Time to produce and emerging market crises

Journal of Monetary Economics 68 (2014) 37–52

Contents lists available at ScienceDirect

Journal of Monetary Economics journal homepage: www.elsevier.com/locate/jme

Time to produce and emerging market crises$ Felipe Schwartzman n Federal Reserve Bank of Richmond, Research Department, Richmond, VA 23220, United States

a r t i c l e i n f o

abstract

Article history: Received 17 July 2012 Received in revised form 21 July 2014 Accepted 21 July 2014 Available online 31 July 2014

After emerging market crises, value added falls more in manufacturing industries that normally exhibit higher inventory/cost ratios. Moreover, the difference in value added between manufacturing industries with different inventory/cost ratios persists years into the recovery. A shock to aggregate TFP cannot by itself match this pattern. In contrast, a persistent increase in the cost of foreign capital can. In the context of a calibrated multisector small open economy model, a shock to the cost of foreign capital consistent with the cross-industry data leads, 3–5 years after the onset of the crisis, to an average reduction of output relative to a trend of 5.4 percent. & 2014 Elsevier B.V. All rights reserved.

Keywords: Working capital Time to build Crises Inventories Emerging markets

1. Introduction A key decision for policymakers in developing countries is whether or not to integrate the national economy in international financial markets, thus turning it into an “emerging market” for foreign investors. Critics of financial globalization argue that emerging economies’ access to international financial markets is unstable and that this leads to macroeconomic instability.1 A major piece of the evidence for that view are recurrent emerging market crises, in which pronounced and persistent output collapses occur in tandem with large scale reversals of foreign capital inflows. Nevertheless, the interpretation of those episodes as being primarily a reflection of interest rate rises has not remained unchallenged. Notably, Aguiar and Gopinath (2007) have argued that the change in the direction of capital flows observed in these episodes is more likely a consequence rather than a cause of the accompanying output collapses, as domestic residents increase their savings in response to a persistent decline in their income. This paper brings cross-industry data to bear on the question of whether the large and persistent drops in output observed after emerging market crises can be traced to an increase in the cost of capital available to domestic firms. In this case, it is also true that industries that take longer to produce and distribute goods, thus normally exhibiting a higher ☆ An earlier and much revised version of the paper is available as a Federal Reserve Bank of Richmond Working Paper, 10-15R. For comments and suggestions I thank Urban Jermann, an anonymous referee and seminar participants and commenters at ISIR 2012, the SCIEAM 2011, the Macro Midwest Meetings 2011, the VII LAFN Workshop 2011, at UC Santa Barbara, UC Santa Cruz, the San Francisco Fed, University of Warwick, NCSU, PUC-Rio, EPGE-FGV, UCL, and the Bank of Spain, as well as several colleagues and professors at Princeton and at the Richmond Fed. I also thank Nobu Kiyotaki and Markus Brunnermeier for advising me in this work, Pierre-Daniel Sarte, Alex Wolman, and Sangeeta Pratap for detailed comments, and Jonathan Tompkins for excellent research assistance. The views expressed in this paper are those of the author and do not reflect the position of the Federal Reserve Bank of Richmond or the Federal Reserve System. n Tel.: þ 1 804 697 4438. E-mail address: [email protected] 1 See Kose et al. (2009) for a comprehensive review of the evidence in favor and against financial globalization. Korinek (2011) reviews the theoretical literature making the welfare case for prudential capital controls aimed at stabilizing foreign capital flows.

http://dx.doi.org/10.1016/j.jmoneco.2014.07.010 0304-3932/& 2014 Elsevier B.V. All rights reserved.

38

F. Schwartzman / Journal of Monetary Economics 68 (2014) 37–52

inventory/cost ratio, should be more intensely affected by the crises. In contrast, there is no a priori reason why an aggregate productivity shock would generate such a cross-industry pattern. The key empirical result of this paper is that, indeed, after emerging market crises, manufacturing industries that normally exhibit larger inventory/cost ratios also take longer to recover. The difference between industries is particularly significant 3–5 years after the episodes. The finding is robust to the inclusion of a large number of controls, and across sub-samples. What does the cross-industry data imply for the impact of a shock to the cost of capital on aggregate output? A quantitative equilibrium model provides an answer. The model features multiple industries, each equipped with a specific time to produce and distribute goods (time to produce, for short) calibrated to match long term averages of industry-specific inventory/cost ratios. The model indicates that the cross-industry variation observed in the data is consistent with a gradual increase in the cost of foreign capital relative to its pre-crisis level. In the context of the model, this increase generates a reduction in output relative to its trend of 5.4 percent, taking the average over those same 3–5 years. This decline accounts for 48 percent of the overall average deviation of GDP relative to trend. In the quantitative model, production is subject to “time to produce” technology, that is, a technology that requires that some of the inputs be used before the final output is produced. This production function is similar to the working capital constraints emphasized in prominent papers in the emerging market business cycle literature such as Neumeyer and Perri (2005), Uribe and Yue (2006), and Mendoza (2010), whereby firms need to acquire variable inputs in advance of production.2 Relative to existing formulations in the literature, it has the added advantage that, as discussed in detail in Section 2, “time to produce” has clear implications for inventory data.3 The results based on cross-industry evidence complement the findings of Garcia-Cicco et al. (2010) and Chang and Fernandez (2013) using time series data, pointing to a prominent role for financial shocks in emerging market business cycles. The result of a persistent rise in the cost of capital after the crisis also echoes findings by Cerra and Saxena (2008) and Reinhart and Rogoff (2009), both of which document persistent drops in output relative to trend after financial crises. The empirical findings are complementary to independent work by Tong and Wei (2009). Using data from firms in emerging markets, they show that higher inventory-to-sales ratios were associated with larger drop in share prices in the aftermath of the Great Recession. Raddatz (2006) presents another set of related empirical findings: industries with high inventory-to-sales ratios are more volatile in countries with a low level of financial development. Lastly, the paper is also related to work by Alessandria et al. (2010), who study the implications of inventory holdings for price dynamics in the aftermath of emerging market crises. Section 2 delineates the relationship between inventory/cost ratios and the response to persistent interest rate and productivity shocks in the context of the decision problem of a single firm, providing the key intuitions for the rest of the paper. Section 3 presents the data analysis. Section 4 introduces a multisector equilibrium model to validate the partial equilibrium intuitions in Section 2 and investigate the implications of the findings in Section 3 for aggregate quantities. The last section concludes.

2. Time to produce and distribute goods, inventories, and the impact of shocks The discussion that follows relies on the analysis of optimal firm decisions given a “time to produce and distribute” (time to produce, for short) model of the production technology. This technology induces the holding of inventories, and underpins the analysis of the relationship between the effects of different kinds of shocks and the inventory/cost ratio. The framework differs from other approaches to modeling inventories. The model provides an intermediate step between the reduced form but highly tractable inventory-in-production-function approach in Ramey (1989) and the tightly microfounded but, for a multi-sector model, prohibitively untractable fixed cost approach in Khan and Thomas (2007). A primary advantage of the approach, for the purposes of the paper, is that it shares with working capital models common in the emerging markets business cycle literature an emphasis on timing restrictions as a key source of borrowing needs. This allows for a more direct comparison with previous work. Firms are endowed with the following Cobb–Douglas production function: S

Y t ¼ ∏ Z t  s ðsÞωðsÞ ; s¼0

S

∑ ωðsÞ ¼ 1;

ð1Þ

s¼0

where Yt is sales at t, Z t  s ðsÞ, s ¼ f0; …; Sg are dated composite inputs with the time subscripts t  s referring to when the inputs are included in the production process, the numbers in brackets refer to the time elapsed between the assignment of 2 Working capital also plays a prominent role in business cycle models designed to study advanced economies, such as Jermann and Quadrini (2012). Such a working capital channel has been the object of empirical investigations, notably by Barth and Ramey (2001). 3 The narrow focus on inventories as opposed to a broader focus that includes other components of working capital such as trade credit and cash, and marketable securities is justified for three reasons: (a) inventories are clearly associated with the use of variable inputs, (b) the opportunity cost of holding inventories is proportional to the real interest rate, rather than the nominal interest rate or some spread and (c) the ranking of industries with respect to inventory/cost ratios is likely to be linked to technological differences between industries. The supplementary materials available online present the argument in more detail. The lack of clear motivation for the focus on real interest rates has been an important source of criticism for the literature on emerging market business cycles (Chari et al., 2005).

F. Schwartzman / Journal of Monetary Economics 68 (2014) 37–52

39

the input to the production process and the sales of the output, and ωðsÞ are constants denoting the share of production costs accounted for by inputs produced on each date. The composite goods Z t  s ðsÞ's are aggregates of individual inputs: Z t  s ðsÞ ¼ K t  s ðsÞ1  αL  αM ðAt  s Lt  s ðsÞÞαL M t  s ðsÞαM ;

sA f0; …; Sg;

ð2Þ

where K t  s ðsÞ, Lt  s ðsÞ and M t  s ðsÞ are, respectively, the fixed capital, the labor and the materials used at date t  s for production of Z t  s ðsÞ, At  s is a time-varying labor-augmenting productivity parameter, αL and αM are parameters between 0 and 1 with αL þ αM o 1, denoting the shares of, respectively, labor and materials in production. As an example, the sale of a sweater in t requires the acquisition and storage of different varieties of wools in t  2 (incorporated in Z t  2 ð2Þ), the labor effort and machinery involved in knitting different components in t  1 (and incorporated in Z t  1 ð1Þ) and the final assembly and the sales efforts in t (incorporated in Z t ð0Þ). For the purpose of the analytical results in this section, we assume that capital used at each stage is a fixed factor owned by the firm, so that K t ðsÞ ¼ K for all t and s. As the analysis in Section 4 will demonstrate, this is not a crucial assumption. The firm's cash flow on a given date t þ r is S

Π t þ r ¼ P t þ r Y t þ r  ∑ ½W t þ r Lt þ r ðsÞ þQ t þ r Mt þ r ðsÞ;

ð3Þ

s¼0

where P t þ r is the price of the finished good at tþr, W t þ r is the wage rate and Q t þ r is the price of the materials. Between periods t þ r 1 and tþr, the firm discounts future profits using the market interest rate Rt þ r , so that at time t profits at tþr are discounted at the rate ∏rz ¼10 Rt þ z . Given perfect foresight, the problem of the firm planning its production at date t is to ð1;SÞ choose sequences for labor input used at each stage fLt þ r ðsÞgð1;SÞ ðr;sÞ ¼ ð0;0Þ , and materials used at each stage fM t þ r ðsÞgðr;sÞ ¼ ð0;0Þ so as to maximize the present discounted value of Πt þ r for all r A f0; …; 1g. The maximization is done subject to the production ð1;SÞ function defined by Eqs. (1) and (2) and with fM t  r ðsÞgð1;SÞ ðr;sÞ ¼ ð1;0Þ , fLt  r ðsÞgðr;sÞ ¼ ð1;0Þ predetermined. With some manipulation, the optimality conditions for labor used s periods in advance of sales are, for all t, r Z 0 and s Z 1,
s1  ∏ Rt þ r þ z  W t þ r Lt þ r ðsÞ ¼ P t þ r þ s ωðsÞαL Y t þ r þ s : ð4Þ z¼0

The same manipulations yield an analogous optimality condition for materials. With a Cobb–Douglas production function, the cost of using any given input is a constant fraction of the value supplied. Because inputs are used in advance, their cost is given by their spot prices multiplied by the accumulated interest between the time that inputs are demanded and the time that output is supplied. 2.1. Time to produce and inventory accounting Standard inventory accounting is based on estimates of production costs.4 The real change in end-of-period inventories is the difference between the cost of production taking place within the period and the cost of the goods sold (COGS) within the same period. Algebraically, S

S

X t ¼ X t  1 þ ∑ ½WLt ðsÞ þ QMt ðsÞ  ∑ ½WLt  s ðsÞ þ QM t  s ðsÞ ; s¼0 s¼0 |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} Current Production Costs

ð5Þ

Cost of Goods Sold ðCOGSÞ

where Xt are real inventory holdings at the end of period t and costs are calculated at constant prices which, for expositional simplicity, are set at their steady-state values, W and Q. The difference between the elements in the two summations is in the date subscripts—whereas all the terms in the first summation are dated t, the terms in the second are dated t s. Iterating backward and applying the boundary condition limt-  1 X t ¼ 0 (firms start out with zero inventories) yield S

s

X t ¼ ∑ ∑ ½WLt  r ðsÞ þ QMt  r ðsÞ;

ð6Þ

s¼1r ¼1

so that inventories are a triangular summation over all the dated inputs. Steady-state inventories satisfy S

X ¼ ∑ s½WLðsÞ þQMðsÞ;

ð7Þ

s¼1

so that, in steady-state, inputs that are used s periods in advance of sales appear in inventories s times. Since in steady-state the current costs and the cost of goods sold are the same, the steady-state inventory/cost ratio, denoted τ, is S

τ¼ ∑

WLðsÞ þ QMðsÞ

s:

S s ¼ 1 ∑s ¼ 0 ½WLðsÞ þ QMðsÞ

4

See Financial Accounting Standards Board (2013, 330-10-30-1).

ð8Þ

40

F. Schwartzman / Journal of Monetary Economics 68 (2014) 37–52

Substituting in the steady-state versions of Eqs. (4) and its analog for materials, and taking the limit as R-1, yield the approximation: S

τ ffi ∑ ωðsÞs:

ð9Þ

s¼1

Given this expression for the steady-state inventory/cost ratio, it is now possible to discuss how cross-industry variation in that ratio is associated with the variation in the responses to interest rate and productivity shocks. 2.2. The steady-state inventory/cost ratio and persistent shocks to A and R This subsection describes the long-run impact of permanent exogenous changes in productivity (A) and the interest rate (R) on industry output in a partial equilibrium setting. Focusing on long-run and partial equilibrium results yields analytical tractability, but as shown in Section 4, is not key for the main results. Consider the firm's input and output choices in a steady-state with Rt ¼ R and At ¼ A. With some algebraic manipulation, it follows that "
αL
αM
S αL þ αM #1=ð1  αL  αM Þ A 1 ∏ ðRs Þ  ωðsÞ ; ð10Þ Y ¼ C  P αL þ αM W Q s¼0 where the multiplicative constant C is a function of the production function parameters αL and ωðsÞ's and of the fixed capital K. Impact of a shock to A: Suppose a firm is in steady-state until a given date t n , when there is a one time permanent and unexpected shock to productivity A. For t between t n and t n þS, sales Yt are constrained by past input choices. For t Zt n þ S onward there is no longer any constraint from past decisions so that the firm reaches a new steady-state. From Eq. (10), it is immediate that, for t Zt n þS the elasticity of sales to a permanent change in aggregate TFP shock at t n is ∂Y A αL ¼ : ∂A Y 1  αL  αM

ð11Þ

The weights on different stages of production ωðsÞ do not enter the expression. Hence, after an initial transition period, the effect of the shock should not, all else equal, be correlated with the pattern of time to produce and distribute the finished good. Eq. (11) provides the elasticity for both sales and output relative to the TFP shock. This is because the difference between the two is given by the change in inventories. Absent trend-growth, after the transition period inventory investment reverts to zero. Thus, aggregate output increases by the same amount as aggregate sales. Using the “double deflation” method, the Laspeyres volume index for value added using t n 1 as the base year is, after the transition period, proportional to V tn þ h p P tn  1 Y tn þ h  Q tn  1 Mtn þ h ;

h ZS;

ð12Þ

n

where V t n þ h denotes value added at t þ h. From the factor demand Eq. (4) it follows that given a shock to A which does not affect prices or the interest rate and given hZ S, V tn þ h is proportional to P t n  1 ð1  ∑Ss ¼ 0 ðωðsÞ=Rs ÞαM ÞY tn þ h . Thus, the longrun elasticity of value added to a TFP shock is, in partial equilibrium, the same as that of output. Effect of a shock to R: Consider now the effect of a permanent increase in the interest rate. Again, there is an initial transition period, as past input decisions constrain current output. Again a new steady-state is reached at t n þS. From Eq. (10), the elasticity of the sales at t Zt n þ S to a permanent increase in the interest rate at t n is ∂Y R αL þ αM S ¼ ∑ ωðsÞs: ∂R Y 1  αL  αM s ¼ 1

ð13Þ

The permanent interest rate shock has an effect on sales that is directly related to the weights on different stages of production ωðsÞ. From Eq. (9), ∂Y R αL þ αM ffi τ: ∂R Y 1  αL  αM

ð14Þ

Hence, all else equal, the elasticity of sales to a persistent change in the interest rate is directly proportional to the steady-state ratio between inventories and the cost of goods. Again, in steady-state, sales are equal to output. The elasticity of the change in the volume index for value added in response to a persistent interest rate cost is approximately:   ∂V R αL þ αM αM ffi  τ: ð15Þ ∂R V 1  αL  αM 1  αM The elasticity of the change in value added to the interest rate shock differs from that of output because in value added accounting inputs are deducted at their spot prices rather than at their opportunity costs, with the latter itself affected by the interest rates. We now have the key elements necessary to interpret the results in the data analysis section that follows.

F. Schwartzman / Journal of Monetary Economics 68 (2014) 37–52

41

3. Data analysis This section discusses the cross-industry impact of emerging market crises. The results point to strong and robust crossindustry differences in the impact of the crises, which correlate with differences in the inventory/cost ratio. 3.1. Events The empirical analysis centers on events in which capital inflows to an emerging market fall by an unusually large amount and where GDP growth is negative either in the same year or in the following year. The period under consideration is the 30 years from 1975 to 2004. These years include the period of rapid financial globalization. I only consider countries with populations above 1 million inhabitants and which were not immersed in civil war during the episode.5 Subject to these filters, I include all countries for which data is available. Most of the data analysis excludes advanced economies, defined as members of the OECD. The focus of the paper on emerging economies follows from the observation in Prasad et al. (2007) that international financial integration may have different effects for less developed economies in comparison to advanced ones. Specifically, it is likely to lead to more macroeconomic instability in countries with lower quality macroeconomic policies and institutions, but lower instability in countries where those are of higher quality. Associated to this difference is the high emphasis that the literature on emerging market business cycles places on interest rate shocks, as compared to developed economies. In Section 3.5 I report results using data from OECD countries which indicate that, to the extent that developed economies have experienced episodes of simultaneous output drop and current account reversals, these do not exhibit the same cross-industry patterns as in less developed economies. This suggests that, among OECD countries, those episodes were not a reflection of instability in the cost of foreign finance.6 The episodes take place in years satisfying the following criteria: (1) There is a large drop in capital inflows. Capital inflows are proxied on a monthly basis by the sum of the trade deficit and declines in international reserves, normalized by the linear trend of GDP.7 A capital outflow event occurs when the difference between capital inflows over a 12-month period and the capital inflows in the previous 12 months is more than two standard deviations below the mean change in capital flows, with both the mean and standard deviation calculated using all data from 1975 to the present but excluding the events themselves.8 (2) Two capital outflow events that either occur in consecutive years or are separated by a single year are part of a single event. (3) There is a drop in GDP either in the year when the capital outflow event starts or in the subsequent year. Including drops that only take place in the following year accommodates episodes that occur late in the year. The episodes are listed in Table 1, and they include most episodes identified by Calvo et al. (2006).9 Fig. 1 shows the average deviation from trend of different aggregate quantities across episodes. The trend is projected taking the year before the episode as a starting point, and using the average growth rate for GDP in the 10 years prior to the corresponding event. One striking fact is that the different variables deviate very persistently from their prior trend. This is consistent with the findings in Cerra and Saxena (2008) and Reinhart and Rogoff (2009) that financial crises have persistent impacts on output. The data also depicts a key finding from Calvo et al. (2006), that investment drops more persistently than output after crises, implying a persistent drop in the investment/output ratio. We will return to these facts when comparing the different models and shock processes in Section 4. 3.2. Empirical specification The results in Section 2 suggest that it is possible to identify interest rate shocks from the behavior of the cross-section of industries. This section lays out a specification for regression equations based on that model. In Section 4, I confirm that the specification continues to perform well in a more general setting. Assuming that within each year value added is well approximated by its steady-state value given the prevailing prices and TFP, take the logarithm of the supply function in Eq. (10), apply the definition of value added (12), the substitutions 5

I use the same dates as Cerra and Saxena (2008) to determine whether a country was at war or not. To answer the more general question about whether the interaction of financial shocks with inventories has played a role in explaining business cycles in developed economies requires a full analysis of the time-series. This is the approach taken by Lubik et al. (2014) for the U.S. economy. 7 The proxy follows from the identity capital inflows þ current account ¼ change in reserves, together with the observation that changes in the current account are mostly due to changes in the trade balance. The methodology follows Calvo et al. (2006). 8 This requires an iterative procedure where events are first calculated given overall mean and standard deviations, then the moments are recalculated excluding the event data, generating possibly new events and repeating the process. 9 The selection criteria are similar to Calvo et al. (2006), except that there is no attempt to select exogenous or unpredictable events. Concretely, because there is no attempt to claim exogeneity or unpredictability, there is no filter for international financial conditions. Also, for the same reason, the selection of the episodes relies on a comparison with both post- and pre-crisis data as opposed to only pre-crisis data as in their paper. Analysis of the episodes in Calvo et al. (2006) presents the same patterns, albeit with larger standard errors because of the smaller sample size. Thus, these events are not treated as natural experiments, but as interesting episodes which generate cross-correlations that merit study with a structural model. 6

42

F. Schwartzman / Journal of Monetary Economics 68 (2014) 37–52

Table 1 Crisis episodes. Country

Year

Country

Year

Country

Year

Argentina Argentina Argentina Bangladesh Bolivia Brazil Brazil Chile Colombia Colombia Costa Rica Costa Rica Cote d'Ivoire Cote d'Ivoire Croatia Dominican Republic Dominican Republic Dominican Republic Ecuador Ecuador Ecuador Egypt Ethiopia

1980 1995 1999 1977 1980 1980 1999 1982 1982 1998 1981 1999 1980 1984 1999 1977 1981 2002 1981 1988 1999 1990 1997

Ghana Guatemala Haiti Honduras Honduras Hong Kong Indonesia Jamaica Jamaica Jordania Kenya Kenya Kenya Kuwait Kuwait Kuwait Malawi Malawi Malawi Malaysia Morocco Mauritius Mauritius

1978 1986 1992 1979 1985 1998 1998 1977 1984 1984 1976 1981 1992 1975 1979 1985 1979 1985 1998 1998 1993 1979 1982

Mexico Mexico Mexico Nicaragua Papua New Guinea Peru Peru Philippines South Africa South Africa South Korea Sudan Thailand Tunisia Turkey Turkey Uruguay Uruguay Venezuela Venezuela Zambia Zambia

1982 1988 1994 1977 1976 1976 1999 1997 1977 1983 1998 1977 1997 1981 1979 1999 1982 2002 1975 1979 1976 1991

List of crisis episodes included in the empirical analysis in Section 3. The episodes start in years in which the capital account balance drops by an unusually large amount and GDP drops either in the same year or in the following one. See text for additional details of how the episodes are identified.

Percentage Deviation from Trend

0

−0.05

−0.1

−0.15

−0.2

−0.25

−0.3 Consumption

−0.35

0

1

2

GDP

3

Investment

4

Manufacturing Output

5

6

7

8

Years after Crisis Fig. 1. Aggregate variables. Average time path of variables following crises, at yearly frequency, with t¼ 0 corresponding to the crisis year. Figures correspond to average deviation from trend calculated using average growth rate of previous 10 years.

discussed in Section 2, and take the difference between values in t n þ h and t n  1 to obtain, for each episode i, industry k and horizon h,     vik;tn þ h  vik;tn  1 ¼  F ik τik r ik;tn þ h  r ik;tn  1 þ F ik pik;t n þ h  pik;tn  1    F ik qik;tn þ h qik;t n  1   αL;ik ~ ik;tn  1 ; ~ n w w ð16Þ  1  αL;ik  αM;ik ik;t þ h ~ ik;t n þ h  wik;tn þ h aik;t n þ h is the cost of effective where lowercase letters denote the logarithm of the upper-case variables, w labor units and F ik 

αL;ik þ αM;ik αM;ik  1  αL;ik  αM;ik 1  αM;ik

is a “flexibility” index capturing the extent to which value added changes in response to price changes.

F. Schwartzman / Journal of Monetary Economics 68 (2014) 37–52

43

Given identified episodes of emerging market crises, a measure of inventory/cost ratios τik and Fik, and controls for the various price changes, it is possible to estimate the time path for r ik;tn þ h r ik;t n  1 . However, price changes are not directly observable. Therefore, I assume that pik;tn þ h  pik;tn  1 ¼ pi;t n þ h  pi;tn  1 þ β p;t n þ h xik þ εpik;t n þ h with pi;tn þ h  pi;tn  1 being an episode-specific component of the change in the price, βp;tn þ h a vector of parameters, xik a vector of industry characteristics, and εpik;tn þ h being an error term, which is independent of all determinants of the change in value added. I make analogous ~ ik;t n þ h  w ~ ik;t n  1 . Given this specification, episode specific dummies capture the part assumptions for qik;tn þ h  qik;tn  1 , and w of price changes which is common to all industries, and the controls capture the industry specific components. From input–output tables, I can calculate the shares of output in each sector, which is dedicated to fixed investment or durable consumer goods, thus providing elements of xik that help predict pik;tn þ h pik;t n  1 . Analogously, the share of imported inputs captures an important dimension of the cross-sectoral differences in the change in intermediate input prices qik;t n þ h  qik;tn  1 . I also include a control for the capital share. This may be relevant since the assumption made in the model in Section 2 that all capital cannot be changed at any horizons is extreme. I check that the specification is adequate using a model with slowly adjusting fixed capital, in Section 4. ~ ik;t n þ h  w ~ ik;tn  1 , I control for the average growth for the industry in To capture sector-specific change in unit labor costs w countries that were not affected by the crisis and for the growth rate for each industry/episode observation in the 6 years preceding the crisis.10 The specification also allows the changes in r ik;tn þ h to vary across sectors as well as across episodes. Such heterogeneity might be important in the presence of financial frictions which operate at the level of firms or sectors and might be correlated with τik. To capture those frictions, I include controls for establishment size as well as the one for dependence on external finance proposed by Rajan and Zingales (1998) and used by Dell'Ariccia et al. (2007) and Kroszner et al. (2006) in their empirical studies of banking crises. As explained in more detail in Section 3.3 below, there is no data available for τik in each country, so that I proxy by an industry specific measure τk. Thus, industry dummies are perfectly collinear with the main variable of interest, precluding their use. The estimation procedure acknowledges the presence of any additional source of commonality in the error terms by allowing for standard errors which are robust to overlapping clusters at both the industry and episode level (Cameron et al., 2006).11 For each horizon h, if we linearize Eq. (16) further to separate out τik from Fik, we obtain, with some manipulation, a linear equation of the form vik;tn þ h vik;t n  1 ¼  γ 0;tn þ h τk þ γ 1;tn þ h F ik þ γ 2;tn þ h þ γ 3;t n þ h xik þ di;tn þ h þuik;tn þ h ;

αik;L

1  αik;L  αik;M ð17Þ

where xik is the set of controls described above and γ 0;t n þ h ¼ E½F ik E½r i;tn þ h r i;tn  1 , where the expectation operator denotes the mean across all observations of the variable in brackets. Thus, γ 0;t n þ h is proportional to the average change in the interest-rate after h years. The supplementary materials discuss in detail how (17) can be obtained from Eq. (16). They also show that, given the assumptions on the relationship between the controls and price changes discussed above, the estimates are consistent so long as, the average across episodes of the product ðr i;t n þ h  r i;t n  1 Þðτi;k  τk Þ varies across industries k independently from τk. 3.3. Data Value added data are from INDSTAT3, a data set compiled by UNIDO. The data originate from official sources in UN member countries. The data-set features time series data for 28 distinct manufacturing industries classified according to three-digit ISIC (Revision 2) for 180 countries between 1963 and 2004. As should be clear from the regression specification, the data analyzed only includes years right before and right after crisis episodes. This is similar to the approach used by Kroszner et al. (2006) to investigate the impact of banking crises. One reason not to use the full dataset is that industry level data is very noisy, thus reducing the power of any estimate. In particular, it includes methodological breaks in the time series that are not properly documented by UNIDO. These breaks are less likely to pose a problem for the time periods included in the analysis, since the sheer size of the driving shock would tend to overwhelm fluctuations in the series due to changes in measurement. In comparison, given a finite sample, those fluctuations could substantially affect estimates based on information from other periods. A further evidence that results are not driven by methodological breaks is that, as we will see, they remain robust across sub-samples. The key regressor is the inventory/cost ratio. Reliable measures of inventories or production costs are not easily available for most countries in the sample. Instead, I follow Rajan and Zingales (1998) and use data from United States firms. If the inventory/cost ratios in the United States reflect the underlying production technology, they should be a reasonable proxy for inventory/cost ratios in other countries. I aggregate firm-level data from COMPUSTAT into multiple sectors. For each firm/year observation I calculate the ratio of Total Inventories (mnemonic: invt) to Cost of Goods Sold (mnemonic: cogs), then I calculate the median inventory/cost ratio for each industry across all firm/year observations. Since COMPUSTAT includes decades' worth of financial reports from listed firms that 10 The supplementary materials available online include a specification that provides further controls for systematic differences in the cyclical variation of different sectors not well captured by the other controls. I do not include them in the main text because these controls for cyclicality are likely to be too stringent, since they use information about the typical cyclicality of different industries, which might include their responses to interest rates. 11 I implement these in STATA using the ivreg2 command. The supplementary materials compare this to other alternatives, including cgmreg.ado file produced by Cameron et al. (2006) and used in previous versions of the paper.

44

F. Schwartzman / Journal of Monetary Economics 68 (2014) 37–52

Time to Produce in US and Korea 4

323

383

COMPUSTAT (United States)

382 361 314

3

332

371 356 354

2

353

385

381

351

313 384 369 311

331

321

390 324

341 352 372

1 355

342

0

0.5

1

1.5

2

2.5

Financial Survey Analysis (South Korea) Fig. 2. Inventory intensity measures. U.S. data are average of Inventories/Cost of Goods Sold per month across firms in industry, Korean data are average of Inventories/Sales per month across firms in industry. Industry classification is ISIC Rev. 2: 311 - Food, 313 - Beverage, 314 - Tobacco, 321 - Textiles, 322 Apparel, 323 -Leather, 324 - Footwear, 331 - Wood Products, 332 - Furniture, 341 - Paper, 342 - Printing and Publishing, 351 - Industrial Chemicals, 352 Other Chemicals, 353 - Refineries, 354 - Products of Petroleum and Coal, 355 - Rubber Products, 356 - Plastic Products, 361 - Pottery, China and Earthenware, 362 - Glass, 369 - Other Non-Metallic Minerals, 371 - Iron and Steel, 372 - Non-ferrous metals, 381 - Metal Products, 382 - Machinery, 383 Electrical Machinery, 384 - Transport Equipment, 386 - Professional and Scientific, 390 - Other.

operate in the US, this procedure smooths business cycle fluctuations, obtaining something close to a steady-state measure.12 Importantly, COMPUSTAT is based on accounting data, so the accounting identity between inventories and costs in Eq. (5) in Section 2 should hold exactly, with whatever inputs omitted from inventories also omitted from the cost of goods sold. The key working assumption is that the technological characteristics of listed US firms are sufficiently informative about the technological characteristics of firms in emerging economies. To validate this assumption, I check the correlation of the inventory/cost ratios calculated using COMPUSTAT against a measure of inventory/sales data from the Korean Financial Survey Analysis. Fig. 2 shows the joint distribution between the two measures. There is an unambiguously positive correlation between the inventory/cost measures from COMPUSTAT and the inventory/sales data from the Korean Survey, above 0.6. The correlation between the two measures is not perfect. This could reflect cross-country differences, but it could also reflect measurement error. Under the assumption that measurement error is classical (i.e., both measures reflect a “true” measure of underlying technological characteristics plus an error which is independent of the actual inventory/cost ratio), the inventory/sales data from Korean firms can be used as an instrument for the inventory/cost data from COMPUSTAT, thus eliminating the measurement error from the data and any attenuation bias that it may generate. The additional controls rely on the UNIDO data (for establishment size and industry trends) and on input–output matrices made available and compiled by OECD, for the various shares.

3.4. Results Fig. 3 shows the scatter plots for the change in value added for each industry averaged across episodes between the year right before the crisis and 1, 3, 5 and 7 years after the beginning of the episode against the inventory/cost ratio using COMPUSTAT data. Three aspects of the data are notable. First, there is a negative correlation between the growth rate and the inventory/cost ratios. Second, the correlation becomes more pronounced with time, peaking at around 5 years after the beginning of the episode. Lastly, the correlation remains negative even as the average output recovers. The raw correlations may reflect alternative sources of heterogeneity that happen to be correlated with the inventory/ cost ratio. Table 2 shows the results of regressions controlling for these alternatives. The general pattern remains the same, with the coefficient under both OLS and IV using Korean inventory/sales ratios as instruments being statistically significant at the 1 percent level in 3 and 5 years after the crisis and at the 5 percent level in 7 years afterwards. The IV regressions yield coefficients which are larger in absolute value than the OLS coefficients. The only other controls that are significant at any level are the investment share in the year of the crisis, average firm size in multiple years and, at a high significance level throughout, industry growth in noncrisis countries. The coefficients on inventory/cost ratio and the confidence intervals are also depicted in Fig. 4. 12 I also correct for seasonality. Most of the observations are from statements in December, so such a correction is helpful. To deal with seasonality, I first take medians over data disclosed in December and June separately. The final number is an arithmetic average of the two.

F. Schwartzman / Journal of Monetary Economics 68 (2014) 37–52

45

Average Change in Value Added x years after crisis

yt+1 -yt-1

yt+3 -yt-1

0.2

0.2

0.1

0.1

0 342 355

-0.1 -0.2

372 353341 371 351 321 352 311 354313 323 324 356 314 369 390 361 331 381 332 385383 -0.395 384

Correlation = E[y|x] = 0.022 -0.026 x 0

1 2 3 Inventory/Cost (months)

382

0 -0.1 -0.2

4

Correlation = -0.649 E[y|x] = 0.131 -0.061 x 0

342 355

-0.1 -0.2

1 2 3 Inventory/Cost (months)

342 355

0.2

353 372 341 356 352 311 354313 351

Correlation = -0.763 E[y|x] = 0.186 -0.064 x 0

361 314 381

323

383 385 382

4

yt+7 -yt-1

385 384 324 371 314 369 361 331 381 390321 323 382383 332

0

332

1 2 3 Inventory/Cost (months)

yt+5 -yt-1 0.2 0.1

353 341 371 311 356 354313 351 352 321 384 369 324 331 390

372 342 355

0.1 0 -0.1 -0.2

4

353 352 351 354313 341 311 356 372 369 324 371 384 390321 331 332

314385 361 381 383 323 382

Correlation = -0.657 E[y|x] = 0.259 -0.062 x 0

1 2 3 Inventory/Cost (months)

4

Fig. 3. Correlations between inventory intensity and output after crises. The horizontal axes are the Inventories/Cost of Goods Sold for each industry calculated using COMPUSTAT data. The vertical axes are the average change in value added (volume) between the year before the onset of the crisis and “h” years after that date. Industry classification is ISIC Rev. 2: 311 - Food, 313 - Beverage, 314 - Tobacco, 321 - Textiles, 322 - Apparel, 323 - Leather, 324 - Footwear, 331 - Wood Products, 332 - Furniture, 341 - Paper, 342 - Printing and Publishing, 351 - Industrial Chemicals, 352 - Other Chemicals, 353 - Refineries, 354 - Products of Petroleum and Coal, 355 - Rubber Products, 356 - Plastic Products, 361 - Pottery, China and Earthenware, 362 - Glass, 369 - Other Non-Metallic Minerals, 371 - Iron and Steel, 372 - Non-ferrous metals, 381 - Metal Products, 382 - Machinery, 383 - Electrical Machinery, 384 - Transport Equipment, 386 - Professional and Scientific, 390 - Other.

3.5. Alternative samples Crisis episodes are concentrated in certain time periods and geographic regions. About half of the crises in the sample took place in the early 1980s and the other half between the mid-to-late 1990s. Geographically, about half of the crises were in Latin America and the rest distributed among all other continents. The point estimate of the inventory/cost ratio is robust to splitting the sample in different ways. The results are in panels 2–5 of Table 3. Given that samples are half the size, it is not surprising that the statistical significance is reduced. It is also notable that the point estimates for the 1990s appear to mean-revert much more quickly than in the 1980s. There are also instances in which advanced economies experienced coincident reduction in the current account deficit and a drop in output. On the face of it, this might suggest that the costs of financial globalization are not restricted to emerging economies. Given the interpretation of inventory/cost ratios as technologically determined, the same methodology can be used to assess this hypothesis.13 The last column in Table 3 depicts the results. Unlike emerging economies, there is no clear correlation between industry level output changes and the inventory/cost ratios. The implication is that international financial flows are not an important source of instability for developed economies.

4. An equilibrium model This section introduces a fully specified, equilibrium, small open economy model. Since the model is less restrictive along many dimensions than the one used to derive the empirical specification in Section 3, it can be used to check whether the interest rate shock remains identified given this less restrictive framework. More importantly, the equilibrium model takes aggregate resource constraints into account. Therefore, it is possible to infer what the empirical results in Section 3 imply for the role of changes in the cost of capital in driving aggregate outcomes. 13 The methodology yields the following episodes: Australia 1983, Austria 1980 and 1993, Canada 1982 and 1989, Spain 1992, Finland 1976 and 1989, France 1992, UK 1979 and 1990, Ireland 1982 and 1986, Italy 1992, Netherlands 1981, New Zealand 1976, 1988, 1991 and 1997, Portugal 1984 and 1993, and USA in 1991.

46

F. Schwartzman / Journal of Monetary Economics 68 (2014) 37–52

Table 2 Regression results. Explanatory variable

yt n þ 1  yt n  1

yt n þ 3  yt n  1

ytn þ 5  yt n  1

yt n þ 7  yt n  1

OLS

IV

OLS

IV

OLS

IV

OLS

IV

Inventory/cost

 0.012 (0.009)

0.000 (0.014)

 0:044 (0.014)

 0:064 (0.025)

 0:056 (0.017)

 0:088 (0.033)

 0.047 (0.018)

 0.086 (0.042)

F

0.00598 (0.011)

0.007 (0.011)

 0.006 (0.010)

 0.007 (0.011)

 0.004 (0.020)

 0.005 (0.022)

 0.011 –

 0.012 (0.007)

αl =ð1 αl  αm )

 0.0308 (0.045)

 0.033 (0.045)

0.020 (0.059)

0.021 (0.064)

 0.009 (0.085)

 0.009 (0.092)

0.002 (0.045)

0.008 (0.054)

Investment share

 0.236 (0.104)

 0:273 (0.081)

 0.067 (0.338)

0.000 (0.349)

0.002 (0.361)

0.116 (0.404)

0.086 (0.385)

0.208 (0.395)

Durable cons. share

 0.238 (0.326)

 0.248 (0.321)

 0.477 (0.485)

 0.467 (0.496)

 0.049 (0.492)

 0.031 (0.523)

 0.812 (0.440)

 0.788 (0.483)

Exports share

0.0958 (0.098)

0.070 (0.093)

0.286 (0.142)

0.325 (0.151)

0.030 (0.288)

0.083 (0.286)

0.315 (0.213)

0.382 (0.242)

Capital share

0.0981 (0.192)

0.118 (0.197)

 0.133 (0.332)

 0.175 (0.345)

 0.513 (0.376)

 0.587 (0.415)

 0.574 (0.536)

 0.602 (0.548)

Imported goods share

 0.157 (0.120)

 0.136 (0.117)

 0.186 (0.107)

 0.218 (0.117)

 0.112 (0.235)

 0.155 (0.232)

 0.195 (0.160)

 0.246 (0.194)

External dependence

 0.0298 (0.027)

 0.033 (0.030)

 0.052 (0.054)

 0.045 (0.057)

 0.035 (0.062)

 0.022 (0.071)

 0.013 (0.060)

0.005 (0.075)

Establishment size

0:081 (0.025)

0:082 (0.025)

0:132 (0.044)

0:131 (0.044)

0.106 (0.063)

0.106 (0.063)

0.087 (0.083)

0.084 (0.082)

Growth non-crisis

0.577 (0.255)

0.619 (0.254)

0:716 (0.241)

0:644 (0.248)

0:552 (0.185)

0.472 (0.213)

0:544 (0.203)

0.443 (0.248)

Previous growth

0.0259 (0.040)

0.027 (0.040)

0.039 (0.039)

0.037 (0.038)

0.008 (0.043)

0.007 (0.044)

0.043 (0.069)

0.042 (0.069)

Observations R-squared

1,095 0.343

1,095 0.341

1,044 0.375

1,044 0.373

970 0.511

970 0.508

729 0.473

729 0.469

Dependent variable is the log change in value added where t is the starting year of a crisis event. IV Regressions use inventory/sales data from Korea as instruments. Regressions include country fixed effects (omitted). Estimated values in italics are significant at a 10 percent level, in bold are significant at 5 percent level, and in both bold and italics are significant at a 1 percent level. Standard errors (in parentheses) are robust to overlapping clusters on country and industry following Cameron et al. (2006). A dash indicates that the Cameron et al. (2006) procedure delivers a negative value for the standard error.

Lastly, the model can also be used to assess how the time-to-produce technology itself alters the response of the aggregate economy to shocks. This is of interest, since the time-to-produce technology is a source of working capital demand, which figures prominently in many treatments of emerging market business cycles and, more recently, of the impact of financial shocks in developed economies. 4.1. Model setup The model is a variant of the small open economy real business cycle model in Mendoza (1991) and Correia et al. (1991) extended to include multiple sectors and time to produce and distribute goods. Otherwise, it is fairly conventional, so as to keep the results comparable to the existing literature. 4.1.1. Firms There are N þ1 distinct sectors each with a representative firm indexed k A f0; …; Ng. The production technology is similar to the one described in Section 2, with the maximum lag between production and sales given by S¼1. Concretely Y k;t ¼ Z k;t ð0Þωk ð0Þ Z k;t  1 ð1Þωk ð1Þ ;

k A f0; …; Ng;

ð18Þ

where Y k;t is total sales of good k and Z k;t ðsÞ denotes a composite of inputs acquired in period t to produce the finished good which is sold s periods ahead. In turn, Z k;t ðsÞ is defined as Z k;t ðsÞ ¼ γ k ðK k;t  1 ðsÞÞ1  αL;k  αM;k ðAt Lk;t ðsÞÞαL;k ðM k;t ðsÞÞαM;k ;

k A f0; …; Ng; s A f0; 1g;

ð19Þ

F. Schwartzman / Journal of Monetary Economics 68 (2014) 37–52

47

Estimated Coefficient

O.L.S. 0 −0.02 −0.04 −0.06 −0.08 −0.1 0

5

10

years after crisis Instrumental Variable 0.05 0 −0.05 −0.1 −0.15 0

5

10

years after crisis Fig. 4. Estimated coefficient for different horizons. The upper panel shows the coefficient on inventories-to-cost ratio in Eq. (17) estimated using OLS for different horizons, and the lower panel shows the same coefficients estimated with IV using inventory/sales from Korean firms as instruments. The bands refer to the 95 percent confidence interval using two-way clustering technique following Cameron et al. (2006) with clusters by industry and by country and implemented in STATA using ivreg2.ado.

Table 3 Sub-samples. Dependent variable

yt n þ 1  yt n  1 yt n þ 3  yt n  1 y t n þ 5  yt n  1 y t n þ 7  yt n  1

1. Benchmark

2. 1980s

2. 1990s

4. LAM

OLS

IV

OLS

IV

OLS

IV

OLS

 0.012 (0.009)  0:044 (0.014)  0:056 (0.017)  0.047 (0.018)

0.000 (0.014)  0:064 (0.025)  0:088 (0.033)  0.086 (0.042)

 0.012 (0.012)  0:053 (0.018)  0:074 (0.025)  0.063 (0.021)

0.002 (0.014)  0.057 (0.028)  0.096 (0.040)  0.114 (0.044)

 0.013 (0.007)  0.027 (0.013)  0.025 (0.014) 0.002 (0.041)

 0.001 (0.024)  0.076 (0.032)  0.081 (0.040) 0.013 (0.079)

 0.009 (0.011)  0:042 (0.016)  0.025 (0.016)  0.024 (0.014)

5. EM 's not in LAM

6. OECD

IV

OLS

IV

OLS

IV

 0.005 (0.024)  0.066 (0.034)  0.050 (0.039)  0.095 (0.053)

 0.015 (0.009)  0.045 (0.019)  0:088 (0.027)  0.071 (0.038)

0.009 (0.012)  0.059 (0.023)  0:122 (0.044)  0.075 (0.059)

 0.005 (0.007)  0.009 (0.008)  0.006 (0.009)  0.009 (0.011)

0.006 (0.016) 0.009 (0.020) 0.006 (0.018)  0.008 (0.020)

Dependent variable is the log change in value added from t n  1 to t n þ h where tn is the starting year of a crisis event. IV Regressions use inventory/sales data from Korea as instruments. The rows refer to different horizons (h). Regressions include country fixed effects (omitted). Estimated values in italics are significant at a 10 percent level, in bold are significant at 5 percent level, and in both bold and italics are significant at a 1 percent level. Standard errors (in parenthesis) are robust to overlapping clusters on country and industry following Cameron et al. (2006).

where γk is a scale parameter, K k;t  1 ðsÞ is the fixed capital stock, set one period in advance, and Lk;t ðsÞ and index of the labor input used in sector k. The variable At is a productivity parameter that fluctuates around a deterministic trend growth g. Labor is a composite of two components, a sector-specific component LSk;t ðsÞ and a general component LGk;t ðsÞ, so that Lk;t ðsÞ ¼ ðLSk;t ðsÞÞθ ðLGk;t ðsÞÞ1  θ ;

k A f0; …; Ng;

s A f0; 1g;

ð20Þ

The assumption of two types of labor is a technical assumption made necessary by the multisector nature of the model. It ensures that sectoral output has a well-defined steady-state (a well-defined steady-state is necessary to make linearization of the model possible, see Schmitt-Grohe and Uribe (2006)). The calibration sets θ to a small number, so that the quantitative impact of this assumption is minimal. 4.1.2. Households There is a representative household. The household supplies labor, accumulates capital and consumes. The utility function of this household is time-separable with discount rate β o 1, and the period utility is uðC t ; Lt Þ ¼

1 1 þ 1=ψ 1  σ ðC t ð1 þ gÞt Lt Þ 1σ

ð21Þ

where Ct is the consumption, Lt is the labor supply and g is the trend growth rate in productivity. This utility function follows Greenwood et al. (1988). It implies that labor supply does not respond to the wealth of the household. As a consequence, small open economy business cycle models equipped with this utility function are more

48

F. Schwartzman / Journal of Monetary Economics 68 (2014) 37–52

successful in generating realistic dynamics for the volatility of consumption and the cyclicality of the balance of trade (Correia et al., 1991). One interpretation of this function is that labor supply is costly because it implies a loss in output from home production. This interpretation also motivates the trend growth rate in labor productivity appearing as part of the utility function (this assumption is also necessary for the economy to exhibit a balanced growth path). 4.1.3. Resource constraints Accumulating capital for any given sector k involves incurring sector-specific adjustment costs. Convex adjustment costs to capital are a common addition to small open economy business cycle models and ensure that aggregate investment is not too volatile relative to the data. Furthermore, adjustment costs at the sectoral level impose a limit on the speed with which resources reallocate across sectors. The capital accumulation equation is   ζ ðK k;t ð1 þ gÞK k;t  1 Þ2 K k;t ¼ 1  δ K k;t  1 þ I k;t  ; 2 K k;t  1

k A f0; …; Ng:

ð22Þ

While capital is sector-specific, there is no impediment to moving capital for production of output to be sold at different dates. This implies the following resource constraint for the capital stock of sector k: 1

∑ K k;t ðsÞ r K k;t ;

k A f0; …; Ng:

s¼0

ð23Þ S

Households supply a fixed amount Lk of sector-specific labor each period. The resource constraint for sector-specific labor is 1

∑ LSk;t ðsÞ r LSk ;

s¼0

k A f0; …; Ng:

ð24Þ

Otherwise, labor is mobile between sectors. This implies the following aggregate resource constraint: N

1

N

∑ ∑ LGk;t ðsÞ þ ∑ LSk r Lt :

k¼0s¼0

ð25Þ

k¼0

To obtain aggregate sales, calculate the sum of sales in each of the sectors: N

Y t ¼ ∑ Y k;t :

ð26Þ

k¼0

Finally, the current account identity is N

N

1

Bnt Rt  1 Bnt ¼ G þ C t þ ∑ I k;t þ ∑ ∑ M k;t ðsÞ þ k¼0

k¼0s¼0




κ Bnt 2 Yt

2 b Y t  Y t ;

ð27Þ

where G is government consumption, assumed constant and included only for calibration purposes, and Bnt is the amount of net foreign debt held by domestic residents. The term κ =2ððBnt =Y t  bÞ2 Y t denotes a transaction cost associated with deviations of foreign/debt to GDP ratio from a long-run target level b. This transaction cost is commonly used in small open economy models to ensure that the foreign debt/GDP ratio has a well-defined steady-state (Schmitt-Grohe and Uribe, 2006). The transaction cost parameter κ is typically set to a small value, so that it has little quantitative relevance over short to medium horizons. 4.1.4. Allocation The allocation is the solution to a planner's problem, where the planner takes the interest rate on foreign debt as given and maximizes the discounted present value of the utility of the representative household denoted by Eq. (21) subject to the technological constraints (18)–(20), and (22) and the aggregate resource constraints (23)–(27). 4.2. Calibration The model has 16 sectors (N ¼15) with sectors 1–15 representing individual manufacturing industries, and sector 0 a residual sector encompassing the rest of the economy, including services, utilities, construction and the public administration. Given that manufacturing accounts for around one third of GDP, allowing for a residual sector is necessary in order to evaluate the aggregate impact of the various shocks. The smaller number of sectors in the model than in the data reflects differences in the industry classification system available in the input–output tables used in the calibration. I use averages of the same input–output tables used in the empirical analysis in Section 3 to calibrate the income share parameters αL;k and αM;k . Allowing for different factor shares is crucial to evaluate whether the specification in the regression Eq. (17) is adequate to identify the size of interest rate shocks.14 The values for each sector are depicted in Table 4. I set the parameter θ governing the share of sector-specific labor in output is 10  5. 14

I thank an anonymous referee for raising this issue.

F. Schwartzman / Journal of Monetary Economics 68 (2014) 37–52

49

The key unconventional parameter for any given industry k is ωk ð1Þ (with ωk ð0Þ ¼ 1  ωk ð1Þ). Eq. (9) in Section 2 with S ¼1 can be used to calibrate this parameter. It follows that, for any industry k, it is the case that

τk ffi ωk ð1Þ:

ð28Þ

For manufacturing industries, the calibration of the ωk ð1Þ's relies on the same values calculated for the empirical Section 3.15 The residual sector (sector 0) includes both construction, which has well-known production lags, and services, which have none. The share of construction in the residual sector is close to 10 percent. Taking the time lag in construction to be equal to four quarters, I set ω0 ð1Þ ¼ 10  4 ¼ 0:4. The values for each sector are depicted in Table 4. The other unconventional production parameter is the quantity of industry specific labor L k . I calibrate this, together with the scale parameter γk, in order to ensure that (a) each industry's relative size conforms to their relative size in the economy and that (b) sector-specific labor is only 1/1000 of the aggregate labor supply. This small share of total labor supply, together with the small value for θ ensures that sector-specific labor has little effect on results. The remaining parameters are more conventional. The target long-run debt/gross output ratio b is 24 percent, which matches the target debt/GDP ratio of 48 percent adopted by Neumeyer and Perri (2005). The parameter governing the transaction costs associated with borrowing more or less than the target debt/output level is set to a small value, 10  5. The annual steady-state growth rate of productivity is 2 percent and the annual rate of depreciation is 10 percent. The curvature of the utility function, σ, is equal to 2. The wage elasticity of labor supply ψ is equal to 53 in the baseline calibration, which corresponds to the calibration in Neumeyer and Perri (2005).16 The discount rate β is equal to 0.9236 in annual terms ensuring that in steady-state the investment/output ratio matches the average ratio calculated from the OECD input–output 1 tables. The steady-state interest rate on debt R is β ð1 þ g Þσ ¼ 1:1264 in annual terms, ensuring that in steady-state the debt/output ratio equals the long-run target level b. Also following Neumeyer and Perri (2005), I set the capital adjustment cost parameter ζ to 25.5.17 Lastly, the OECD data implies that government expenditures are 16.2 percent of GDP in steadystate, which pins down G. 4.2.1. Shock processes To calibrate the shocks, I search for paths for Rt and At which allow the time series generated by the model to match the paths for (i) the average for the observed deviations of GDP from trend depicted in Fig. 1 and (ii) the OLS coefficients of the change in value added on the Inventory/Cost ratio estimated in Section 3. I use data up to 9 years after the crisis. These paths allow me to calculate nine values for beginning of the year TFP and interest rates. I use linear interpolation for the quarters in between. After 9 years, I assume that the shocks decay very slowly, at a rate of 0.999 per quarter. Since I calculate these paths based on statistics derived from a broad sample of episodes, I take them to be typical of such crises and, hence, perfectly forecasted at the time of the shocks. The use of quantity data to infer interest rate shocks has precedents in the emerging market business cycle literature, notably Garcia-Cicco et al. (2010). Compared with using interest rate data, one motivation for this procedure is that it relies on information derived from the full set of crises episodes analyzed in the empirical section, as opposed to the small number of episodes for which appropriate interest rate data can be found. For example, the data-set compiled by Uribe and Yue (2006) to study the dynamics of interest rate spreads paid by sovereign bonds includes only nine episodes out of the 68 identified in the paper. Furthermore, the episodes for which data is available are not representative of the sample, since they are concentrated in the late 1990s. A second reason to use quantity data is that, while the data on interest rates paid by sovereign bonds may be a reliable guide for the cost of capital faced by firms in normal periods, in crisis times they are less so. This is because on one hand, to the extent that domestic banks intermediate between foreign lenders and domestic firms, the cost of sovereign debt may understate the true cost of capital faced by firms, since the domestic banking system is often impaired during emerging market crises. On the other hand, it may overstate the cost of capital faced by the sovereign since the probability of sovereign default can increase substantially in times of crises. 4.3. Results Panel (a) in Fig. 5 shows the implied time paths for deviations of the interest rate (in annual terms) and TFP from their steady-state values. The interest rate rises on impact by around 1.5 percent in annual terms, and then increases significantly reaching peak levels around 4.5 percent 12–24 quarters after the onset of the crisis. TFP falls by close to  1.4 percent on impact and subsequently hovering around that level. Panel (b) in Fig. 5 decomposes time path for the coefficient of sectoral value added on inventory/cost ratios obtained from estimating the regression Eq. (17) using model generated data.18 The red dashed line represents the coefficients estimated in Section 3, the solid blue line with squares represents what the coefficients would have been if the economy was only 15

For consolidated industries, I take the weighted average, with weights given by the value produced in each industry. Neumeyer and Perri (2005) calibrate the exponent on employment in the utility function ν ¼ 1 þ 1=ψ . The numbers above refer to ψ. The main results are robust to alternative values for ζ. 18 The regression only includes the controls which depend on the factor shares, including the “flexibility index” Fi. The remaining controls are irrelevant since in a small open economy the relative prices of different sectors and of imported input remain constant. Furthermore, the model does not 16 17

50

F. Schwartzman / Journal of Monetary Economics 68 (2014) 37–52

Table 4 Calibration of sectoral parameters. Name

Y i =Y (%)

ωð1Þ

αk

αl

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

Residual Food Textiles Wood Paper Oil Chemicals Rubber Non-metallic minerals Basic metals Metal products Machinery n.e.c. Electrical machinery Medical and precision Transportation equipment Manufacturing n.e.c.

67.22 7.26 3.02 0.66 1.85 2.11 3.45 1.20 1.17 2.62 1.23 2.22 1.62 0.14 3.18 1.05

0.40 0.64 0.79 0.65 0.32 0.45 0.51 0.41 0.69 0.55 0.89 0.95 1.00 0.94 0.52 0.71

0.23 0.14 0.08 0.12 0.12 0.21 0.12 0.08 0.18 0.10 0.08 0.04 0.07 0.10 0.07 0.15

0.37 0.14 0.22 0.22 0.23 0.07 0.17 0.24 0.24 0.16 0.25 0.24 0.21 0.26 0.18 0.22

αm 0.39 0.73 0.70 0.66 0.65 0.72 0.71 0.68 0.59 0.73 0.66 0.72 0.72 0.64 0.75 0.63

0

4

Estimated value

6 R

2 0

R R+TFP Data

−0.02

−0.04

TFP −2

0

10 20 30 Quarters after shocks

0

−5

−10

−15 0

2 4 6 Years after shocks

−0.06

40

8

Percentage deviation from S.S.

Percentage deviation from S.S.

Percentage deviation from S.S.

Sector

0

2 4 6 Years after shocks

8

0

2 4 6 Years after shocks

8

0

−10

−20

−30

Fig. 5. Simulation results. Simulated paths for aggregate variables after an interest rate and a TFP shock as percentage deviations from the deterministic trend. Yearly values are calculated by time aggregating quarterly impulse response functions implied by the calibrated model and averaging over paths implied by different starting quarters for the crisis (see supplementary materials available online for details). (For interpretation of the references to color in this figure caption, the reader is referred to the web version of this paper.)

affected by the interest rate shock and the green line with circles represents the total impact of the two shocks on the coefficients. Given the linearization procedure used to solve the model, the effects of the shocks on the coefficients are

(footnote continued) feature sector specific financial frictions or sector specific trends, so that I also omit the controls for establishment size and external financial dependence and trends.

−0.5

−1

−1.5

0

2

4

6

8

10

R shock consistent with data

0 −2 −4 −6 −8

0

2

4 6 Years after shock

8

10

Percentage deviation from S.S.

Transitory R shock 0

Percentage deviation from S.S.

Percentage deviation from S.S.

Percentage deviation from S.S.

F. Schwartzman / Journal of Monetary Economics 68 (2014) 37–52

51

Transitory TFP shock 0 with time to produce −2

without time to produce

−4 −6 −8 0

2

4

6

8

10

TFP shock consistent with data

0 −2 −4 −6 −8

0

2

4 6 Years after shock

8

10

Fig. 6. Counter-factual simulations. Simulated paths for aggregate variables after an interest rate and a TFP shock as percentage deviations from the deterministic trend. Yearly values are calculated by time aggregating quarterly impulse response functions implied by the calibrated model and averaging over paths implied by different starting quarters for the crisis (see supplementary materials available online for details).

additive, so that the effect of the TFP shock is just the difference between the total effect and that of the interest rate shock. The three lines are virtually indistinguishable. This means that the estimated coefficients are almost solely explained by the interest rate shock, validating the empirical approach in Section 3. Panel (c) of Fig. 5 shows the same decomposition for GDP. In the year of the crisis, virtually all of the drop in GDP is explained by the TFP shock. As time progresses the interest rate shock accounts for an increasing share of the deviation of GDP from trend. Between 3–5 years after the onset of the crisis, the deviation of GDP from trend explained by the interest rate shock is on average equal to  5.4 percent, accounting for, again on average, about 48 percent of its deviation from trend on average in those same years. Panel (d) shows the model implied behavior of fixed investment as compared to the data. Since investment data has not been used in the calibration of the shock paths, it serves as an external validation of the results. The paths for investment in the model and in the data are strikingly similar, with an initial large drop in investment relative to trend of close to 27 percent in the year of the crisis followed by a persistent deviation from trend which is much larger than the deviation of GDP. Notably, and unlike GDP, most of the drop in investment is accounted for by the interest rate shock. The inability of the TFP shock to account for most of the drop in investment is further evidence for an increase in the cost of capital following emerging market crises. Finally, Fig. 6 investigates the implications of the time to produce technology itself for the amplification and propagation of shocks. To that end I compare the responses of the economy to shocks in the baseline parameterization and in an alternative, counterfactual parameterization with ωk ð0Þ ¼ 1 for all k A f0; …; Ng. The upper panels compare the responses of GDP to transitory TFP and interest rate shocks.19 The time to produce technology does not generate a discernible difference in the response of GDP to the productivity shock. However, there is significant difference in the response to the interest rate shock. The lower panels show the responses given the shocks obtained from the calibration. Again, the only discernible difference is in the response to the interest rate shock.

19 For the interest rate shock, I assume a 10 percent increase in annual terms with a mean-reversion parameter of 0.78, consistent with Neumeyer and Perri (2005). For the TFP shock I assume a 3 percent drop and a mean reversion parameter of 0.95.

52

F. Schwartzman / Journal of Monetary Economics 68 (2014) 37–52

5. Conclusion Emerging markets were subject to frequent episodes in which output dropped around the same time foreign capital stopped flowing into the country. These episodes were associated with persistent declines in output, calling into question the case for financial globalization insofar as less developed economies were concerned. The paper contributes to this debate by showing that an interest rate shock can account for the persistent cross-industry differences observed following those episodes, whereas a productivity shock cannot. Furthermore, the interest rate shock necessary to account for the cross-section of industries is large enough to account for a deviation of GDP from its trend of 5.4 percent or about 48 percent of the overall deviation in the data 3–5 years after crisis. In fact, the results in the paper may understate the role of the interest rate shock since, as shown by Meza and Benjamin (2009) and Pratap and Urrutia (2011), the interaction of working capital constraints with other frictions can lead to significant fluctuations in measured productivity and in output. Appendix A. Supplementary materials Supplementary data associated with this article can be found in the online version at http://dx.doi.org/10.1016/j.jmoneco. 2014.07.010.

References Aguiar, M., Gopinath, G., 2007. Emerging market business cycles: the cycle is the trend. J. Polit. Econ. 115 (1), 69–102. Alessandria, G., Kaboski, J., Midrigan, V., 2010. Inventories, lumpy, trade and large devaluations. Am. Econ. Rev. 100, 2304–2339. Barth, M.J., Ramey, V.A., 2001. The cost channel of monetary transmission. NBER Macroecon. Annu. 16, 199–240. Calvo, G., Izquierdo, A., Talvi, E., 2006. Phoenix miracles in emerging markets: recovering without credit from systemic financial crises. National Bureau of Economic Research Working Paper series (12101). Cameron, A.C., Gelbach, J.B., Miller, D.L., 2006. Robust inference with multi-way clustering. National Bureau of Economic Research Working Paper series (T0327). Cerra, V., Saxena, S.C., 2008. Growth dynamics: the myth of economic recovery. Am. Econ. Rev. 98 (1), 439–457. Chang, R., Fernandez, A., 2013. On the sources of fluctuations in emerging economies. Int. Econ. Rev. 54, 1265–1293. Chari, V.V., Kehoe, P., McGrattan, E., 2005. Sudden stops and output drops. Am. Econ. Rev. 95 (2), 381–387. Correia, I., Neves, J., Rebelo, S., 1991. Business cycles in a small open economy. Eur. Econ. Rev. 81 (4), 797–818. Dell'Ariccia, G., Detragiache, E., Rajan, R., 2007. The real effect of banking crises. J. Financ. Intermed. 17 (1), 89–112. Garcia-Cicco, J., Pancrazi, R., Uribe, M., 2010. Real business cycles in emerging countries? Am. Econ. Rev. 100, 2510–2531. Greenwood, J., Hercowitz, Z., Huffman, G., 1988. Investment, capacity utilization and the real business cycle. Am. Econ. Rev. 78 (3), 402–417. Jermann, U., Quadrini, V., 2012. Am. Econ. Rev. 102 (1), 238–271. Khan, A., Thomas, J., 2007. Inventories and the business cycle: an equilibrium analysis of (s,s) policies. Am. Econ. Rev. 97 (4), 1165–1188. Korinek, A., 2011. The new economics of prudential capital controls. IMF Econ. Rev. 59 (3). Kose, M.A., Prasad, E., Rogoff, K., Wei, S.-J., 2009. Financial Globalization: A Reappraisal. IMF Staff Papers. Kroszner, R., Laeven, L., Klingebiel, D., 2006. Financial Crises, Financial Dependence, and Industry Growth. Center for Economic Policy Research 5623. Lubik, T., Sarte, P.-D., Schwartzman, F., 2014. What Inventory Behavior Tells us About How Business Cycles Have Changed. Federal Reserve Bank of Richmond Working Paper series 14(06). Mendoza, E.G., 1991. Real business cycles in a small open economy. Am. Econ. Rev. 81 (4), 797–818. Mendoza, E.G., 2010. Sudden stops, financial crises and leverage. Am. Econ. Rev. 100 (5), 1941–1966. Meza, F., Benjamin, D., 2009. Total factor productivity and labor reallocation: the case of the korean 1997 crisis. B.E. J. Macroecon. 9 (1). Neumeyer, P., Perri, F., 2005. Business cycles in emerging economies: the role of interest rates. J. Monet. Econ. 52 (2), 345–380. Prasad, E.S., Rogoff, K., Wei, S.-J., Kose, M.A., 2007. Trade credit: theories and evidence. In: Harisson, A. (Ed.), Globalization and Poverty, University of Chicago Press, Chicago, pp. 457–516. Ch. 11. Pratap, S., Urrutia, C., 2011. Financial frictions and total factor productivity: accounting for the real effects of financial crises. Rev. Econ. Dyn. 15 (3), 336–358. Raddatz, C., 2006. Liquidity needs and vulnerability to financial underdevelopment. J. Financ. Econ. 80, 677–722. Rajan, R., Zingales, L., 1998. Financial dependence and growth. Am. Econ. Rev. 88 (3), 559–586. Ramey, V., 1989. Inventories as factors of production and economic fluctuations. Am. Econ. Rev. 79 (3), 338–354. Reinhart, C., Rogoff, K., 2009. This time is different: eight centuries of financial folly. Princeton University Press, Princeton, NJ. Schmitt-Grohe, S., Uribe, M., 2006. Closing small open economy models. J. Int. Econ. 61, 163–185. Tong, H., Wei, S.-J., 2009. The Composition Matters: Capital Inflows and Liquidity Crunch During a Global Economic Crisis. National Bureau of Economic Research (15207). Uribe, M., Yue, V., 2006. Country spreads and emerging countries: who drives whom? J. Int. Econ. 69 (1), 6–36.