The transmission of US monetary policy shocks to EMEs: an empirical analysis

Transcription

1 The ransmission of US moneary policy shocks o EMEs: an empirical analysis Manuel Ramos-Francia and Saniago García-Verdú 1 Absrac We consider he exen o which hree channels ha ransmi US moneary policy shocks o emerging marke economies (EMEs) may have changed in imporance and relaive srengh ie srucurally changed since he hird quarer of We run linear regression models and esimae a facor-augmened vecor auoregression (FAVAR) for he Unied Saes and each EME in our daabase, using macroeconomic variables from each economy. We find ha he possibiliy of srucural change in he policy rae, exchange rae, and long-erm ineres rae channels generally depends on he EME in quesion. Also, in he case of some ess, accouning for unconvenional moneary policy (UMP) does seem o make a difference in his resul. However, EMEs seem o have experienced some srucural changes more uniformly in wha can be inerpreed as (i) second-round effecs in he channels or, more likely, (ii) as changes in channels ha we have no explicily modelled bu ha are noneheless being capured. Such changes highligh he poenial for a renewed inerdependence beween he moneary policies in mos EMEs and US moneary policy above and beyond he unprecedened sance of he laer. They also underscore he imporance for mos EMEs of aking appropriae policy measures, in paricular as he Unied Saes begins ighening is moneary sance. Our resuls should be inerpreed wih some cauion given he limied lengh of he ime series. Keywords: Moneary policy, cenral banking JEL classificaion: E4, E5 1 Bank of Mexico. The opinions expressed in his paper are exclusively he responsibiliy of he auhors and do no necessarily reflec he poin of view of he Bank of Mexico. BIS Papers No

2 Inroducion I is well undersood ha during recen decades he world s economies and financial sysems have grown more inegraed hrough a relenless globalizaion process (eg Friedman (2005)). However, wha is perhaps less well undersood is he exen o which he process has made moneary policy more globally inegraed, hereby ighening he mechanisms by which moneary policy shocks are ransmied inernaionally. Despie some significan effors owards undersanding his opic in general (eg Galí and Gerler (2009)) as well as wih regard o emerging marke economies (EMEs) (eg Kim and Yang (2009)), he issues peraining o EMEs have received relaively lile aenion. In his conex, several relevan quesions arise: Wha is he naure of he mechanism by which US moneary policy effecs are ransmied o EMEs? Wha is he srengh and relaive imporance of he ransmission channels? Have hese channels changed afer he recen financial crisis and, if so, o wha exen? Wha are he implicaions, if any, for EME economic policies? Agains his backdrop, we underake wo sudies. Firs, we quaniaively explore he policy rae, exchange rae, and long-erm ineres rae channels by which US moneary policy shocks are ransmied o EMEs. 2 Of course, hese are bu hree among oher channels presen in he ransmission mechanism. Second, we assess he exen o which he srengh and relaive imporance of hese channels migh have changed since recen global financial crisis; we hereafer refer o such changes as srucural changes. We apply hese enquiries o a se of 15 EMEs (someimes fewer, depending on daa availabiliy) ha have varying exchange rae arrangemens, moneary regimes, degrees of financial openness, and policy responses (he Appendix includes he lis of EMEs along wih descripions of heir exchange arrangemens and moneary regimes). 3 We esimae a se of linear regression models and a facor-augmened vecor auoregression (FAVAR) for he Unied Saes and each EME in our daabase, using macroeconomic variables from each economy. This allows us o incorporae a wide range of ime series daa in our esimaion. We assess he exen o which such channels migh have changed beginning wih he hird quarer of To do so, we run a se of regressions wih daa for he 22 quarers preceding he crisis (Q o Q2 2008) and, in a separae esimaion, wih daa for he 22 subsequen quarers (Q o Q4 2013). We use an auxiliary dummy variable o accoun for possible changes in he coefficiens measuring he impac of he US policy rae (ie he federal funds rae) on relevan macroeconomic variables. Accordingly, we run a 2 3 For a specific emerging marke economy, we explore hese channels by examining how changes in he US policy rae affec (i) he EME s policy rae; (ii) he EME s exchange rae; and (iii) he EME s long-erm ineres raes. As he ransmission effec is no limied o hese hree variables, we aim o capure oher effecs hrough a vecor auoregression. One migh divide he EMEs in our daabase ino he following groups: (i) Souh American EMEs, mos of which have benefied from he commodiies super cycle and he significan increase in demand, in paricular from China; (ii) Asian EMEs, characerized by an efficien indusrial block; (iii) easern European EMEs, whose siuaion is grealy affeced by he EU; (iv) Mexico, which is closely relaed o he US economy; and (v) ohers. Specific paerns wihin hese groups do no appear in our exercises. 364 BIS Papers No 78

3 baery of saisical ess on he relevan coefficiens. I is imporan o clarify ha while in he regressions we use daa from he quarers preceding he crisis and, in a separae esimaion, we use daa from he subsequen quarers; in he case of he FAVAR models we use he whole sample. The key difference is ha in one FAVAR we allow for a srucural change in some coefficiens, while in he oher FAVAR we do no allow for a srucural change, as laer explained in deail in he ex. In his conex, a naural challenge is o measure he US moneary policy sance once he US federal funds rae reaches he zero lower bound. A ha poin, we sidesep he issue by using he rae proposed by Wu and Xia (2013) (hereafer, he Wu and Xia rae). Tha rae essenially coincides wih he US federal funds rae as long as he policy rae is posiive; bu once he policy rae his he zero lower bound, he Wu and Xia rae can become negaive. When negaive, is disance from zero is a quaniaive measure of he effeciveness of unconvenional moneary policy. Having a beer undersanding of he inernaional ransmission mechanism of moneary policy is relevan for many reasons. Moneary policy in advanced economies has played an essenial role in he recovery from he global financial crisis. In fac, i has been referred o as he only game in own (Rajan (2013)). 4 Also, i is crucial for invesors and policymakers alike o undersand he implicaions of global moneary policy for EMEs, where capial flows can be direcly influenced by changes in he moneary policy sance in advanced economies. A beer undersanding of he inernaional ransmission of US moneary policy can herefore provide criical assisance o policy planning wihin EMEs. Our main findings are as follows. We obain iniial evidence ha he exisence of a srucural change in he policy rae, exchange rae, and long-erm ineres rae channels depends on he EME in quesion. 5 However, EMEs seem o have experienced some srucural changes more uniformly. These changes can be hough of as second-round effecs of he hree channels modelled; bu we believe hey are more likely o be arising from oher channels ha our model is implicily capuring. Here, we will discuss our resuls in a broader perspecive. Much of he aenion given o he inernaional ransmission mechanisms of US moneary policy sared wih he global financial crisis. In responding o ha crisis, he US cenral bank lowered he policy rae esenially o he zero lower bound, afer which i began implemening unconvenional moneary policies. Under hose measures, significan capial flows swifly enered and exied EMEs, depending o a grea exen on diverse episodes and siuaions in he advanced economies. Moreover, here are possibly several facors influencing each of he hree channels we sudy here. For he policy rae channel, we could hardly have expeced uniform resuls since: (i) he iniial condiions of each EME before he global financial crisis differed subsanially; (ii) we have observed diversiy among EMEs in heir managemen and policy responses o unconvenional moneary policies; and (iii) excep for he wors 4 5 Moreover, some researchers have cauioned abou is limis and risks (eg Eichengreen e al (2011)). As will be discussed shorly, srucural change in he long-erm ineres rae channel appears o be more uniform across EMEs if he linear regression models accoun for unconvenional moneary policies in he Unied Saes. BIS Papers No

4 par of he crisis, very few EMEs have shared he same business cycle phase wih he main advanced economies, in paricular he Unied Saes. Moreover, business cycles have no been synchronized even among EMEs. Our mixed resuls are in line wih he significan differences among EMEs in he facors considered above. For he exchange rae channel, he resuls depend on he specific EME as well. Ye, arguably, here is less evidence of possible srucural changes in his channel across he baery of ess. We could have expeced his for wo reasons. These EMEs have, for a relaively long ime now, mainained a saionary inflaionary process (Noriega, Capisrán and Ramos-Francia (2013)). Indeed, hey have no moneized heir fiscal deficis, ending fiscal dominance. This has produced a much lower pass-hrough. Finally, he long-run ineres rae channel also offers mixed evidence. Ye, a specific se of exercises (he linear regression models) give some evidence of a more generalized change when one accouns for unconvenional moneary policies. We believe ha here are several facors behind hese resuls, some of which are above and beyond he global financial crisis. On he one hand, he secular decrease in long-erm raes in he Unied Saes ha began in he early 1980s acceleraed, given he search-for-yield phenomenon, and was furher suppored by elemens such as US demographic dynamics and he implemenaion of unconvenional moneary policies. In fac, he fall in he long-erm raes in EMEs has been faser han heir corresponding individual deflaionary processes. In addiion, hese US facors have played a reinforcing role abroad. For example, he aggressive search for yield has cerainly affeced he ineres raes in EMEs. In his conex, a key quesion is wha would be he scenario once normalizaion of he US federal funds rae sars. We believe ha, alhough some bous of volailiy will quie possibly ake place given he secular decrease in US long-erm raes, he fierce compeiion among asse managemen companies for he highes yields in EMEs will coninue. In he long-run rae channel, oher facors have surely played a role as well and hus have led o differing resuls among EMEs. For insance, some of hese economies have implemened capial conrols, esablished macroprudenial policies, allowed credi booms, or had more flexibiliy in heir fiscal policies. All in all, i would have been a surprise o observe uniform resuls for he long-erm ineres rae channel. However, join ess (ie ess joinly assessing a srucural change in various channels) end o rejec he null hypohesis of no srucural change. We believe here are channels ha may have changed bu ha we are no explicily modelling. For insance, he global financial crisis and he associaed financial regulaory response have significanly affeced he balance shees of banks and, o an exen, also of he so-called shadow banks. As is well known, heir aim has been o improve heir risk pricing. These facors are likely o have affeced oher channels, for insance, he credi channel. In sum, we find more uniform evidence of a srucural change in channels we do no observe bu ha we hink our models are capuring. A cavea abou our resuls is ha only a few years have passed since he global financial crisis. Thus, he pos-crisis period offers a relaively small number of observaions for each ime series, which makes saisical relaionships difficul o deec. In addiion, he relaively shor ime series limi he number of model specificaions we can enerain. Thus, our resuls represen an exploraion raher han robus saisical findings. 366 BIS Papers No 78

5 In he following secions we describe our daa, model, and esimaions, discuss our main resuls and offer some concluding remarks. Daa, model and esimaions As one would expec, here are more ime series available for he Unied Saes han for any EME. In addiion, daa are no uniform across EMEs. Prominenly, for some EMEs we do no have he ime series of he long-erm ineres rae, which in hose cases prevens us from sudying ha channel. Noneheless, we have ried o keep daabases as uniform as possible. All ime series have been ransformed o a quarerly frequency and rendered saionary, he laer a needed condiion for VAR model esimaion. Le z be a generic variable. In general, excep for ineres raes, he percenage growh is aken (ie log ( z) log ( z 1 )). Noneheless, if a ime series has negaive values, hen he difference is aken (ie z z 1 ). To obain evidence on saionariy, a sabiliy es is performed on each of he VARs we esimae. 6 These ess are assessed bu no repored. Daa US daa We use a relaively large number of ime series for he Unied Saes. The US federal funds rae is an indicaor of he US moneary policy sance. To accoun for changes in he use of unconvenional moneary policies or, equivalenly, for hose periods in which he US federal funds rae essenially his he zero lower bound, we use he Wu and Xia rae in separae esimaions. Tha rae coincides wih he US federal funds rae if he laer is nonnegaive. If he US federal funds rae his he zero lower bound, he Wu and Xia rae can urn negaive, and is disance from zero is a measure of he effeciveness of unconvenional moneary policy. 7 The following lis shows each of he US variables we have used and heir unis of measure. 8 For convenience, we have divided hem ino hree groups financial, moneary and real wih no direc implicaions for our models. Financial 3-monh US ineres rae (percenage); 10-year US ineres rae (percenage); Morgan Sanley Capial Inernaional Index (MSCI) (percenage growh); Dow Jones 30 Indusrial (percenage growh); and This es verifies ha all of he norms of he marix eigenvalues are sricly less han 1. Wu and Xia (2013) show ha heir model can be used o convey he macroeconomic effecs of unconvenional moneary policy a he zero lower bound. The source for each series is Haver Analyics excep as oherwise noed. BIS Papers No

6 morgage rae (percenage). Moneary US federal funds rae (percenage); Wu and Xia rae (percenage); moneary base (percenage growh); M1 (percenage growh); M2 (percenage growh); currency in circulaion (percenage growh); Federal Reserve balance shee asses (percenage growh); and Federal Reserve balance shee liabiliies (percenage growh). Real real GDP, seasonally adjused (SA), billions of chained (2009) dollars (percenage growh); curren accoun (per cen of GDP) (difference); expors (percenage growh); impors (percenage growh); consumer credi (percenage growh); micro-finance indusry lending o privae secor (percenage growh); privae secor credi, over GDP (difference); cenral governmen budge, over GDP (difference); deb ousanding: domesic economy, over SA GDP (difference); households and non-profi insiuions serving households, over quarerly GDP a a seasonally adjused annual rae (SAAR) (difference); non-financial corporaions, over quarerly SAAR GDP (difference); US financial corporaions/insiuions, over quarerly SAAR GDP (difference); gross capial formaion (percenage growh); corporae gross operaion surplus (percenage growh); gross disposable income (percenage growh); gross savings (percenage growh); earnings (percenage growh); manufacuring (percenage growh); manufacuring, excluding consrucion (percenage growh); housing prices (percenage growh); 368 BIS Papers No 78

7 housing sars (percenage growh); housing permis (percenage growh); housing compleions (percenage growh); shipmens (percenage growh); reail value (percenage growh); wholesale (percenage growh); consumer confidence (percenage growh); consumer expecaions (percenage growh); capaciy (percenage growh); employmen (percenage growh); labour force (percenage growh); and uni labour cos (percenage growh). EME daa We divided he EME ime series ino he same hree groups. We have aken as given he policy rae saus provided by our main daa source, Haver Analyics. The variables and unis of measure are as follows. Financial exchange rae (percenage growh); Morgan Sanley Capial Inernaional, MSCI (percenage growh); 3-monh ineres rae (percenage); and long-run (10-year) ineres rae (percenage). Moneary policy ineres rae (percenage); M1 (percenage growh); M2 (percenage growh); and M3 (percenage growh). Real real GDP (quarer percenage growh); curren accoun, over GDP (difference); manufacuring (percenage growh); rade balance, over GDP (difference); privae consumpion expendiure (percenage growh); public consumpion expendiure (percenage growh); and gross capial formaion (percenage growh). BIS Papers No

8 Exchange rae arrangemens, moneary policy frameworks and financial openness Each economy varies in he way i ses and conducs is exchange rae and moneary policies. In a broader conex, i is relevan o undersand he policy framework being implemened in he EME of ineres, including is moneary and fiscal policies, capial flow managemen, and macro-prudenial policies. Presumably, he sae of hese policies is parially capured by he macroeconomic variables we have incorporaed ino our VAR models. Ye, i is plausible ha some aspecs no capured explicily may noneheless be relevan in inerpreing our resuls. In his conex, we firs provide an IMF (2013) classificaion of he de faco exchange rae arrangemens and moneary policy framework for each of he economies in our daabase (Table 1). We have omied hose classificaions in which none of our economies appear. Hong Kong SAR is a one end of he exchange rae arrangemens, having a currency board. A he oher end, we find counries, such as Chile and Mexico, wih an independenly floaing regime. For he moneary policy framework, Hong Kong SAR is again disincive, having he US dollar as is anchor; mos EMEs in our daabase mainain an inflaion-argeing regime. IMF de faco exchange rae arrangemens and moneary policy regimes 1 Table 1 Moneary policy framework Exchange rae arrangemen Currency board arrangemen Oher convenional fixed peg arrangemen Exchange rae anchor USD Hong Kong SAR Inflaion argeing framework Oher Crawling peg China Managed floaing wih no predeermined pah for he exchange rae Colombia Indonesia Peru Thailand Brazil Chile Czech Republic Israel Souh Korea Mexico Poland Souh Africa India Independenly floaing 1 See he Appendix for a more deailed explanaion of he moneary policy regimes. Source: IMF (2013). Second, we presen he Chinn and Io index (2006) for each counry (Figure 1). This index is a de jure measuremen of he financial openness in an economy. We are ineresed in knowing he level of financial openness in each EME and wheher i has changed in he sample period. Noable cases of an increase in financial openness are Colombia and Korea, whereas Thailand is a case in which financial 370 BIS Papers No 78

9 openness diminished. In general, however, economies ha began he sample period eiher financially open, such as Hong Kong SAR and Israel, or relaively closed, such as China and India, have remained ha way. Chinn-Io index 1 Figure 1 1 Averages for he and periods. The index is a de jure measure of financial openness. A larger and posiive (negaive) value means greaer (lesser) financial openness. Values based on he resricions on cross-border financial ransacions in he IMF s Annual Repor on Exchange Arrangemens and Exchange Resricions. Source: Chinn and Io (2006). Third, in he Appendix, we provide he IMF s definiions of de faco exchange arrangemens shown in Table 1 as well as brief descripions of exchange arrangemens and moneary regimes, as conained in HSBC (2011), for he EMEs in our daabase. The HSBC informaion complemens, alhough in some respecs does no exacly coincide wih, he IMF s classificaions. In sum, a horough assessmen of our resuls requires an undersanding of hese arrangemens and regimes. As noed, we do no include his informaion in he model; hey are for inerpreaion purposes only. Model We implemen an approach similar o ha in Bernanke e al (2005) and Boivin e al (2009) by using a facor-augmened VAR (FAVAR). There are a leas hree imporan aspecs o heir approach. Firs, moneary policy decisions involve he analysis of large amouns of informaion. The use of principal componen analysis (PCA) allows for a sysemaic use of a wide range of daa. Second, some of he variables on which moneary policy depends are unobservable o agens, eg poenial oupu growh. Presumably, he use of PCA parially capures laen variables. Third, we could have aemped o calibrae, say, a dynamic sochasic general equilibrium (DSGE) model. Indeed, here are close relaionships beween BIS Papers No

10 DSGE and VAR models (eg Fernandez-Villaverde e al (2007)). However, DSGE models require srong idenifying resricions. Moreover, alhough he VAR approach has no been free of criicism, i provides enough flexibiliy and has few idenificaion assumpions. There are several seps o our model s esimaion. The firs one enails he exracion of he main principal componens from our ime series se. On one hand, consider x, an nx 1 vecor ha conains all of he ime series (a ime ) in he EME daabase of ineres, excep for he variables in quesion, such as he policy rae and US exchange rae of he EME a hand. On he oher hand, consider x, an nx 1 vecor conaining all he ime series (a ime ) in he US daabase, excep for he federal funds rae. Nex, consider he following approximaions: x E x + υ c ( ) us us us us ( ) x E x + υ c where c is an mx 1 vecor wih he m h firs principal componens a ime us associaed wih he PCA decomposiion of x, wih m << n. Similarly, c is an mx 1 us vecor wih he ( m ) h firs principal componens associaed wih he PCA us us us decomposiion of he ime series x, wih m << n. 9 The vecor υυ ( us ) conains he facor loadings associaed wih he PCA decomposiion of he ime series in us x x. ( ) Second, once he c us and c ime series have been obained, he following vecor is consruced by sacking he vecors obained above along wih he economic variables of ineres, specifically: US y = c ff c lrii i dfx, us where we have he ( m ) h h and m firs principal componens ( c US and c ), he percenage change in he foreign exchange rae ( dfx ), he EME s long-run ineres rae ( lri ) and is policy rae ( i ), and he US federal funds rae ( ff ). The long-run rae is no available for all he economies in our daabase. In a differen exercise, we subsiue he US federal funds rae for he Wu and Xia rae. Thus, he VAR model for y is posied as: ( ) 1 y = ψ L y + ε (1) where ψ ( L ) is a lag polynomial, and ε has mean 0 and variance-covariance marix Σ. Third, he VAR model is esimaed under some idenifying assumpions, for which we provide more deails in he nex subsecion. 9 x E x c. Similarly, if If m equals n, hen ( ) = υ us m equals us us us us x E x = c. us n, hen ( ) υ 372 BIS Papers No 78

11 Our main ess are based on he following exension of model (1): where [ ] ( ) 1 y = ψ L y + u (2) y = y ff d and d is a dummy variable defined as follows: 0 if < 3 Q.2008 d = 1 if 3 Q.2008 Similarly, ψ ( L ) is a lag polynomial ha accommodaes he coefficiens associaed wih he ff d variable. The error erm u has mean 0 and variance-covariance marix Ω. To see how he dummy capures a possible srucural change in our original VAR model (1), consider he following simplified version of he model. In i, we have i, included only he variaion in he exchange rae ( ) and he US federal funds rae ( ) dfx, he EME policy rae ( ) ff. The fourh variable is an auxiliary ( ) defined above. Hence, we have: dfx = a11dfx 1+ a12i 1+ a13ff 1+ a14ff 1 d 1+ u1, i = a21dfx 1+ a22i 1+ a23ff 1+ a24ff 1 d 1+ u2, ff = a31dfx 1 + a32i 1 + a33 ff 1 + a34 ff 1 d 1 + u3, ff d = a dfx + a i + a ff + a ff d + ε, , which can be rewrien as: dfx = a11dfx 1 + a12i 1 + ( a13 + a14 d 1 ) ff 1 + u1, i = a21dfx 1 + a22i 1 + ( a23 + a24 d 1 ) ff 1 + u2, ff = a31dfx 1 + a32i 1 + ( a33 + a34 d 1 ) ff 1 + u3, n ff d = a dfx + a i + a + a d ff + u ( ) , ff d, as Our firs es considers wheher he coefficien a 24 is saisically significanly differen from zero. If so, ha would provide us wih some evidence ha he policy rae channel had a srucural change afer he hird quarer of Tha is, he coefficien, which measures he conemporaneous effec of ff on i, would have changed from a 23 o a 23 + a 24. This happens provided ha a 24 is saisically differen from zero. Thus, we posi as a null hypohesis ha a 24 = 0. Similarly we can es for a srucural change in he exchange rae channel by saisically assessing wheher a 14 = 0. In addiion, we can joinly es wheher here has been a change in boh channels by considering he null hypohesis: a14 = a 24 = An underlying assumpion is ha we know a priori he period in which he srucural change could have aken place. BIS Papers No

12 In sum, we perform each of hese ess as linear resricions on he following sysem: 11 ψ ( ) 1 y = L y + u. As an imporan exercise, we plo he impulse-response funcions (IRFs) given a 25 basis poin orhogonal shock o he US federal funds rae under model (1) and again under model (2). This allows for a visual comparison of, on one hand, he IRFs from he model in which i is assumed ha here has been no srucural change in he channels, wih, on he oher hand, he IRFs from he model ha allows for a possible srucural change beween he Q Q period and he Q Q period. 12 Compared wih he linear ess, he IRFs depend on he specific Cholesky decomposiion considered. An underlying assumpion is ha he res of he coefficiens do no go hrough a srucural change. Alhough i would be desirable o consider a more general hypohesis in order o capure a possible srucural change in all of ψ ( L ), his exercise is limied by he relaively shor lengh of he ime series. 13 Esimaion For esimaion purposes, we firs divide our daabase ino wo ses, as suggesed by our noaion. In one se, we have he US ime series, excep for he US federal funds rae (or he Wu and Xia rae). The second se conains an EME s ime series excep for is policy rae, exchange rae, and long-erm ineres rae. Second, by decomposing he US variables using PCA and also decomposing he EME s variables using PCA, we obain a pair of ses of principal componens. The componens associaed wih he US daa are denoed by c US, and hose associaed wih he EME s daa by c. Third, we consruc he vecor y = [ US ] c ff c lri i dfx, having used he firs hree principal componens from each se. 14 Accordingly, he c US and c enries in y sand for vecors, and he res sand for scalars for each These linear resricions, under he null hypohesis, follow an F disribuion, wih he number of resricions ( v ), and he number of observaions minus he number of parameers being 1 esimaed ( v ), as degrees of freedom; ie is disribuion is F ( ν, ν ). 2 We have esimaed models (1) and (2) under he idenifying assumpions of orhogonal shocks, as will be explained laer. 2 In he VAR model, if one has n variables, n + n + n( n 1) 2 parameers have o be esimaed. Adding one more variable implies ha such number increases by 4n 1. Assessing a possible srucural change in all of ψ ( L ) calls for addiional coefficiens plus he ransiion probabiliies n addiional coefficiens. Regime swiching calls for We normally choose hree componens, for he following reasons: (i) On average, hree componens explain more han 50% of he accumulaed variance of he whole daa se; in general, he incremen of accumulaed variance ended o markedly drop by he fourh componen. And (ii) using he same number of lags for each EME has he advanage of mainaining a level of comparabiliy across EMEs. 2 n 374 BIS Papers No 78

13 Fourh, we assess he number of lags by means of Schwarz s Bayesian informaion crierion (SBIC). In mos cases, SBIC poins o a lag of 1. This is reasonable given ha we are using principal componens, which can proxy he lagged variables dynamics. 15 For hose cases in which he SBIC indicaes a greaer lag, we insead increase he number of EME and US componens unil he es poins o a lag of Fifh, he sysems y = ψ ( L) y 1 + ε and y = ψ ( ) L y 1 + u are esimaed. As a final sep, ess and IRFs are calculaed and assessed. Also, we redo he seps above wih he Wu and Xia rae insead of he US federal funds rae o allow he model o capure he effecs of unconvenional moneary policies. A VAR esimaion has some idenifying resricions. In paricular, one has o make a choice regarding he idenificaion of shocks. In our case, we assume orhogonal shocks by applying he Cholesky decomposiion o he variancecovariance marices. To his end, we order our VAR variables by heir speed of adjusmen, as suggesed by our noaion y = [ US ] c ff c lri i dfx. Thus, on impac, we have assumed ha he adjusmen of he exchange rae is followed by ha of ineres raes and hen of he EME s componens. Then he US federal funds rae adjuss, followed by he US componens. 17 Thus, in general, he EME s variables respond on impac o changes in US variables bu no he oher way around. For idenificaion purposes, we could have esimaed a srucural VAR (SVAR). We have noneheless preferred a Cholesky decomposiion, for wo reasons: (i) i is equivalen o a jus-idenified SVAR and hus generally implies a less resricive scheme; and (ii) we find i more suiable, a leas iniially, as a SVAR would generally require sronger idenificaion assumpions and consequenly more a priori knowledge abou he variables relaionships. Resuls To se he sage, we firs consider he cross-correlaions beween he US federal funds rae and each of our hree variables of ineres for each EME he policy rae (Figure 2), he exchange rae (Figure 3) and long-erm ineres raes (Figure 4) for he Q Q and he Q Q periods. The auocorrelaions have In fac, Sock and Wason (1999) have referred o (1) as a dynamic model. Specifically, for Chile, Peru and China, he SBIC calls for a lag greaer ha 1. The issue is ha, given our ime series shor lengh, increasing he lag migh no be feasible because of he large number of coefficiens ha would need o be esimaed. Adding lags makes esimaed coefficiens grow exponenially, and adding componens makes hem grow linearly; efficiency calls for he laer approach. Thus, o miigae he problem, we increase he number of componens, which ends o reduce he lag indicaed by SBIC. Such order migh be beer suied o some EMEs han ohers. We noneheless keep he same order o mainain comparabiliy across EMEs. BIS Papers No

14 lags and leads of up o four quarers. These saisics provide us wih iniial clues on possible changes in he relaionships beween he variables. 18 Auocorrelaions of he US federal funds rae wih each EME s policy rae 1 Figure 2 1 The maximum lead and lag ime is four quarers. The blue (dashed) lines correspond o he Q Q period, he red (solid) lines o he Q Q period. Firs, he auocorrelaions wih he policy raes (which end o peak when he lag is se o zero) sugges ha some economies, including Brazil, could have experienced a change in such relaionships beween he wo periods. Second, hey also indicae ha some economies, including Colombia and Mexico, migh have mainained such relaionships over he wo periods. Second, he auocorrelaions wih he exchange rae are generally smaller in magniude han hose wih policy raes. In he Q Q period, heir values ended o be less sable hrough he leads and lags. Ineresingly enough, some of he auocorrelaions in his period share a similar paern in several economies, peaking a around wo lags and hen dropping owards he zero lag mark They of course capure uncondiional momens. This is noable for Chile, Colombia, Mexico and Peru; Indonesia and Thailand; he Czech Republic and Poland; and Israel and Souh Africa. 376 BIS Papers No 78

15 Auocorrelaions of he US federal funds rae wih variaions in each EME s exchange rae 1 Figure 3 1 The maximum lead and lag ime is four quarers. The blue (dashed) lines correspond o he Q Q period, he red (solid) lines o he Q Q period. Third, he auocorrelaions wih he long-erm raes on average increased in he second period, eg for he Czech Republic, Indonesia, Korea, Mexico, Poland and Souh Africa. In conras, wo economies have seen hese auocorrelaions decrease: India and, excep for he zero lag, Thailand. As a nex sep, we analyse he auocorrelaions beween he Wu and Xia rae and he hree variables: he policy rae (Figure 5), he exchange rae (Figure 6) and long-run ineres raes (Figure 7). Noe ha for he firs period, Q Q2 2008, he auocorrelaions coincide wih hose in our las exercise, as he US federal funds rae and he Wu and Xia rae are essenially he same during he period. Firs, regarding he auocorrelaions wih he policy raes, hose wih he Wu and Xia rae have decreased in magniude. This suggess ha he policy raes channel could have los some imporance for he second period, Q Q Second, he auocorrelaions wih he exchange raes do no have a sable paern in he second period. Third, he auocorrelaions wih he long-run raes show an overall increase in value, boh for he leads and lags; he wo excepions are India and Thailand, which seem o have swapped signs. This resul suggess ha he long-run rae channel migh have gained a greaer role in he period afer he crisis. BIS Papers No

16 Auocorrelaions of he US federal funds rae wih each EME s long-run ineres rae 1 Figure 4 1 The maximum lead and lag ime is four quarers. The blue (dashed) lines correspond o he Q Q period, he red (solid) lines o he Q Q period. Auocorrelaions of he Wu and Xia rae wih each EME s policy rae 1 Figure 5 1 The maximum lead and lag ime is four quarers. The blue (dashed) lines correspond o he Q Q period, he red (solid) lines o he Q Q period. See Wu and Xia (2013) for he Wu and Xia rae. 378 BIS Papers No 78

17 Auocorrelaions of he Wu and Xia rae wih variaions in each EME s exchange rae 1 Figure 6 1 The maximum lead and lag ime is four quarers. The blue (dashed) lines correspond o he Q Q period, he red (solid) lines o he Q Q period. See Wu and Xia (2013) for he Wu and Xia rae. Auocorrelaions of he Wu and Xia rae wih each EME s long-run ineres rae 1 Figure 7 1 The maximum lead and lag ime is four quarers. The blue (dashed) lines correspond o he Q Q period, he red (solid) lines o he Q Q period. See Wu and Xia (2013) for he Wu and Xia rae. BIS Papers No

18 In sum, alhough auocorrelaions are only a firs sep, hey provide some iniial evidence ha he erupion of he global financial crisis and he policy responses could have affeced o some exen he channels by which US moneary policy shocks are ransmied o EMEs. Linear regression models As a nex sep, we assess he following linear regression models. Specifically, we posi he following hree benchmark regressions for each of he corresponding economies: 20 i = β + β c + β ff + e dfx = β + β c + β ff + e lri = β + β c + β ff + e , , , where, as above, c is a vecor wih he firs hree principal componens of he EME s macroeconomic variables, ff is he US federal funds rae, β.,0 are consans and e., are he error erms. We separaely esimae hem for Q Q and for Q Q Our focus is on he saisics of β.,2 beween hose wo periods, in paricular analysing he associaed -saisics and 2 R (Table 2). The firs regression can be seen as a rule ha is more general han he Taylor rule. We depar from he Taylor rule for wo reasons. Firs, a given counry may no sricly follow a Taylor rule o se is policy rae. Second, in he same vein, using principal componens allows us o have a model ha is similar across economies and he variables of ineres, and hus allowing us o make closer comparisons. For he policy rae, he resuls are as follows. For hree EMEs, he coefficiens los heir saisical significance; for four ohers, hey aained saisical significance; for he res, he coefficiens mainained heir saus. Moreover, while he explained variance generally increased beween periods, he -saisics decreased on average. In sum, he relevance of he policy rae channel seems o depend on he economy in quesion. Second, he coefficiens associaed wih he exchange rae channel are in general no saisically significan in eiher period. The excepions o his seem o be he coefficien for China, which mainained saisical significance; for Indonesia and Poland, which aained saisical significance; and for Souh Africa, which has swapped is sign. Hence, his es indicaes ha, in he majoriy of cases, he exchange rae channel does no seem o have had a srucural change. Third, for long-erm raes, hree economies (ou of he eigh EMEs for which we have he requisie rae daa) have a significan coefficien in he firs period India, Mexico and Thailand. Mexico and India lose saisical significance in he second period, while Thailand mainains i. Souh Africa s coefficien aains saisical significance in he second period. For he period in which he US federal funds rae his he zero lower bound, we consider he following model: (3) 20 The long-erm rae model is esimaed only for he eigh economies for which we have daa on long-erm raes. 380 BIS Papers No 78

19 i = γ + γ c + γ xx + ε , dfx = γ + γ c + γ xx + ε , lri = γ + γ c + γ xx + ε , (4) where, insead of he US federal funds rae, we use he Wu and Xia rae, denoed by xx. Hence, our focus is now on esimaes γ.,2. We repor he associaed -saisics 2 and he R in Table 3. Coefficien saisics for he US federal funds rae 1 Table 2 EME macroeconomic variable and regression period Policy Rae Exchange Rae Long-run ineres rae Period 1 Period 2 Period 1 Period 2 Period 1 Period 2 Brazil R Chile R Colombia R Mexico R Peru R China R Hong Kong SAR R India R Indonesia R Souh Korea R Thailand R Czech Republic R Poland R Israel R Souh Africa R For equaion (3). Period 1 = Q Q Period 2 = Q Q T-saisics are for he coefficiens β ; hose in bold.,2 are for saisically significan coefficiens a a 90% confidence level. BIS Papers No

20 For he policy rae channel, he coefficien for only wo economies Indonesia and Souh Africa aained saisical significance. For he res of he EMEs, he coefficien eiher los or mainained significance. Coefficien saisics for he Wu and Xia rae 1 Table 3 EME macroeconomic variable and regression period Policy Rae Exchange Rae Long-run ineres rae Period 1 Period 2 Period 1 Period 2 Period 1 Period 2 Brazil R Chile R Colombia R Mexico R Peru R China R Hong Kong SAR R India R Indonesia R Souh Korea R Thailand R Czech Republic R Poland R Israel R Souh Africa R For equaion (4). Period 1 = Q Q Period 2 = Q Q T-saisics are for he coefficiens γ.,2 are for saisically significan coefficiens a a 90% confidence level. See Wu and Xia (2013) for he Wu and Xia rae. ; hose in bold 382 BIS Papers No 78

21 For he exchange rae channel, only wo EMEs have a saisically significan coefficien in he firs period China, which aained i, and Colombia, which mainained i. Souh Africa s coefficien los is significance beween he wo periods. In hese exercises, he majoriy of EMEs do no presen srong evidence of a srucural change in he exchange rae channel. For he long-erm rae channel, five of he eigh EMEs for which we have daa gained a significan coefficien in he second period. This performance conrass somewha wih he resuls obained wih he US federal funds rae. The resul, however, does no sricly hold in he ess we perform, as discussed nex. Tess The hypoheses we es are posied in erms of linear resricions in (2), as follows: H : a 0. A direc es of no srucural change in he policy rae channel i, = ff d H : a 0. A direc es of no srucural change in he exchange rae channel dfx, = ff d H : a 0. A direc es of no srucural change in he long-erm rae channel lri, = ff d H : a 0 and a lri, = 0 ff d. A join direc es of no srucural change in he policy rae, exchange rae, and long-run rae channels (he join all raes es ) i, = ff d H : a 0 for all rows r. An indirec es of no srucural change in any 5. 0 r, ff = d channel (he join all variables es ). This es involves oher channels whose mechanisms are, in our view, embedded in he principal componens. By consrucion, ess 4 and 5 are more sringen han ess 1, 2 and For he US policy rae, we run he ess using he US federal funds rae and, on a separae se of FAVARs, he Wu and Xia rae. The laer accouns for he possible effecs of unconvenional moneary policy and allows us o compare he effecs of solely radiional moneary policy wih hose of unconvenional policy. We presen our esimaes in ha order. The p-value resuls of our ess (Table 4) show ha, firs, eigh EMEs rejec null hypohesis 1 ( a, = 0 ), wo a a 10% confidence level and six a a 5% confidence i ff d level, providing some evidence of a change in he policy rae channel. Second, seven EMEs rejec null hypohesis 2 ( a dfx, = 0 ff d ), one a a 10% confidence level and six a a 5% confidence level, which indicaes a possible change in he exchange rae channel. Third, for he long-run rae channel, four of eigh EMEs rejec null hypohesis 3 a a 5% confidence level. 21 Under he null hypohesis, all of he hem have an F disribuion wih (i) he number of resricions and (ii) he number of observaions minus he number of parameers being esimaed as degrees of freedom. These ess are invarian o he specific Cholesky decomposiion one ops o use. BIS Papers No

22 Fourh, 11 EMEs rejec null hypohesis 4 (he join all raes es), one a a 10% confidence level and en a a 5% confidence level. Ineresingly enough, some economies, including he Czech Republic and Indonesia, rejec some individual ess bu fail o rejec he join all raes es. Conversely, some ohers, including Poland, fail o rejec rae ess 1, 2 and 3 bu rejec he join all raes ess. These resuls are a consequence of he behaviour of he confidence regions in he join ess compared wih ha of he confidence inervals in he individual ess. P-values of es resuls for he US federal funds rae 1 Table 4 Policy Rae Exchange Rae EME macroeconomic variable Long-run ineres rae Join all raes Joins all variables Brazil Chile Colombia Mexico Peru China Hong Kong SAR India Indonesia Souh Korea Thailand Czech Republic Poland Israel Souh Africa For he F-ess of he linear resricions se on each FAVAR, as in equaion (2). Darkes shading indicaes a p-value of less han 0.05; ligher shading, a p-value beween 0.05 and 0.1. Fifh, all EMEs rejec null hypohesis 5 (he join all variables es) a a 5% confidence level. Our presumpion is ha hese ess capure oher channels we are no measuring direcly. Hence, we ake hese resuls as evidence of a possible srucural change in channels we are no explicily modelling bu ha we are noneheless capuring. In sum, wheher an EME experienced a srucural change in any of he hree channels we are assessing depends on he EME in quesion. We nex consider he same baery of ess using he Wu and Xia raes in place of he US federal funds rae (Table 5). Firs, for he policy rae, five EMEs fail o rejec null hypohesis 1, one a a 10% confidence level and four a a 5% confidence level. Thus, wheher a policy rae channel changed or no depends on he EME in quesion. Second, for he exchange rae, four EMEs rejec null hypohesis 2, one a a 10% confidence level and hree a a 5% confidence level. Here oo, alhough less 384 BIS Papers No 78

23 han one hird of EMEs rejec his es, i seems ha a change in he channel generally depends on he specific economy being considered. Third, of he eigh EMEs esed for he long-run rae channel, hree rejec null hypohesis 3, one (Thailand) a a 10% confidence level and wo (India and Korea) a a 5% confidence level. Again, hese resuls indicae a counry-dependen change. P-values of es resuls for he Wu and Xia rae 1 Table 5 Policy Rae Exchange Rae EME macroeconomic variable Long-run ineres rae Join all raes Joins all variables Brazil Chile Colombia Mexico Peru China Hong Kong SAR India Indonesia Souh Korea Thailand Czech Republic Poland Israel Souh Africa For he F-ess of he linear resricions se on each FAVAR, as in equaion (2). Darkes shading indicaes a p-value of less han 0.05; ligher shading, a p-value beween 0.05 and 0.1. See Wu and Xia (2013) for he Wu and Xia rae. Fourh, for he join all raes es, eigh EMEs rejec null hypohesis 4, wo a a 10% confidence level and six a a 5% confidence level. Fifh, for he join all variables es, all EMEs excep Poland rejec null hypohesis 5. Again, his resul underscores he possible srucural changes in wha we hink are some of he channels we have no modelled explicily. In sum, our ess provide some evidence ha a share of economies seem o have experienced a change in heir policy rae, exchange rae, or long-run ineres rae channels. Moreover, evidence of possible srucural changes appears for channels ha we are no explicily modelling. Anoher plausible inerpreaion is ha such changes are second-round effecs of he respecive channels. Tha is, variaions in he US federal funds rae affec he US componens, which in urn affec EMEs hrough one or more of he hree channels we es. Imporanly, he resuls for he Wu and Xia rae do no markedly differ from hose for he US federal funds rae. This conrass wih he es of he linear regression models. BIS Papers No

24 Impulse-Response Funcions The ess jus presened are a saisical way o assess evidence of a possible srucural change in he seleced ransmission channels. The VAR s IRFs provide an addiional perspecive: They show possible changes in he pahs and magniudes of he complee variables beween he periods we have analysed. And, of course, IRFs capure variaions above and beyond changes in each of he specific coefficiens we have associaed wih each channel. There are a leas wo key IRF feaures o wach. The firs is he relaionship beween he responses of he policy rae and hose of he exchange rae. In general, hey should mainain some degree of consisency. One can hink of his consisency in erms of uncovered ineres rae pariy, as explained in furher deail below. Second, and perhaps mos imporan, is he comparison of he response dynamics under (i) he assumpion in model (1) of no srucural change afer he crisis and (ii) he assumpion in model (2) of a possible srucural change afer he crisis, ha is, in he Q Q period. Impulse response funcions of each EME s policy rae o a US federal funds rae shock 1 Figure 8 1 The impulse is a 25 basis poin orhogonal shock o he US federal funds rae. The blue lines are obained from model (1), which assumes no change in he policy rae channel. The red lines are obained from model (2), which allows for a change in he policy rae channel and poenially in oher channels; any such srucural changes are assumed o have occurred in he Q Q period. Confidence inervals are se a he 70% level. 386 BIS Papers No 78

25 In he case of he policy rae channel (Figure 8), he IRFs arise from an orhogonal shock of 25 basis poins o he US federal funds rae. The responses from model (1) and model (2) are boh shown. A bird s-eye view of Figure 8 suggess ha, given a posiive shock o he US federal funds rae, he policy rae response srongly depends on he counry in quesion. For insance, he magniude of he responses across EMEs varies somewha. In mos cases, he shock evenually leads o an increase in he policy rae. Impulse response funcions of variaions in each EME s exchange rae o a US federal funds rae shock 1 Figure 9 1 The impulse is a 25 basis poin orhogonal shock o he US federal funds rae. The blue lines are obained from model (1), which assumes no change in he exchange rae channel. The red lines are obained from model (2), which allows for a change in he exchange rae channel and poenially in oher channels; any such srucural changes are assumed o have occurred in he Q Q period. Confidence inervals are se a he 70% level. Firs, he difference in he effec of he US shock beween he pre-crisis and pos-crisis periods varies across some economies. Moreover, if we inerpre our confidence inervals lierally (a he 70% level), he channels for Indonesia and Poland aain saisical significance only in he second period. Second, in a some cases including Colombia, Hong Kong SAR, India, Indonesia and Thailand he IRFs sugges ha he policy rae responses became more persisen in he second period. For hese EMEs, he change in he policy rae channel seems o make a difference along all of he responses. BIS Papers No

26 Third, in conras, various EMEs appear o have mainained similar responses in each period. In sum, hese EME-dependen resuls are broadly in line wih our es resuls. Impulse response funcions of each EME s long-erm ineres rae o a US federal funds rae shock 1 Figure 10 1 The impulse is a 25 basis poin orhogonal shock o he US federal funds rae. The blue lines are obained from model (1), which assumes no change in he long-erm ineres rae channel. The red lines are obained from model (2), which allows for a change in he long-erm ineres rae channel and poenially in oher channels; any such srucural changes are assumed o have occurred in he Q Q period. Confidence inervals are se a he 70% level. Moreover, some responses do no sricly conform wih wha one could expec under he assumpion of uncovered ineres rae pariy. Specifically, he responses of Brazil and Souh Africa are negaive and saisically significan. 22 In he periods considered, hese economies have some of he larges average inflaion raes wihin our sample. Moreover, boh economies score low in he Chinn and Io index of financial openness, a facor ha could play a role in he daa and hus in our esimaions. In he case of he exchange rae channel (Figure 9), he IRFs again arise from an orhogonal shock of 25 basis poins o he US federal funds rae. The responses are 22 In addiion, heir exchange rae responses are no saisically significan. 388 BIS Papers No 78

27 from model (1), which assumes no change in he channel, and from model (2), which allows for a change in he pos-crisis period. The resuls indicae, firs, ha only a very few economies may have faced a change in heir exchange rae channel. Overall, he poenial srucural changes in he exchange rae channel do no seem very imporan. Moreover, many of he responses are no saisically significan in any of he periods. Impulse response funcions of each EME s policy rae o a Wu and Xia rae shock 1 Figure 11 1 The impulse is a 25 basis poin orhogonal shock o he Wu and Xia rae. The blue lines are obained from model (1), which assumes no change in he policy rae channel. The red lines are obained from model (2), which allows for a change in he policy rae channel and poenially in oher channels; any such srucural changes are assumed o have occurred in he Q Q period. Confidence inervals are se a he 70% level. See Wu and Xia (2013) for he Wu and Xia rae. Second, he magniude of he responses varies widely across economies. For example, he exchange rae of Hong Kong SAR essenially says pu, while Souh Africa responds wih 100 basis poins of currency depreciaion on impac. Third, only hree EMEs undergo a saisically significan depreciaion of he currency a some poin. Moreover, in some economies, he possible srucural change in he exchange rae channel has implicaions for heir responses o he shock, bu only afer several quarers have elapsed. In general, he prevalen exchange rae arrangemen and moneary policy regime are relevan in explaining he effecs we have measured wih he IRFs. For example, Hong Kong SAR, which anchors is currency o he US dollar, experiences only a negligible variaion in is exchange rae. In conras, inflaion argeers end BIS Papers No

28 o have more ample and lagged variaions in heir exchange rae responses, as he exchange raes serve as buffers o exernal shocks. For he IRFs involving he long-run rae, he evidence of a change in he relaed channel is no conclusive (Figure 10). In general, he pahs are mainained. Moving on o he IRFs ha use he Wu and Xia rae in place of he US federal funds rae (Figure 11), we find ha, firs, he evidence of a possible srucural change in he policy rae channel depends on he EME in quesion. For mos cases, unconvenional moneary policies seem o aenuae he differences in he responses before and afer he crisis. Impulse response funcions of variaions in each EME s exchange rae o a Wu and Xia rae shock 1 Figure 12 1 The impulse is a 25 basis poin orhogonal shock o he Wu and Xia rae. The blue lines are obained from model (1), which assumes no change in he exchange rae channel. The red lines are obained from model (2), which allows for a change in he exchange rae channel and poenially in oher channels; any such srucural changes are assumed o have occurred in he Q Q period. Confidence inervals are se a he 70% level. See Wu and Xia (2013) for he Wu and Xia rae. Second, excep for a few cases, he exchange rae channel does no srucurally change for mos EMEs. The responses o shocks o he Wu and Xia rae are in general very similar, and in some cases he response s pahs are idenical (Figure 12). Third, in he case of he long-run rae channels, srucural changes in he periods considered (Figure 13) are also counry dependen. This is broadly in line 390 BIS Papers No 78

29 wih he resuls we have previously shown, in paricular in Table 3. Two economies Indonesia and, especially, Souh Africa have, on impac, a sronger response; and surprisingly, he effecs have he opposie sign from he one expeced. We se all IRFs confidence inervals a a 70% level. Had we considered a higher level, confidence inervals would have been much wider, and evidence of no srucural change in he ransmission channels would have been sronger. This commen is applicable o all he IRFs we consider in his paper. Thus, he ess are more likely o rejec he null hypohesis of no change, whereas he confidence inervals in he IRFs are more likely o be supporive of he null hypohesis. The IRF confidence inervals are no direcly comparable o hose of he ess because he former capure he uncerainy associaed wih he shocks o he economies, and he laer he uncerainy relaed o he esimaed coefficiens. Thus, in his vein, he ess on he linear regression and he FAVAR coefficiens are more comparable o each oher han eiher is o he IRFs. Impulse response funcions of each EME s long-erm ineres rae o a Wu and Xia rae shock 1 Figure 13 1 The impulse is a 25 basis poin orhogonal shock o he Wu and Xia rae. The blue lines are obained from model (1), which assumes no change in he long-erm ineres rae channel. The red lines are obained from model (2), which allows for a change in he long-erm ineres rae channel and poenially in oher channels; any such srucural changes are assumed o have occurred in he Q Q period. Confidence inervals are se a he 70% level. See Wu and Xia (2013) for he Wu and Xia rae. Moreover, i is relevan o assess he degree of consisency beween he policy rae responses and he exchange rae responses and he exen o which i migh have changed afer he hird quarer of As menioned, i is useful o consider BIS Papers No