curvature factor appears to be influenced by domestic financial variables. We also show that macroeconomic/financial factors captures significant predictive information over yield curve factors by running individual country factor-augmented predictive regressions and variable selection algorithms such ridge regression, LASSO and Elastic Net. Our findings have important implications for policymakers and fund managers by explaining the underlying forces of movements in the yield curve and forecasting accurately dynamics of yield curve factors. ¨ Ozet Bu makalede, Ocak 2006-Nisan 2019 arası o¨ rneklem kullanılarak, 10 adet gelis¸mekte olan u¨ lke ic¸in devlet tahvili getiri e˘grisi fakt¨orleri ile makroekonomik fakt¨orler arasındaki ilis¸ki aras¸tırılmaktadır. Bu amac¸la, d¨ort kategori (k¨uresel de˘gis¸kenler, enflasyon, yurtic¸i finansal de˘gis¸kenler ve ekonomik aktivite) altında elde edilen dif¨uzyon endeksleri ile verim e˘grisi fakt¨orleri arasındaki ilis¸kiyi dinamik panel veri regresyonları olus¸turularak incelenmektedir. Ayrıca, getiri e˘grisi ile makroekonomik / finansal fakt¨orler arasındaki dinamik etkiles¸imi yakalamak ic¸in sistem GMM yaklas¸ımına dayalı bir panel VAR analizi kullanılmaktadır. Ampirik sonuc¸lar, seviye fakt¨or¨un¨un enflasyon, yurtic¸i finansal de˘gis¸kenler ve k¨uresel de˘gis¸kenlerden kaynaklanan s¸oklara yanıt verdi˘gini g¨ostermektedir. Di˘ger yandan, e˘gim fakt¨or¨u k¨uresel de˘gis¸kenlerdeki s¸oklardan etkilenirken e˘grilik fakt¨or¨u u¨ zerinde ise yurtic¸i finansal de˘gis¸kenler etkili olmaktadır. Son olarak, makroekonomik / finansal fakt¨orlerin, o¨ rneklemde yer alan u¨ lkelerin getiri e˘grisi fakt¨orlerini tahmin etmede o¨ nemli bir o¨ ng¨or¨u performansına sahip oldu˘gu fakt¨or eklenmis¸ regresyonlar ve Ridge, LASSO ve Elastik Net gibi de˘gis¸ken sec¸im algoritmaları kullanılarak g¨osterilmektedir. Bulgularımızın, getiri e˘grisindeki hareketlerin altında yatan etkenlerin anlas¸ılmasına ve getiri e˘grisi fakt¨orlerinin do˘gru tahmin edilmesine katkı sunarak politika yapıcılar ve fon y¨oneticileri ic¸in fayda sa˘glayaca˘gı d¨us¸u¨ n¨ulmektedir. 1
4. Keywords : Yield Curve, Macroeconomic Factors, Nelson Siegel Model, Panel VAR, Forecasting, Variable Selection. JEL Classification: C1, C5, F2, F3, F4, G1. ∗ The views expressed in this paper are only those of the authors and should not be interpreted as reflecting those of the Central Bank of the Republic of Turkey. Email addresses : Oguzhan Cepni (Oguzhan.Cepni@tcmb.gov.tr), I. Ethem Guney (Ethem.Guney@tcmb.gov.tr), Doruk Kucuksarac (Doruk.Kucuksarac@tcmb.gov.tr) and M. Hasan Yilmaz (Muhammed.Yilmaz@tcmb.gov.tr) Central Bank of the Republic of Turkey, Haci Bayram Mah. Istiklal Cad. No:10 06050 Ulus, Altındag, Ankara, Turkey. 2
5. Non-Technical Summary In conformity with the experienced wave of globalization and financial development in emerging markets (EM) for last decades, local bond markets of those economies have gained prominence regarding the financing of economic entities. In recent years, monetary policies of advanced economies have been designed on the expansionary front coupled with unconventional measures leading to abundant liquidity in financial system. This leads to a swift decline in policy rates and put downward pressure on the yields of financial assets. Given the “search for yield” behavior of global investors and relatively higher yields offered by EM financial assets, there have been observed voluminous capital inflows to EM economies through debt securities. Thus, amplified interest in local EM bonds has brought the modelling of sovereign yield curve dynamics into attention. Given the relevance of macroeconomic fundamentals with yield curve modelling parameters, recently, the academic literature has focused on extensions of yield curve models that incorporate macroeconomic variables. Building on these views, our aim is to extend the empirical evidence addressing the question of how yield curve factors are related to macroeconomic factors for ten emerging sovereign bond markets using the large set of macroeconomic and financial variables. To this end, firstly, we extract the yield curve factors by employing NS methodology. Secondly, we estimate dynamic panel data regressions between yield curve factors and latent factors that are constructed on a relatively large set of economic indicators, namely global variables, domestic financial variables, economic activity and inflation related variables. Furthermore, we provide a characterization of the dynamic interactions between the yield curve and macroeconomics factors via estimating panel VAR model and computing impulse response functions. It is observed that a shock coming to inflation factor is being transmitted to a positive response in the level factor of yield curve. Second, one unit shock of innovation to the factor representing economic activity brings about an increase in the slope component leading to steepening in EM countries’ yield curve formation. Moreover, the response of slope factor to the shocks in global factor is negative long lasting effect on the slope factor. 3
6. 1 . Introduction In conformity with the experienced wave of globalization and financial development in emerging markets (EM) for last decades, local bond markets of those economies have gained prominence regarding the financing of economic entities. In recent years, monetary policies of advanced economies have been designed on the expansionary front coupled with unconventional measures leading to abundant liquidity in financial system. This leads to a swift decline in policy rates and put downward pressure on the yields of financial assets. Given the “search for yield” behavior of global investors and relatively higher yields offered by EM financial assets, there have been observed voluminous capital inflows to EM economies through debt securities. Furthermore, differentiation of local economic factors in EM countries from developed counterparts creates potential diversification benefits and strong appetite for this asset class (Miyajima et al. (2015)). Apart from this, broadening and deepening of cross-border financial links reinforced by the behavior of global market participants, financial liberalization attempts of local authorities and economic integration actions of supranational organizations have resulted rising momentum of local EM bond markets (Wooldridge et al. (2003), Garcia-Herrero and Wooldridge (2007)). Accordingly, prominent foreign interest in EM debt securities have brought financial deepening in the market structure. International bond issuances from selected EM group appears to increase dramatically since the beginning of 2000s (Figure 1). On the other hand, domestic bond markets mostly consisted of local currency issuances also grew in size1 . Although, these markets were historically dominated by sovereign entities, non-financial corporations also started to obtain financing through local debt markets (Figure 1). − Insert Figure 1 about here. − Amplified interest in local EM bonds has brought the modelling of sovereign yield curve dynamics into attention. One part of the literature analyzes the yield curve dynamics via affine term structure models. In this context, single-factor model, Cox et al. (1985) model and multi-factor model are vastly utilized which impose restrictions ruling out arbitrage opportunities (Baz and Chacko (2004), Ahokpossi et al. (2016)). On the other hand, the other part of the literature has tried to associate yield curve dynamics with few unobservable latent factors. The most popular parametric technique among academia, practitioners and 1 This data corresponds to the summation of notional outstanding amount of selected EM countries comprising the sample of this study. Details about countries are provided in the Section 2. 4
7. central bankers is founded by the study of Nelson and Siegel (1987) which disentangles the entire yield curve into a three unobservable (latent) factors as well as their related factor loadings (Paccagnini (2016)). In this context, level, slope and curvature factors are thought to govern the yield curve formations. From theoretical point of view, the macro-finance literature establishes strong connections between yield curve factors and macroeconomic forces(Sarno et al. (2007), Ang et al. (2008)). Given the fact that level factor represents long-term rates, natural consequence is that it should be in association with inflationary tendencies. As stated by Hannikainen (2017), monetary authorities tend to raise short term rates in the face of overheating and inflationary pressures. That kind of policy action is typically followed by moderation in economic growth. Moreover, if economic agents expect that monetary policy will be eased once inflationary pressures are contained, short term rates are likely to increase more than the possible movements in long term rates which would result in flattened or even inverted yield curve. Hence, many studies in the literature hypothesize that slope parameter tends to be countercyclical and carry predictive power for future economic activity (Stock and Watson (2003), Ang et al. (2006), Argyropoulos and Tzavalis (2016) among others). On the other hand, curvature factor is thought to capture monetary policy actions related to short or medium term adjustments of the current stance of monetary policy (Bekaert et al. (2010), Dewachter and Lyrio (2006), M¨onch (2012)). Given the relevance of macroeconomic fundamentals with yield curve modelling parameters, recently, the academic literature has focused on extensions of yield curve models that incorporate macroeconomic variables. Diebold et al. (2006) combine the Nelson-Siegel (NS) model with inflation, output and the policy rate to provide new insights into the relationship between the term structure of interest rates and the US economy. They show that the level factor is somewhat correlated with inflation whereas the slope factor is related to economic activity. Unlike the level and the slope factors, the curvature factor appears to be unrelated with any of the main US macroeconomic variables. Lu and Wu (2009) examine the interactions between the US Treasury yield curve and 17 macroeconomic data and find that shocks on inflation-related variables such as consumer prices and producer prices have sizeable positive impact on the yield curve, resulting in a parallel shift across different maturities. On the other hand, shocks on real activity variables such as GDP growth, industrial production and capacity utilization have broader impacts on the short end than on the long end of the yield curve, thus resulting in a flatter or a steeper yield curve. Hannikainen (2017) evaluates the predictive content of yield curve factors for US real activity in a data-rich environment. It is shown that while predictive power is subject to alterations from a historical perspective, slope emerges 5
8. as a key parameter in understanding the real economic fluctuations . By covering the advanced economies such as US, Germany, Canada and UK, Argyropoulos and Tzavalis (2016) provide evidence supporting the importance of slope and curvature for future changes in economic activity. As a recent study, Paccagnini (2016) interpolates term structure of the US Treasury Rates for the period 1984-2007 with the help of three yield curve parameters. It is founded that term structure informs policymakers about how macroeconomic shocks are related to yield curve dynamics. While there exists a profound literature about developed markets, few studies directly examine EM yield curve parameters and their interaction with macroeconomic environment. Kanjilal (2013) examines the debt market in India over the period 1997-2011. By applying NS methodology, sovereign yield curve is estimated and almost all of the movements across yield curve is found to be explained by latent level, slope and curvature factors. Alves et al. (2011) utilize the dynamic version of NS model to reproduce stylized facts of Brazil’s term structure and find that the model fits the data well. Kaya (2013) conducts the same exercise for Turkish economy and find similar results. In a more recent study, Sowmya and Prasanna (2017) analyze the contemporaneous relation between macroeconomic factors and yield curve movements in nine Asian sovereign bond markets. They indicate that increases in the policy rate and inflation affect the slope of the term structure while the output growth has a significant influence on the long-term rates in the region. The purpose of this paper is to extend the empirical evidence adressing the question of how yield curve factors are related to macroeconomic factors for ten emerging sovereign bond markets using the large set of macroeconomic and financial variables. To this end, firstly, we extract the yield curve factors by employing NS methodology. Secondly, we estimate dynamic panel data regressions between yield curve factors and latent factors that are constructed on a relatively large set of economic indicators, namely global variables, domestic financial variables, economic activity and inflation related variables. Furthermore, we provide a characterization of the dynamic interactions between the yield curve and macroeconomics factors via estimating panel VAR model and computing impulse response functions. Finally, utilizing an individual factor augmented predictive regressions and variable selection algorithms for each emerging markets, we investigate whether macroeconomic/financial factors have predictive ability for yield curve factors. Our findings can be summarized as follows. First, it is observed that a shock coming to inflation factor is being transmitted to a positive response in the level factor of yield curve. Second, one unit shock of innovation to the factor representing economic activity brings about an increase in the slope component leading to steepening in EM countries’ yield curve formation. Third, the response of slope factor to the shocks in 6
9. global factor is negative and significant after around first month which have a lasting effect on slope factor . This finding implies that global factors have power to explain movements in the yield curves of emerging markets, in addition to the local factors. Fourth, macroeconomic/financial factors captures significant predictive information over yield curve factors and employing variable selection methods improves forecast accuracy further. This paper is organized as follows: Section 2 provides detailed information about utilized data sets. Section 3 covers the dynamic factor model to summarize macroeconomic and financial forces, NS methodology to estimate yield curve factors of individual countries as well as dynamic panel regressions and panel VAR model constructed to assess the interaction between yield curve factors and macroeconomic forces. Section 4 presents empirical results, and Section 5 concludes the discussion. 2. Data We have a balanced panel dataset of monthly observations between January 2006 and April 2019 for ten emerging markets2 . The dataset includes a large set of indicators that are selected to represent a broad range of macroeconomic variables that can be classified into the following four categories: • Real Economic activity: unemployment, industrial production, balance of payments statistics, retail trade, vehicle productions, completed buildings recorded and new orders; • Prices: producer prices and consumer prices; • Domestic Financial variables: interest rates, exchange rates, implied volatility, money supply and stock prices; • Global variables: Economic activity and financial market variables on global scale such as US term premium, US economic policy uncertainty, EU industrial production, ISM manufacturing PMI, MSCI emerging markets indices and Nomura China stress indicator; Table 1 shows the final number of series in each category, as well as the total number of series for each country. The final selection of the variables for each country is determined based on the data availability. We consider the series of indicators that are available for each economy from Bloomberg that are followed most closely by market participants. It is thus fairly comprehensive as the data include both supply-side 2 Brazil, Hungary, India, Malaysia, Mexico, Poland, Russia, South Africa, Thailand and Turkey 7
10. Table 1 : Number of indicators by type for selected emerging markets. Categories Brazil Hungary India Malaysia Mexico Poland Russia South Africa Thailand Turkey Real economic activity 63 47 48 20 60 28 31 61 31 44 Prices 16 13 11 15 17 20 17 17 14 16 Domestic financial variables 29 38 37 26 37 32 32 37 26 40 Global variables 48 48 48 48 48 48 48 48 48 48 Total 156 146 144 109 162 128 128 163 119 148 and demand-side indicators. All variables are subject to preliminary transformations to induce stationarity as needed. In addition to the above set of macroeconomic and financial indicators, which are used in our construction of local and global factors, we collect monthly zero coupon yields of maturities 3, 6, 12, 24, 36, 48, 60, 72, 84, 96, 108 and 120 months to estimate the yield curve factors for each emerging markets in our sample. The zero coupon bond yields were also obtained from the Bloomberg terminal. 3. Empirical Methodology 3.1. Extraction of common factors using dynamic factor model In our analysis, we separately extract potentially useful common factors from our four datasets (i.e., real economic activity, prices, domestic variables and global variables) for each country. To do this, we utilize the widely used dynamic factor model (DFM) of Giannone et al. (2008). As is typical in such models, individual variables are represented as the sum of components that are common to all variables in the economy (i.e., the factors) and an orthogonal idiosyncratic part. Formally, the DFM can be written as a system of equations: a measurement equation (i.e., Eq. (1)) that links the observed variables to the unobserved common factor to be estimated, and transition equation (Eq. (2)) that describe the dynamics of the common factor. Once Eqs. (1)-(2) are written in state space form, we employ the Kalman filter and smoother in order to extract the common factors and generate projections for all of the variables in the model. We start by characterizing the dynamics for the monthly data. Let Xi,t denote panel of observable economic variables where i shows the cross-section unit of macroeconomic variables, i = 1, ..., N and t indicates the monthly time index, t = 1, ..., T . We assume that Xi,t has the following factor model representation: Xt = ΛFt + ξt , ξt ∼ N(0, Σe ), (1) p Ft = Ψi Ft−i + ut , ut ∼ N(0, Q), (2) i=1 8
11. where Ft is an r × 1 vector of unobserved common factors with zero mean and unit variance, that reflect “most” of the co-movements in the variables, Λ is a corresponding N × r factor loading matrix, and the idiosyncratic disturbances, ξt , are uncorrelated with Ft at all leads and lags, and have a diagonal covariance matrix, Σe . The common factors are modeled as a stationary vector autoregressive (VAR) process of order p driven by the common shocks, ut ∼ N(0, Q), and that the Ψi are r × r matrices of autoregressive coefficients. Also, the common shocks, ut , and the idiosyncratic shocks, t , are assumed to be serially independent and independent of each other over time. We estimate the model using the two-step approach proposed by Giannone et al. (2008), see Doz et al. (2011) and select the first factor that explains the highest variation in each dataset3 . The lags of the factors are chosen via use of Schwarz information criteria. In particular, four diffusion indexes (i.e., factors) are constructed. While three of the four factors that are separately extracted using the datasets belonging to real economic activity, prices, domestic variables are called local factors, the factor extracted from the set of global variables is called global factor. 3.2. Estimation of yield curve factors: Nelson-Siegel Model The yield curve factors are obtained using NS model where the zero rates can be described explicitly by the following functional form:    m  m    − − m     1 − e τ  −  τ  1 − e  yt (m) = β1 + β2  m  + β3  m − e τ      τ τ (3) Accordingly, yt (m) denotes the continuously compounded zero-coupon nominal yield at time t of a bond with maturity m, and β1 , β2 , β3 and τ are NS parameters to be estimated. Eq.(3) represents a four-component approximation to the cross-section of yields at any time. Diebold and Li (2006) interpret the NS parameters as the level (β1 ), slope (β2 ) and curvature (β3 ). The coefficient τ, which is frequently referred to as the shape parameter, determines both the steepness of the slope factor and the location of the hump (Annaert et al. (2013)). The parameters are estimated using non-linear least squares where the objective function is to minimize the squared difference between duration-inverse weighted actual and fitted prices. However, employing the non-linear least squares optimization leads to non-smooth parameter estimates, especially for the slope and curvature parameters. Therefore, we estimate level, slope and curvature parameters by Ordinary Least Squares (OLS) fixing τ parameter to reduce the volatility of these parameters as 3 Explanatory powers of extracted first three factors are presented in Table A1 of the Appendix. The expanatory power of first factors range between 53% - 22%. 9
12. proposed by Diebold and Li (2006). We run a grid search to find the optimal τ parameter, which gives us the smallest Root Mean Squared Error (RMSE) for each emerging markets in our sample4 . 3.3. Dynamic Panel Data Estimations Before undertaking dynamic interaction between yield curve factors and macroeconomic/financial determinants, an initial empirical investigation is performed by using an estimation technique exploiting the longitudinal nature of the data which also incorporates the timewise autoregressive structure of yield curve factors. In this context, difference generalized method of moments (GMM) approach developed by HoltzEakin et al. (1988) and Arellano and Bond (1991) is implemented where explained variables are dynamic (meaning they are being dependent on their own past realizations) and explanatory variables are not strictly exogenous (meaning they are correlated with the past and present realizations of the error term). Arellano-Bond estimation involves a transformation of regressors (mostly by differencing) and an application of GMM. Modeling through fixed effects, despite the fact that underlying data generating process is dynamic by nature, creates a correlation between error term and regressors because of the demeaning attempt of dependent and independent variables in fixed effects estimation. Since demeaning operation creates a set of regressors which are not distributed independently of the disturbance term, coefficient estimator for lagged dependent variable is inconsistent (Nickell, 1981). The solution to this evident problem is to apply a transformation to the model. First differencing to the original model is mostly used in practice to remove the unobserved individual effect. When model is transformed, then it becomes eligible for instrumental variable estimation. Difference GMM method is doing this by establishing a system of equations (for each time period) and by economizing internal instruments (lagged values of instrumented variables) to make the estimation. Hence, our methodological framework entails the use of one-step difference GMM method of Arellano and Bond (1991). For this study, we utilize following series of specifications in which yield curve components are defined as dependent variables and static factors describing inflation, economic activity, local financial conditions and global financial outlook are added as explanatory variables in an incremental manner. As the most comprehensive specification, the final model includes all the macroeconomic/financial factors. 2 YCit−s + γ1 In f lationit + ui + eit YCit = ρ (4) s=1 4 For this purpose, the estimations are iterated for more than one million times. 10
13. 2 YCit = ρ YCit−s + γ2 Activityit + ui + eit (5) YCit−s + γ3 Financialit + ui + eit (6) YCit−s + γ4Globalit + ui + eit (7) YCit−s + γ1 In f lationit + γ2 Activityit + γ3 Financialit + γ4Globalit + ui + eit (8) s=1 2 YCit = ρ s=1 2 YCit = ρ s=1 2 YCit = ρ s=1 where YCit refers to the yield curve components which are level, slope and curvature. ρ stands for the autoregressive parameters obtained from the first and second lags of yield curve components included as covariates. γ coefficients measure the impact of macroeconomic and financial dynamics on yield curve formation. 3.4. Dynamic Common Correlated Effects Dynamic panel data models with system GMM estimations have advantages such as accounting from dynamic structure in the variable of interest and capability to handle endogeneity problems (Roodman (2006, 2009), Labra and Torrecillas (2018)). However, as noted by Ruiz-Porras (2012), applying this methodology on data structures with longer time dimensions (T) and shorter cross-sectional dimension (N) could result in the over-identification of the model.5 Furthermore, it does not account for unobserved dependencies between cross-sectional units in the examined data set. Despite the fact that in system GMM estimations we tried to mitigate over-identification problem by limiting the number of lags of instruments in level and difference equations, specifications described above are also estimated by utilizing dynamic common correlated effects for robustness.6 In this context, estimation procedure conceptualized by Chudik and Pesaran (2015) and operationalized by Ditzen (2018) is implemented. Following empirical identification is considered: P Yi,t = λi Yi,t−1 + βi Xi,t + γi,k Z¯t−k + εi,t , Z¯t = Y¯ t , X¯ t k=0 5 6 We thank the anonymous referee for pointing out this issue and suggesting alternative estimation technique. As it does not vary over cross-section units, the variable termed “global” is excluded from these estimations. 11 (9)
14. where Y and X describe dependent and independent variables , whereas βi = β + vi , vi ∼ IID (0, Ωv ) and λi = λ + ζi , ζi ∼ ID 0, Ω f represent heterogeneous coefficients which are randomly distributed around a common mean. As quoted in Ditzen (2018), Pesaran (2006) in static models with no lagged dependent variable terms as additional explanatory variables, estimations will be consistent by approximating the unobserved common factors with cross-section averages Y¯ t and X¯ t under strict exogeneity. On the other √3 hands, in dynamic models, Chudik and Pesaran (2015) show that estimator gains consistency if P = T lags of the cross-sectional averages are incorporated into the specification. We follow this methodology to obtain coefficient estimations for level, slope and curvature factors. In addition to this, we also implement Pesaran (2015) test for cross-sectional dependence to evaluate dependencies across countries7 . 3.5. Panel VAR model using a system GMM approach In the following part of our empirical setting, we exploit the informative content of yield curve parameters and macroeconomic factors within the context of panel VAR model. This class of modelling framework has been increasingly utilized to study interdependencies, particularly in the fields of macroeconomics and finance such as economic activity, business cycle tendencies, and transmission of financial shocks among many others (Canova and Ciccarelli (2013)). To assess the dynamic relation between yield curve components and macro-factors, we utilize a panel VAR model using generalized method of moments (GMM) approach as described by Abrigo and Love (2016). The estimated system of equations referring for the panel VAR model of order p with countryspecific fixed effects can be specified as the following: Yit = A1 Yit−1 + A2 Yit−2 + ... + A p Yit−p + ui + eit E[eit ] = 0, E[eit eit ] = Σ, (10) E[eit eis ] = 0, f ort > s where Yit is a (1xm) vector of endogenous variables (prices, economic activity, domestic financial, global, level, slope, curvature), ui and eit represent the (1 × m) dependent variable specific panel fixed effects and idiosyncratic errors, respectively. The idiosyncratic disturbances eit have a diagonal covariance matrix, Σ. As mentioned before, variables are obtained from yield curve factors by utilizing Nelson and 7 The test results are presented in Table A4 of the Appendix. 12
15. Siegel (1987) methodology and the common factors related to each category (prices, economic activity, domestic financial, global) are extracted by applying dynamic factor model based on the large sets of variables included in each category. The ordering in panel VAR is chosen to reflect the transmission channel for EM in which originated global shocks are propagated to local financial conditions and, in the next step, they are incorporated in the formulation of yield curve dynamics to characterize the influence on economic activity and pricing behavior. In this framework, ordering of the variables does not alter the coefficient estimates for the panel VAR, while it is expected to affect the impulse-response functions (IRFs). However, it is found that IRFs are not subject to alterations when order of the variables is changed. Empirical analysis with panel VAR model is multifaceted for which the initial results are obtained for the stationarity of variables to make reliable inferences. In this context, we benefit from the panel unit root testing procedures of Im et al. (2003) and Levin et al. (2002). Additionally, consistent moment and model selection criteria of Andrews and Lu (2001) as well as the Hansen (1982) J-statistics of over-identifying restrictions are reviewed to decide on the optimal lag length of the mode. While IRFs are utilized to gain deeper insight about the dynamic inter-relation of yield curve factors with macro-forces in the EM countries, the stability conditions of panel VAR estimates are also checked by calculating the modulus of eigenvalues of the estimated model. 3.6. Out-of-sample forecasting exercise for individual countries Apart from investigating dynamic interdependencies, we employ factor augmented predictive regressions commonly used in the empirical finance studies, for investigating the predictability of yield curve factors separately for each countries. Specifically, we construct our predictive regressions of the following form: yt+1 = α0 + β Zt + εt+1 (11) where yt+1 is the yield curve factors in period t + 1 and Zt includes factors (prices, economic activity, domestic financial) extracted using the dynamic factor model approach of Giannone et al.(2008). We select the benchmark model as random-walk (RW) model since comparing our model results with this model will tell us whether macroeconomic factors add value to forecasting of yield curve factors. Out-of-sample forecasting exercise over the period January, 2012 to April, 2019, with an in-sample period of January, 2006 to December, 2011, is employed recursively to provide insight into the predictive ability of macroeconomic 13
16. and financial factors for yield curve factors . For each month, we produce a sequence of six h-month-ahead forecast for h = 1, 2, 3, 4, 5, 6. To assess the statistical significance of forecast performance of different models compared to our benchmark model, the Diebold and Mariano (2002, DM) test is utilized using quadratic loss function. 4. Empirical Results 4.1. Relation between yield curve factors and extracted common factors Before moving into dynamic panel estimations, it is informative to visually investigate the co-movements between macroeconomic factors and yield curve parameters. By pooling cross-sectional dimensions of sample countries with historical time series, Figure 2 depicts the scatterplots of yield curve factors with macro-forces that are theoretically known to be relevant. Here, it is seen that there exists a positive correlation between level factor of EM sovereign yield curves and inflation factor extracted from CPI and PPI series of sample countries. Hence, we suspect that price pressures entailing high inflation rates can be preemptively associated with higher level of the yield curve. While the degree of association is lower compared to level-inflation case, there is a negative linear relation between slope factor and local macroeconomic activity component. In other words, steepening in yield curves can be relevant to the loss of momentum in growth tendencies. Lastly, as a striking finding, we demonstrate a relation between curvature factor and local financial factor as a common pattern in EM countries. − Insert Figure 2 about here. − 4.2. Dynamic Panel Estimation Results First of all, dynamic panel estimations reveal that, as expected, autoregressive dependence on time dimension is evident for yield curve components. When level, slope and curvature factors are defined as dependent variables, corresponding regressions support the expectation that lags of the explained variables are statistically significant. For the univariate specifications for level factor, it is seen that inflation factor is an important determinant of long-term component of the yield curve. In particular, upward movements in inflation factor tracking the price pressures in EM countries create significant and positive impact on level factor. In addition to this, local financial factors, for which increases are corresponding to deterioration in financial indicators and volatility in financial markets, turn out to be associated with level factor as well. 14
17. In univariate cases , economic activity and global factors are found to be somewhat significant supporting the expectation that growth outlook and global forces might have an influence the formation of slope component. However, when multivariate case is considered, while economic activity retains its significance, local financial conditions emerge as a significant determinant of slope factor. In terms of curvature factor, unlike most of the studies in the previous literature, local financial conditions is found to be an important driver. − Insert Table 2 about here. − − Insert Table 3 about here. − − Insert Table 4 about here. − 4.3. Dynamic Common Correlated Effects Results Results obtained from dynamic common correlated effects estimations are vastly in line with system GMM results. Lag dependence structure of level factor is still evident. Univariate specifications for level equation shows that inflation and economic activity factors are significantly associated with level component of EM yield curves in which former is positively and latter is negatively related with long-term yield curve factor entailing long-term interest rates. When we change the estimation technique, previously documented significant role of domestic financial conditions in explaining level dynamics disappear. In full specification, as expected, only significant explanatory variable for level equation is inflation outlook. In the second step, similar estimations are conducted for slope component. Here, in contrast to previous estimations, the predictive power of the equations regarding in-sample context is improved, as manifested by statistically significant effects of inflation and financial factors, on top of theoretically and empirically suggested economic activity factor. Significance is also retained in the broadest specification. Lastly, curvature factor is considered. As it is not widely observed in the empirical literature, dynamic common correlated effects estimations indicate that curvature component of EM yield curves is significantly driven by the course of domestic financial conditions. Hence, any movements in sub-components of domestic financial conditions including credit growth, exchange rates, stock market outlook and capital flows will have implications on the yield curve formations. For each yield curve factor, Pesaran (2015) test results in broadest specifications show that null hypothesis of weak cross-sectional dependence is vastly rejected pointing out the fact that common correlated 15
18. effects estimations controlling for unobserved dependencies across EM countries are reliable in this setting8 . − Insert Table 5 about here. − − Insert Table 6 about here. − − Insert Table 7 about here. − 4.4. Panel VAR results As stated in Section 3.4, the stationarity behavior of variables utilized in panel VAR model are evaluated with IPS and LLC panel unit root test. When level values are assessed with these tests (with only constant and with both constant and trend terms in test specifications), there appears to be some evidence pointing out non-stationarity. Hence, we proceed with transformation of variables into first differences yielding stationarity before conducting estimations9 . In terms of the selection of lag length, the informative content of Hansen’s J-statistic and information criterion are considered. In this case, overwhelming evidence is the use of one lag in the specification of panel VAR model10 . To analyze the interaction among common factors and yield curve factors, we perform impulse response functions. Figure 3 presents the cumulative orthogonalized impulse-response functions from the estimated panel VAR. 95% confidence interval to analyze the statistical significance are created by using 1000 Monte Carlo simulation draws. The forecast horizon is determined as 12 months. It is observed that a shock coming to inflation factor is being transmitted to a positive response in the level factor of yield curve. The impact that occurs following the inflationary shock lasts almost 6 months, whereas it losses the significance after 3rd month. Thus, it supports the argument that level is the long term factor in the yield curve formulation which reflects the inflationary dynamics as well as inflation expectations. − Insert Figure 3 about here. − Impulse-response functions also reveal the statistically significant relation between economic growth and slope component of the yield curve. Following one unit shock to the factor representing the economic activity, the slope component increases leading to steepening in EM countries’ yield curve formation. The 8 The test results are reported in Table A5 of the Appendix. Results of panel unit root tests are provided in Table A2 in the appendix. 10 Information criterion results are given in Table A3 in the appendix. 9 16
19. impact seems to peak around 6-months horizon . It is interesting to note that local financial conditions in EM countries are found to be strongly associated with curvature component of yield curve. In particular, one unit impulse given to the factor summarizing local financial dynamics is tracked to have a statistically significant influence on curvature factor, while the majority of the impact occurs within a shorter period of time. The response of slope factor to shocks in global factor is negative and significant after around 1st month and also these shocks have a lasting effect on the slope factor. This result supports the findings of Jotikasthira et al. (2015), that U.S. yield factor have power to explain movements in the curves of other countries. The inverse relation between global and slope factor might indicate that a shock to the global factor may increase the expectations of raising the short-term policy rate by central banks. Hence, this situation puts upward pressure on short-term government bonds, thereby resulting in a negative relationship. This is often seen as a bag sign for the economy since the yield spread is historically narrowed ahead of recessions. We also examine the impulse-response among yield curve factors themselves from the estimated panel VAR model (Figure 4). It could be seen that shocks coming to level factor significantly and negatively affect slope and curvature factors in the examined horizon. Impulse-response function also depicts the strong influence of slope factor on level factor, whereas no significant result is obtained for the impact of slope on curvature. Lastly, impulse-response functions display that shocks coming to curvature does not produce statistically significant responses for level factor. On the other hand, impulses occurred to curvature is tracked to create slightly significant responses on slope parameter. Overall, our results highlight the relevance of local and global factors for better understanding the movements of yield curves in emerging markets. − Insert Figure 4 about here. − 4.5. Out-of-sample forecasting results The ratios of root mean squared errors (RMSEs) for our set of forecasting models are presented in Table 5 for each of the forecast horizons. Models that yield the lowest RMSE values at each horizon are denoted in bold. Overall, the entries in Table 5 in general are less than unity, which reveals that the factoraugmented predictive regressions usually produce better forecasts than the benchmark RW model. This finding is further supported by the DM test, indicating statistically significant improvements in forecast accuracy compared to the RW model. Our results also suggest that the RMSE values generally increase with the forecast horizon, confirming the out-of-sample predictive power of macro and financial factors 17
20. especially for short term horizons . In particular, the predictive power of macro/financial factors is notable for Brazil, Hungary, India, Mexico, Poland and Thailand. Surprisingly, the factor augmented predictive regressions yield better forecasts for curvature factor in 7 countries out of 10 with a limited number of exceptions. This can be seen from Table 5 by noting that the lowest RMSEs are denoted in bold. However, while the RMSEs of factor type predictions are lower more than 40% compared to those of RW model for Hungary and Poland, their forecast performances are relatively poor in Russia and South Africa particularly for level and slope factors. As suggested by Bai and Ng (2008), Kuzin et al., (2011), Cepni and Guney (2019), and Cepni et al. (2018, 2019), it is important to choose appropriate predictors prior to estimation of predictive regressions. The reason is that model and parameter uncertainty may adversely affect the explanatory variables’s marginal predictive content. In this respect, as an robustness check, we investigate alternative variable selection methods namely, the Elastic-Net, the Least Absolute Shrinkage Operator (LASSO), and the Ridge regression in order to pre-select variables prior to the predictions. Accordingly, for each month, we recursively choose predictors from the set of our four macroeconomic and financial factors, instead of using all of them. As presented in Table A6-A8 of appendix, machine learning algorithms are useful for selecting predictors when constructing predictions. Put differently, variable selection methods results in predictive gains by providing sparsity for model estimation compared to the predictive regressions utilizing all macroeconomic and financial factors simultaneously for each month. This can be seen from Table A3-A5 of appendix by noting that the entries in general are less than unity. − Insert Table 5 about here. − 5. Conclusion This paper investigates the relative importance of the local and global factors in driving movement in term structure of interest rates in emerging markets. For this purpose, initially, the yield curve factors are extracted using the NS methodology for the 2006:01- 2019:04 period. Rather than analyzing the effect of macroeconomic variables by using a few empirical proxies for price developments, growth and monetary policy stance, the macroeconomic and financial variables are classified as global variables, economic activity, domestic financial developments and inflation. Then, a panel VAR model is employed to explore the dynamics of the yield curve factors and macroeconomics factors. 18
21. Empirical results suggest that the level factor responds positively to the shocks originating from inflation developments as well as financial variables . However, the effect of domestic financial variables on the level factor tends to be larger in size compared to inflationary shocks. Whereas slope factor is affected by shocks in global variables, curvature factor appears to be influenced by domestic financial variables. Besides, utilizing an individual factor augmented predictive regressions and variable selection algorithms also confirm thst macroeconomic/financial factors have predictive power for yield curve factors. Our findings indicate that macroeconomic and global financial variables are informative in terms of explaining changes in yield curve of emerging markets countries. Given the unconventional monetary policy implementations and low-rate environment in developed countries, the emerging market domestic bond rates tend to be exposed to the swings in global financial conditions, which weakens the monetary policy transmission mechanism in emerging markets. Hence, policymakers should take into account the possible implications of shocks stemming from global financial framework as well as economic activity and local financial variables. Additionally, deciphering the relation among macroeconomic forces and each particular factor of yield curve enables to anticipate the changes in the yield curve through the evolvement in those forces and creates a better environment for producing accurate forecasts. 19
22. References Abrigo , M. R., and Love, I. (2016). Estimation of panel vector autoregression in Stata. The Stata Journal, 16(3), 778-804. Ahokpossi, M. C., Martinez, P. G., and Kemoe, L. (2016). Monetary and Fiscal Policies and the Dynamics of the Yield Curve in Morocco. International Monetary Fund Working Papers, No:16/103. Andrews, D. W., and Lu, B. (2001). Consistent model and moment selection procedures for GMM estimation with application to dynamic panel data models. Journal of Econometrics, 101(1), 123-164. Ang, A, G Bekaert and M Wei (2008). The term structure of real rates and expected inflation. Journal of Finance, vol 63, 797–849. Ang, A., Piazzesi, M., and Wei, M. (2006). What does the yield curve tell us about GDP growth?. Journal of econometrics, 131(1-2), 359-403. Annaert, J., Claes, A. G., De Ceuster, M. J., and Zhang, H. (2013). Estimating the spot rate curve using the Nelson–Siegel model: A ridge regression approach. International Review of Economics and Finance, 27, 482-496. Arellano, M., and Bond, S. (1991). Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations. The Review of Economic Studies, 58(2), 277-297. Argyropoulos, E., and Tzavalis, E. (2016). Forecasting economic activity from yield curve factors. The North American Journal of Economics and Finance, 36, 293-311. Bai, J., and Ng, S. (2008). Forecasting economic time series using targeted predictors. Journal of Econometrics, 146, 304-317. Baz, J., and Chacko, G. (2004). Financial derivatives: Pricing, Applications, and Mathematics. Cambridge University Press, ISBN: 9780521815109. Bekaert, G., Cho, S., and Moreno, A. (2010). New Keynesian macroeconomics and the term structure. Journal of Money, Credit and Banking, 42(1), 33-62. Canova, F., and Ciccarelli, M. (2013). Panel Vector Autoregressive Models: A Survey. Emerald Group Publishing Limited, 205-246. Cepni, O., and Guney, I. E. (2019). Nowcasting emerging market’s GDP: the importance of dimension reduction techniques. Applied Economics Letters, 1-5. Cepni, O., Guney, I. E., and Swanson, N. R. (2018). Forecasting and nowcasting emerging market GDP growth rates: The role of latent global economic policy uncertainty and macroeconomic data surprise 20
23. factors . Journal of Forecasting. Cepni, O., G¨uney, I. E., and Swanson, N. R. (2019). Nowcasting and forecasting GDP in emerging markets using global financial and macroeconomic diffusion indexes. International Journal of Forecasting, 35(2), 555-572. Chudik, A., and Pesaran, M. H. (2015). Common correlated effects estimation of heterogeneous dynamic panel data models with weakly exogenous regressors. Journal of Econometrics, 188(2), 393-420. Cox, J. C., Ingersoll Jr, J. E., and Ross, S. A. (1985). An intertemporal general equilibrium model of asset prices. Econometrica: Journal of the Econometric Society, 363-384. Dewachter, H., and Lyrio, M. (2006). Macro factors and the term structure of interest rates. Journal of Money, Credit, and Banking, 38(1), 119-140. Diebold, F. X., and Li, C. (2006). Forecasting the term structure of government bond yields. Journal of econometrics, 130(2), 337-364. Diebold, F. X., and Mariano, R. S. (2002). Comparing predictive accuracy. Journal of Business and Economic Statistics, 20(1), 134-144. Diebold, F. X., Rudebusch, G. D., and Aruoba, S. B. (2006). The macroeconomy and the yield curve: a dynamic latent factor approach. Journal of econometrics, 131(1-2), 309-338. Ditzen, J. (2018). Estimating dynamic common-correlated effects in Stata. The Stata Journal, 18(3), 585-617. Doz, C., Giannone, D., and Reichlin, L. (2011). A two-step estimator for large approximate dynamic factor models based on Kalman filtering. Journal of Econometrics, 164(1), 188-205. Garc´ıa-Herrero, A., and Wooldridge, P. D. (2007). Global and regional financial integration: progress in emerging markets. BIS Quarterly Review, September. Giannone, D., Reichlin, L., and Small, D. (2008). Nowcasting: The real-time informational content of macroeconomic data. Journal of Monetary Economics, 55(4), 665-676. H¨annik¨ainen, J. (2017). When does the yield curve contain predictive power? Evidence from a data-rich environment. International Journal of Forecasting, 33(4), 1044-1064. Hansen, L. P. (1982). Large sample properties of generalized method of moments estimators. Econometrica: Journal of the Econometric Society, 1029-1054. Holtz-Eakin, D., Newey, W., and Rosen, H. S. (1988). Estimating vector autoregressions with panel data. Econometrica: Journal of the Econometric Society, 1371-1395. 21
24. Im , K. S., Pesaran, M. H., and Shin, Y. (2003). Testing for unit roots in heterogeneous panels. Journal of econometrics, 115(1), 53-74. Kanjilal, K. (2013). Factors causing movements of yield curve in India. Economic Modelling, 31, 739-751. Kaya, H. (2013). Forecasting the yield curve and the role of macroeconomic information in Turkey. Economic Modelling, 33, 1-7. Kuzin, V., Marcellino, M., and Schumacher, C. (2011). MIDAS vs. mixed-frequency VAR: nowcasting GDP in the euro area. International Journal of Forecasting, 27, 529–542. Jotikasthira, C., Le, A., and Lundblad, C. (2015). Why do term structures in different currencies comove?. Journal of Financial Economics, 115(1), 58-83. Labra, R., and Torrecillas, C. (2018). Estimating dynamic Panel data. A practical approach to perform long panels. Revista Colombiana de Estad´ıstica, 41(1), 31-52. Levin, A., Lin, C. F., and Chu, C. S. J. (2002). Unit root tests in panel data: asymptotic and finite-sample properties. Journal of econometrics, 108(1), 1-24. Lu, B., and Wu, L. (2009). Macroeconomic releases and the interest rate term structure. Journal of Monetary Economics, 56(6), 872-884. Miyajima, K., Mohanty, M. S., and Chan, T. (2015). Emerging market local currency bonds: diversification and stability. Emerging Markets Review, 22, 126-139. M¨onch, E. (2012). Term structure surprises: the predictive content of curvature, level, and slope. Journal of Applied Econometrics, 27(4), 574-602. Nelson, C., and Siegel, A. (1987). Parsimonious Modeling of Yield Curves. The Journal of Business, 60(4), 473-489. Nickell, S. (1981). Biases in dynamic models with fixed effects. Econometrica: Journal of the Econometric Society, 1417-1426. Paccagnini, A. (2016). The macroeconomic determinants of the US term structure during the Great Moderation. Economic Modelling, 52, 216-225. Pesaran, M. H. (2006). Estimation and inference in large heterogeneous panels with a multi-factor error structure. Econometrica, 74(4), 967-1012. Pesaran, M. H. (2015). Testing weak cross-sectional dependence in large panels. Econometric Reviews, 34(6-10), 1089-1117. 22
25. Prasanna , K., and Sowmya, S. (2017). Yield curve in India and its interactions with the US bond market. International Economics and Economic Policy, 14(2), 353-375. Roodman, D. (2006). How to do xtabond2: An introduction to difference and system GMM in Stata. Roodman, D. (2009). A note on the theme of too many instruments. Oxford Bulletin of Economics and Statistics 71(1), 135-158. Ru´ız-Porras, A. (2012). Econometric research with panel data: History, models and uses in mexico, MPRA-paper 42909, University Library of Munich, Germany. Sarno, L., Thornton, D. L., and Valente, G. (2007). The empirical failure of the expectations hypothesis of the term structure of bond yields. Journal of Financial and Quantitative Analysis, 42(1), 81-100. Stock, J. H., and W Watson, M. (2003). Forecasting output and inflation: The role of asset prices. Journal of Economic Literature, 41(3), 788-829. Wooldridge, P. D., Domanski, D., and Cobau, A. (2003). Changing links between mature and emerging financial markets. BIS Quarterly Review, September. 23
26. Table 2 : Dynamic panel estimation results : Level factor Dependent Variable : Level (1) (2) (3) (4) (5) L.Level 1.133*** 1.196*** 1.153*** 1.178*** 1.097*** (0.064) (0.063) (0.054) (0.059) (0.068) L2.Level -0.288*** -0.313*** -0.272*** -0.300*** -0.277*** (0.054) (0.057) (0.052) (0.054) Inflation 0.0794*** (0.029) Activity (0.053) 0.0642*** (0.023) -0.0114 -0.0173 (0.016) (0.013) Financial 0.0548** 0.0382 (0.027) (0.027) Global -0.0152 -0.0015 (0.010) (0.006) Observations 1,340 1,340 1,340 1,340 1,340 Number of country 10 10 10 10 10 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1. 24
27. Table 3 : Dynamic panel estimation results : Slope factor Dependent Variable : Slope L.Slope L2.Slope Inflation (1) (2) (3) (4) (5) 1.186*** 1.137*** 1.176*** 1.177*** 1.110*** (0.052) (0.044) (0.062) (0.047) (0.051) -0.352*** -0.308*** -0.353*** -0.327*** -0.307*** (0.051) (0.047) (0.056) (0.042) (0.049) 0.005 -0.009 (0.019) (0.024) Activity 0.038*** 0.048*** (0.014) (0.017) Financial 0.025 0.045*** (0.022) (0.015) Global 0.014* -0.005 (0.008) (0.009) Observations 1.340 1.340 1.340 1.340 1.340 Number of country 10 10 10 10 10 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1. Table 4: Dynamic panel estimation results : Curvature factor Dependent Variable : Curvature L.Curvature L2.Curvature Inflation (1) (2) (3) (4) (5) 1.065*** 1.046*** 1.050*** 1.058*** 1.041*** (0.083) (0.088) (0.084) (0.091) (0.084) -0.347*** -0.343*** -0.345*** -0.346*** -0.341*** (0.083) (0.083) (0.082) (0.087) (0.083) 0.007 -0.029 (0.045) (0.048) Activity 0.037 0.022 (0.030) Financial (0.019) 0.087** 0.095** (0.037) Global (0.040) 0.008 -0.005 (0.026) (0.025) Observations 1.340 1.340 1.340 1.340 1.340 Number of country 10 10 10 10 10 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1. 25
28. Table 5 : Dynamic common correlated effects estimation results : Level factor Dependent Variable : Level L.Level L2.Level Inflation (1) (2) (3) (4) 1.088*** 1.105*** 1.101*** 1.024*** (0.044) (0.048) (0.041) (0.033) -0.227*** -0.236*** -0.229*** -0.215*** (0.054) (0.051) (0.049) (0.049) 0.029*** 0.025** (0.007) (0.012) Activity -0.027** -0.015 (0.011) Financial (0.011) 0.002 0.001 (0.011) (0.014) Observations 1.340 1.340 1.340 1.340 Number of country 10 10 10 10 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1. Table 6: Dynamic common correlated effects estimation results : Slope factor Dependent Variable : Slope (1) (2) (3) (4) L.Slope 1.116*** 1.124*** 1.123*** 1.170*** (0.056) (0.046) (0.053) (0.052) L2.Slope -0.244*** -0.267*** -0.246*** -0.256*** (0.045) (0.048) (0.046) Inflation 0.022*** 0.023** (0.008) (0.011) Activity (0.046) 0.015*** 0.015*** (0.004) (0.005) Financial 0.016*** 0.018*** (0.006) (0.006) Observations 1.340 1.340 1.340 1.340 Number of country 10 10 10 10 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1. 26
29. Figure 1 : International bond issuances and EM domestic bond markets total size International Bond Issues (EM Group, billion USD, notional outstanding amount) 3000 2500 2000 1500 1000 500 07.18 08.17 09.16 10.15 11.14 12.13 01.13 02.12 03.11 04.10 05.09 06.08 07.07 08.06 09.05 10.04 11.03 12.02 01.02 02.01 03.00 0 Source:BIS EM Domestic Bond Markets-Selected Group (General Government, billion USD, notional outstanding amount) EM Domestic Bond Markets-Selected Group (All Issuers, billion USD, notional outstanding amount) 5500 4500 3500 2500 1500 Source:BIS 27 11.18 03.18 07.17 11.16 03.16 07.15 11.14 03.14 07.13 11.12 03.12 07.11 11.10 03.10 07.09 11.08 03.08 07.07 11.06 03.06 500
30. Figure 2 : Scatterplots of relation between yield curve factors and estimated common components 28
31. Table 7 : Dynamic common correlated effects estimation results : Curvature factor Dependent Variable : Curvature (1) (2) (3) (4) L.Curvature 1.048*** 1.063*** 1.124*** 1.107*** (0.050) (0.047) (0.100) (0.099) L2.Curvature -0.218*** -0.221*** -0.242** -0.255** (0.054) (0.055) (0.111) Inflation 0.033 (0.110) 0.049** (0.021) Activity (0.021) -0.004 0.028 (0.007) Financial (0.017) 0.031** 0.028** (0.012) (0.011) Observations 1.340 1.340 1.340 1.340 Number of country 10 10 10 10 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1. 29
32. Table 8 : Factor augmented predictive regressions : out-of-sample forecasting results Brazil h=1 h=2 h=3 h=4 h=5 h=6 RW 0.0187 0.0190 0.0193 0.0194 0.0195 0.0196 Level 0.817*** 0.864** 0.909** 0.951* 0.986* 1.011 Slope 0.798*** 0.847*** 0.886** 0.913** 0.928** 0.933** Curvature 0.970* 1.007 1.049 1.094 1.140 1.189 Hungary RW 0.0167 0.0169 0.0172 0.0173 0.0175 0.0177 Level 0.732*** 0.761*** 0.787*** 0.815*** 0.841*** 0.862** Slope 0.469*** 0.480*** 0.493*** 0.506*** 0.517*** 0.527*** Curvature 0.527*** 0.540*** 0.556*** 0.566*** 0.574*** 0.582*** India RW 0.0147 0.01473 0.01477 0.01478 0.01461 0.01442 Level 0.778*** 0.804*** 0.831** 0.850** 0.867** 0.875** Slope 0.982* 1.011 1.034 1.046 1.054 1.067 Curvature 0.726*** 0.751*** 0.778*** 0.803*** 0.825*** 0.846** Malaysia RW 0.0040 0.0041 0.0041 0.0041 0.0042 0.0042 Level 0.880** 0.904** 0.917** 0.942** 0.969* 0.993* Slope 1.432 1.475 1.506 1.547 1.587 1.629 Curvature 1.090 1.133 1.170 1.203 1.225 1.250 Mexico RW 0.0116 0.0117 0.0118 0.0119 0.0119 0.0120 Level 0.849*** 0.888** 0.919** 0.949** 0.979* 1.022 Slope 0.929** 0.969* 0.995 1.009 1.021 1.034 Curvature 0.873*** 0.909** 0.941** 0.960** 0.979* 0.990 Poland RW 0.0134 0.0136 0.0137 0.0139 0.0141 0.0142 Level 0.501*** 0.521*** 0.540*** 0.559*** 0.572*** 0.583*** Slope 1.042 1.068 1.090 1.107 1.123 1.139 Curvature 0.834** 0.883** 0.929* 0.976* 1.018 1.058 RW 0.0106 0.0107 0.0109 0.0110 0.0112 0.0112 Level 1.271 1.276 1.275 1.293 1.302 1.298 Slope 1.315 1.361 1.397 1.427 1.454 1.480 Curvature 0.890** 0.921** 0.940* 0.947* 0.939* 0.914** RW 0.0065 0.0065 0.0065 0.0066 0.0066 0.0066 Level 1.144 1.161 1.169 1.175 1.179 1.184 Slope 1.079 1.105 1.129 1.149 1.177 1.208 Curvature 0.833*** 0.851** 0.872** 0.888** 0.904** 0.919** Russia S.Africa Thailand RW 0.0097 0.0097 0.0097 0.0098 0.0098 0.0099 Level 0.781*** 0.784*** 0.788*** 0.791*** 0.795*** 0.797*** Slope 1.144 1.212 1.259 1.296 1.318 1.327 Curvature 0.950** 0.972* 0.990 0.999 1.003 0.999 RW 0.0269 0.0268 0.0268 0.0270 0.0273 0.0278 Level 1.156 1.192 1.227 1.258 1.270 1.284 Slope 0.869*** 0.889*** 0.901** 0.904** 0.916** 0.932* Curvature 1.038 1.058 1.067 1.064 1.056 1.037 Turkey Entries in the first row of the table are point RMSEs based on the benchmark random walk (RW) model, while the rest are relative RMSEs. Hence, a value of less than unity indicates that a particular model and estimation method is more accurate than that based on the RW model, for a given forecast horizon. Models that yield the lowest MSFE for each forecast horizon are denoted in bold. Entries superscripted with an asterisk (*** = 1% level; ** = 5% level; * = 10% level) are significantly superior than the RW model, based on the DM predictive accuracy test. 30
33. Figure 3 : Cumulative orthogonalized impulse-response functions 31
34. Figure 4 : Accumulated impulse-responses among yield curve factors 32
35. Appendix Supplementary Tables and Figures Table A1 : Explanatory powers of extracted factors Country Economic Activity Domestic Financial F1 F1 F2 F3 F1 F2 F3 F1 F2 F3 F2 F3 Inflation Global Brazil 44.7 12.2 5.5 44.0 19.8 10.9 39.8 20.8 9.8 38.7 15.8 8.7 Hungary 39.1 16.7 8.1 26.0 19.3 14.2 41.0 16.3 12.7 38.7 15.8 8.7 India 27.5 11.3 6.0 24.4 20.0 9.6 53.0 18.0 16.2 38.7 15.8 8.7 Indonesia 21.8 14.0 9.0 37.4 15.4 8.8 30.1 19.0 11.7 38.7 15.8 8.7 Malaysia 47.0 9.2 7.0 25.3 20.0 15.2 40.8 20.4 13.3 38.7 15.8 8.7 Mexico 44.1 10.6 7.2 23.0 20.2 16.9 42.3 19.6 12.8 38.7 15.8 8.7 Poland 36.0 18.7 8.3 31.9 26.4 8.7 51.5 10.7 8.6 38.7 15.8 8.7 Russia 49.5 13.4 5.5 45.3 17.6 8.6 48.2 18.6 8.4 38.7 15.8 8.7 S. Africa 22.1 14.7 8.8 30.6 23.6 11.5 25.9 21.6 14.8 38.7 15.8 8.7 Thailand 42.2 11.7 7.5 24.3 23.0 15.4 29.8 19.7 15.8 38.7 15.8 8.7 Turkey 37.8 11.5 7.6 40.9 19.2 10.4 50.2 13.8 12.3 38.7 15.8 8.7 Table A2: Panel Unit Root Test Results Im-Pesaran-Shin Test Level Levin-Li-Chu Test First Difference Level First Difference Variables Constant Constant&Trend Constant Constant Constant&Trend Constant Inflation -1.45* -0.69 -20.38*** -5.76 -6.80 -22.64*** Activity -2.57*** -0.13 -29.88*** -6.61 -7.31 -32.30*** Curvature -5.76*** -5.49*** -31.31*** -8.44* -9.32* -26.21*** Slope -4.01*** -3.30*** -25.83*** -5.65 -7.80 -23.94*** Level -4.49*** -3.96*** -28.34*** -6.92* -7.61 -25.82*** Local Financial -3.45*** -1.31* -24.76*** -8.19 -8.55 -21.98*** Global -5.71*** -4.03*** -24.78*** 33
36. Table A3 : Lag Length Selection in Panel VAR Lag Length J-statistic MBIC MAIC MQIC 1 305.16 -1439.74 -184.84 -656.81 2 147.92 -1248 -244.08 -621.65 3 75.93 -971.01 -218.07 -501.25 4 36.79 -661.17 -159.21 -347.99 5 14.84 -334.14 -83.16 -177.55 Table A4: Pesaran (2015) Panel Unit Root Test in the Presence of Cross-Section Dependence Variables Level First Difference Inflation -1.425 -6.086*** Activity -2.469** -6.100*** Curvature -3.356*** -6.190*** Slope -2.303* -6.190*** Level -2.578*** -6.190*** Local Financial -1.986 -6.025*** Table A5: Pesaran (2015) Test for Weak Cross Sectional Dependence Level Equation Slope Equation Curvature Equation (Full Specification) (Full Specification) (Full Specification) CD Test Statistic 22.32 11.71 4.79 P-value 0.00 0.00 0.00 34
37. Table A6 : Out-of-sample forecasting exercise based on machine learning algorithms : Level Brazil h=1 h=2 h=3 h=4 h=5 h=6 Ridge 1.039 1.007 0.986 0.965 0.948 0.936 LASSO 1.025 1.000 0.984 0.972 0.964 0.950 Elastic Net 1.019 0.995 0.980 0.971 0.956 0.944 Hungary Ridge 1.020 1.015 1.014 1.013 1.010 1.010 LASSO 1.019 1.019 1.016 1.016 1.016 1.016 Elastic Net 1.028 1.023 1.021 1.021 1.020 1.019 India Ridge 0.999 0.996 0.994 0.993 0.994 0.991 LASSO 0.988 0.986 0.982 0.983 0.983 0.983 Elastic Net 0.992 0.990 0.989 0.986 0.985 0.983 Ridge 0.992 0.975 0.968 0.979 0.965 0.947 LASSO 0.966 0.978 0.997 0.990 0.995 0.986 Elastic Net 0.989 0.978 0.952 0.933 0.931 0.940 Malaysia Mexico Ridge 0.997 0.993 0.992 0.988 0.985 0.984 LASSO 1.002 1.001 1.005 1.000 1.000 1.003 Elastic Net 0.997 0.998 0.999 1.000 1.002 0.999 Ridge 1.004 1.008 1.008 1.012 1.011 1.011 LASSO 1.003 1.004 1.004 1.008 1.004 1.001 Elastic Net 1.000 1.001 1.001 1.003 1.000 1.001 Ridge 0.920 0.915 0.921 0.921 0.906 0.898 LASSO 0.935 0.923 0.924 0.914 0.906 0.891 Elastic Net 0.929 0.901 0.901 0.901 0.899 0.886 Poland Russia S.Africa Ridge 0.995 0.993 0.992 0.992 0.992 0.991 LASSO 0.993 0.996 0.996 0.996 0.995 0.995 Elastic Net 0.992 0.996 0.995 0.996 0.994 0.994 Ridge 1.009 1.011 1.012 1.012 1.010 1.010 LASSO 1.021 1.022 1.021 1.023 1.022 1.020 Elastic Net 1.016 1.018 1.021 1.023 1.015 1.014 Thailand Turkey Ridge 0.966 0.958 0.952 0.950 0.949 0.943 LASSO 0.989 0.982 0.975 0.970 0.961 0.957 Elastic Net 0.985 0.975 0.965 0.955 0.947 0.938 Dependent variable is level factor. Entries are relative RMSEs based on the model that includes all factors (prices, economic activity, domestic financial) extracted using the dynamic factor model approach of Giannone et al.(2008). Hence, a value of less than unity indicates that a 35 particular variable selection method yields more accurate forecasts than those of the model that do not utilize variable selection methods recursively, for a given forecast horizon. Models that yield the lowest RMSE for each forecast horizon are denoted in bold.
38. Table A7 : Out-of-sample forecasting exercise based on machine learning algorithms : Slope Brazil h=1 h=2 h=3 h=4 h=5 h=6 Ridge 1.078 1.059 1.066 1.061 1.057 1.040 LASSO 1.124 1.127 1.109 1.081 1.061 1.046 Elastic Net 1.064 1.073 1.070 1.064 1.067 1.067 Hungary Ridge 0.995 0.993 0.993 0.992 0.990 0.990 LASSO 0.999 0.998 0.998 0.998 0.998 0.997 Elastic Net 1.001 1.001 1.002 1.002 1.001 1.001 India Ridge 0.984 0.975 0.976 0.979 0.983 0.993 LASSO 0.984 0.974 0.973 0.968 0.976 0.982 Elastic Net 1.003 1.008 0.997 0.993 0.980 0.979 0.890 0.910 0.910 0.914 0.923 0.920 Malaysia Ridge LASSO 0.864 0.854 0.868 0.862 0.856 0.859 Elastic Net 0.862 0.843 0.846 0.848 0.858 0.857 Mexico Ridge 0.997 0.997 0.994 0.994 0.994 0.992 LASSO 0.988 0.989 0.989 0.990 0.993 0.993 Elastic Net 0.990 0.989 0.987 0.989 0.989 0.990 Ridge 0.988 0.991 0.990 0.992 0.992 0.994 LASSO 0.998 0.999 0.999 0.999 1.000 0.998 Elastic Net 0.993 0.997 0.994 0.996 0.995 0.996 Ridge 0.947 0.954 0.958 0.960 0.965 0.965 LASSO 0.976 0.976 0.981 0.981 0.984 0.979 Elastic Net 0.970 0.971 0.974 0.972 0.975 0.982 Poland Russia S.Africa Ridge 0.999 0.996 0.996 0.995 0.993 0.991 LASSO 1.001 0.999 1.001 0.999 0.998 1.000 Elastic Net 1.004 1.003 1.002 1.001 0.999 0.998 Ridge 0.986 0.988 0.983 0.975 0.968 0.975 LASSO 0.971 0.954 0.942 0.944 0.946 0.950 Elastic Net 0.979 0.984 0.967 0.956 0.958 0.968 Thailand Turkey Ridge 1.011 1.000 0.995 0.991 0.984 0.978 LASSO 1.010 1.008 1.000 0.993 0.981 0.975 Elastic Net 1.014 1.008 0.998 0.990 0.981 0.978 Dependent variable is slope factor. Entries are relative RMSEs based on the model that includes all factors (prices, economic activity, domestic financial) extracted using the dynamic factor model approach of Giannone et al.(2008). Hence, a value of less than unity indicates that a 36 particular variable selection method yields more accurate forecasts than those of the model that do not utilize variable selection methods recursively, for a given forecast horizon. Models that yield the lowest RMSE for each forecast horizon are denoted in bold.
39. Table A8 : Out-of-sample forecasting exercise based on machine learning algorithms : Curvature Brazil h=1 h=2 h=3 h=4 h=5 h=6 Ridge 0.984 0.981 0.976 0.968 0.965 0.959 LASSO 0.986 0.980 0.970 0.969 0.964 0.958 Elastic Net 0.987 0.980 0.972 0.967 0.961 0.958 Hungary Ridge 1.009 1.010 1.009 1.007 1.006 1.004 LASSO 1.007 1.005 1.004 1.004 1.004 1.004 Elastic Net 1.003 1.004 1.005 1.004 1.004 1.002 India Ridge 0.996 0.994 0.997 0.994 0.997 0.993 LASSO 0.989 0.990 0.988 0.991 0.987 0.986 Elastic Net 0.988 0.988 0.991 0.992 0.990 0.988 0.978 0.973 0.961 0.955 0.959 0.962 Malaysia Ridge LASSO 0.978 0.967 0.958 0.949 0.950 0.945 Elastic Net 0.976 0.967 0.951 0.949 0.948 0.937 Mexico Ridge 0.991 0.990 0.988 0.990 0.990 0.988 LASSO 0.995 0.995 0.997 0.998 0.997 0.994 Elastic Net 0.994 0.993 0.993 0.996 0.995 0.995 Ridge 1.012 0.987 0.965 0.952 0.934 0.925 LASSO 1.012 0.997 0.983 0.971 0.960 0.951 Elastic Net 1.013 0.996 0.987 0.973 0.964 0.956 Ridge 0.977 0.943 0.933 0.932 0.944 0.966 LASSO 1.002 0.994 1.002 1.014 1.013 1.015 Elastic Net 0.960 0.953 0.955 0.968 0.986 1.002 Poland Russia S.Africa Ridge 0.994 0.994 0.992 0.993 0.992 0.992 LASSO 0.996 0.994 0.993 0.993 0.994 0.996 Elastic Net 0.991 0.993 0.994 0.993 0.995 0.995 Ridge 0.967 0.961 0.948 0.939 0.945 0.942 LASSO 0.947 0.939 0.936 0.937 0.929 0.934 Elastic Net 0.959 0.952 0.961 0.958 0.954 0.938 Ridge 0.990 0.983 0.978 0.973 0.972 0.972 LASSO 0.987 0.992 0.993 0.992 0.983 0.981 Elastic Net 0.990 0.988 0.986 0.978 0.972 0.974 Thailand Turkey Dependent variable is curvature factor. Entries are relative RMSEs based on the model that includes all factors (prices, economic activity, domestic financial) extracted using the dynamic factor model approach of Giannone et al.(2008). Hence, a value of less than unity indicates 37 that a particular variable selection method yields more accurate forecasts than those of the model that do not utilize variable selection methods recursively, for a given forecast horizon. Models that yield the lowest RMSE for each forecast horizon are denoted in bold.
40. Central Bank of the Republic of Turkey Recent Working Papers The complete list of Working Paper series can be found at Bank ’s website (http://www.tcmb.gov.tr) The Role of Imported Inputs in Pass-through Dynamics (Dilara Ertuğ, Pınar Özlü, M. Utku Özmen, Çağlar Yüncüler Working Paper No. 20/03, February 2020) Nowcasting Turkish GDP with MIDAS: Role of Functional Form of the Lag Polynomial (Mahmut Günay Working Paper No. 20/02, February 2020) How Do Credits Dollarize? The Role of Firm’s Natural Hedges, Banks’ Core and Non-Core Liabilities (Fatih Yılmaz Working Paper No. 20/01, February 2020) Hidden Reserves as an Alternative Channel of Firm Finance in a Major Developing Economy (İbrahim Yarba Working Paper No. 19/36, December 2019) Interaction of Monetary and Fiscal Policies in Turkey (Tayyar Büyükbaşaran, Cem Çebi, Erdal Yılmaz Working Paper No. 19/35, December 2019) Cyclically Adjusted Current Account Balance of Turkey (Okan Eren, Gülnihal Tüzün Working Paper No. 19/34, December 2019) Term Premium in Turkish Lira Interest Rates (Halil İbrahim Aydın, Özgür Özel Working Paper No. 19/33, December 2019) Decomposing Uncertainty in Turkey into Its Determinants (Emine Meltem Baştan, Ümit Özlale Working Paper No. 19/32, December 2019) Demographic Transition and Inflation in Emerging Economies (M. Koray Kalafatcılar, M. Utku Özmen Working Paper No. 19/31, December 2019) Facts on Business Dynamism in Turkey (Ufuk Akçiğit, Yusuf Emre Akgündüz, Seyit Mümin Cılasun, Elif Özcan Tok, Fatih Yılmaz Working Paper No. 19/30, September 2019) Monitoring and Forecasting Cyclical Dynamics in Bank Credits: Evidence from Turkish Banking Sector (Mehmet Selman Çolak, İbrahim Ethem Güney, Ahmet Şenol, Muhammed Hasan Yılmaz Working Paper No. 19/29, September 2019) Intraday Volume-Volatility Nexus in the FX Markets: Evidence from an Emerging Market (Süleyman Serdengeçti, Ahmet Şensoy Working Paper No. 19/28, September 2019) Is There Asymmetry between GDP and Labor Market Variables in Turkey under Okun’s Law? (Evren Erdoğan Coşar, Ayşe Arzu Yavuz Working Paper No. 19/27, September 2019) Composing High-Frequency Financial Conditions Index and Implications for Economic Activity (Abdullah Kazdal, Halil İbrahim Korkmaz, Muhammed Hasan Yılmaz Working Paper No. 19/26, September 2019) A Bayesian VAR Approach to Short-Term Inflation Forecasting (Fethi Öğünç Working Paper No. 19/25, August 2019) Foreign Currency Debt and the Exchange Rate Pass-Through (Salih Fendoğlu, Mehmet Selman Çolak, Yavuz Selim Hacıhasanoğlu Working Paper No. 19/24, August 2019) Two and a Half Million Syrian Refugees, Tasks and Capital Intensity (Yusuf Emre Akgündüz, Huzeyfe Torun Working Paper No. 19/23, August 2019) Estimates of Exchange Rate Pass-through with Product-level Data (Yusuf Emre Akgündüz, Emine Meltem Baştan, Ufuk Demiroğlu, Semih Tümen Working Paper No. 19/22, August 2019) Skill-Biased Occupation Growth (Orhun Sevinç Working Paper No. 19/21, August 2019) Impact of Minimum Wages on Exporters: Evidence From a Sharp Minimum Wage Increase in Turkey (Yusuf Emre Akgündüz, Altan Aldan, Yusuf Kenan Bağır, Huzeyfe Torun Working Paper No. 19/20, August 2019)