Global Trade and GDP Co-Movement

K.7 Global Trade and GDP Co-Movement de Soyres, François and Alexandre Gaillard Please cite paper as: de Soyres, François and Alexandre Gaillard (2020). Global Trade and GDP Co-Movement. International Finance Discussion Papers 1282. https://doi.org/10.17016/IFDP.2020.1282 International Finance Discussion Papers Board of Governors of the Federal Reserve System Number 1282 May 2020

Board of Governors of the Federal Reserve System International Finance Discussion Papers Number 1282 May 2020 Global Trade and GDP Co-Movement François de Soyres and Alexandre Gaillard NOTE: International Finance Discussion Papers (IFDPs) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the International Finance Discussion Papers Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers. Recent IFDPs are available on the Web at www.federalreserve.gov/pubs/ifdp/. This paper can be downloaded without charge from the Social Science Research Network electronic library at www.ssrn.com.

lobal rade and o movement G T GDP C - * F rançois de S oyres† A lexandre G aillard‡ Federal Reserve Board Toulouse School of Economics 2020 This version: January Abstract We revisit the association between trade and GDP comovement for 135 countries from 1970to2009. Guidedbyasimpletheory,weintroducetwonotionsoftradelinkages: (i)the usual direct bilateral trade index and (ii) new indexes of common exposure to third countriescapturingtheroleofsimilarityintradenetworks. Bothmeasuresareeconomicallyand statistically associated with GDP correlation, suggesting an additional channel through which GDP fluctuations propagate through trade linkages. Moreover, high income countries become more synchronized when the content of their trade is tilted toward inputs while trade in final goods is key for low income countries. Finally, we present evidence that the density of the international trade network is associated with an amplification of the association between global trade flows and bilateral GDP comovement, leading to a significantevolutionofthetradecomovementslopeoverthelasttwodecades. Keywords: International trade, international business cycle comovement, networks, inputoutput linkages JEL Classification: F 15 , F 44 , F 62 *Wethankthe2020WorldDevelopmentReportteamaswellasseminarandconferenceparticipantsforhelpful comments. The views in this paper are solely the responsibility of the authors and should not necessarily be interpretedasreflectingtheviewsoftheBoardofGovernorsoftheFederalReserveSystemorofanyotherperson associatedwiththeFederalReserveSystem. †Email: francois.m.desoyres@frb.gov; Corresponding author. Address: Board of Governors of the Federal ReserveSystem,2051ConstitutionAvenueNW,Washington,DC ‡Email: alexandre.gaillard@tse-fr.eu. 1

1 Introduction Over the past decades, both import and export flows have increased much faster than GDP for almost all countries in the world. This march toward more open economies has been accompanied by a reorganisation of he world’s production across different locations, with both trade in intermediate inputs and in final goods trade representing an increasing share of 1970 worldGDP,nowreachingaroundthreetimestheshareobservedinthe s. Invalued-added 5 terms, trade increased at an average annual growth rate of more than percent during the 1990 2009 - period, with the share of trade in intermediate inputs roughly constant at around 70 % of total trade. During the same period, the average GDP co-movement across all pairs of 6 38 countries rose from % to %. The general surge in trade-over-GDP also implies more complex patterns for international propagation: when two countries are increasingly connected to the same direct or indirect trade partners, the associated surge in "third country" exposure can create systemic interdependence that operates over and above direct trade linkages. The consequences of these changes in trade patterns for the synchronization of economic activity are an important is- 1 sue because they can have implications for macroeconomic policies. In light of these global trends, several questions arise: did the rise of Global Value Chains (GVCs) have a specific effect on the correlation of GDP and its association with both direct and indirect trade flows? Didtheriseinproductionfragmentationhavethesameeffectacrossincomegroups? Aredirect trade linkages more important than common exposure to third markets? Did the sensitivity of GDP co-movement to an increase in bilateral trade flows evolve over time? 1998 Since the seminal paper by Frankel and Rose ( ), hereafter FR, a large empirical literaturehasstudiedthedeterminantsofcross-countrybusinesscycleco-movement, showingthat bilateraltradeisanimportantandrobustelementassociatedwithchangesinGDPcorrelation whilemeasuresoffinanciallinkagesorcountries’sectoralsimilarityarenotstatisticallyassoci- 2 atedwithhigherbilateralsynchronization. Inthispaperwere-assesstheassociationbetween 135 global trade and cross-country business cycle correlation using a large sample of coun- 1970 2009 tries from to , including high and low income countries. Using constructed panel data and controlling for both observed and unobserved heterogeneity between countries and 1 Forexample,theextenttowhichtheEuroZonecanbeconsideredasanoptimalcurrencyarea(and,therefore a common monetary policy could be optimal) largely depends on the synchrony of business cycles among the membercountries. 2 Among many others, see Frankel and Rose (1998), Clark and van Wincoop (2001), Imbs (2004), Baxter and Kouparitsas(2005),Calderonetal.(2007),Inklaaretal.(2008),DiGiovanniandLevchenko(2010),Ng(2010),Liao and Santacreu (2015), di Giovanni et al. (2016) and Duval et al. (2015). The literature mostly focused on high incomecountries,withthenotableexceptionofCalderonetal.(2007),andsetupestimationequationsthatunveil asingletime-invariantvaluefortheassociationbetweenbilateraltradeflowsandbusinesscyclecorrelation. 2

over time, we estimate the trade co-movement slope (TC-slope) across different income groups and unveil a series of new determinants of GDP co-movement, including the different role of the content of trade flows for each income group as well as the presence of network effects and how they interact with bilateral proximity. Moreover, we also uncover important time variations in the TC-slope, which suggests that the sensitivity of GDP correlation to changes in trade proximity is not akin to a time-invariant deep parameter but is a function of other elements that evolve over time. Building on earlier literature, this paper makes several contributions. First, starting with the role of bilateral trade flows, we update previous analysis by separating trade flows into trade in intermediate inputs and trade in final goods and investigate separately their specific role forGDPsynchronizationforhighandlowincomecountries. AsshownindeSoyresandGaillard ( 2019 ) and confirmed in this paper, trade in intermediate inputs plays a particular role in the TC-slope for OECD countries. However, this finding is complemented and nuanced here by a novel insight regarding low income countries. Using only within country-pair variations and controlling for several factors including changes in the similarity of industrial structure across country pairs, we show that economies at the lower end of the income distribution experience an increase in the correlation of their GDP with their trade partners when the contentoftheirtradeflowsismoretiltedtowardfinalgoodstrade. Tounderstandthisdifference, we use disaggregated trade data and show that country-pairs with a large TC-slope in intermediate inputs are also characterized by high proximity in the sectoral composition of their trade flows. All told, our analysis suggests that trade in inputs is associated with higher GDP 3 correlation when countries have a similar industrial structure. Second, guided by recent debates on the role of Global Value Chains and the systemic interdependence that can arise from worldwide input-output linkages, we move beyond bilateral trade linkages and construct new indices of network proximity for all country pairs. We argue that changes in GDP synchronization between two countries can be the result of an increased common exposure to third markets, which can happen either at the first order when two countries have similar trade partners or at the second order when countries’ direct partners have similar partners. On the whole, our results reveal that first order common exposure is particularly strong for high-income countries, while second-order proximity, a measure of more indirect propagation, is more prevalent for low income economies. Moreover, we show theoretically and empirically that the marginal increase in GDP comovement associated with the increase in any trade link is itself increasing in the overall density of the network. As such, this amplification aspect linked with overall network density helps rationalize the wide array 3 To the extend that such similarity is in turn associated with a higher degree of input specificity, then this findingisfullyconsistentwithresultsinBarrotandSauvagnat(2016). 3

of TC-slopes found in the literature since any estimate depends on both the time and country coverage. Interestingly, this result challenges the usual assumption of a single time-invariant relationship between trade and GDP comovement. While the complementarity between networkandbilateraltradecouldrationalizeourfindingthattheTC-slopesignificantlyincreased in the last two decades, we cannot rule out the possibility that other factors weighed into this evolution. In particular, the growth of price distortion could have also have played a role. Finally,weprovidevariousrobustnesschecks,usingdifferentcontrols,measuresandsample selection. For instance, controlling for bilateral financial interconnection of the banking sector or foreign direct investment does not affect our main findings (although it reduces our sample due to data coverage). Overall, our results are robust to a wide range of specifications and trade indexes and highlight important disparities among country groups and over time. Relationship to the literature. Starting with Frankel and Rose ( 1998 ), a large number of papers have studied and confirmed the positive association between trade and GDP comove- 4 ment in the cross-section. This paper is mostly related to a few recent contributions. First, di Giovanni et al. ( 2016 ) uses a cross-section of French firms and presents evidence that international input-output linkages at the micro level are an important driver of the value added comovementobservedatthemacrolevel. Theirevidenceisinlinewiththefindingsofthispaper and supports the role of Global Value Chains in the synchronization of GDP fluctuations 5 2015 acrosscountries. Second,LiaoandSantacreu( )isthefirsttostudytheimportanceofthe extensive margin for GDP and TFP synchronization and shows that changes in the number of products traded across countries (rather than the average shipment per product) plays an im- 2019 portant role in the synchronization of GDP. Huo et al. ( ) uses a more structural approach and proposes a perfectly competitive production framework to measure technology and nontechnology shocks and subsequently analyzes their cross-sectional properties. In this setup, international transmission through trade accounts for a third of total comovement. Third, 2007 Calderon et al. ( ) investigate the relationship between trade and business cycle comovement for both developed and developing countries. Based on cross sectional estimates, they find that the impact of trade integration on business cycles is higher for industrial countries than for developing countries. Fourth, our paper is related to a recent series of papers developing accounting and theoretical frameworks to measure GVC participation, including Bems 4 Seepaperscitedforinstanceinfootnote2. 5 Relatedly,Bursteinetal.(2008)usesacrosssectionoftradeflowsbetweenUSmultinationalsandtheiraffiliates as well as trade between the United States and Mexican maquiladoras to measure production-sharing trade and its link with the business cycle. Moreover, Ng (2010) uses cross-country data from 30 countries and shows that bilateral production fragmentation has a positive effect on business cycle comovement. The concept of bilateral production fragmentation used is different from this paper as it takes into account only a subset of trade in intermediates, namely imported inputs that are then further embodied in exports. Moreover, the cross-sectional natureoftheanalysisallowsforneitherdyadicnortimewindowsfixedeffects. 4

2011 et al. ( ) and others. If the empirical association between bilateral trade and GDP comovement has long been known,theunderlyingeconomicmechanismleadingtothisrelationshipisstillunclear. Using 2006 the workhorse IRBC with three countries, Kose and Yi ( ) have shown that the model can explain at most 10 % of the slope between trade and business cycle synchronization, leading to what they called the TradeComovementPuzzle (TCP). Since then, many papers including John- 2014 2015 son ( ) or Duval et al. ( ), have refined the puzzle, highlighting different ingredients 6 that could bridge the gap between the data and the predictions of standard models. The rest of the paper is organized as follows. We first provide a simple trade network 2 model highlighting the role of trade in the global GDP-comovement. We then turn to our 3 empiricalcontribution. Section presentsthedataandthedifferentconstructedvariablesused 4 throughout the paper. Section investigates the global TC-slope not only across countries in different income groups, but also over time. We discuss the main implications of our results 5 6 insection and,insection ,testseveralpossibleexplanationsforsomeofthekeydifferences 7 between the results relative to high and low income countries. Finally, section concludes. 2 A simple trade network model To motivate our empirical work and formalize our intuition, we begin by writing a parsimonious static model of international trade with multiple countries and sectors. Our main goal is to illustrate through a series of example several mechanisms through which GDP in two countries can be correlated. In particular, we show that GDP comovement is the result of a combinationofmanyfactors,includingthecorrelationstructureofshockshittingeverycountry in the world, bilateral trade linkages between countries as well as their indirect exposure totherestofthetradenetwork,andtheassociationbetweengrossoutputandGDPwhichcan 7 be time varying. For simplicity, our framework abstracts from other relevant considerations such as the presence of financial linkages or the possibility of common (or coordinated) monetarypolicy. Note,however,thatwewillcontrolfortheseandotherelementsinourempirical 6 For a quantitative solution to the Trade Comovement Puzzle, see de Soyres and Gaillard (2019), where it is shownthatproductionlinkagesalonearenotsufficientforamacromodeltodeliveratradeco-movementslopein linewiththedata. 7 As discussed in Johnson (2014), comovement in intermediate input, and the resulting comovement in gross output,doesnotnecessarilytranslateintorealvalue-addedcomovement. Buildingonthisinsight,deSoyresand Gaillard(2019)showsthattheintroductionofmarkupsandextensivemarginadjustmentscancreateamechanical linkbetweeninputcorrelationandGDPcorrelation. Wesimplifythediscussionherebyintroducingasimplead hocproportionaltransformationbetweenoutputandrealvalue-addedthatillustratesthefactthatthesensitivity of GDP comovement to trade proximity is a function of other elements – which could include the prevalence of markupsforexample. 5

investigations in subsequent sections. 2.1 Basic setup Production and pricing. Consider a world with many countries (i,j ∈ {1,...,N}) and many sectors(s,s(cid:48) ∈ {1,...,S}). Incountryiandsectors,grossoutputistheresultofaCobbDouglas combinationofthreemainelements: ( 1 )anexogenoustechnologyshock(Z ),( 2 )intermediate i,s inputs from all other sector-countries in the world (X j,s(cid:48) ), and ( 3 ) inelastic domestic factors of i,s production (L ). i,s (cid:32) (cid:33) Y i,s = Z i,s · ∏ (X i j , , s s(cid:48) )α i j , , s s(cid:48) ·L i γ ,s i,s, ( 1 ) j,s(cid:48) with ∑ α j,s(cid:48) +γ = 1. The production cost of a representative firm in each country i and j,s(cid:48) i,s i,s sector s is a function of the price charged by its input suppliers and the suppliers of its suppliers. For simplicity we assume that there are no trade costs. Moreover, we also assume that firms’ markups (µ ) are completely exogenous and independent of the destination market i,s which further implies that prices are equal across all destination markets. Denoting p the i,s priceofoutputproducedbycountry-sector(i,s)andw thepriceofdomesticfactorincountry i i, standard cost minimization conditions imply that the price in (i,s) is given by: p i,s = µ i,s ·MC i,s = µ i,s · Z c i,s ×w i γi,s · ∏ (p j,s(cid:48) )α i j , , s s(cid:48) ( 2 ) i,s j,s(cid:48) With MC the marginal cost in (i,s) and c a constant depending only on parameters. 8 As i,s i,s is usual in all models with input-output linkages, the price in a given sector-country is a direct function of all other prices in the economy. To simplify notation, we stack prices in all countries and sectors into an (N×S,1) vector P, where the first S rows contain the prices of all sectors in country 1 , subsequent S rows contain all prices in country 2 , etc... Taking the Ω log and denoting by the cross-country input-output matrix of the economy, prices are the 8 Thevariablec i,s isdefinedas: c i,s =γ i − ,s γi,s∏ j,s(cid:48)α i j , , s s(cid:48)−α i j , , s s(cid:48) 6

9 solution of a simple linear system:   k −log(Z )+γ log(w ) 1,1 1,1 1,1 1 P = (I NS −Ω)−1   . . .   ( 3 )   k −log(Z )+γ log(w ) N,S N,S N,S N Clearing conditions. Gross output is used both as an intermediate input in production and to produce a composite final good used by consumers. With Cobb Douglas production function, the representative firm in country j and sector s(cid:48) spends a fraction α j,s(cid:48) on goods i,s coming from (i,s), so that: p i,s X i j , , s s(cid:48) = α i j , , s s(cid:48) p j,s(cid:48)Y j,s(cid:48) , for all i,j,s,s (cid:48) ( 4 ) 10 j Aggregatedemandineachcountry jisdenotedby D . Country jaddressesafraction β j i,s of its total demand to country-sector (i,s), so that market clearing in the final goods market can be written as: p y j = β j D , for all i,j,s ( 5 ) i,s i,s i,s j where y j is the amount of good produced in (i,s) that are absorbed as final demand in i,s country j. We store all shares β j into a (NS,N) matrix B where each row corresponds to i,s a sector-country (i,s) and the columns correspond to all countries. 11 Finally, the resource constraint condition is given by Y = ∑ y j + ∑ X j,s(cid:48) , for all i,s ( 6 ) i,s i,s i,s j j,s(cid:48) 9 We denote k = log(µ ·c ) and the Input-Output matrix Ω can be defined by block using country-pair i,s i,s i,s input-outputmatricesΩ as: i,j  Ω Ω ...  1,1 1,2 Ω=  . . . . . . . . .   , with (Ω i,j ) s,s(cid:48) =α i j , , s s(cid:48) Ω ... Ω N,1 N,N 10 A natural general equilibrium closing of the model would be to assume that total demand D equals total i income of domestic production factor w L as well as domestic profits. We keep things more general here and i i solveforgrossoutputforanyleveloffinaldemand, whichmakesitpossible, inprinciple, tostudybothsupply shocks(throughshockstotechnologyZ )aswellasdemandshocksifweretointroduceshockstoD. i,s i 11 MatrixBisdefinedas:  β1 β2 ... βN  1,1 1,1 1,1  β1 β2 ... βN   1,2 1,2 1,2  B= . . . .   . . . .   . . . .  β1 β2 ... βN N,S N,S N,S 7

4 5 6 Combining ( ), ( ) and ( ), we can solve for nominal output in each country and sector:     p Y D 1,1 1,1 1   . . .   = (cid:16) I NS − (cid:16) ΩT (cid:17)(cid:17)−1 ·B·   . . .   ( 7 )     (cid:124) (cid:123)(cid:122) (cid:125) p Y D N,S N,S =T N In this stylized framework, solving for gross output in each sector-country amounts to jointly 3 7 solving for prices using ( ) and nominal output using ( ). Defining Real Value Added. Measuring real value added in this framework is not straightforward. Statisticalagenciesmeasurerealvalueaddedineachsectorasthedifferencebetween gross output and intermediate input, measured using base period prices. As discussed in Ke- 2008 2014 hoe and Ruhl ( ) or in Johnson ( ), in a perfect competition setting, this procedure amounts to measuring changes in domestic factor supply (i.e. changes in labor L ). Hence, i,s withoutmarkups,ourassumptionthatdomesticfactorsarecompletelyinelasticwouldleadto constantmeasuredrealvalueadded. However,BasuandFernald( 2002 ),deSoyresandGaillard 2019 ( ) and others note that things differ markedly when one introduces markups. By introducing a wedge between marginal cost and marginal revenue product of inputs, the presence of markups creates a proportional relationship between gross output and profits fluctuations. In such a case, even with inelastic domestic factor supply, real value added can still fluctuate owing to movements in profits. We parsimoniously account for such a channel by positing a reduced form relationship between gross outputY = ∑ Y and measured real GDP, so that RGDP = ∑ L +κ Y. With i s i,s i s i,s i i fixed domestic factor supply, changes in real GDP come only from gross output fluctuations. In the rest of this section, we show how correlation of gross output fluctuations can emerge from a variety of different channels, which we then formally test in the rest of the paper. 2.2 Propagation channels Considering technology as the only source of shocks, the proportional change in gross output in any country-sector is a function of the vector of shocks and the Leontieff inverse: Y(cid:98) = [I−Ω]−1 Z(cid:98) ( 8 ) Intherestofthissection,wepresentstylizedexampleswithspecificInput-Outputmatricesto illustrate several determinants of bilateral comovement. In particular, we show that, absent of anybilateraltradebetweentwocountries,and,indeed,eveninsituationswheretwocountries 8

donotexportatalltheglobaltradenetworkcouldgiverisetoendogenousoutputcorrelation. Consider a world with six countries and only one sector per country. We choose a specific structure of input-output linkages in order to show how (i) bilateral trade, (ii) direct common trade exposure, and (iii) indirect common trade exposure all play a role in bilateral output 1 (and ultimately GDP) comovement. The structure is described in figure and the associated Ω matrix is given by:   0 α2 α3 α4 0 0 1 1 1    α1 0 α3 0 α5 0   2 2 2    Ω =   0 0 0 0 0 0    0 0 0 0 0 α6   4     0 0 0 0 0 α6   5  0 0 0 0 0 0 Figure1. Networkrepresentationofinput-outputlinkages 3 1 2 4 5 6 8 1 Using equation ( ), we can write the proportional change in gross output in country as 12 a function of all shocks and trade linkages: 1 (cid:16) (cid:17) Y(cid:98) 1 = | Ω | Z(cid:99)1 +α2 1 Z(cid:99)2 +(α3 1 +α2 1 α3 2 )Z(cid:99)3 +α4 1 Z(cid:99)4 +α2 1 α5 2 Z(cid:99)5 +(α4 1 α6 4 +α2 1 α5 2 α6 5 )Z(cid:99)6 ( 9 ) 1 (cid:16) (cid:17) Y(cid:98)2 = | Ω | α1 2 Z(cid:99)1 +Z(cid:99)2 +(α3 2 +α1 2 α3 1 )Z(cid:99)3 +α1 2 α4 1 Z(cid:99)4 +α5 2 Z(cid:99)5 +(α5 2 α6 5 +α1 2 α4 1 α6 4 )Z(cid:99)6 ( 10 ) j where we recall that α is the spending share in country i on goods coming from country j. i The multifaceted effect of global trade. Let us consider the case where technology shocks 12 Thevariable|Ω|isthedeterminantofmatrixΩ 9

are uncorrelated, so that Cov(Z,Z ) = 0 for all i and j. In such a case, correlation between i j Y(cid:98) 1 and Y(cid:98)2 is solely due to global trade linkages. Using equations ( 9 ) and ( 10 ), we can write a simple expression for corr(Y(cid:98) 1 ,Y(cid:98)2 ): 13 (cid:18) corr(Y(cid:98) 1 ,Y(cid:98)2 ) = λ α2 1 +α1 2 +(α3 1 +α2 1 α3 2 )(α3 2 +α1 2 α3 1 )+α1 2 (α4 1 )2+α2 1 (α5 2 )2 ( 11 ) (cid:19) +(α4α6+α2α5α6)(α5α6+α1α4α6) 1 4 1 2 5 2 5 2 1 4 11 Equation ( ) reveals that several types of trade linkages can give rise to endogenous output co-movement. We will now examine three cases that provide economic intuition for the empirical exercise we perform in the next sections. 1 . Only bilateral trade. Consider a situation where countries 1 and 2 both import inputs from one-another but do not trade with the rest of the world. In other words, α2 and 1 α1 are strictly positive but all other input-output shares are zero. Following ( 11 ), the 2 correlation between Y(cid:98) 1 and Y(cid:98)2 is simply a function of bilateral trade shares: (cid:16) (cid:17) corr(Y(cid:98) 1 ,Y(cid:98)2 ) = λ α2 1 +α1 2 ( 12 ) 2 . Nobilateraltradeandonlyfirstorder“thirdcountry”exposure. Thissituationhappenswhen countries 1 and 2 import intermediates from country 3 (meaning that α3 and α3 are 1 2 strictly positive) but all other input-output shares are zero. This situation is interesting 1 2 because neither country nor country exports any value added to any other country. 1 2 With uncorrelated technology shocks, the only reason countries and co-move is that 3 11 they are commonly exposed to country . Using ( ), we then get a simple expression for bilateral correlation of output: corr(Y(cid:98) 1 ,Y(cid:98)2 ) = λ (cid:0) α3 1 α3 2 (cid:1) ( 13 ) 3 . Nobilateraltradeandonlysecondorder“thirdcountry”exposure. In this case, the only trade 6 4 5 flows are as follows: country exports inputs to countries and , which themselves 1 2 1 2 export to countries and respectively. In such a configuration, countries and have neither bilateral ties nor any first order network proximity since there is no overlap 6 between their trade partners. However, they are both indirectly exposed to country . 13 whereλisdefinedby (cid:18) (cid:113) (cid:19)−1 λ= |Ω|2 Var(Y(cid:99)1 )Var(Y(cid:99)2 ) 10

11 Equation ( ) then yields the following expression of bilateral output correlation: (cid:16) (cid:17) corr(Y(cid:98) 1 ,Y(cid:98)2 ) = λ α4 1 α6 4 α5 2 α6 5 ( 14 ) For simplicity, we chose here a sparse and symmetric second order network exposure, butothertypesofindirectexposurewillnaturallyariseinthedatagiventhehighdensity of the actual network of trade linkages. For instance, indirect exposure could arise if 6 1 214 country is both directly linked to country and indirectly linked to country . 12 14 An inspection of equations ( ) to ( ) reveals that GDP co-movement is the result of several 1 type of linkages, including direct bilateral trade links (case ) as well as common exposure to 2 3 third countries using first- or second-order partners (cases and , respectively). The role of network density. So far, we have considered situations where different types of linkages were analyzed in isolation from one another. In practice, these mechanisms do not operate independently, and the density of the global trade network can act as a powerful 1 amplification factor. We illustrate this point by considering a situation where countries and 2 trade with each-other (α2 and α1 are non-zero) and are also commonly exposed to country 3 1 2 (α3 and α3 are non-zero). Using equation ( 11 ), we can write bilateral output correlation as: 1 2 (cid:18) (cid:19) (cid:18) (cid:19) corr(Y(cid:98) 1 ,Y(cid:98)2 ) = λ α2 1 +α1 2 +(α3 1 +α2 1 α3 2 )(α3 2 +α1 2 α3 1 ) > λ α2 1 +α1 2 +α3 1 α3 2 ( 15 ) 15 The inequality in equation ( ) reveals that the correlation stemming from the combination of bothbilateraltradeandcommonexposureislargerthanthesumofeachchannelindividually. As such, it shows the complementarity that arises from these channels that amplify one another. Morebroadly,thestrengthofeachchannelincreaseswiththepresenceofotherlinkages inthetradenetwork,whichmeansoneshouldnotexpectthatthemarginaleffectofincreasing 1970 any given link in the sparse network of the s is the same as the effect of increasing a link in today’s network. In the empirical exercise below, we test and provide support for such amplification through network density. The role of sectoral composition. We slightly modify our setup and consider a world with only two countries and two sectors. To streamline the discussion, we also assume that there are no trade flows at all, implying that countries do not have any link with one another and technology shocks do not propagate across countries. Furthermore, technology shocks are sector specific and do not embed any country-specific component, in the sense that sectors s are hit by the same shock Z in both countries. This assumption creates a link between s 14 Formally,thiscasehappenswhenα6,α6 andα5 areallstrictlypositive. 1 5 2 11

sectoral specialization and bilateral comovement even in the absence of any trade flows. We assume these shocks are uncorrelated across both sectors and follow a distribution with a common variance σ2. Proportional changes in output in each country i and sector s are given by Y(cid:99)i,s = (cid:99)Z s . Introducing π as the share of sector s in country i’s gross output, we can write the i,s aggregate change in country i as a function of sectoral changes: Y(cid:98)i = π i,1 Y(cid:99)i,1 +π i,2 Y(cid:99)i,2 = π i,1 Z(cid:99)1 +π i,2 Z(cid:99)2 . Usingourassumptionsonshocksorthogonalityandthefactthatπ i,2 = 1−π i,1 15 for both countries, we can write the following: π π +(1−π )(1−π ) corr(Y(cid:98) 1 ,Y(cid:98)2 ) = (cid:113) 1,1 2,1 (cid:113) 1,1 2,1 ( 16 ) π2 +(1−π )2· π2 +(1−π )2 1,1 1,1 2,1 2,1 16 Fromequation( ),itisapparentthatbilateraloutputcorrelationequalszerowhenevercountries are fully specialized in different sectors, while it equals one if and only if π = π . 1,1 2,1 In other words: bilateral co-movement increases with countries’ similarity in their sectoral composition. 2.3 Key Takeaways from the Model To summarize, the model developed in this section gives rise to four testable predictions: 1 . Bilateral GDP correlation increases with direct bilateral trade. 2 . Bilateral GDP correlation increases with common exposure to third countries through direct as well as indirect linkages (this prediction is what we call the network channel). 3 . Thepreviouslymentionedchannelscomplementoneanotherinthesensethatthemarginal effect of increasing bilateral trade linkages or increasing common exposure to other countries depends on the density of the overall network of trade linkages. Owing to increasing trade linkages over the past four decades, this result also implies that both trade- and network-comovement slopes are expected to increase over time. 4 . Bilateral GDP correlation increases with the similarity of sectoral composition between two countries.. In the rest of the paper, we will test for the presence of all these channels in the data as well as other related aspects of the relationship between global trade and GDP correlation. 15 Notethatthe(common)varianceσ2disappearsfromthisequationsinceitappearsinboththenumeratorand thedenominator. 12

It is worth noting that, on top of the forces discussed in the framework developed in this section, an obvious additional source of bilateral comovement is simply the correlation of country-specific shocks. Since this source is unlikely to be fully captured by our index of sectoral similarity discussed below, we will add country-pair fixed effects in our specification that effectively control for any time invariant factor affecting bilateral correlation. 3 Data sources and construction of our main variables One of the objectives of this paper is to investigate the heterogeneity of the TC-slope across different levels of development as well as across different time periods. To be able to do so, webuildonandexpandpreviousstudiesbybroadeningbothtimeandgeographicalcoverage 40 135 andwebuildasamplecontaining yearsofdataandatotalof countries,whichaccounts for almost the totality of world trade flows and world GDP. To investigate the role of income level in the determinants of bilateral GDP correlation, we create four types of country-pairs: (i) pairs where both countries belong to the OECD, (ii) pairs where both countries are high income (defined as HH pairs) according to the World Bank definition of income group, (iii) pairs where one country is high income and the other is not (defined as HL pairs), and (iv) 16 pairs where no country is categorized as high income (defined as LL). Note that for clarity of exposition we do not separate middle and low income countries, and only investigate the differences between high income and other countries. Moreover, the first sub-sample (constructed based on OECD membership) is not informed by income level but is designed to capture possible specificities related to being part of what is usually considered as an “rich countries’ club.” Our analysis will reveal that results in the OECD and HH sub-samples turn out to be qualitatively similar but quantitatively different. As will be clear below, all of our specifications are designed to control for unobserved country-pair heterogeneity by using only within country-pair time series variations. Hence, 40 1970 2009 wedivideour yearsoftimecoverage,stretchingfrom to ,intofournon-overlapping 1 timewindows. Intable ,wereporttheshareoftotaltradeflowsofeachincomegroupinour sample, relative to total world trade flows. The extent to which countries have correlated GDP can be influenced by many factors beyond international trade, including correlated shocks, financial linkages, common monetary policies, and so on. Because those other factors can themselves be correlated with the index of trade proximity in the cross section, using cross-sectional identification could yield biased 16 Theclassificationofhigh,middleorlowincomecountriesistakenfromtheWorldBankclassification: http: //databank.worldbank.org/data/download/site-content/OGHIST.xls. 13

Table.1. Tradeflowsinthedifferentincomegroupsb Share of total trade (%) Period Total Flows a OECD HH HL LL 19701979 303 607 652 324 21 : . . . . 19801989 881 645 706 294 19 : . . . . 19901999 1864 619 643 349 26 : . . . . 20002009 3972 481 478 465 62 : . . . . a inbillionsofUSdollars. b selectedincomegroupsarenotexclusive. Somecountriesamongthe LLgroupalsoappearinOECD.Forinstance,thisisthecaseforMexico. results. Indeed, in their seminal paper, FR use cross-sectional variations to evaluate whether bilateral trade intensity correlates with business cycle synchronization, but their specification does not rule out omitted variable bias such as, for example, the fact that neighboring countries have at the same time more correlated shocks and larger trade flows. By constructing a paneldatasetandcontrollingforbothcountry-pairandtimewindowsfixedeffects,thispaper 17 relatestorecentstudiesthattrytocontrolforunobservedcharacteristics. Therefore,inorder toseparatetheeffectoftradelinkagesfromotherunobservableelements,weconstructapanel 18 dataset by creating four periods of ten years each. Within each time window, we compute GDP correlation as well as the average trade intensities defined below. 3.1 Trade Proximity and GDP-comovement GDP. We use annual GDP data from the Word Development Indicators (WDI) of the World 2010 19 Bank, measured using constant prices in US dollars. For our analysis, GDP series need to be filtered in order to extract the business cycle component from the trend. Our main andbenchmarkfilteristhestandardHodrick-Prescott(HP)filterwithasmoothingparameter 100 of which is consistent with the yearly frequency of our data. Such a transformation allows us to capture the standard business cycle fluctuations. With this setting, we mostly 8 32 6 keep fluctuations that have a frequency between and quarters. In section , we provide 20 robustness checks using a Baxter and King (BK) filter and a simple log-first difference. With 17 DiGiovanniandLevchenko(2010)includescountrypairfixedeffectsinalargecross-sectionofindustry-level datawith55countriesfrom1970to1999inordertotestfortherelationshipbetweensectoraltradeandoutput(not value-added)comovementattheindustrylevel. Duvaletal.(2015)includescountrypairfixedandyeareffectsina panelof63countriesfrom1995to2013andtesttheimportanceofvalueaddedtradeinGDPcomovement. 18 AddingtimewindowsfixedeffectscontrolsfortherecentriseofworldGDPcorrelationsincethe1990s,which couldbeunrelatedtotradeintensity. 19 Weusedthedataseriescalled“NY.GDP.MKTP.KD”. 20 We use a Baxter and King (BK) filter to isolate medium-term fluctuations in the spirit of Comin and Gertler (2006). Wekeepfluctuationsbetween32and200quarters,followingCominandGertler(2006). Asimplelog-first 14

the filtered GDP, we compute the GDP correlation for each country-pair (i,j) within each time-window t of 10 years, denoted Corr GDP . ijt Trade Proximity. We collect data on bilateral trade flows from the Observatory of Economic 215 1962 2014 Complexity(MIT).Thisdatabasecovers countriesovertheperiod - . Thedataare 4 classified according to the -digit Standard International Trade Classification (SITC), Revision 2 . Only products and commodities are considered. To classify trade flows into final goods 2 and intermediate inputs, we use a concordance table from SITC Rev. to Broad Economic Categories (BEC). 21,22 Finally, we exclude country-pairs with less than two time-windows for which trade proximity is available. We then aggregate trade flows in each category at the country-pair level. For each type of flow d ∈ {total,inter, final} (for total trade flows, trade in intermediate inputs and trade in final goods respectively) we construct an index for bilateral trade proximity of a country-pair (i,j) in a given time-window t, as follows: Td Traded = i↔j,t ∀d ∈ {total,I,F} ( 17 ) ijt GDP +GDP it jt where Td = Td +Td is total trade flows between countries i and j, defined as the sum i↔j,t i→j,t j→i,t 23 of exports from i to country j and exports from j to country i. In the result tables below, we refer to Total≡ Ttotal, Inter≡ Tinter and Final≡ Tfinal for simplicity. 3.2 Network Effects A key contribution of this paper is to provide evidence that the association between GDP comovement and trade linkage operates not only through bilateral trade intensity, but also through common exposure to third countries, which we refer to as network effects. First order network index. In a world with many countries, the bilateral index of trade 24 proximity is not a sufficient measure of trade linkages. We first complement the above codifferenceisamore“agnostic”transformationthataccountsforboththecyclicalandthetrendcomponentsembodiedinanyyear-to-yearfluctuation,butitissometimesconsideredaslesssensitivetoresearcher’sassumptions andpreferencesregardingtheparametersofthefilteringmethod. 21 The concordance table from SITC Rev2 to BEC can be found on the UN Trade Statistics webpage: https: //unstats.un.org/unsd/trade/classifications/correspondence-tables.asp. 22 We merge capital goods and intermediate inputs as a single bundle of intermediate inputs. Trade in capital goodsisroughly14%to15%oftotaltradeflows. Forrobustness,wealsoconsidertradeincapitalgoodsseparately. Themainresultsremainunchanged. 23 This specification is widely used in the literature. As a robustness check, we also adopt an alternative used (cid:110)Td +Td Td +Td (cid:111) index: Traded =max i→j,t j→i,t, i→j,t j→i,t . 24 Theimp i o jt rtanceofthir G d D c P o it untrye G ff D e P c j t t isalsomentionedinKoseandYi(2006)andDuvaletal.(2015)analyzes 15

variate with an index of first order network proximity that is constructed to reflect the fact that two countries might experience a surge in their GDP correlation if their exposure to a common third country increases. In other words, over and above changes in bilateral trade flows, two countries that are increasingly linked to similar partners are likely to become more synchronized. To account for such a common exposure mechanism, we construct a third country index aiming to capture the first order component of a trade network, such that: (cid:12) (cid:12) network1st = 1− 1 ∑(cid:12) (cid:12) T i↔k,t − T j↔k,t(cid:12) (cid:12) ( 18 ) ijt 2 (cid:12) T T (cid:12) k i,t j,t where T represents the total trade flows between country i and country k and T denotes i↔k,t i,t the total trade flows of country i vis-a-vis all of its partners. This index effectively measures the geographical overlap in two countries’ trade partners. Note that country-pairs with very similartradepartnershaveanindexclosetoonewhiletwocountriestradingwithcompletely different partners have an index of zero. Second order Network effect. As a measure of 2 nd order network proximity for any pair (i,j),webuildanindexmeasuringtowhatextendcountryi’spartnersarelinkedwithcountry j’s partners, weighted by the importance of the partners in terms of total trade flows of the two countries i and j: (cid:32) (cid:33) network2nd = 1 ∑ ∑ (w (i,z)+w (i,y)+w (j,z)+w (j,y))∗network ( 19 ) ijt 4 t t t t zyt z∈P(i)y∈P(j) where w (i,z) = Ti↔z,t. Under this specification, the more the partners of my partner are t Ti,t similar to my partners in terms of 1 st order network, the higher the index network2nd. ijt Cross-Network effect. In the robustness tests (section 6 ), we go a step further and construct anotherindexcapturingnon-symmetricsituationswhereacountry’sdirectpartnersarelinked withanothercountrysecondorderpartners. Werefertothissituationasacross-networkeffect of trade proximity, denoted (crossnetwork ) and defined as follows: ijt (cid:32) (cid:12) (cid:12) (cid:12) (cid:12)(cid:33) crossnetwork ijt = 1− 1 ∑ w t (j,z) ∑(cid:12) (cid:12) T i↔k,t − T z↔k,t(cid:12) (cid:12)+ ∑ w t (i,z) ∑(cid:12) (cid:12) T j↔k,t − T z↔k,t(cid:12) (cid:12) 4 (cid:12) T T (cid:12) (cid:12) T T (cid:12) z∈P(j) k i,t z,t z∈P(i) k j,t z,t (cid:32) (cid:33) = 1 ∑ w (j,z)network + ∑ w (i,z)network ( 20 ) t izt t jzt 2 z∈P(j) z∈P(i) the role of indirect trade linkages between two countries using a value-added approach. Our approach differs fromDuvaletal.(2015)becausecommonexposuretothirdcountriescanhappenevenwhentwocountriesdonot exchangeanyvalueaddedwithone-another. 16

Figure2. Illustrationoffirstorder,secondorderandcrossnetworkproximityindexes. 1st Order Network 2nd Order Network Cross-Network C D A B A B E C D C D E E F Note: dashed areas represent 1st order network. The second order effect and the cross-network effect canberepresentedasacombinationof1stordernetworkeffects. where w (i,z) = Ti↔z,t. Theindexmeasurestheextenttowhichacountry i inthecountry-pair t Ti,t (i,j) is similar in terms of trade partners (i.e. in terms of direct network index) to all countries z ∈ P(j) trading with its partner j, weighted by the importance of z in the total trade of j. 2 As an illustration, we combine the three network representations in figure . Finally, we summarize in table 2 the evolution of our three network indexes in our sample. Interestingly, the first order network effect is much larger in OECD countries relative to the other group considered. Table.2. Networkindexinthedifferentincomegroups a 100 Network index * First order Second order Cross-network Period OECD HH HL LL OECD HH HL LL OECD HH HL LL 7079 536 474 471 501 464 465 475 478 421 422 409 400 : . . . . . . . . . . . . 8089 557 480 473 501 480 479 488 490 424 421 409 403 : . . . . . . . . . . . . 9099 565 496 466 486 485 493 494 492 427 427 417 410 : . . . . . . . . . . . . 0009 550 472 445 467 484 493 493 492 433 433 426 423 : . . . . . . . . . . . . a Numbersreportedaretheaverageoverallcountry-pairs. 3.3 Proximity in sectoral composition 2 Asdiscussedinsection ,ifshockshaveasectoralcomponentthentwocountrieswithincreasing similarity in sectoral specialization could experience a corresponding surge in business 17

cycle co-movements even in the absence of any trade linkages. In order to account for such a mechanism, we build two bilateral indexes of proximity in sectoral composition. The first index is based on countries’ proximity in terms of sector share in GDP while the second focuses on 4 the proximity in traded goods, at the -digit SITC level or ISIC level, as proxy for domestic specialization in exported goods. Data for sector shares in GDP come from the World Bank’s WDI. We use the share in value added of nine main sectors composed of service, agriculture and seven manufacturing sectors (textile, industry, machinery, chemical, high-tech, food and 25 tabacco, and other). Such an index is a direct measure of two countries’ specialization, but its usefulness is somewhat limited by the high level of sectoral aggregation which allows us to capture only specialization in broad sectors. Moreover, data are available only for a subset of all countries. prox We define the sectoral proximity index in terms of traded goods denoted export for a ijt given country-pair (i,j) in time-window t as: export prox = 1− 1 ∑ (cid:12) (cid:12) (cid:12) EX i,t (s) − EX j,t (s)(cid:12) (cid:12) (cid:12) ( 21 ) ijt 2 (cid:12) EX EX (cid:12) s∈S EX i,t j,t where EX (s) refers to total export of country i in sector s ∈ S , with S being the set of i,t EX EX 4 sectors(each -digitSITCcodeorISICcode,dependingonthedefinitionadopted). Wedefine prox the sectoral proximity index in terms of sector shares in GDP, denoted sector , for a given ijt country-pair (i,j) in time-window t as: sector prox = 1− 1 ∑ (cid:12) (cid:12) (cid:12) Y i,t (s) − Y j,t (s)(cid:12) (cid:12) (cid:12) ( 22 ) ijt 2 (cid:12) Y Y (cid:12) s∈S i,t j,t where Y (s) refers to total value-added of country i in sector s ∈ S, with S being the set of i,t sectors. For both indexes, country pairs with very similar sectoral/trade composition have an 1 index close to , while countries that completely specialize in different sectors have an index 0 3 of . We provide in table the evolution of sectoral proximity and export proximity over time 4 for the income groups considered. In section , we use the export proximity constructed at 4 6 the -digit SITC level and leave the ISIC specification as a robustness exercise in section . Looking at indices based on exports as well as GDP, we note that country-pairs in the OECDaresignificantlymoresimilarthanthoseinothergroups. Moreover, thetimeevolution of these indices also reveals a higher convergence, in terms of economic structure, among OECD countries compared with other sub-samples. 25 Dataareavailablehere: https://databank.worldbank.org/data/source/. 18

Table.3. Sectoralandexportproximityindexinthedifferentincomegroups a 100 100 Export proximity* Sectoral proximity* 4 -digit SITC ISIC b WDI c Period OECD HH HL LL OECD HH HL LL OECD HH HL LL 7079 299 219 113 142 461 354 287 368 852 833 753 784 : . . . . . . . . . . . . 8089 326 212 120 150 484 346 268 337 888 845 786 815 : . . . . . . . . . . . . 9099 373 245 140 156 526 386 268 300 895 881 788 816 : . . . . . . . . . . . . 0009 381 260 161 171 533 394 287 307 896 835 755 782 : . . . . . . . . . . . . a Numberreportedistheaverageoverallcountry-pairs. b We classify goods and products at the ISIC level following the correspondence table https://unstats.un.org/unsd/tradekb/Knowledgebase/50054/ Correlation-between-ISIC-and-SITC-codes-or-Commodity-and-Industry. c WDIreferstosectoralproximityintermsofshareofWDIsectorsGDPintotalGDP.Dataisavailablehere: https://databank.worldbank.org/data/source/. 4 The Global Trade-Comovement Slope Inthissection,werevisittheseminalFRanalysisanduseallvariablesdefinedintheprevious section as well as additional controls to investigate the determinants of business cycle correlation for different income groups and time periods. We proceed step-by-step and gradually introduce our variables. 4.1 The initial Frankel and Rose (1998) specification We first review the FR results by extending the analysis to a large sample of countries separated into different income groups. Following the more recent literature, we use a panel fixedeffectinordertocontrolforunobservedheterogeneitybetweencountry-pairs,aswellas 26 changesineconomicconditionsovertimethatarenotrelatedtotrade. Asafirststep,weestimate a panel with country-pair (CP) and time-window (TW) fixed effects with the following specification: Corr GDP = β ln(Tradetotal)+X +CP +TW +(cid:101) , ( 23 ) ijt 1 ijt ijt ij t ijt where X is a vector of additional control variables that includes dummies for URSS ijt countries, the euro area, and the different waves of the European Union. On the one hand, the introduction of CP fixed effects means that we are using only within country-pair time variations for the identification. These dummies effectively control for time invariant factors 26 Inordertodiscriminatebetweenfixedorrandomeffects,werunaHausmantestwhichdisplayasignificant difference(p<0.001),andwethereforerejecttherandomeffectmodel. 19

thatcaninfluenceGDPcomovementbetweentwocountries,suchasdistance,commonborder, commonlanguage,etc. Ontheotherhand,TWfixedeffectscaptureaggregatechangesinGDP comovement for all country-pairs in the world that could be due to aggregate shocks. In this specification as well as all subsequent analysis, standard errors are robust to clustering at the country-pair level, which accounts for serial correlation across time. That is, we allow for the error term to have a fixed country-pair component common to all (i,j) observations. In a second step, we introduce our network indexes (first and second order), which aim to capture the network effect of trade on GDP comovement stemming from both direct and indirect exposure to third countries. For this exercise, we use the following specification: Corr GDP = β ln(Tradetotal)+γγγnetwork +X +CP +TW +(cid:101) ( 24 ) ijt 1 ijt ijt ijt ij t ijt In equation ( 24 ), network defines a vector composed of the first and second order netijt 4 work measures discussed above. The results are gathered in table . Twomainresultsemerge. First,aspreviouslyhighlightedintheliterature,tradeproximity using total trade flows is significantly associated with more GDP correlation, for all considered groups. However, the strength of this association is very heterogeneous. Using the point 25 75 estimate obtained with all country pairs, we find that moving from the th to the th per- 50 centiles of log total trade is associated with an increase in GDP correlation of . percentage points. The same number increases up to 16 . 7 percentage points for OECD country pairs, 7 . 1 percentage points for pairs in the HH group, 5 . 9 percentage points for the HL group and 9 . 3 percentage for the LL sub-sample. Second, the effect of trade through the network effect is high and significant. According to 25 75 our point estimate, moving from the th to the th quantiles of the direct network index 73 implies an increase in GDP correlation of about . percentage points for all country-pairs, again with stark differences across sub-samples. For pairs in the OECD group, moving from 25 75 309 the thtothe thpercentilesisassociatedwithanimpressive . percentagepointincrease in bilateral GDP correlation, while it is 14 . 6 percentage point for pairs in the HH group and only 6 . 2 percentage points for pairs in HL. Interestingly, the strength of a marginal increase in the direct network indexes is decreasing as the sample includes countries at the lower end of the income distribution, with the latter effect becoming statistically insignificant for the LL group. Interestingly, the second order network index also plays a significant role for GDP comovement when using the whole sample. Concerning the classification in terms of income group, the point estimates increase as we move to the low income group. For example, for 20

theLLgroup, whenmovingfromthe 25 thtothe 75 thquantilesofthisindex, GDP-correlation 88 increases . %, while the first order network effect is insignificantly correlated with more GDP-comovement. This result highlights the possible strong dependence of those countrypairs to the global network as opposed to their more direct bilateral network. When focusing on the particular OECD group, we find the second order network effect is particularly high, 213 with an increase in GDP correlation of . % when moving from the first to the last quartiles. Alltold,thisfirstexerciserevealsasubtleassociationbetweentradeandGDPco-movement. While previous investigations highlighted the role of either direct or indirect bilateral trade, the economic and statistical significance of our network indices sheds light on an additional channel stemming from increasing exposure to other countries. As we will further show below, the strength of this new channel is increasing over time, which makes it all the more relevantforunderstandingrecentandfuturechangesincross-countrybusinesscyclesynchronization. Table.4. TradeComovementslopewithtotaltradeindex corrGDP All All OECD OECD HH HH HL HL LL LL ln(Trade) 0.023∗∗∗ 0.021∗∗∗ 0.083∗∗∗ 0.090∗∗∗ 0.038∗∗∗ 0.028∗ 0.024∗∗∗ 0.023∗∗∗ 0.032∗∗∗ 0.038∗∗∗ (0.005) (0.006) (0.030) (0.033) (0.013) (0.015) (0.006) (0.006) (0.012) (0.006) network1st 0.293∗∗∗ 1.082∗∗∗ 0.540∗∗∗ 0.254∗∗∗ −0.044 (0.067) (0.262) (0.163) (0.073) (0.073) network2nd 0.457∗∗ 2.914∗∗∗ −0.201 0.567∗∗ 1.143∗∗∗ (0.221) (0.913) (0.555) (0.243) (0.243) CP+TWFE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Observations 13,079 13,079 1,224 1,224 2,541 2,541 10,538 10,538 2,745 2,745 R2 0.006 0.009 0.068 0.094 0.033 0.039 0.002 0.005 0.005 0.009 Notes: ∗p<0.1;∗∗p<0.05;∗∗∗p<0.01. 4.2 Accounting for trade in intermediate inputs and final goods Expanding on our results using total trade flows, we now refine the analysis and decompose total trade flows into two sub-categories: trade in intermediate inputs and trade in final goods. 2019 As discussed in de Soyres and Gaillard ( ) for the case of high income countries, trade in intermediate inputs is significantly correlated with GDP comovement, while trade in final 27 goods is not. In this section, we test the differential relationship between business cycle 27 IndeSoyresandGaillard(2019),wealsoshowtheoreticallyhowinternationalinput-outputlinkages,coupled withmarketpowerandextensivemarginadjustments,canquantitativelygenerateastronglinkbetweentradein 21

synchronization and both trade in intermediate inputs and trade in final goods in a larger sample covering different income groups. We run the following specification (with and without network effects), where we disaggregate total trade flows (Tradetotal) into trade in intermediate inputs (Tradeinter) and trade in ijt ijt final final goods (Trade ): ijt Corr GDP = β ln(Tradeinter)+β ln(Trade final)+X +CP +TW +(cid:101) ( 25 ) ijt 1 ijt 2 ijt ijt ij t ijt Corr GDP = β ln(Tradeinter)+β ln(Trade final)+γγγnetwork +X +CP +TW +(cid:101) ijt 1 ijt 2 ijt ijt ijt ij t ijt 26 ( ) Theresultsareshownintable 5 . Whenfocusingoncountry-pairsinOECDandHH,wesee that the TC slope is significantly driven by trade in intermediate inputs as opposed to trade 28 infinalgoods Turningtocountry-pairsintheHLandLLgroups, wefindanoppositeresult: the TC slope is significantly related to more trade in final goods while trade in intermediate inputs is not significantly associated with higher GDP comovement. These findings are also strongly economically significant: according to the point estimate obtained when controlling 25 75 for network effects, moving from the th to the th quantiles of log trade in intermediate 217 inputs is associated with a . percentage points increase in GDP correlation for pairs in the OECD group and a 12 . 9 percentage point increase for pairs in the HH group. For pairs in the HL and LL groups, moving from the 25 th to the 75 th quantile of log trade in final goods 49 51 increases respectively GDP comovement by . and . percentage points respectively. In 6 section ,weshowtheseresultsarerobusttoanumberofalternativespecifications,including financial controls (FDI and constructed financial interconnection BIS indexes), different GDP filters and different measures of trade intensities. There are several possible explanations for the difference between high and low income countries. Forexample,onecouldarguethatintermediateinputstradedbylowincomecountriesaremorestandardizedthantheheavilycustomizedproductstradedbetweenhighincome 29 countries. Wefurtherinvestigatethisissuebelowandarguethatatleasttwochannelsmight beatplay: (i)similarsectoralspecializationbetweentwocountriesseemstoamplifytheeffect of intermediate input trade on GDP correlation ; and (ii) there is a different time evolution of market power in low versus high income countries. intermediateinputsandGDP-comovement,resolvingtheTrade-ComovementPuzzle 28 Notice that we combine trade in capital goods with trade in intermediate inputs. Separating those flows to theregressionprovidessimilarresultsasshowninthesensitiveanalysis. 29 Notethatintermediateinputsincludemanycommoditiessoldontheglobalmarketandlittlebilateralstickiness in the buyer-supplier relationship. For such goods, it is not surprising that direct bilateral trade is not associatedwithbilateralGDPcomovement. 22

Table.5. TradeComovementslopewithdisaggregatedtradeindex corrGDP All All OECD OECD HH HH HL HL LL LL ln(inter) 0.011∗∗ 0.009 0.106∗∗∗ 0.103∗∗∗ 0.043∗∗∗ 0.034∗∗ 0.008 0.007 0.014 0.018∗∗∗ (0.005) (0.005) (0.030) (0.030) (0.013) (0.014) (0.006) (0.006) (0.011) (0.006) ln(final) 0.012∗∗∗ 0.012∗∗ −0.023 −0.009 −0.011 −0.009 0.017∗∗∗ 0.016∗∗∗ 0.015 0.017∗∗∗ (0.005) (0.005) (0.024) (0.025) (0.012) (0.012) (0.005) (0.005) (0.010) (0.005) network1st 0.305∗∗∗ 1.046∗∗∗ 0.521∗∗∗ 0.264∗∗∗ −0.040 (0.067) (0.259) (0.163) (0.073) (0.073) network2nd 0.416∗ 2.906∗∗∗ −0.170 0.518∗∗ 1.076∗∗∗ (0.221) (0.938) (0.550) (0.243) (0.243) CP+TWFE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Observations 13,079 13,079 1,224 1,224 2,541 2,541 10,538 10,538 2,745 2,745 R2 0.006 0.009 0.074 0.099 0.035 0.040 0.003 0.005 0.004 0.008 Notes: ∗p<0.1;∗∗p<0.05;∗∗∗p<0.01. 4.3 The role of sectoral proximity and export proximity We complement the analysis with additional controls that aim to capture the proximity in 2 termsofsectoralcomposition. Asshowninsection ,iftwocountrieshavesimilarsectors,they are more likely to experience GDP comovement. We test this prediction using two measures prox of proximity: (i) proximity in terms of exported goods using our index export and (ii) ijt prox prox proximity in terms of GDP sector shares captured by the index sector . While export ijt ijt prox is available for all considered country-pairs, sector is constrained by data availability. We ijt run the following two specifications: Corr GDP = β ln(Tradeinter)+β ln(Trade final)+γγγnetwork +α export prox ijt 1 ijt 2 ijt ijt 1 ijt +X +CP +TW +(cid:101) ( 27 ) ijt ij t ijt Corr GDP = β ln(Tradeinter)+β ln(Trade final)+γγγnetwork +α export prox ijt 1 ijt 2 ijt ijt 1 ijt +α sector prox+X +CP +TW +(cid:101) ( 28 ) 2 ijt ijt ij t ijt 6 The results are gathered in table . On the one hand, focusing on the first five columns, we find that similarity in the type of traded goods has a surprisingly negative effect on GDP- 2 comovement, which is not captured by our simple model in section . This result could be explainedbythepossiblesubstitutabilitybetweentradedgoods,whichimpliesthatapositive technology shock in one country decreases the market share of producers in other countries, leading to fluctuations in the opposite direction when they trade similar goods. On the other 23

prox hand,otherestimatedcoefficientsarenotsensitivetotheadditionofexport . Turningtothe ijt effect of sectoral proximity (last five columns) on the sub-sample for which data are available, prox we do not find a statistically significant effect of sector on GDP-comovement. ijt Table.6. Theroleofsectoralproximityandexportproximity corrGDP All OECD HH HL LL All OECD HH HL LL ln(inter) 0.010∗ 0.102∗∗∗ 0.037∗∗∗ 0.008 0.018 0.010 0.142∗ 0.089∗∗ 0.009 0.044∗∗ (0.005) (0.030) (0.014) (0.006) (0.011) (0.014) (0.082) (0.040) (0.014) (0.022) ln(final) 0.013∗∗∗ −0.008 −0.007 0.017∗∗∗ 0.017∗ −0.002 −0.046 −0.189∗∗∗ 0.013 −0.010 (0.005) (0.025) (0.013) (0.005) (0.010) (0.012) (0.050) (0.042) (0.013) (0.021) network1st 0.307∗∗∗ 0.967∗∗∗ 0.445∗∗∗ 0.271∗∗∗ −0.027 0.321 0.703 0.287 0.324 −0.345 (0.066) (0.269) (0.162) (0.073) (0.156) (0.201) (0.660) (0.504) (0.221) (0.368) network2nd 0.446∗∗ 2.745∗∗∗ −0.302 0.554∗∗ 1.081∗∗ 2.103∗∗∗ 3.216∗ −1.107 2.471∗∗∗ 2.396∗∗ (0.221) (0.955) (0.550) (0.243) (0.469) (0.607) (1.885) (1.544) (0.656) (0.976) exportprox −0.394∗∗∗ −0.381 −0.754∗∗∗ −0.318∗∗∗ −0.075 −0.091 0.916∗ −0.928∗∗ 0.024 0.333 (0.083) (0.237) (0.182) (0.092) (0.146) (0.188) (0.473) (0.439) (0.206) (0.295) sectorprox 0.296 −0.249 0.430 0.172 −0.235 (0.189) (0.705) (0.430) (0.213) (0.409) CP+TWFE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Observations 13,079 1,224 2,541 10,538 2,745 4,499 655 893 3,606 1,091 R2 0.012 0.101 0.049 0.007 0.008 0.025 0.165 0.166 0.017 0.022 Notes: ∗p<0.1;∗∗p<0.05;∗∗∗p<0.01. 4.4 The evolution of the TC Slope from 1970 to 2009 Having established the link between global trade flows and GDP comovement for different income groups, we now turn to the issue of the potential time evolution of the association between a marginal increase in trade and business cycle synchronization. More precisely, we 1970 2009 provideevidenceregardinganoticeableevolutionoftheTCslopebetween and . This 1990 2009 evidence is of particular importance since GDP-comovement surged between and , formanycountries,includingthoseinthelowincomegroups. Moreover,establishingthatthe associationbetweentradeandbusinesscyclesynchronizationnotonlyvariesbyincomegroup but is also time varying yields two additional benefits. First it helps reconcile different values of the TC-slope that are found in the literature and which rely on different geographic and time coverages. Second, it would lend support to the hypothesis that the marginal effect of increasing(eitherdirectorindirect)tradeflowsbetweentwocountrieschangeswitheconomic 24

conditions. Such a hypothesis can then be investigated further in order to uncover what are the factors that enable and amplify the relationship between trade and GDP comovement. We introduce a dummy variable LTW which equals to 1 for the last two time-windows t 19901999 20002009 0 in our sample – that is for the periods : and : – and otherwise. This “Late Time Window” dummy is then interacted with the determinants of GDP comovement, allowing us to formally test for the difference between the TC-slope in earlier time windows 30 and the slope observed toward the end of the time coverage. Formally, we now test the change in the slope using the following specifications: Corr GDP = β ln(Tradeinter)+β LTW ×ln(Tradeinter)+β ln(Trade final) ijt 1 ijt 2 t ijt 3 ijt +β LTW ×ln(Trade final)+γγγnetwork +X +CP +TW +(cid:101) ( 29 ) 4 t ijt ijt ijt ij t ijt Corr GDP = β ln(Tradeinter)+β LTW ×ln(Tradeinter)+β ln(Trade final) ijt 1 ijt 2 t ijt 3 ijt +β LTW ×ln(Trade final)+γγγ network +γγγ LTW ×network 4 t ijt 111 ijt 222 t ijt +X +CP +TW +(cid:101) ( 30 ) ijt ij t ijt where X refers again to controls. By adding these interaction terms, we specify that coijt efficients β , β and γγγ indicate whether the TC slope estimated with respect to trade in 2 4 222 intermediateinputsandfinalgoodsandthecoefficientsassociatedwithnetworkeffectsinthe 1990 2009 1970 1989 period - are different from the coefficients estimated using the period - . 7 Wepresentourresultsintable . Lookingatnon-interactedpointestimates,weseethatthe findings are consistent with previous specifications, with trade in intermediate inputs being significantly correlated with more GDP comovement for country pairs in the OECD and HH sub-sample, while pairs in the other groupings feature a more prominent role for trade in final goods. Moreover, the results show that the TC-slope changed significantly over time for somecountrygroups. Amongdevelopedcountries(thatis, intheOECDandHH groups), the 1970 1989 TC slope in intermediate inputs significantly increased over time. Between and , the estimates indicate a significant TC-slope of about 0 . 066 for country-pairs among the OECD 0156 1990 2009 group, while it rises to . for the period to . In terms of magnitude, moving 25 75 from the th to the th percentile of log trade in intermediate inputs (for all time-windows) 130 would have implied an increase in GDP comovement of about . percentage point from 1970 1989 312 to , which is much lower than the . percentage point increase corresponding to the slope estimated using the 1990 to 2009 period. For HH, the TC-slope is statistically 30 NotethatwithCPfixedeffectsweareonlyusingwithincountry-pairtimevariationsintradeproximityand GDPcorrelation. Hence,itisimportantforourLateTimeWindowdummytocover(atleast)twotime-windowsso thattherearetimevariationswithinthelatesub-sample. 25

1990 2009 different from zero only for the to period. In turn, it is also worth noting that the TC-slope in final goods also increased over time for the HL group, implying that an increase in final goods trade between high and low income countries is associated with a stronger increase in GDP comovement in recent time windows. Finally, as regards the time evolution of the association between network proximity and GDP comovement, we find a significant positive increase in the network-comovement slope in 1990 2009 1970 1989 theperiod to relativetotheperiod to foralmostallsub-samples. Thehigh pointestimatesoftheseinteractedtermsrevealthatthesurgeintheassociationbetweentrade network and business cycle synchronization is very large. Altogether, these findings show that the association between international trade linkages and GDP correlation experienced a strongincreaseovertime,eitherdirectly(throughbilateraltrade)orindirectly(viathenetwork effect). Table.7. TimeevolutionoftheTradeComovementslope corrGDP All All OECD OECD HH HH HL HL LL LL ln(inter) 0.002 0.004 0.074∗∗ 0.066∗∗ 0.013 0.008 0.003 0.005 0.012 0.015∗∗ (0.006) (0.006) (0.033) (0.033) (0.016) (0.016) (0.007) (0.007) (0.013) (0.007) LTW*ln(inter) 0.010∗∗ 0.006 0.074∗∗ 0.090∗∗∗ 0.028∗∗∗ 0.025∗∗ 0.006 0.002 0.012 0.007 (0.005) (0.005) (0.029) (0.031) (0.011) (0.011) (0.006) (0.006) (0.013) (0.006) ln(final) 0.007 0.008 0.019 0.019 −0.021 −0.022 0.011∗ 0.011∗ 0.019∗ 0.020∗∗∗ (0.006) (0.006) (0.028) (0.028) (0.014) (0.014) (0.006) (0.006) (0.011) (0.006) LTW*ln(final) 0.008 0.007 −0.034 −0.045 0.009 0.004 0.010∗ 0.010∗ −0.006 −0.009 (0.005) (0.005) (0.032) (0.033) (0.011) (0.011) (0.006) (0.006) (0.012) (0.006) network1st 0.317∗∗∗ 0.211∗∗∗ 0.800∗∗∗ 0.699∗∗ 0.608∗∗∗ 0.482∗∗∗ 0.273∗∗∗ 0.189∗∗ −0.033 −0.133 (0.066) (0.073) (0.269) (0.273) (0.165) (0.176) (0.073) (0.082) (0.156) (0.082) network2nd 0.375∗ −0.030 2.830∗∗∗ 1.944∗ −0.637 −1.154∗∗ 0.509∗∗ 0.102 1.101∗∗ 1.302∗∗∗ (0.221) (0.227) (0.984) (1.012) (0.552) (0.553) (0.244) (0.252) (0.469) (0.252) LTW*network1st 0.236∗∗∗ 0.211 0.369∗∗∗ 0.180∗∗∗ 0.154∗∗ (0.058) (0.158) (0.117) (0.068) (0.068) LTW*network2nd 0.661∗∗∗ 1.056∗∗∗ 0.846∗∗∗ 0.643∗∗∗ −0.157 (0.129) (0.404) (0.246) (0.150) (0.150) CP+TWFE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Observations 13,079 13,079 1,224 1,224 2,541 2,541 10,538 10,538 2,745 2,745 R2 0.015 0.019 0.110 0.121 0.062 0.071 0.010 0.013 0.008 0.009 Notes: ∗p<0.1;∗∗p<0.05;∗∗∗p<0.01. 26

4.5 Network density as an amplification channel 2 In section , we discussed a mechanism that could explain the recent increase in the strength of the association between global trade and business cycle synchronization – namely, the role of overall network density. Our model highlights how network density interacts with the channels discussed above and strengthens the connection between individual links and bilateral GDP comovement. 8 To show the change in network density over time, we first display in table the average share of total worldwide trade flows over total worldwide GDP in each of our four time windows, where we normalize with respect to the value in the first time window. Since the 1970 s, the average trade flow over GDP has more than tripled. As a result of such an increase in network density, we expect a corresponding surge in the correlation between trade and GDP comovement. Table.8. TotaltradeflowsoverworldwideGDP 7079 8089 9099 0009 Period : : : : 10 181 213 310 Trade flows / GDP . . . . We then move on to formally testing our intuition and implement a bilateral measure of overall network connectiveness that reflects how much countries in a given country-pair are connected to the rest of the trade network. We compute the average bilateral network density for a given country-pair, as follows: (cid:16) (cid:17) w(i,j)∑ Tradetot +w(j,i)∑ Tradetot NetworkDensity = z∈P(i) −j i↔z,t z∈P(j) −i i↔z,t ( 31 ) ijt w(i,j)+w(j,i) where P(i) defines the set of i-partners except the country j and w (i,j) = Ti↔j,t. This index −j t Ti,t measures the average trade volume over GDP of the two countries within the country-pair (i,j)whenbilateraltradeflowsarenottakenintoaccount,anditaimstomeasuretheconnectedness of two countries to the rest of the network. In this sense, it should be interpreted not as a measure of overall network density, but rather as a measure of trade proximity between 9 the pair at hand and the rest of the world. Table presents the time evolution of this measure for each sub-sample. 27

Table.9. Evolutionofourbilateralmeasureofnetworkdensitya Income Group Period All OECD HH HL LL 7079 03 05 05 03 02 : . . . . . 8089 07 10 10 07 04 : . . . . . 9099 10 16 15 09 06 : . . . . . 0009 16 24 22 14 10 : . . . . . a Reportednumbersareaverageoverallcountry-pairs. We test the following specification: Corr GDP = β ln(Tradeinter)+β ln(Trade final)+β density ×ln(Tradeinter) ijt 1 ijt 2 ijt 3 ijt ijt +β density ×ln(Trade final)+γγγ network +γγγ density ×network 4 ijt ijt 1 ijt 2 ijt ijt +X +CP +TW +(cid:101) ( 32 ) ijt ij t ijt Notice that our measure of bilateral density is directly linked to the first order network index as the later measures the intensive margin of the first order trade network, while the 31 10 former can be interpreted as measuring similarity in the first order trade network. Table summarizes the findings. The results present interesting differences across income groups. Regarding the first col- 10 umn of table which presents the results using the whole sample, the interaction between ourbilateralmeasureofdensityandtradenetworkeffectsispositive,suggestingthatcountrypairs that are more connected to the rest of the world feature a higher marginal effect of network proximity. We also find that the interaction between bilateral trade and bilateral density is significant with different patterns in different sub-samples: bilateral density acts as an amplifier of intermediate trade proximity for the OECD and HH groups, while it strengthens the role of final goods trade in the HL group. Overall, our findings imply that the TC-slope usually measured in the literature is a function of overall connectivity between a country-pair and the rest of the world. In other words, bilateral trade flows have a higher marginal effect on GDP-comovement when two countries trade more with each other. 31 As a robustness, we also used total trade flows over worldwide GDP as a measure of network density and interacted it our first order network effect. Results are similar in this case, although the logic of the estimation differs markedly: using world trade over world GDP as a measure of density means means the index is not bilateral and mostly measure an increasing trend for the whole sample. We see this exercise as confirming the findingsinsection4.4inthesensethatthereisaworldwideincreaseintheassociationbetweenglobaltradeand bilateralGDPco-movement. 28

Table.10. NetworkdensityandTradeComovementSlope corrGDP All OECD HH HL LL ln(inter) 0.004 0.033 0.005 0.009 0.015 (0.006) (0.035) (0.017) (0.007) (0.014) ln(final) 0.003 0.042 −0.007 0.003 0.024∗∗ (0.006) (0.031) (0.015) (0.006) (0.012) density*ln(inter) 0.004 0.095∗∗∗ 0.030∗∗∗ −0.006 −0.005 (0.004) (0.021) (0.008) (0.005) (0.013) density*ln(final) 0.011∗∗∗ −0.030∗ 0.000 0.016∗∗∗ −0.020 (0.004) (0.017) (0.007) (0.004) (0.012) network1st 0.230∗∗∗ 0.931∗∗∗ 0.506∗∗∗ 0.165∗ −0.212 (0.075) (0.298) (0.173) (0.085) (0.175) network2nd 0.213 2.588∗∗∗ −0.671 0.302 1.480∗∗∗ (0.228) (0.999) (0.573) (0.250) (0.499) density*1stnet. 0.085∗∗ −0.076 0.003 0.106∗∗ 0.244∗∗ (0.035) (0.093) (0.061) (0.044) (0.104) density*2ndnet. 0.162∗∗ 0.607∗∗ 0.272∗∗∗ 0.158∗ −0.582∗∗ (0.063) (0.260) (0.099) (0.083) (0.233) CP+TWFE Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Observations 13,079 1,224 2,541 10,538 2,745 R2 0.016 0.122 0.066 0.012 0.016 Notes: ∗p<0.1;∗∗p<0.05;∗∗∗p<0.01. 4.6 Summary of empirical evidence The preceding sections provided empirical support for the theoretical results discussed in 2 section and offered novel insights on the complex association between global trade flows and bilateral GDP comovement. In particular, our findings can be summed up as follows: 1 . ThecorrelationbetweentradeinintermediateinputsandGDPcomovementissignificant and positive for countries in the OECD and HH groups, suggesting a specific role for Global Value Chains. Interestingly, trade in final goods is significantly correlated with higher business cycle synchronization for the low income groups. 2 . Common exposure to third countries, measured using our network indices is significantly positively correlated with more GDP comovement. First order network effects decay as we move to low income group while second order network effects increase for those in the lowest income group. 29

3 . SimilarityinexportedgoodscompositionissignificantlynegativelycorrelatedwithGDP comovement while there is no statistically significant effect of sectoral composition on GDP-comovement. 4 . The correlation between GDP comovement and both bilateral trade and network effects tends to increase over-time. As suggested by our simple model, this increase could be rationalized by a surge of trade network density which can amplify the association between global trade and bilateral comovement. In the next section, we also discuss a complementary channel that could rationalize the time evolution of point estimates – namely, an increase in the sensitivity of GDP to foreign shocks due to higher markups. 5 Discussion 1 We now dig further into two particular findings: ( ) the differential role of sector versus trade 2 proximity, and ( ) the time evolution of the marginal association between trade and GDP comovement. First, weshowthattheassociationbetweentradeandbusinesscyclecorrelation strongly depends on the sectoral composition of the economy, which creates another link between our explanatory variables. Second, we show that market powers have substantially 1090 risen in the s, with the possible implication that international shock transmission could have strengthened even in the absence of increased trade flows. 5.1 Global Value Chain and Sectoral Composition A number of studies document the strong influence of global value chain (GVC) in transmitting shocks across countries. Indeed, to the extent that inputs tend to be more customized and less substitutable than final goods, bilateral production fragmentation can be associated with higher complementarity between production factors, which would then lead to higher GDP comovement. Conversely, when two countries trade final goods while producing very similar goods, they can be seen as competing in the same market, which implies that a positive supply side shock (such as a technology innovation) in one country leads to an increase in this country’s market share at the expanse of its trade partner, leading to opposite GDP movements. In this section, we test the hypothesis that two economies with similar sectors and similar exported goods are more likely to comove when trade in intermediate inputs is high, and that they are less likely to comove when trade in final goods is high. To do so, we run the 30

following two specifications: Corr GDP = β ln(Tradeinter)+β ln(Trade final)+β export prox ijt 1 ijt 2 ijt 3 ijt +β export prox×ln(Tradeinter)+β export prox×ln(Trade final) 4 ijt ijt 5 ijt ijt +X +CP +TW +(cid:101) ( 33 ) ijt ij t ijt Corr GDP = β ln(Tradeinter)+β ln(Trade final)+β sector prox ijt 1 ijt 2 ijt 3 ijt +β sector prox×ln(Tradeinter)+β sector prox×ln(Trade final) 4 ijt ijt 5 ijt ijt +X +CP +TW +(cid:101) ( 34 ) ijt ij t ijt where X refers to controls, including first and second order network indices. The results ijt 11 are shown in table and can be stated as follows. Country-pairs with similar export compositionco-movemorewhentheytrademoreinintermediateinputsandco-movelesswhenthey 3 trade more final goods. Given the mean export proximity index in table and the estimates 11 25 75 in table , moving from the th to the th quantiles of the log trade index in intermediate inputs would imply an increase in GDP comovement of 0 . 20 for the HH group against 0 . 012 for the HL group. On the opposite side, moving from the 25 th to the 75 th quantiles of the log 001 trade index in final goods would imply an increase in GDP comovement of negative . for the HH group against 0 . 05 for the HL group. Regarding the similarity in GDP sector shares, the results are significant and consistent for OECD group only. For this sub-sample, given the mean sectoral proximity index, moving 25 75 from the th to the th percentiles of the log intermediate inputs (respectively final goods) 029 011 distributionwouldleadtoanincreaseinGDPcomovementof . (respectivelynegative . ). 25 75 Notice, however, that moving from the th to the th quartile of the sectoral composition 023 index would imply an average decrease in GDP comovement of . . 5.2 Market Power 2019 In de Soyres and Gaillard ( ), we argued that the rise in market power (and hence a greater share of profits in overall GDP) can help solve for the Trade Comovement Puzzle by generating a disconnect between movements in production factors (capital and labor) and GDP. The mechanism can be stated as follows: firms that charge a markup have a disconnect betweenthemarginalcostandthemarginalrevenueproductoftheirinputs(whicharepartly imported). The difference between these two is accounted as value added in the form of 31

Table.11. Theroleofsectoralcompositionasanamplifierofintermediateinput’srole corrGDP All OECD HH HL LL All OECD HH HL LL ln(inter) −0.012 0.044 −0.007 −0.012 −0.007 −0.059 −1.025∗ 0.247 −0.082 −0.025 (0.008) (0.062) (0.019) (0.008) (0.017) (0.081) (0.545) (0.255) (0.088) (0.160) ln(inter)*exportprox 0.155∗∗∗ 0.227 0.266∗∗∗ 0.131∗∗ 0.147∗∗ (0.044) (0.176) (0.070) (0.052) (0.073) ln(inter)*sectorprox 0.085 1.324∗∗ −0.210 0.109 0.082 (0.097) (0.614) (0.294) (0.106) (0.188) ln(final) 0.023∗∗∗ 0.026 0.029∗ 0.020∗∗∗ 0.035∗∗ 0.006 1.705∗∗∗ 0.143 −0.053 −0.035 (0.007) (0.051) (0.017) (0.008) (0.014) (0.082) (0.578) (0.314) (0.088) (0.174) ln(final)*exportprox −0.084∗∗ −0.094 −0.229∗∗∗ −0.037 −0.112 (0.042) (0.161) (0.066) (0.048) (0.073) ln(final)*sectorprox −0.010 −1.998∗∗∗ −0.355 0.082 0.027 (0.102) (0.650) (0.347) (0.110) (0.209) exportprox 0.135 0.462 −0.775 0.489∗ 0.074 (0.249) (0.765) (0.487) (0.287) (0.566) sectorprox 0.979 −8.253∗∗∗ −4.987∗ 2.124∗∗ 0.856 (0.953) (2.859) (2.936) (1.021) (1.723) CP+TWFE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Observations 13,079 1,224 2,541 10,538 2,745 4,499 655 893 3,606 1,091 R2 0.015 0.105 0.054 0.011 0.012 0.025 0.194 0.180 0.020 0.021 Notes: ∗p<0.1;∗∗p<0.05;∗∗∗p<0.01. profits. Therefore, any change in input usage leading to a change in profits triggers a change invalueadded. SincehighermarkupsareassociatedwithhigherprofitsharesinGDP(ceteris paribus), the higher the market power, the larger the association between foreign shocks and domestic GDP movements. 1980 Since , market powers have been increased substantially. For instance, Diez et al. 2018 2018 ( )andDeLoeckerandEeckhout( )haveshownthatmarketpowershaveconsiderably 3 increased over time in almost all developed countries. Figure shows that markups have increased over time in Europe, North America, Asia and Oceania, while they seem to have been more subdued in Africa and South America. This evolution suggests that countries that experienced higher markup increases over time should also have experienced surges in the sensitivityoftheirGDPtoforeignshocks. Admittedly,thisphenomenoncouldalsocontribute to the time evolution of the TC-slope uncovered in the previous section, though studying this mechanism is beyond the scope of this paper. 32

Figure3. GlobalMarketPowers. Source: figure3inDeLoeckerandEeckhout(2018). 6 Robustness and additional exercises This section provides additional results corroborating the findings that there is a significant correlation between bilateral trade and the trade network with GDP-comovement. We first investigate if the addition of financial integration (FI), such as FDI and flows of assets significantly affects our results. We then separate intermediate inputs and capital trade flows and then include cross network effects. Finally, we provide sensitive results with respect to additional controls, alternative measures, sets of countries, and time periods. 6.1 Financial Integration: role of FDI and flows of assets Previous studies found that financial interconnection is significantly (and negatively) associ- 2013 atedwithGDPcomovements. Kalemli-Ozcanetal.( )identifiesastrongnegativeeffectof bankingintegrationonoutputsynchronization,conditionalonglobalshocksandcountry-pair heterogeneity. Such a result is consistent with a resource shifting hypothesis where an integration of capital market between two countries means that global savings are invested in the countries with the highest marginal productivity of capital – at the expense of investment in 33

32 the rest of the world. Bilateraldataonfinancialintegration(FI)isscarceforpairswithtwolowincomecountries, but it is relatively widespread for other pairs. Hence, we focus our attention on the OECD and the HL groups for this exercise and account for the role of financial flows by using the consolidated banking statistics from the Bank for International Settlement and construct an 33 index of financial proximity (FP). We use the total bilateral cross-border claims (including bank and non-bank sectors for all maturities) between countries i and j to construct an index of financial proximity (FP), such that FP = Ci→j,t +Cj→i,t , where here C refers to total ijt GDPit +GDPjt i→j,t cross-border claims from country i to country j in period t. Additionally, we control for 34 FDI which might affect GDP co-movement independently of trade proximity. We use up- 206 to-date and systematic FDI data for economies around the world from the UNCTAD’s Bilateral FDI Statistics, covering inflows, outflows, inward stock and outward stock by region 35 and economy. We use the inflows and outflows in order to construct a bilateral financial integration (FI) controls, such that: FI = FDIi→j,t +FDIj→i,t, where here FDI refers to total ijt GDPit +GDPjt i→j,t FDI from country i to country j in period t. For this measure, we report only the effect of including this control for the whole sample. The results of the benchmark specification with additional controls are shown in table 12 . The bilateral trade comovement slope and the trade network comovement slope are not affected by the inclusion of financial variables, suggesting that the link between trade and GDP comovement remains unaffected by the inclusion of these controls. 6.2 Separating capital and intermediate inputs Wecombinedinourbenchmarkspecificationtradeinintermediateinputsandtradeincapital goods. Wenowrelaxthisassumptionandtestseparatelytheeffectofbilateraltradeincapital 15 goodsandinintermediateinputs. Tradeincapitalgoodsaccountforonlyaround %oftotal trade, and some country-pairs do not trade capital goods. We construct the index relative to 32 Inotherwords,ifsavingscanbeallocatedacrossborders,apositivetechnologyshockinonecountryrelative toitspartnerscreatesaninflowofcapitalintothiscountryattheexpenseofothereconomies. 33 Thedatasetisavailablehere: https://stats.bis.org/. 34 AccordingtoFontagné(1999),tradeandFDIarepositivelycorrelated,whichimpliesthatfailingtocontrolfor FDIislikelytobiasourestimatesoftherelationshipbetweentradeandGDPcorrelation. 35 Data are in principle collected from national sources. In order to cover the entire world, where data are not available from national sources, data from partner countries (also called mirror data) as well as from other internationalorganizationshavealsobeenused. DatacanbedownloadedontheUNCTADwebsite. 34

Table.12. Effectoffinancialintegration corrGDP All All All All OECD OECD HL HL ln(inter) 0.169∗∗ 0.164∗∗ −0.006 −0.006 0.405∗∗∗ 0.402∗∗∗ −0.013 −0.015 (0.078) (0.078) (0.018) (0.018) (0.093) (0.094) (0.023) (0.023) ln(final) −0.091∗ −0.088 0.008 0.007 −0.059 −0.054 0.023 0.021 (0.054) (0.054) (0.015) (0.015) (0.046) (0.047) (0.018) (0.018) network1st 0.802 0.832 0.855∗∗∗ 0.855∗∗∗ −0.558 −0.573 0.941∗∗∗ 0.942∗∗∗ (0.634) (0.636) (0.189) (0.189) (0.807) (0.804) (0.218) (0.218) network2nd −1.620 −1.650 −0.112 −0.111 −0.616 −0.681 0.404 0.378 (1.735) (1.731) (0.568) (0.569) (1.936) (1.876) (0.650) (0.649) BISindex −23.453 (27.939) FDIindex 0.531 −6.283 15.823∗∗ (3.573) (5.380) (6.547) CP+TWFE Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Observations 846 846 2,947 2,947 492 492 2,030 2,030 R2 0.098 0.099 0.026 0.026 0.229 0.233 0.021 0.024 Notes: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01. Notethatcolumn1showstheresultsusingthe sub-samplecontainingdataontheBISindexwhilecolumn3showstheresultusing onlycountry-pairswithdataontheFDIindex. capital 3 capital goods Trade as we did in section . We then test the following specification: ijt Corr GDP = β ln(Tradeinter)+β ln(Trade capital)+β ln(Trade final) ijt 1 ijt 2 ijt 3 ijt +γγγnetwork +X +CP +TW +(cid:101) ( 35 ) ijt ijt ij t ijt 13 The results are gathered in table , where we also show the benchmark specification without separating trade flows. Overall, our previous results are still valid except that now intermediate inputs seem to also play a role for HL pairs. Trade in capital goods is, however, not significant for the OECD and LL groups. Interestingly, trade in capital goods seems to be positively correlated with more GDP comovement for HH, while we find the opposite for HL, suggesting again possible heterogeneity in the effect of bilateral trade flows along the development stage for capital goods. 35

Table.13. TradeComovementslopewithdisaggregatedtradeindexandcapitalflows corrGDP All All OECD OECD HH HH HL HL LL LL ln(inter) 0.009 0.103∗∗∗ 0.034∗∗ 0.007 0.018∗∗∗ (0.005) (0.030) (0.014) (0.006) (0.006) ln(inter) 0.014∗∗ 0.117∗∗∗ 0.026∗∗ 0.016∗∗∗ 0.017 (0.006) (0.031) (0.013) (0.006) (0.012) ln(capital) −0.010∗∗∗ −0.006 0.028∗∗∗ −0.015∗∗∗ −0.012 (0.004) (0.019) (0.009) (0.004) (0.008) ln(final) 0.012∗∗ 0.016∗∗∗ −0.009 −0.013 −0.009 −0.010 0.016∗∗∗ 0.021∗∗∗ 0.017∗∗∗ 0.024∗∗ (0.005) (0.005) (0.025) (0.024) (0.012) (0.013) (0.005) (0.005) (0.005) (0.011) network1st 0.305∗∗∗ 0.349∗∗∗ 1.046∗∗∗ 1.026∗∗∗ 0.521∗∗∗ 0.603∗∗∗ 0.264∗∗∗ 0.301∗∗∗ −0.040 0.034 (0.067) (0.069) (0.259) (0.258) (0.163) (0.163) (0.073) (0.076) (0.073) (0.159) network2nd 0.416∗ 0.490∗∗ 2.906∗∗∗ 2.971∗∗∗ −0.170 0.119 0.518∗∗ 0.613∗∗ 1.076∗∗∗ 1.199∗∗ (0.221) (0.226) (0.938) (0.938) (0.550) (0.556) (0.243) (0.249) (0.243) (0.492) CP+TWFE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Observations 13,079 12,702 1,224 1,222 2,541 2,508 10,538 10,194 2,745 2,614 R2 0.009 0.011 0.099 0.103 0.040 0.047 0.005 0.009 0.008 0.009 Notes: ∗p<0.1;∗∗p<0.05;∗∗∗p<0.01. 6.3 Cross network effects In the baseline specification, we have estimated two different network effects: a first order network effect and a second order network effect. We now test the implication of a third network index which aims to capture how a country i is similar in terms of trade partners to the partners of a country j. This measure is described in section 3 , and table 14 provides the results and show that the inclusion of such a measure does not change the estimated coefficients regarding bilateral trade flows and first order network effects. However, it influences point estimates associated with the second order network measure, which is likely to be the case, as those two indexes measure trade network effects taking place further down in the trade network. 6.4 Other robustness exercises 15 Ourresultsarerobusttoanumberofotheralternativespecificationsthatwegatherintable . We first confirm that among non-OECD pairs, trade in final goods is significantly associated with more GDP-comovement. When considering only the last two time windows, we find 36

Table.14. TradeComovementSlopewithcrossnetworkeffects corrGDP All All OECD OECD HH HH HL HL LL LL ln(inter) 0.009 0.008 0.103∗∗∗ 0.107∗∗∗ 0.034∗∗ 0.035∗∗ 0.007 0.006 0.018∗∗∗ 0.017 (0.005) (0.005) (0.030) (0.031) (0.014) (0.014) (0.006) (0.006) (0.006) (0.011) ln(final) 0.012∗∗ 0.012∗∗ −0.009 −0.009 −0.009 −0.009 0.016∗∗∗ 0.016∗∗∗ 0.017∗∗∗ 0.017∗ (0.005) (0.005) (0.025) (0.025) (0.012) (0.012) (0.005) (0.005) (0.005) (0.010) network1st 0.305∗∗∗ 0.294∗∗∗ 1.046∗∗∗ 1.084∗∗∗ 0.521∗∗∗ 0.523∗∗∗ 0.264∗∗∗ 0.246∗∗∗ −0.040 −0.035 (0.067) (0.067) (0.259) (0.264) (0.163) (0.165) (0.073) (0.074) (0.073) (0.153) network2nd 0.416∗ 0.364∗ 2.906∗∗∗ 3.019∗∗∗ −0.170 −0.160 0.518∗∗ 0.432∗ 1.076∗∗∗ 0.866∗ (0.221) (0.221) (0.938) (0.961) (0.550) (0.554) (0.243) (0.242) (0.243) (0.504) networkcross 0.166 −0.330 −0.056 0.245∗ 0.343 (0.115) (0.409) (0.258) (0.126) (0.279) CP+TWFE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Observations 13,079 13,079 1,224 1,224 2,541 2,541 10,538 10,538 2,745 2,745 R2 0.009 0.009 0.099 0.099 0.040 0.040 0.005 0.006 0.008 0.009 Notes: ∗p<0.1;∗∗p<0.05;∗∗∗p<0.01. consistent results except for sub-samples with the lowest number of observations (the LL and HH groups), for which bilateral trade flows turn out to be statistically insignificant. Network effects, however, are in line with previous results. We then find that under alternative sectoral proximity indexes, the results remain very 2 similar when using ISIC export proximity or SITC -digit export proximity. Finally, we confirm the general pattern using three additional analysis: (i) an alternative measure of trade proximity built using the mean of log trade intensities instead of the log of the mean trade intensities within time-windows, (ii) an alternative measure using the max (cid:16) (cid:17) operator when computing the trade intensities: ln(Trade) = max T /GDP,T /GDP , i↔j i i↔j j 36 (iii) BK filtered GDP instead of HP filtered GDP, (iv) First difference instead of HP-filtered GDP. For the alternative bilateral trade measures using the max operator or the mean log, the results are very close to the benchmark specification. With the alternative filters, all results areconsistentexceptfortheLLgroupforwhichthecorrelationbetweentradeinintermediate inputsandGDP-comovementturnsouttobepositiveandsignificantinthecaseofBK-filtered GDP. 36 Inthisexercise,wekeepfluctuationsbetween32and200quarters. 37

Table.15. Sensitiveanalysis: TradeandGDP-comovement Estimatedcoefficient ln(input) ln(final) network1st network2nd Sample Pairs|Obs. (i)Sampleselection AlternativeGroup 0.007 0.012** 0.25*** 0.23 Non-OECD 4094|11830 Period(1990:2009) 0.206*** -0.107* 1.38* 1.93 OECD 316|616 Period(1990:2009) 0.025 0.018 0.98** 0.56 HH 754|1434 Period(1990:2009) 0.044*** 0.034*** 0.49*** 3.80*** HL 3598|6582 Period(1990:2009) 0.027 0.004 0.64*** 4.09*** LL 1140|1972 (ii)Alternativecontrols SITC2-digitexportprox 0.097*** -0.007 0.96*** 2.75*** OECD 320|1224 SITC2-digitexportprox 0.036*** -0.008 0.53*** -0.30 HH 764|2533 SITC2-digitexportprox 0.009 0.017*** 0.27*** 0.49** HL 3650|10521 SITC2-digitexportprox 0.018* 0.017* -0.01 1.08** LL 1165|2744 ISICexportprox 0.103*** -0.010 1.02*** 2.77*** OECD 320|1224 ISICexportprox 0.038*** -0.011 0.56*** -0.29 HH 764|2533 ISICexportprox 0.008 0.016*** 0.28*** 0.49** HL 3650|10521 ISICexportprox 0.018 0.017* -0.02 1.09** LL 1165|2744 (iii)AlternativeMeasures meanln(index) 0.111*** -0.034 1.05*** 2.68*** OECD 320|1224 meanln(index) 0.037*** -0.016 0.58*** -0.28 HH 764|2533 meanln(index) 0.007 0.016*** 0.26*** 0.45* HL 3650|10521 meanln(index) 0.017 0.016* -0.04 1.03*** LL 1165|2744 Maxtradeindexc 0.103*** -0.013 1.06*** 2.92*** OECD 320|1224 Maxtradeindexc 0.038*** -0.011 0.57*** -0.18 HH 764|2533 Maxtradeindexc 0.009 0.018*** 0.26*** 0.47* HL 3650|10521 Maxtradeindexc 0.018 0.017* -0.05 1.12** LL 1165|2744 BKfilter 0.115*** -0.010 1.05*** 3.24*** OECD 324|1260 BKfilter 0.047*** -0.017 0.57*** -0.10 HH 320|1224 BKfilter 0.010* 0.014*** 0.26*** 0.52** HL 3650|10521 BKfilter 0.026** 0.015 -0.05 1.11*** LL 1165|2744 Firstdifference 0.078*** 0.004 0.67** 2.74*** OECD 320|1224 Firstdifference 0.017 -0.025* 0.59*** -0.38 HH 764|2533 Firstdifference 0.009 0.023** 0.25*** 0.62** HL 3650|10521 Firstdifference 0.021** 0.020** 0.14 1.03** LL 1165|2744 Notes: ∗p<0.1;∗∗p<0.05;∗∗∗p<0.01. Inparenthesis: std. deviation. a We provide the results using EU and USSR dummies since adding those controls substantially reducethesignificanceoftradeinfinalgoods. b FDI and BIS are mostly available for OECD and HH groups. We report only the robustness with thosetwogroups. (cid:16) (cid:17) c Wedefinemaxtradeindexasthemeasureusingmax T i↔j /GDP i ,T i↔j /GDP j . 38

7 Conclusion Thispapertakesafreshlookintoanoldquestion: whatistheassociationbetweentradeflows and GDP comovement at business cycle frequencies? Guided by a simple theory, we provide novel evidence on the role of both bilateral and global trade flows and emphasize the strong interactionarisingbetweenbilaterallinkagesandtheglobaltradenetwork,whichimpliesthat thepreviouslystudiedTrade-Comovementslopeshouldnotbe–andindeedisnot–constant over time. Taking a closer look at different income groups, we also present new facts on the role of sectoral composition and on the type of trade that seems to be associated with GDP correlation. Looking ahead, we believe the paper provides interesting scope for future research. Most notably, it becomes crucial to understand why the TC-slope seems to be mostly driven by trade in intermediate inputs in developed countries, while it is mostly driven by trade in final goods in developing countries. Moreover, the TC-slope has significantly increased over time. While the literature has documented the possible role of the global rise of markups, it seems important to investigate further the channels that could explain this pattern. 39

References 2016 Barrot, J.-N. and Sauvagnat, J. ( ). Input Specificity and the Propagation of Idiosyncratic Shocks in Production Networks *. The Quarterly Journal of Economics, 131 ( 3 ): 1543 – 1592 . Basu,S.andFernald,J.G.( 2002 ). Aggregateproductivityandaggregatetechnology. European Economic Review, 46 ( 6 ): 963 – 991 . 2005 Baxter, M. and Kouparitsas, M. A. ( ). Determinants of business cycle comovement: a robust analysis. Journal of Monetary Economics, 52 ( 1 ): 113 – 157 . 2011 Bems, R., Johnson, R. C., and Yi, K.-M. ( ). Vertical linkages and the collapse of global trade. American Economic Review, 101 ( 3 ): 308 – 12 . 2008 Burstein, A., Kurz, C., and Tesar, L. ( ). Trade, production sharing, and the international transmission of business cycles. Journal of Monetary Economics, 55 ( 4 ): 775 – 795 . 2007 Calderon, C., Chong, A., and Stein, E. ( ). Trade intensity and business cycle synchronization: Are developing countries any different? Journal of international Economics, 71 ( 1 ): 2 – 21 . Clark, T. and van Wincoop, E. ( 2001 ). Borders and business cycles. Journal of International Economics, 55 ( 1 ): 59 – 85 . Comin, D. and Gertler, M. ( 2006 ). Medium-term business cycles. American Economic Review, 96 3 523 551 ( ): – . 2018 24768 De Loecker, J. and Eeckhout, J. ( ). Global market power. Working Paper , National Bureau of Economic Research. 2019 de Soyres, F. and Gaillard, A. ( ). Value added and productivity linkages across countries. FRB International Finance Discussion Paper No. 1266. 2010 Di Giovanni, J. and Levchenko, A. A. ( ). Putting the parts together: Trade, vertical linkages, and business cycle comovement. American Economic Journal: Macroeconomics, 2 ( 2 ): 95 – 124 . 2016 di Giovanni, J., Levchenko, A. A., and Mejean, I. ( ). The Micro Origins of International 21885 Business Cycle Comovement. NBER Working Papers , National Bureau of Economic Research, Inc. Diez, M. F., Leigh, M. D., and Tambunlertchai, S. ( 2018 ). Global market power and its macroeconomic implications. International Monetary Fund. 40

2015 Duval, R., Li, N., Saraf, R., and Seneviratne, D. ( ). Value-added trade and business cycle synchronization. Journal of International Economics, pages 251 – 262 . 1999 Fontagné, L. ( ). Foreign Direct Investment and International Trade: Complements or 1999 3 Substitutes? OECD Science, Technology and Industry Working Papers / , OECD Publishing. 1998 Frankel,J.A.andRose,A.K.( ). TheEndogeneityoftheOptimumCurrencyAreaCriteria. Economic Journal, 108 ( 449 ): 1009 – 25 . 2019 Huo, Z., Levchenko, A. A., and Pandalai-Nayar, N. ( ). The Global Business Cycle: Mea- 669 surement and Transmission. Working Papers , Research Seminar in International Economics, University of Michigan. Imbs, J. ( 2004 ). Trade, finance, specialization, and synchronization. Review of Economics and Statistics, 86 ( 3 ): 723 – 734 . 2008 Inklaar, R., Jong-A-Pin, R., and De Haan, J. ( ). Trade and business cycle synchronization in oecd countries: A re-examination. European Economic Review, 52 ( 4 ): 646 – 666 . Johnson, R. C. ( 2014 ). Trade in intermediate inputs and business cycle comovement. American Economic Journal: Macroeconomics, 6 ( 4 ): 39 – 83 . 2013 Kalemli-Ozcan, S., Papaioannou, E., and Peydro, J.-L. ( ). Financial regulation, financial globalization, and the synchronization of economic activity. The Journal of Finance, 68 3 1179 1228 ( ): – . 2008 Kehoe, T. J. and Ruhl, K. J. ( ). Are shocks to the terms of trade shocks to productivity? Review of Economic Dynamics, 11 ( 4 ): 804 – 819 . 2006 Kose,M.A.andYi,K.-M.( ). Canthestandardinternationalbusinesscyclemodelexplain the relation between trade and comovement? Journal of International Economics, 68 ( 2 ): 267 – 295 . 2015 Liao, W. and Santacreu, A. M. ( ). The trade comovement puzzle and the margins of international trade. Journal of International Economics, 96 ( 2 ): 266 – 288 . Ng, E. C. ( 2010 ). Production fragmentation and business-cycle comovement. Journal of International Economics, 82 ( 1 ): 1 – 14 . 41