Full text of Working Papers (Federal Reserve Bank of Richmond) : Time to Produce and Emerging Market Crises, Working Paper 10-15R

View original document

The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.

Working Paper Series

Time to Produce and
Emerging Market Crises

WP 10-15R

This paper can be downloaded without charge from:
http://www.richmondfed.org/publications/economic_
research/working_papers/index.cfm

Felipe Schwartzman
Federal Reserve Bank of Richmond

Time to Produce and Emerging Market Crises∗
Working Paper No. 10-15R
Felipe Schwartzman
July 12, 2012

Abstract
The inventory/cost ratio is a measure of the time to produce and distribute goods (“time to
produce”) and, therefore, an important determinant of working capital demand. In the aftermath
of emerging market crises, manufacturing industries with higher inventory/cost ratios experienced a
larger drop in output, a drop which persisted multiple years into the recovery. This observation is
shown to emerge following a persistent foreign interest rate shock in an environment where time to
produce in different sectors matches inventory/cost ratios in the data. In such an environment, the
interest rate shock is able to account for up to 25% of the deviation of output from its previous trend.
In contrast, without time to produce, the interest rate shock generates a boom in the year of the crisis
and cannot account for cross-sectoral differences. Likewise, it is impossible to generate cross-sectoral
differences of the required magnitude in response to a productivity shock. These results underscore
first the importance of working capital constraints as a transmission channel for financial shocks,
and second, the importance of persistent interest rate shocks as a driving force of business cycles in
emerging economies.
JEL classification: E22, E23, F41
Keywords: Working Capital, Time to Build, Crises, Inventories, Emerging Markets

∗

The views expressed in this paper are those of the author and do not necessarily reflect the position of the Federal
Reserve Bank of Richmond or the Federal Reserve System. I would like to thank Nobu Kiyotaki and Markus Brunnermeier
for advising me in this work. I also thank the participants in various workshops and seminars, as well as colleagues and
professors at Princeton and colleagues at the Richmond Fed for helpful advice. I also thank Jonathan Tompkins for excellent
research assistance. All remaining errors and omissions are my responsibility entirely.

Introduction

Working capital is a key link between finance and production. It consists of the (net) current assets
that firms hold, many of which are tightly linked to short-term production processes. Increased cost
of working capital financing is a leading explanation for the recessionary impact of capital flights
in small open economies (Neumeyer and Perri (2005), Uribe and Yue (2006) among many others)
and of financial shocks in closed economies (Jermann and Quadrini (2012), Quadrini and Perri (2011),
Christiano et al. (2010)). In spite of its significance the quantitative business cycle literature has given
little attention to the appropriate motivation, measurement, calibration and exploration of alternative
implications of the working capital channel. This has considerably diminished its appeal (see Chari
et al. (2005) for a critique). This paper aims to be a first step in filling this gap.
The starting point is to acknowledge that, at a minimum, working capital costs of production are a
determined by the length of time that it takes to produce and distribute goods. This time is reflected
in one particular item in working capital: the sum of raw material, work-in-process and finished goods
inventories. A natural measure of this length of time is the steady-state inventory/cost ratios. With
this measurement, students of financial crises can interpret patterns in the data and add discipline to
quantitative models.
The paper applies this idea to emerging market crises, an arena where working capital has been
frequently invoked as a key factor (see, for example, Mendoza (n.d.) and Meza and Benjamin (2009)).
Those episodes feature simultaneous capital flights and output drops, thus constituting extreme examples of Neumeyer and Perri’s (2005) finding that current-account balances are pro-cyclical in emerging
economies. As such, they constitute a large departure from the international economics textbook prediction that countries should import capital in bad times. These episodes are also interesting because,
unlike major recessions in advanced economies, they often featured large rises in inflation and drops
in real wages, thus straining explanations that rely on nominal frictions (see Calvo et al. (2006) for
these facts).1 In the opposite direction, there is no a priori reason why the working capital mechanism
should be specific to emerging economies and, as such, these episodes provide a laboratory to better
understand the impact of financial crises in the US and Europe.
An empirical contribution of this paper is to document a new stylized fact of emerging market crises,
defined as episodes in which there is a marked coincidence of a drop in output and capital outflows: It
shows that manufacturing industries with large inventory/cost ratios take relatively longer to recover
from these episodes, with the difference between industries being particularly salient three to five years
after the onset of the crisis.
This stylized fact can be used to discipline models of emerging market crises. Using a quantitative
model with time to produce and distribute goods (time to produce, for short) calibrated to match
inventory/cost ratios, I find that the cross-sectional finding is readily explained if a) the inventory/cost
ratios are good measures of working capital demand, b) foreign savings stop flowing in because they
1

This was not the case in all episodes. See for example Gertler et al. (2007) for an account of the Korean crisis that allows
for inflation and nominal frictions play a key role.

are scarcer or require a higher interest rate (as in Calvo (1998) and Mendoza (n.d.)) and c) the cost of
foreign capital increases persistently relative to the pre-crisis level. With or without working capital
demand, the cross-sectional finding is impossible to explain if the sole shock affecting the economy
is a persistent productivity shock (as in Aguiar and Gopinath (2007)). The paper is therefore a
contribution to the literature attempting to identify the relative importance of interest rate shocks
versus productivity shocks in driving emerging market business cycles (see Garcia-Cicco et al. (2010)
and Chang and Fernandez (2012))
With the calibrated model we can revisit the role for working capital in transforming financial
shocks into output drops. In an open economy, time to produce allows for a strong negative reaction
to foreign interest rate shocks even without assuming any market friction. With instantaneous production, the reduction in capital inflows only has a delayed impact, as firms fail to replace the capital
stock, which then slowly depreciates. Also, with time to produce, the shock generates an immediate
drop in tradable production, rather than the export boom commonly found in the literature (see
Kehoe and Ruhl (2008) for an example).
The quantitative model is a generalization of the Long and Plosser (1982) model for an open
economy with heterogeneity in time to produce across sectors. As such, this paper is also a contribution
for the literature on multi-sector growth models (Horvath (1998), Horvath (2000), Dupor (1999),
Foerster et al. (2011)). Different papers in the literature take different stands on the proper calibration
of the lead-lag relationship between the different sectors, which potentially has important consequences
for the propagation and amplification properties of the model. The inventory/cost ratio provides a
source of discipline that so far this literature has not explored.
Finally, the empirical results in the paper are contribution to the literature on inventory investment.
One puzzle in that literature is that in US data inventory investment does not seem to react to changes
in the federal funds interest rate unless the changes are very persistent (Maccini et al. 2004). To the
extent that the episodes under analysis are associated with persistent increases in interest rates, the
cross-sectional pattern in the data confirms that these increases do interact with inventory holdings
in a way which is consistent with theory.
The focus on the time to produce contrasts with a more common alternative in the quantitative
business cycle literature that, rather than emphasizing the opportunity cost of holding inventories,
emphasizes the cost of other components of working capital, such as cash and trade credit. In contrast,
I take the step back of assuming away the credit and payment frictions that motivate looking at these
components and focus solely on the size of the time gap that needs to be financed. This has some
important pay-offs:
i) The opportunity cost of waiting for production to be completed and distributed is proportional
to the real interest rate. This is not necessarily the case for other short-term assets and liabilities
included in working-capital such as account payables, since these generate interest income of their
own. It is also not the case for cash, whose opportunity cost is tied to the nominal interest rate, and,
hence, to inflation.

ii) Conceptually there is a direct relationship between steady state inventory turnover ratios and
the cost of variable inputs. This is the relationship which is important for the quantitative business
cycle literature that emphasizes working capital as a channel. The same is not true for cash holdings,
which can be held either to facilitate day to day input purchases or to allow firms to take advantage
of sporadic investment opportunities (see, for example, Kiyotaki and Moore (2008) for a model of
cash-holding centering on this latter motive).
iii) Steady state inventory turnover ratios are more likely than other components of working capital
to be associated with technological characteristics of the production process as opposed to highly institutional specific payment or credit practices.2 This robustness to institutional context considerably
facilitates cross-country comparisons.
The empirical findings in this paper are complementary to independent work by Tong and Wei
(2009). Using data from firms in emerging markets, the authors show that a high inventory-tosale ratio was associated with a relatively large drop in share prices in the aftermath of the latest
global crisis. Their price evidence complements the quantity evidence presented here. Raddatz (2006)
presents another set of related empirical findings. He finds that industries with high inventory-to-sale
ratio are relatively more volatile in countries with a low level of financial development.3 His findings
emphasize the role of the financial system in providing liquidity insurance. It 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. Finally, within the United States, Maccini et al. (2004) show
that inventory holdings react to the interest rate as long as the shock is a large break with previous
patterns, and Jones and Tuzel (2010) find a robust relationship between inventory investment and
risk premia.
Time to produce provides a channel through which financial shocks can affect output, but it leaves
open the determination of these shocks. This task is undertaken by Mendoza (n.d.). The author
demonstrates how a business cycle model with an occasionally binding constraint on foreign debt is
able to generate alternating periods of normal business cycle variations and sudden stops.4 In itself,
these frictions only imply a slow drop in output after a sudden stop, and Mendoza relies on a paymentin-advance constraint similar to Neumeyer and Perri (2005) to generate a stronger short run impact
on output. On the empirical side, a similar point applies to studies by Kroszner et al. (2006) and
Dell’Ariccia et al. (2007). They find that, in banking crises, industries that rely more on external
finance to fund long term investment lose relatively more of their output. Their empirical results
illuminate the effect of banking crises on the cost of capital, but they leave open how this exposure
2

For a very interesting but highly context-specific example of how changes in payment arrangements amplified an emerging
market crisis, see Kim and Shin’s (2007) discussion of the use of bills of exchange by Chaebols in South Korea before and
after the crisis.
3
Markups vary less than inventories across firms, so that inventories/cost and inventories/sale ratios are normally closely
linked.
4
Like in this paper, the marginal cost of capital in Mendoza (n.d.) is the same for all firms. Implicitly, the balance
sheet constraints in his model are the ones faced by a representative domestic financial intermediary as opposed to individual
domestic productive units, as suggested by Krishnamurthy (2003).

translates into a drop in output.
Section 2 discusses in some detail the different components of working capital and the motivation
for the focus on inventories. Section 3 gives empirical results for the cross-section of industries in
the aftermath of emerging market crises. Section 4 uses a simple two-sector model to lay out the
main qualitative implications of the empirical results to the understanding of emerging market crises.
Section 5 lays down a detailed quantitative model to allow for an evaluation of the empirical strategy
in section 3 as well as a quantitative assessment of the implications derived in section 4. The last
section concludes.

Components of Working Capital, Costs and Impact

on Production
I start with a discussion of the different components of working capital, the cost of holding them and
how they are linked to the production process. My main purpose is to motivate the focus on inventories
throughout the paper as opposed to other components of working capital, which have received more
emphasis in the literature on financial propagation.
Working capital is defined as the difference between current assets and current liabilities. This is
essentially the sum of 1) inventories, 2) cash and cash-like securities and 3) trade-receivables net of,
4) trade payables.
To fix the notation, suppose that, in order to produce output Yt firms need to use dated inputs
Zt (0), Zt−1 (1), ... , Zt−V (V ) which are purchased at spot prices qt (0), qt−1 (1), ...., qt−V (V ). The
time subscripts t − v refers to when the input is purchased and the numbers in brackets {1, .., V } refer
to the time elapsed between the use of the input and the sale of the output. Importantly, the input
Zt−v (v) is used to produce Yt and not, say Yt+1 or Yt−1 . This explains why Zt (0) and Zt (1) are
different entities, just as are Zt (1) and Zt−1 (1).
This formulation accommodates linear storage technology. It is sufficient to assume that the
various Zt (v)’s are perfect substitutes in production so that, at any period, firms can move resources
to the future by reducing Zt (0) and increasing some Zt (v) with v > 0. To allow for depreciation, it
is possible to redefine the Zt ’s to be the quantity of the input available after depreciation, with the
depreciation costs conflated in the prices.
1) Inventories
Inventories exist because of technological constraints on the timing of production and distribution
of goods. Examples of such constraints are fixed costs of transportation and the need for insurance
against stock-outs. There is a vast literature that examines the details of inventory dynamics given
alternative motivations for inventory holdings (see for example Ramey (1989), Blinder and Maccini
(1991), Ramey and West (1999), Bils and Kahn (2000), Humphreys et al. (2001), Kahn and Thomas
(2007) and Iacoviello et al. (2011), Wen (n.d.) among many others).
Standard inventory accounting uses information on the cost of goods sold at each date (as opposed

to its sales value). The change in end-of-period inventories is the difference between the (spot) cost
of current production and the (spot) cost of the goods sold (COGS). Algebraically,

Invt = Invt−1 +

V
X

qt (v) Zt (v) −

qt−v (v) Zt−v (v)

v=0

V
X

}

Current Production Costs

}

Cost of Goods Sold (COGS)

The difference between the elements in the two summations is in the dates. Whereas all the terms
in the first summation are dated t, the terms in the second are dated t − v. Iterating backwards and
applying the boundary condition limt→−∞ Invt = 0 (firms start out with zero inventories),
Invt =

V X
v−1
X

qt−j (v) Zt−j (v)

v=1 j=0

Inventories are a triangular summation over the dated inputs. In steady state:
V
X

Inv =

vq (v) Z (v)

v=0

Inputs that are used v periods in advance of sales appear in inventories v times.
The ratio of steady state inventory to cost of goods sold (COGS) is:
V

X
q (v) Z (v)
Inv
=
v PV
COGS
v=0 q (v) Z (v)
v=0

Inventories are denominated in dollars and the cost of goods sold is denominated in dollars per
unit of time. Therefore, the ratio is denominated in time – it is the number of months worth of cost
held in inventories. This ratio also shares the formula for the duration of debt, with the cost paid
at each period substituting for coupon payments. This is a second sense in which the steady state
inventory/cost ratio has an interpretation as the average time that it takes to produce and distribute
goods.
Steady state inventories are informative about the steady state opportunity cost of production
in terms of foregone interest. To see this, note that, absent any payment or financial frictions, the
economic (as opposed to accounting) cost of producing Yt is (recall qt (v) are spot prices):
Economic Costt =

V
X

Rt−v,t qt−v (v) Zt−v (v) ,

v=0

where Rt−v,t is the interest rate faced by the firm between times t − v and t. The cost of working
capital finance is the difference between the economic cost and the cost of purchasing the various
inputs in the spot market:

W orking Capital Costt =

V
X

(Rt−v,t − 1) qt−v (v) Zt−v (v)

v=0

In steady state:

W orking Capital Cost =

V
X

(Rv − 1) q (v) Z (v)

v=0

∼
=

V
X

v (R − 1) q (v) Z (v)

v=0

Where the last expression follows from the fact that, for R close to 1, Rv − 1 ∼
= v (R − 1). It then
follows that close to steady state:
W orking Capital Cost
Inv
= (R − 1)
COGS
COGS
Thus, close to steady state, working capital costs are proportional to inventories. Persistent changes
in the real rate of interest that do not alter the inventory/cost ratio by much will have an impact on
the ratio of working capital costs to the costs of goods sold proportional to the inventory/cost ratio.5
Within this context, the cost of working capital is intimately linked to production at a very short
time range. This is the kind of link that allows financial shocks to have an immediate impact on
output.
2) Cash and Marketable Securities
Firms hold cash and marketable securities for a variety of reasons. These securities may facilitate
day-to-day payments of variable inputs, they may serve as cushions to allow firms to insure against
negative cash flow shocks (Bates et al. (2009)), they can help firms take opportunity of fleeting investment opportunities (Kiyotaki and Moore 2008), or they can help them with their tax management
(Foley et al. 2007). Of those motives, the relevant business cycle literature focuses on the first, which
is the payments for variable inputs.
The simplest way to think about the use of cash for payments is through a cash-in-advance model.
This corresponds to the interpretation of Christiano et al. (1992) and is the subject of empirical
investigation by Ramey and West (1999).
A simple form of cash in advance model is as follows:6 Each period is subdivided in two, morning
and afternoon. Firms buy a fraction ψ of their inputs in the morning, but sell their output in the
afternoon. Transactions in the morning have to be paid with cash, but financial transactions and
settlements only occur in the afternoon. Thus, in order to pay for “morning” inputs firms have to
hold cash overnight and forego interest payments. This generates an opportunity cost associated with
5

This will be the case per Shepard’s lemma if the change in the interest rate is small and dated inputs are transformed
into output through a neo-classical production function. A proof is available upon request.
6
The exposition follows Neumeyer and Perri (2005).

production.
P
If, as above, firms buy Vv=0 qt (v) Zt (v) dollars of inputs within the period, they must set aside
P
ψ Vv=0 qt (v) Zt (v) dollars of cash in the end of the previous afternoon. When doing this, they forego
P
(It − 1) ψ Vv=0 qt (v) Zt (v) dollars, where It is the nominal interest rate paid on bonds.
Since the cost of holding cash depends on the nominal interest rate, it increases vastly during
hyper-inflations. This has led the emerging market business cycle literature to abandon the pure
cash-in-advance model of working capital in order to avoid large shifts in the business cycle properties
of the model in the transitions between high- and low-inflation regimes. Instead it has favored a cashin-advance constraint where the opportunity cost of cash is given by the real rate of interest. One
motivation is that firms have often access to a large variety of interest yielding securities that they
can use to perform much of their payments. The opportunity cost of holding these securities is not
the nominal interest rate, but their liquidity premium. The maintained assumption in the literature
is that this premium varies one for one with the real interest rate paid by the governments of these
countries but does not depend on inflation.
Calibrating these models requires first deciding which assets can be used and are in fact used for
day-to-day payment for variable inputs into production. Since cash and marketable securities can have
many different uses, using all of their holdings as a measure of working capital costs may overstate
the short-run link between financial or interest rate shocks and short-term production decisions.
3) Accounts Receivables
The third component of working capital is accounts receivables. This has not been a focus of papers
in the quantitative business cycle literature, although accounts receivables do figure as part of some
of the empirical tests of the working capital channel (see Barth and Ramey (2001) and Gaiotti and
Secchi (2006)). Accounts receivables are generated whenever firms extend credit to their customers.
Since it is extended at the sale of goods, the opportunity cost of accounts receivables is intimately
linked to short-term production decisions.
It is easy to see why accounts receivables that do not pay interest should be accounted for when
calculating the cost of working capital. Under such generous lending standards, waiting for an additional week to receive a payment for a good sold is exactly the same as waiting an additional week to
sell that same good.
In reality, firms do charge interest on their accounts receivables and, in fact, the literature on trade
credit points out that interest rates can be very high (Petersen and Rajan (1997), but see Giannetti
et al. (n.d.)). Also, even if there is evidence that firms with easier access to credit extend more trade
credit to other firms, (Petersen and Rajan (1997)), it is not clear how trade credit extension varies
with aggregate credit conditions.
4) Accounts Payables
The last component of working capital is accounts payables which, contrary to the other components, enter on the liability side. Some papers in the emerging market business cycle literature have
used measures of short-term liabilities held by firms as a way to calibrate working capital demand

(Mendoza (n.d.) and Meza and Benjamin (2009)).
Accounts payables are the relevant quantity to measure the impact of fluctuations of the interest
rate on the cost of variable inputs if they are secured by near-term production processes (are “selfliquidating”) and if alternative sources of funding for short-term assets are not sensitive to the interest
rate.
As an example, suppose firms are owned by entrepreneurs with linear utility and deep pockets.
The marginal value of a dollar for these entrepreneurs is given by their discount rate β. Suppose
entrepreneurs can borrow short term to purchase variable inputs v periods in advance of production
at a rate Rt,t+v < β −1 , so that they choose to borrow as much as they can. Their borrowing limit
is a fraction θ of the value of the inputs that they purchase. Therefore, the opportunity cost of
buying inputs at t for sale v periods in the future is R̃t,t+v = θRt,t+v + (1 − θ) β v . By analogy to the
calculations for the cost of holding inventories, close to steady state the working capital cost of goods
sold at t is:

W orking Capital Cost =
∼
=

V
X
v=0
V
X

R̃v − 1 q (v) Z (v)

v θ (R − 1) + (1 − θ) β −1 − 1

q (v) Z (v)

v=0

It follows that close to steady state:

W orking Capital Cost
COGS

Inv
θInv
+ (1 − θ) β −1 − 1
COGS
COGS

AP
Inv
= (R − 1)
+ (1 − θ) β −1 − 1
COGS
COGS
= (R − 1)

Where AP now are account payables. Thus, fluctuations in the interest rate affect the cost of
working capital to the extent that firms are exposed to it by financing this working capital with
short-term liabilities.
The approach is appealing to the extent that i) the opportunity cost of internal funds held by an
entrepreneur is insensitive to the cost of outside credit, ii) short-term financing is used primarily for
working capital finance, and iii) outside finance for working capital is primarily funded short term.
Whether these are good or bad assumptions depend on specifics about the environment in question
and are unlikely to be robust for a large range of countries with widely diverse financial institutions.

2.1

Why Focus on Inventories

The review above should highlight why inventories are a good starting point to measure the working
capital costs of production. It is the only component of working capital which a) is intimately linked

to short production cycles, b) has an opportunity cost which is clearly proportional to the real interest
rate and c) is robust to the precise institutional environment.

Data Analysis

During emerging market crises, sectors with high inventory/cost suffer a persistent drop in their output
relative to other sectors. The data analysis in this section shows that the differences do not reflect
other plausible explanations, such as alternative sources of cross-industry heterogeneity. Also, sectors
with a high inventory/cost ratio take a long time to catch up with the other sectors, if they catch up
at all. This suggests that the results are not a simple reflection of short-run inventory adjustment
dynamics as in Alessandria et al. (2010), whereby high inventory firms reduce production temporarily
in order to reduce their inventory holdings. Lastly, the results are a robust feature of emerging market
crises, having occurred in the ’80s and the ’90s, in Latin America and in other continents, so that
they do not reflect the specificities of a particular historical or geographical circumstance.
The empirical analysis centers on events where capital inflows to an emerging market drop by an
unusually large amount and where there is a drop in G.D.P. either in the same year or in the following
year (see Appendix A for details). The selection criteria are similar to Calvo et al. (2008) and Calvo
et al. (2006), except that there is no attempt to select clearly exogenous or unpredictable events.7
Thus, these events are not treated as natural experiments, but as interesting episodes which generate
cross-correlations that merit study with a structural model. The episodes are listed in table 1, and
they include most episodes identified by Calvo et al. (2006).
Figure 1 shows the average deviation from trend of different aggregate quantities across episodes.
The trend is projected from data 10 years prior to the crisis. 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), Reinhart and Rogoff (2009) and the International Monetary Fund (2009) that
financial crises have persistent impacts on output. The data replicates a key finding from Calvo et
al. (2008), that investment drops much more than output, implying a drop in the investment/output
ratio. Another interesting fact is that manufacturing output drops strongly after the episode, even
though it is mostly tradable and hence less constrained by drops in the domestic demand components.
As we will see, the co-movement between total output and tradable output is hard to explain if the
only important change in the economic environment is a sudden scarcity of foreign savings (see Kehoe
and Ruhl (2008). Also, see Alessandria et al. (2012) for an attempt to address that same issue). We
will return to these fact when comparing the different models and shock processes in section 5.
I test whether time to produce is associated with lower performance at different horizons after the
crisis. For horizons spanning from the year of the crisis to 8 years afterwards, I run the following
7

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.

Country
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

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

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

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

Country
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

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

Table 1: Crises Episodes
The episodes start in years in which the capital account balance drops by an unusually large amount and
G.D.P. drops either in the same year or in the following one.

Figure 1: Aggregate Variables
Time path of variables at yearly frequency. Figures correspond to average deviation from trend calculated
using average growth rate of previous 10 years.
regression:
yi,k,t∗ +h − yi,k,t∗ −1 = βτk + γXi,k,t∗ + αi + εik
where yi,k,t∗ +h is log value added in industry k, episode i, h years after the start of the event, τk is
inventory/cost ratio for industry k and Xi,k,t∗ is a vector of controls. I also allow for episode fixed
effects αi . The hypothesis is that β < 0, so that firms with longer production time do relatively worse.
The controls account for a large range of sources of cross-industry variation. Most important of all
are the extent to which various industries are producing fixed investment goods or durable consumer
goods since durable goods are notably more cyclical. I also include measures of input shares to pick
up the effect of changes in the relative price of inputs following the episodes, including the share of
imported inputs. A third set of controls captures trend growth for the industry in countries that were
not affected by the crisis. Finally, I include variables that aim to capture the role of financial frictions,
such as average firm size in each sector and the measure of financial dependence 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. The data-appendix (Appendix A) includes a detailed discussion of how all
of these controls are calculated.

3.1

The Inventory/Cost Ratio

Before turning to the results, I discus the measurement inventory/cost ratio in detail, since this is
part of the main contribution of the paper.
Reliable measures of inventory/cost ratios are not easily available for most countries in the sample. Instead, I follow Rajan and Zingales (1998) and use data from US firms. If the inventory/cost
ratios in the US reflect the underlying production technology, they should be a reasonable proxy for
inventory/cost ratios in other countries. To the extent that the proxy is imperfect, the measurement
error will tend to attenuate the results.
For American firms, I aggregate firm-level data from COMPUSTAT into multiple sectors. COMPUSTAT includes decades’ worth of financial reports from listed firms that operate in the US. The
long time series allows me to smooth business cycle fluctuations when calculating the inventory/cost
ratios, obtaining something close to a steady state measure. Importantly, because COMPUSTAT is
based on balance sheet data, the accounting identity between inventories and costs described in section
2 is exact, with whatever inputs that are omitted from inventories also omitted from the cost.8
The working assumption is that technological characteristics of listed US firms are sufficiently informative about the technological characteristics of firms in crisis countries. To validate this assumption,
I check the correlation of the inventory/cost ratios calculated using COMPUSTAT against a measure
of inventory-to-sales data from the Korean Financial Survey Analysis. This survey is representative
of all Korean businesses and inventory-to-sales should be strongly correlated with inventory/cost as
long as markups and the share of fixed inputs are not too variable across industries. The data from
this survey will also be useful for robustness exercises further along.9
There is an unambiguous amount of correlation between the inventory/cost measures from COMPUSTAT and the inventory-to-sales data from the Korean Survey, above 0.6, (table 7). This suggests
that the inventory to cost ratios do reflect underlying technological characteristics.
Table 6 has the descriptive statistics for the two variables. The average inventory/cost ratio
is 2.5 months of costs held in inventories in the COMPUSTAT sample, ranging from less than a
month’s worth of inventories held in printing and publishing to almost four months in manufacture of
electrical machinery. The average inventory/sales ratio is slightly lower in the Korean data, which is
to be expected, given that sales are larger than costs. Figure 2 shows the joint distribution between
the two measures.
The correlation 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 plus an error which is independent of the actual inventory/cost ratio), I
can use the inventory sales data from Korean firms as an instrument for the inventory cost data
8

This contrasts, for example, with the manufacturing survey data, which is only available for a much shorter time span and
does not impose any accounting consistency between inventories and data. The main results are robust to using manufacturing
data and are available upon request.
9
COMPUSTAT data suggest that inventory/cost and inventory/sales are very closely correlated. The calculations are
available upon request.

from COMPUSTAT, thus eliminating the measurement error and any attenuation bias that it may
generate.10 In what follows I report results both of the O.L.S. regressions and the I.V. regressions
correcting for the measurement error.

Time to Produce in US and Korea
323

383
382

COMPUSTAT (United States)
1
2
3

361
314

332

371
356
354
353

351

313
384
369
311

385

381

331

321

390
324

341
352
372

355

342

1
1.5
2
Financial Survey Analysis (South Korea)

2.5

Figure 2: Inventory Intensity Measures
U.S. data are average of Inventories/Cost of Goods Sold across firms in industry, Korean data are average
of Inventories/Sales 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.

3.2

Results

Figure 3 shows the scatter plots for the change in value added for each industry averaged across
episodes against τk , the inventory/cost ratio using COMPUSTAT data for 0, 2, 4 and 6 years after the
beginning of the episode (the growth rates are with respect to the year immediately before). Three
10

The procedure works even if a substantial part of the difference is due to cross-country variation, so long as the measurement error is not correlated with that.

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 4 years after the beginning of the episode. Last, the correlation remains negative even as the
average output increases.

Average Change in Output x years after crisis

.1
0
−.2 −.1

−.2 −.1

Two

Zero

1
2
3
Inventory/Cost (months)

.1
0
−.2 −.1

−.2 −.1

Six

Four

1
2
3
Inventory/Cost (months)

Figure 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 “t” years after that date
These raw correlations may reflect other characteristics of the data which happen to be correlated
with inventory/cost ratio. Table 3 shows O.L.S. regression results and table 2 shows the I.V. regression
results. The general pattern remains the same, with the coefficient under O.L.S. being statistically
significant at the 5% level in years 3-5 and at the 10% level up to year 7, and the coefficient in the
I.V. regressions significant at the 1% level in years 2-4, significant at the 5% level in years 5, 6, 8 and
9 and significant at the 10% level in year 1. Note that in the I.V. regression the coefficients are nearly
twice as large as in the O.L.S. regressions, averaging 6.7% as compared to 3.7%.
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

16
1,097
0.323

1,095
0.343

1,066
0.306

1 year
2 years
-0.0121
-0.0256
(0.0122) (0.0160)
-0.224
-0.126
(0.141)
(0.336)
-0.244
-0.403
(0.372)
(0.530)
0.0991
0.186
(0.121)
(0.138)
-0.193
-0.290
(0.359)
(0.379)
0.0853
0.0200
(0.204)
(0.252)
-0.164
-0.194
(0.149)
(0.165)
-0.0293
-0.0383
(0.0350) (0.0498)
0.0808** 0.0892**
(0.0357) (0.0413)
0.574** 0.593**
(0.290)
(0.288)
0.0257
0.0438
(0.0494) (0.0567)
1,044
0.375

3 years
-0.0436**
(0.0190)
-0.0755
(0.385)
-0.469
(0.556)
0.286*
(0.172)
0.0534
(0.467)
-0.0655
(0.355)
-0.180
(0.149)
-0.0521
(0.0676)
0.131**
(0.0544)
0.716**
(0.280)
0.0386
(0.0502)
1,018
0.437

4 years
-0.0435**
(0.0174)
-0.0464
(0.468)
-0.152
(0.599)
0.156
(0.250)
0.0460
(0.455)
-0.231
(0.398)
-0.141
(0.225)
-0.0663
(0.0745)
0.115*
(0.0650)
0.795***
(0.249)
0.00620
(0.0544)
815
0.486

5 years
-0.0562**
(0.0218)
-0.0155
(0.421)
-0.0400
(0.569)
0.0343
(0.332)
-0.0695
(0.473)
-0.357
(0.396)
-0.0992
(0.295)
-0.0359
(0.0768)
0.104
(0.0745)
0.556***
(0.215)
0.00776
(0.0589)
970
0.511

6 years
-0.0474*
(0.0243)
-0.0889
(0.423)
0.0882
(0.623)
0.269
(0.224)
-0.705
(0.663)
-0.156
(0.384)
-0.254
(0.189)
0.00795
(0.0843)
0.124
(0.0939)
0.542*
(0.282)
0.0154
(0.0711)
729
0.473

721
0.459

7 years
8 years
-0.0464* -0.0427
(0.0274) (0.0264)
0.0784
0.0302
(0.462)
(0.494)
-0.786
-0.721
(0.526)
(0.548)
0.324
0.410
(0.253)
(0.259)
-0.392
-0.443
(0.443)
(0.452)
-0.328
-0.246
(0.557)
(0.514)
-0.185
-0.216
(0.218)
(0.247)
-0.00950 -0.00714
(0.0800) (0.0752)
0.0837
0.0508
(0.103)
(0.106)
0.538** 0.593***
(0.250)
(0.207)
0.0420
0.0593
(0.0907) (0.105)

Table 2: O.L.S. Regressions: Change in Value added from t∗ − 1 to t∗ + h
where t∗ is the starting year of a crisis event. The dependent variable is the log change in value added in the three years
leading to the trough of G.D.P.. The columns refer to different horizons (h). Regressions include country fixed effects (omitted).
Standard errors are robust to overlapping clusters on country and industry following (Cameron et al. 2006).

Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

Observations
R-squared

Previous growth

Growth in Non-Crisis Countries

Size

External Dependence

Imported Inputs Share

Capital Share

Labor Share

Exported Share

Durable Consumption Share

Investment Share

Inventory/Cost (U.S. firms)

0 years
-0.00849
(0.00755)
-0.417**
(0.170)
0.0988
(0.190)
0.187
(0.129)
-0.175
(0.237)
0.119
(0.199)
-0.221
(0.150)
-0.0298
(0.0196)
0.0238
(0.0244)
0.870***
(0.296)
-0.000180
(0.0446)

17
1,097
0.261

1,068
0.254

1,046
0.3

1 year
2 years
-0.0212* -0.0660***
-0.0119
-0.0247
0.281
0.438
-0.312
-0.391
-0.503*
-0.553
-0.3
-0.375
0.00981
0.0837
-0.11
-0.185
-0.172
0.0557
-0.265
-0.363
-0.0777
-0.233
-0.223
-0.34
0.0212
0.0303
-0.0635
-0.124
-0.00577
-0.00765
-0.0329
-0.0533
0.0688*
0.112*
-0.0407
-0.0576
0.697***
0.543*
-0.203
-0.305
-0.0347
-0.0658*
-0.0385
1,020
0.363

3 years
-0.0746***
-0.0237
0.481
-0.47
-0.231
-0.407
-0.0214
-0.271
0.0601
-0.348
-0.393
-0.39
0.0585
-0.245
-0.0247
-0.0614
0.0953
-0.0665
0.698***
-0.253
-0.108*
-0.0588
972
0.451

4 years
-0.0899***
-0.0331
0.503
-0.448
-0.0788
-0.416
-0.132
-0.334
-0.0346
-0.374
-0.507
-0.378
0.0935
-0.302
0.000369
-0.0695
0.0935
-0.0738
0.492**
-0.224
-0.118*
-0.0653
817
0.434

5 years
-0.0779**
-0.0339
0.479
-0.495
-0.176
-0.475
0.127
-0.193
-0.605
-0.483
-0.286
-0.308
-0.101
-0.127
0.0482
-0.0739
0.0881
-0.0834
0.418
-0.256
-0.0893
-0.0628
731
0.425

723
0.418

6 years
7 years
-0.0896** -0.0634*
-0.0417
-0.0366
0.682
0.593
-0.485
-0.444
-0.993** -0.968**
-0.419
-0.386
0.185
0.217
-0.249
-0.241
-0.35
-0.458
-0.322
-0.296
-0.532
-0.464
-0.461
-0.443
-0.08
-0.0665
-0.179
-0.211
0.0458
0.042
-0.0755
-0.0615
0.0507
0.0189
-0.0858
-0.0876
0.354
0.468**
-0.227
-0.197
-0.0684
-0.0401
-0.0705
-0.0848

∗

Table 3: Instrumental Variables Regressions: Change in Value added from t∗ − 1 to t∗ + h
where t is the starting year of a crisis event. The dependent variable is the log change in value added in the three years leading
to the trough of G.D.P.. The columns refer to different horizons (h). Inventory/Cost is calculated using COMPUSTAT data
and is instrumented using data from the South Korean Financial Survey Analysis. Regressions include country fixed effects
(omitted). Standard errors are robust to overlapping clusters on country and industry following (Cameron et al. 2006).

Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

Observations
R-squared

Previous Growth

Growth in Non-Crisis Countries

Size

External Dependence

Imported Inputs Share

Capital Share

Labor Share

Exported Share

Durable Consumption Share

Investment Share

Inventory / Cost

0 years
0.00427
(0.00985)
0.154
(0.189)
-0.333
(0.226)
-0.0997
(0.0791)
-0.0299
(0.223)
-0.0124
(0.128)
0.0734
(0.0762)
-0.00513
(0.0195)
0.0544***
(0.0192)
0.621***
(0.236)
-0.00113
(0.0266)

723
0.426

8 years
-0.0908**
-0.0401
0.68
-0.434
-1.138**
-0.449
0.15
-0.293
-0.427
-0.344
-0.552
-0.432
0.0464
-0.246
0.0776
-0.0731
0.0436
-0.093
0.311
-0.195
-0.0381
-0.0797

in non-crisis countries.
Figure 2 provides a cleaner view of the coefficient of interest estimated for each of the time horizons,
together with the 95% confidence interval. The upper panel is a graphical representation of the first
line in table 3. The lower panel is the same regression coefficient for the I.V. regression. The use of
instrumental variables to correct for the measurement error has a substantial impact on the result.
The coefficients are almost twice as large and statistically significant at the 5% level for almost all of
the years.
The lack of a strong correlation in the first few years is somewhat puzzling and could reflect slow
sectoral reallocation. Alternatively, the early data might be noisy due to imprecisions in the criterion
to select the crisis date.11

3.3

Robustness

3.3.1

Sub-samples

Crisis episodes were heavily concentrated in certain time periods and geographic regions. About half
of the crises in the sample took place in the early ’80s and the other half between the mid-to-late
’90s. Geographically, about half of the crises were in Latin America and the rest distributed among
all other continents. I assess whether the coefficient on the inventory/cost ratio is robust to splitting
the sample in different ways.
The results are in figures 5 and 6. The point estimate of the coefficients is remarkably robust, as is
the overall negative effect. Given that samples are half the size, it is not surprising that the statistical
significance is reduced. The point estimates for the ’90s do seem to mean-revert much more quickly
than in the ’80s.

3.3.2

A Stringent Control: Expected Cyclical Behavior

One concern is that industries are exhibiting their normally expected behavior given that aggregate
output is depressed. To factor this out, I add a control to pick up whether the demand or costs
of particular goods are more or less cyclical for reasons that are not well captured by the different
controls. The control is calculated by running regressions of change in industry output on change in
G.D.P. for periods outside of the crises episodes and using the forecast from this regression for the
crisis period.
This control is very stringent, perhaps overly so. It assumes that non-crisis business cycles are
qualitatively different from crisis episodes. If the results are not robust to adding the control, then
all this implies is that there is no strong evidence that these large scale episodes are fundamentally
11

One case in point is Argentina. According to the criteria, the Argentinean crisis starts in 1999, since this is the first
year in which Argentinean output drops significantly and its current account surplus becomes significantly positive. But the
worst of the crisis only occurs in 2001.

Estimated Coefficient

−.15 −.1 −.05

.05

O.L.S.

5
years after crisis

−.2 −.15 −.1 −.05 0 .05

Instrumental Variable

5
years after crisis

Figure 4: Estimated coefficient for different horizons.
The upper panel shows the coefficient on inventories-to-cost ratio showed in table 3, and the lower panel
shows the coefficients on the I.V. regressions. The bands refer to the 95% confidence interval using two-way
clustering technique following Cameron et al. (2006) with clusters by industry and by country.

Estimated Coefficient
O.L.S.
’90s

−.2 −.1

−.15 −.1 −.05 0

.05

O.L.S.
’80s

5
years after crisis

Instrumental Variable
’90s

−.3 −.2 −.1

−.3 −.2 −.1 0

Instrumental Variable
’80s

5
years after crisis

Figure 5: Estimated coefficient for different horizons and decades
The panels show the results for different decades. The upper panel shows the coefficient on inventories-tocost ratios using COMPUSTAT data, and the lower panel shows the coefficients on the I.V. regressions using
inventories-to-sales ratios from the Korean Financial Survey Analysis as an instrument for the inventoriesto-cost ratio calculated using COMPUSTAT. The left panels show the estimates using crises that started
in the ’80s and the right panels show the estimates from crises that started in the ’90s. The bands refer to
95% confidence interval using a two-way clustering technique following Cameron et al. (2006) with clusters
by industry and by country.

Estimated Coefficient
O.L.S.
Others

−.1

−.05

−.3 −.2 −.1

.05

O.L.S.
Latin America

5
years after crisis

Instrumental Variable
Latin America

−.3 −.2 −.1

−.2−.15−.1−.05 0 .05

Instrumental Variable
Latin America

5
years after crisis

Figure 6: Estimated coefficient for different horizons and continents
The panels show the results for different continents. The upper panel shows the coefficient on inventories-tocost ratios using COMPUSTAT data, and the lower panel shows the coefficients on the I.V. regressions using
inventories-to-sales ratios from the Korean Financial Survey Analysis as an instrument for the inventoriesto-cost ratio calculated using COMPUSTAT. The left panels show the estimates using crises in Latin
American countries and the right panels show the estimates from crises in other continents. The bands
refer to a 95% confidence interval using a two-way clustering technique following Cameron et al. (2006)
with clusters by industry and by country.

different from more regular business cycles. However, to the extent that they are, they pass a very
stringent test.
The result from adding this control is depicted in figure 7. Now the estimated coefficients are
smaller in absolute value but still negative. Furthermore, for some of the years they cease to be
significant at a 95% level, but it is still significant in one year in the O.L.S. regression and four years
in the I.V. regression.

A Small-Scale Model

A small scale model clarifies the role of time to produce in generating cross-sectional dispersion and
a drop in output in response to a shock to the foreign interest rate. The model is a variant of
Mendoza (1991) and Correia et al.’s (1991) small open economy real business cycle model with two
sectors instead of one and time to produce in each of the sectors. The small model allows for a
direct comparison with other papers in the literature. Section 5 below uses a much more detailed
quantitative model to provide a closer link to the data analysis in section 3, as well as explores the
interaction of time to produce with other mechanisms of interest.

4.1
4.1.1

Model Setup
Households

There is a representative household, who is able to borrow and lend abroad at an exogenous riskless gross interest rate Rt . It can also borrow and lend to the firm. The household supplies labor,
accumulates capital and consumes. The utility function of this household is time-separable with
discount rate β < 1 and the period utility is
1
u (Ct , Lt ) =
1−σ

1−σ

1
1+ ψ
t
Ct − (1 + g) χLt

where Ct is consumption and Lt is labor supply and g is the growth rate of the economy.
This utility function follows Greenwood et al. (1988). One motivation is that labor supply is
costly because it implies a loss in output from home production or self employment. This motivation
also accounts for why the growth rate is part of the utility function (this is also necessary in order
to achieve a balanced growth path). Such an interpretation is particularly suitable for developing
countries, where a large fraction of the population is self-employed in an informal sector whose output
is less likely to be well measured. An important implication of this utility function is that labor
supply does not respond to the wealth of the household, only to wages. As a consequence, Small Open
Economy business cycle models equipped with this utility function are more successful in generating
realistic dynamics for the volatility of consumption and the cyclicality of the balance of trade (Correia
et al. 1991).

Estimated Coefficient
Control for Cyclicality

−.1

−.05

.05

O.L.S.

5
years after crisis

−.2−.15−.1−.05 0 .05

Instrumental Variable

5
years after crisis

Figure 7: Estimated coefficient for different horizons, controlling for expected cyclical effect.
The panels shows results controlling for the expected behavior of each industry in non-crisis periods. The
upper panel shows the coefficient on the inventories-to-cost ratios using COMPUSTAT data, and the
lower panel shows the coefficients on the I.V. regressions using inventories-to-sales ratios from the Korean
Financial Survey Analysis as an instrument for the inventories-to-cost ratio calculated using COMPUSTAT.
The bands refer to a 95% confidence interval using two-way clustering technique following (Cameron et
al. 2006) with clusters by industry and by country.

Consumption is a composite of the goods produced in the two sectors and imports:
1 1 1−γC∗
∗
2
2
C2,t
Ct = ΛC C1,t
(C∗,t )γC
Since this is a small open economy, the relative price of the consumption goods is set abroad. We
normalize them to be equal to 1 and pick Λ so that total consumption expenditures are identical to
Ct .
The household’s problem is to maximize the expected discounted sum of utility functions subject
to the budget constraint and the capital accumulation constraint

BtH

−

H
Rt−1 Bt−1

2
X

ρi,t Ki,t + wt Lt − Ct −

i=1

2
X

Ii,t −

i=1

Ki,t = (1 − δ) Ki,t−1 + Ii,t −

κ 2
B
2 t

ζ (Ki,t − (1 + g)Ki,t−1 )2
, i ∈ {1, 2} ,
2
Ki,t−1

where BtH is household financial assets including net foreign assets and firm debt, wt is the wage rate
received by the household, ρi,t is the rental price of capital used in sector i and Ii,t is the amount
invested in capital of sector i. Capital is sector specific and is produced by combining old capital
and investment goods. There is a quadratic adjustment cost to capital, parameterized by ζ. The
parameter κ denotes a small financial friction that keeps the debt level stationary.
As with consumption, investment is a composite of goods produced in the different sectors and
imports.
Ii,t

1 1 1−γI∗
∗
2
2
I2i,t
(I∗i,t )γI
= ΛI I1i,t

The price of investment is also normalized to 1 so that It corresponds to total expenditures in
investment goods.

4.1.2

Firms

The production function of each sector is
Yi,t = Zi,t (0)1−ωi Zi,t−1 (1)ωi
where Yi,t is the production of the input i in t. Zi,t (v) denotes a composite of inputs acquired in
period t to produce the finished good which is sold v periods ahead. Zi,t (v) is defined as:
Zi,t (v) = γi (Ki,t−1 (v))αi (1 + g)t At Li,t (v)

1−αi

where Ki,t−1 (v) is capital stock, Li,t (v) is labor and, as before, g is the growth rate of the economy. As
before, the index v implies that the input is meant for production of the output that will be available
v periods in the future.
Firms maximize expected discounted profits, defined as:
∞
X

s=0

Rt,t+s

"
Yi,t −

2
X

wt Li,t (v) −

v=1

2
X

#
ρi,t Ki,t (v)

v=1

subject to the dynamic budget constraint:
F
Bi,t+1

−

F
Rt−1,t Bi,t

= Yi,t −

2
X

wt Li,t (v) −

v=1

2
X

ρi,t Ki,t (v) ,

v=1

F are financial assets held by the firm in sector i, w is the wage rate and ρ is the rental rate
where Bi,t
t
i,t

of capital for sector i. Note that, in general, asset accumulation by firms is not equal to zero. The
difference corresponds to the working capital financing need of the firm. Financial asset accumulation
is also closely related to inventory accumulation, since it is determined by the difference between sales
and current production costs.

4.1.3

Resource Constraints

While capital is sector specific, there is no limit in reshuffling it for production of output at different
dates. This implies the following resource constraint for the capital stock of sector i:12
2
X

Ki,t (v) ≤ Ki,t .

v=1

Labor is perfectly mobile between sectors. This implies the following constraint:
2 X
2
X

Li,t (v) ≤ Lt .

i=1 v=1

Exports from sector i are identical to the total output from sector i minus the sum of the domestic
demands for this good:

EXPi,t = Yi,t − Ci,t −

2
X

Iij,t .

j=1

Total imports are the sum of all foreign produced goods demanded for different uses:

IM Pt = C∗,t +

2
X

I∗j,t ,

j=1
12

The allocation procedure only determines ui,t (v) Ki,t (v) but not the individual components. This is irrelevant for any
of the results, since they only depend on ui,t (v) Ki,t (v).

The prices of the foreign goods are constant over time and are normalized to be equal to 1. The
current account identity is
Bt∗ − Rt−1 Bt∗ =

2
X

EXPi,t − IM Pt .

i=1

where Bt∗ is the amount of net foreign assets that domestic households and firms hold. From Walras’s
law, it follows that:
Bt∗ = BtH + BtF

4.1.4

Equilibrium

Given paths for the exogenous variables {Rt , At }, an equilibrium is given by quantities {Ct , Ci,t , Ii,t , Iij,t , Ki,t (v) ,
Li,t (v) , EXPi,t , IM Pt , BtH , BtF , Bt∗ } and prices {wt , ρi,t } so that both households and firms optimize
and the various resource constraints are satisfied.

4.2

Calibration

As written, the model has few parameters that need calibration, most of which are fairly standard.
I follow Neumeyer and Perri’s (2005) baseline calibration wherever the two models share common
parameters.
We set the import shares to 20%. This is inconsequential since, as written, the model implies no
movement in exchange rates and hence no interesting substitution between home and foreign goods.
The most important unconventional parameter is ω. The same inventory/cost ratios used in the
empirical section provide the discipline.
To see this, note that steady state inventories are identical to the cost of paying for the early inputs
that enter Zi (1). This follows from the inventory account equation:
Invi,t = Invi,t−1 + Current Costsi,t − COGSi,t ,
where as in section 2, Invi,t are inventories held in the end of period t and COGSi,t are the cost of
goods sold. In this environment, the costs are:

Current Costst = wt (Li,t (0) + Li,t (1)) + ρi,t (Ki,t (0) + Ki,t (1))
COGSt = wt Li,t (0) + wt−1 Li,t−1 (1) + ρi,t Ki,t (0) + ρi,t−1 Ki,t−1 (1) .
Note that all the terms in (0) cancel out so that:
Inventoriesi,t = Inventoriesi,t−1 + wt Li,t (1) − wt−1 Li,t−1 (1) + ρi,t Ki,t (1) − ρi,t−1 Ki,t−1 (1) .

Iterating backward and applying the boundary condition limt→−∞ Inventoriesi,t = 0,
Inventoriesi,t = wt Li,t (1) + ρi,t Ki,t (1) .
The steady state inventory to costs ratio is equal to:
1
wLi (1) + ρi Ki (1)
Invi
R ωYi
=
=
ω
COGSi
wLi (1) + wLi (0) + ρi Ki (1) + ρi Ki (0)
1−ω+ R
Yi

where the second equality follows from the first order conditions of the firm’s problem. If R ∼
= 1, it
follows that
Invi ∼
= ω.
Costi
Note that we get the same result if the rental cost of capital is not accounted for in inventories
and in the cost of goods sold. In that case, we have:
1
αi ) Yi
wLi (1)
Inventoriesi
R ω (1 −
∼

=
=
=ω
ω
Costi
wLi (1) + wLi (0)
1 − ω + R (1 − αi ) Yi

The 1 − αi cancels out and we have the same equation as before. The crucial assumption is that
the input shares do not vary with the time lag in which they are used. In the complete quantitative
model with intermediate inputs, a similar set of assumptions will ensure that the calibration is correct
if only intermediate goods are accounted for in inventories.13
The inventory-cost ratios calculated from the COMPUSTAT data suggest a range between 0 and
1 quarter, with the median close to half a quarter. Thus, two values of ω that are consistent with
that range are ω1 = 0.25 and ω2 = 0.75. In the quantitative model below I will use the actual sector
by sector calculations of the inventory to cost ratios to calibrate each different sector.

4.3

Simulations

The solid blue line in figure 4.3 shows the reaction of different variables in the model to a unit shock
to the foreign interest rate that mean reverts at a rate of 78% per quarter. This coefficient of mean
reversion is consistent with Neumeyer and Perri’s (2005) estimates for the process of interest rate
shocks. The results coincide with their findings, with consumption dropping more than output and
investment even more so. With consumption and investment dropping more than output, net exports
rise. The model also generates a prediction for sectoral reallocation, with the difference in output
between sectors 1 and 2 increasing on impact. Recall that sector 1 is the one with a smaller share of
“early” inputs so that the result is consistent with it suffering a smaller impact.
The green line marked with circles shows the same variables in the model without time to produce
13

This is not generally the case. Work-in-process inventories include the direct labor costs of production, as do the finished
goods inventories.

Consumption

GDP
0.2

−0.2

0.1

−0.3

−0.4

−0.1

−0.5

−0.2

−0.6

−0.3

−0.7

−0.4

−0.8

−0.5

−0.9

Value Added1 − Value Added2

Investment
−0.5

−1

0.8

−1.5

0.6

−2
0.4
−2.5
0.2

−3

−3.5
−4

Time to Produce

−0.2

12
No Working Capital

Payment in Advance

Figure 8: Impulse Response Function for R shock - Simple Model
Impulse response functions to a unit shock to the foreign interest rate with mean reversion parameter
0.78. The blue solid line (“Time to Produce”) is the model proposed in the paper. The green line marked
with circles (“No Working Capital”) is the standard model where there is no working capital need and the
dashed red line (“Payment in Advance”) corresponds to the setup in Neumeyer and Perri (2005). See text
for calibration. .

(ω1 = ω2 = 0) or any other form of working capital. It makes it clear why some form of working
capital financing cost of production is essential for the interest rate shock to generate an immediate
drop in output. Without such a cost, output drops only slowly as fixed capital is gradually adjusted
downward.
The red dashed line in figure shows how the model compares to Neumeyer and Perri’s (2005)
payment in advance constraint setup. Taken literally, the payment in advance setup does not imply any
end-of-period inventories. Since all production takes place within the period, it cannot be disciplined
by inventory holdings. Still, to facilitate the comparison, I set the share of “cash” inputs in the
payment-in-advance setup to be the same as the share of early inputs in the time to produce setup.
In this setup firms only learn about the shocks after they have incurred the opportunity cost of
holding “cash” inputs between periods. Hence, the effect of increased working capital costs on output
is delayed. In contrast, the time-to-produce model generates a smoother initial response. The reason
is that even if current sales are largely set before the shock, firms immediately reduce their production
of early inputs associated to sales in the following period. Past this short-term adjustment, the two
models behave essentially the same way.
Figure 9 shows the implications of a T.F.P. shock to output and sectoral reallocation in the three
variants (with time to produce, without time to produce and with payment in advance constraint).
There is very little difference except that the response of output is a little bit more delayed in the
presence of the time to produce and that there is some initial oscillation in sectoral allocation of
output. Importantly, the T.F.P. shock is unable to generate any persistent sectoral reallocation when
the only source of heterogeneity is in the demand for working capital.

4.4

Main Lessons

The time-to-produce model has very similar implications to Neumeyer and Perri’s (2005) cash in
advance model. The finding confirms that the way the quantitative literature has been modeling
working capital demand should not in and of itself have strong implications for the model dynamics.
The simple model also makes it clear that working capital is only important for the dynamics of
the aggregate economy in the presence of a foreign interest rate shock. In contrast, the effects of a
T.F.P. shock largely independent on whether or not there is some form of working capital. The same
is true for the cross-sectional dynamics, so long as the only difference between sectors is given by the
intensity of working capital demand.
The major benefit from using time to produce as opposed to Neumeyer and Perri’s (2005) payment
in advance constraint is that it allows to use inventories to discipline the model. This is significant because it allows for important discipline in the role of working capital. However, in order to fully exploit
this benefit, it is necessary to have a more realistic model that allows for a variety of amplification
and feedback mechanism. We turn to this in the next section.

GDP

Consumption

−1.5

−2.6

−1.6

−2.7

−1.7

−2.8

−1.8
−2.9
−1.9
−3
−2
−3.1

−2.1

−3.2

−2.2
−2.3

−3.3

0.5

−2.1

0.4

−2.2

0.3

−2.3

0.2

−2.4

0.1

−2.5

−2.6

−0.1
0

GDP1 − GDP2

Investment
−2

−2.7

Time to Produce

−0.2

12
No Working Capital

Payment in Advance

Figure 9: Impulse Response Function for T.F.P. shock - Simple Model
Impulse response functions to a unit shock to the total factor productivity with mean reversion parameter
0.78. The blue solid line (“Time to Produce”) is the model proposed in the paper. The green line marked
with circles (“No Working Capital”) is the standard model where there is no working capital need, and
the dashed red line (“Payment in Advance”) corresponds to the setup in Neumeyer and Perri (2005). See
text for calibration.

Quantitative Model

The simple model above clarified the mechanisms at play. The next step is to set up a model which
is rich enough to accommodate a range of mechanisms relevant in the literature. A more complete
model also allows for a meaningful effect of a range of sources of cross-sectoral heterogeneity.
The purpose is two-fold. First, the richer model, with more detailed cross-sectoral heterogeneity
allows for an assessment of the empirical strategy adopted in section 2, specifically the choice of
controls that make sure that cross-sectional differences reflect differences in inventory holdings and
not other differences unrelated to working capital demand. Second, the model allows for a quantitative
assessment of the role of time to produce in generating a sharp drop in output in response to a shock
to foreign capital flows.
These additions make the model setup quite involved, so that to simplify the exposition, details
are relegated to appendix B. The model is, at its core, a generalization of Long and Plosser (1982) to
an open economy with time to produce calibrated to match average inventory/cost ratios. In addition
to the stripped down model in section 4, the quantitative model features a number of extensions
that, over time, the research on emerging market business cycles has found to be meaningful (see for
example Kehoe and Ruhl (2008), Meza and Benjamin (2009), Gertler et al. (2007) among others):
• General CES utility function over goods from different sectors, as well as a general CES production function over inputs produced in the different sectors.
• Non-Tradable Production: Most production, including services and construction is non-tradable.
With non-tradables the economy has fewer potential margins of adjustment since a sizeable
part of output has to be used domestically. This means that non-tradables are potentially
an important source of general equilibrium effects. Furthermore, non-tradables introduce an
additional way to evaluate the model, which is its ability to generate co-movement between
tradables and non-tradables.
• Input-Output linkages: Allowing for input-output linkages brings the model closer to a small
open economy version of Long and Plosser (1982). Relative to that classic paper, one innovation
is that the time lags in production that Long and Plosser (1982) assume to be identical to one
quarter are in fact calibrated to match inventory holdings in different sectors. The extension
allows for an assessment of the extent to which the kind of sectoral linkages that Long and Plosser
(1982) argued for could account for co-movement in the presence of idiosyncratic shocks would
also serve a similar role in the presence of an interest rate shock with potentially very different
implications for production costs in different sectors.
• Time to build in construction. Unfinished structures are typically not accounted for as inventories, but as part of the stock of physical capital. Thus, the calibration procedure used in the

paper does not properly capture all the relevant production lags in the economy. Allowing for
time to build in the construction sector adds that feature.
• Household demand for durable goods: The main purpose of this addition is to allow for an additional degree of heterogeneity across manufacturing sectors, thus allowing for a more complete
assessment of the empirical strategy in section 3.
• Capacity Utilization Margin: This has an important effect on the quantitative results of the
model as it allows for a further margin of adjustment to shocks. In particular, as highlighted by
King and Rebelo (1999), capacity utilization significantly increases the ability of small shocks to
generate large business cycle fluctuations.

5.1
5.1.1

Calibration
Sectors and Shares

The model has 15 tradable sectors and 3 non-tradable sectors. The tradable sectors seek to match the
manufacturing sectors in the empirical section (see appendix B for details). The three non-tradable
sectors allow for differentiation between 1) construction, which has a long time to produce, and 2)
services and retail, which have a short time to produce. The third non-tradable sector is residential
services. Its sole purpose in the model is to absorb the residential capital built in the construction
sector, thus allowing for a realistic calibration of the share of construction in manufacturing investment.
To calibrate the model, I use an average of the same input-output tables used in the data analysis
section 3. The resulting input-output table has 46 sectors which I then aggregate into 18. In order
to find the shares, I need to take into account that non-tradable sectors do in fact trade and that
the steady state of the model features balanced trade. Appendix B discusses the assumptions and
re-balancing procedures needed to accommodate them.

5.1.2

Time to Produce

The calibration process for time to produce in the manufacturing sectors follows exactly the same
procedure as the one exposed in section 4 above, with the ωi ’s chosen to match inventory to cost
ratios. Specifically, I choose ωi for each sector so that the inventory/cost ratio matches the numbers
constructed from COMPUSTAT.14
The production function for the services and residential services sectors is not subject to any time
to produce, so that it is simply
Yi,t = Zi,t (0) .
14

For sectors which encompass more than one sector in my data set, I pick the number corresponding to the one with the
relatively larger value added.

Finally, the construction sector requires inputs to be accumulated over four quarters, just as in
Kydland and Prescott (1982):
1

4
4
4
4
YConst,t = ZConst.,t
ZConst.,t−1
ZConst.,t−2
ZConst.,t−3
.

(1)

Counter-Factual Calibration: For the sake of comparison, I also include a counter-factual
calibration. I assume that all manufacturing sectors produce instantaneously, so that
Yi,t = Zi,t ∀i ∈ {1, ..., NT } .

5.1.3

Preferences, Capital Accumulation and Capacity Utilization Cost

For the capital accumulation equations, I set the depreciation of fixed capital δi equal to 10% annually
for all sectors except residential services. I set the depreciation rate for capital in the residential
services sector to 2.3% annually. Finally, I set the depreciation of durable consumer goods δH to 17%.
These values approximate the rate used by the Bureau of Economic Analysis to depreciate stocks of
physical assets in the US economy.
I model capacity utilization costs as stemming from greater need for maintenance. Maintenance
in the model requires the use of a composite good which is identical to investment, except that I
exclude output from the construction sector. This is a reasonable assumption given that commercial
and industrial real estate is less likely to have its depreciation depend on the speed of production. It
is also quantitatively relevant, since it keeps firms from deciding to ramp up production in response
to a drop in real estate prices. I calibrate the sensitivity of maintenance costs to capacity utilization,
to 0.5. It is close to the estimate by Burnside and Eichenbaum (1996). I also present results when
the capacity utilization margin is shut down.
I experiment with the same levels of capital adjustment cost used by Neumeyer and Perri (2005),
namely 8, 25.5 and 40, with 25.5 being the benchmark, and with the same levels of elasticity labor
supply, 31 , 35 , and 5, with

5
3

as the benchmark.15 I also set the curvature of the utility function σ, to

2. The household discounts the future at a rate β. I set that to

1
R(1+g)σ

in steady state, where g is

the steady state growth rate of labor productivity, set to 2%. This discount rate ensures that there
is no tendency for the household to accumulate or decrease assets, and choose R̄ to match the steady
state investment to output ratio.
Lastly, I set the elasticity of substitution between goods produced in different sectors to 2.

5.1.4

Shock Processes

Two exogenous variables drive the economy: total factor productivity and interest rate on foreign
borrowing.
Neumeyer and Perri (2005) calibrate the exponent on employment in the utility function ν = 1 + ψ1 . The numbers above
refer to ψ.
15

ρ
σ
ψ
β
ζ
δi
δreal estate
δH
ξ
r̄
κ
χ
g

Parameter
Elasticity of substitution between home sectors
Relative risk aversion
Elasticity of labor supply
Discount rate
Quadratic Adjustment Cost Parameter
Depreciation of fixed capital
Depreciation of capital in residential services sector
Depreciation of durables
Elasticity of maintenance cost to utilization
Steady state interest rate
Cost of holding debt
Coefficient on Employment in utility
Steady State Growth Rate

Value
2
2
5
3

0.9274% (per year)
25.5
10% (per year)
2.3% (per year)
17% (per year)
0.5
12.22% (per year)
10−5
2.48
2% (per year)

Table 4: Benchmark Calibration
The interest rate shock as modeled here has two important characteristics: It is not insurable and
it is not correlated with any change in the level or composition of foreign demand for domestic output.
These two assumptions are consistent with a model, such as Mendoza (n.d.), Chari et al. (2005)
and Kehoe and Ruhl (2008), where there is a sudden quantitative limit on the amount of resources
in excess of total exports that can be transferred from abroad. Per definition, this limit cannot be
insured against, and, in a decentralized setup, the imposition of this limit induces a wedge between
the interest rate faced by domestic agents and that faced by foreigners. Allowing for an interest rate
shock instead of a quantitative limit also accommodates models where the quantitative limit is not
rigid, but the interest rate is increasing in the amount of capital inflows. A flexible enough interest
rate shock could replicate the wedge between domestic and foreign interest rates induced by any such
model.
I consider the impact of an interest rate shock which is part transitory and part permanent. The
transitory component follows the process estimated by Neumeyer and Perri (2005) for the risk premium
paid by the Argentinean government on its foreign debt to calibrate the transitory component of the
interest rate shock.
The permanent component has a mean reversion parameter of 0.999. A very persistent increase
in the cost of foreign borrowing is consistent with two pieces of evidence. First, after emerging
market crises, credit aggregates and current account balances do not return to their pre-crisis level.
Second, the investment drops much more than output, to a level which cannot be easily explained by
a persistent drop in productivity (see Calvo (1998) for all of these facts).
I set the size of the foreign interest rate increase to a 10% increase in the annualized rate, consistent
with the increase in the real interest rate in Argentina in 1995 and in 2001, as well as the increase in
Mexico in 1995 (I calculate this using the data made available by Neumeyer and Perri (2005)), with

the permanent part accounting for 1 percentage point of the initial increase.
I model the productivity shock as a permanent shift in the level of productivity, as suggested by
Aguiar and Gopinath (2007). Such a permanent shift is consistent with the aggregate data discussed
in the beginning of section 3, as well as with findings by the International Monetary Fund (2009) for
the evolution of G.D.P. in the aftermath of financial crises in emerging economies. I calibrate the size
of the productivity shock so that the economy settles at a level of G.D.P. 15% below its trend, as it
appears to do on average in the data presented in figure 1 shown in section 3,
I look at the impulse response functions to individual shocks. This, rather than full-sample correlations is more likely to isolate the way the economy accommodates the extreme events analyzed in
the data section. To solve the model, I log-linearize the model equations around the non-stochastic
steady state and then solve for the rational expectations equilibrium. The linear approximation is
necessary because of the large number of state variables.16

5.2

Impulse Response Functions

Figure 10 shows the response of aggregate variables to the interest rate and the productivity shock.
As with the simple model, consumption responds to the interest rate shock by dropping more than
output. With time to produce, the drop in output is immediate and sizeable, with a 10% shock to the
interest rate generating a 1.9% drop in output relative to trend in the first quarter, whereas without
time to produce the initial impact is slightly positive, becoming increasingly negative overtime, as
fixed capital disperses. The intuition is the same as in the simple model: without time to produce,
the interest rate shock does not have an immediate impact on production costs.
The model also generates separate responses for tradable G.D.P. Without time to produce there is
an initial boom in tradable production, which grows by 8.3% on impact and does not enter negative
territory until more than three years after the initial impact. The reason is that non-tradable output
is constrained by domestic demand, which is depressed by the lower interest rate. The inputs freed
by non-tradable production flow into tradable output, which can then be exported.
With time to produce, tradable output has close to no reaction on impact, and then slowly drops.
Thus, the model is successful in generating at least some co-movement between the tradable and
non-tradable sectors over the first two years after the crisis.
Finally, consumption and investment drop somewhat less in the model with time to produce than
in the model without. This is because with time to produce there is less of an incentive for resources
to flow from the production of consumption and investment goods (which have a high non-tradable
share) to the production of exports.
The reactions to the productivity shock are largely independent of whether or not the model allows
for time to produce. I investigate the response of output to a 15% permanent drop in T.F.P.. This
very high number matches the average deviation from trend in the data after 9 years. Even though
16

I solve the model using Dynare.

GDP

Consumption

0.02

−0.02

0
−0.03
−0.02
−0.04

−0.04
−0.06

−0.05

−0.08
−0.06

−0.1
−0.12

−0.07

−0.14
0

−0.08

Investment

Value Added in Manufacturing

−0.08

0.1

−0.1

0.05

−0.12
0
−0.14
−0.05
−0.16
−0.1

−0.18

−0.2

Time to Produce, R shock
−0.15
No time12
to produce, R shock 0
Time to Produce, TFP shock
No Time to Produce, TFP shock

Figure 10: Impulse Response Function for Interest Rate and T.F.P. shock - Quantitative Model
Impulse response functions to a 10% rise in the foreign interest rate and a 15% permanent drop in productivity. The blue solid line (“Time to Produce”) is the model proposed in the paper. The green line
marked with circles (“No Time to Produce”) is the model without time to produce in the manufacturing
sectors. See text for calibration and discussion of the underlying models..

productivity drops rapidly, output declines slowly because of the adjustment cost to capital. Because
the shock is persistent, permanent income leads consumption to drop very strongly, consistent with
Aguiar and Gopinath’s (2007) insight. All variables co-move.
One important difference between the interest rate shock and the productivity shock is that the
interest rate shock generates a drop in investment which is disproportionately large compared to
output. In fact, even though the drop in G.D.P. in response to the interest rate shock is comparatively
muted, the investment/output ratio drops by 9.2% between years 1 and 6 after the crisis, which is
more than half of the 17% drop in the data. In contrast, the 15% drop in T.F.P. generates a drop in
investment to G.D.P. ratio of only 2.7%.17

5.3

Cross Section: The Coefficient on Inventory/Cost Ratio

I repeat the regressions in section 3 in the model. I include as explanatory variables all the variables
for which there is a model counterpart. This includes all the demand and cost shares as well as the
steady state inventory/cost ratios. It excludes average firm size, external dependence, world trend
and previous trend.
Figure 12 shows the coefficients on inventory/costs estimated from the model for different time
horizons, with and without time to produce and with both the foreign interest rate shock and the productivity shock. The figure highlights that both time to produce and a persistent shock to the interest
rate are required in order to generate sizeable coefficient of industry output on the inventory/sales
ratio. The coefficient is very close to the I.V. estimates, averaging -7.6% between years 1 and 6 after
the shock, compared to -6.98% in the data.

5.4

Extension: Adjustment Costs to Labor

The model with time to produce and interest rate shock fails to the extent that it implies a very large
coefficient in the years immediately after the crisis, when in the actual data the sectoral differences
increase much more slowly. One way to correct this is to add adjustment costs to labor (see appendix
B for details). I add the same adjustment costs to labor as the ones added to capital. They are
quadratic and sector specific. I also set the adjustment cost parameter in labor to be the same as in
capital, 25.5.
The coefficients in figure 12 below are the ones implied by the model with adjustment costs to
labor. Now there is a steady increase of the effect over time, just as in the data. Contrary to the data,
there is no sign of mean reversion in the coefficient, suggesting that the simulation may overstate the
degree of persistence in the interest rate shock. Again, both a persistent interest rate shock and time
to produce are necessary.
17

This particular result echoes the findings in Otsu (2007).

0.02

−0.02

−0.04

−0.06

−0.08

−0.1

Time to Produce, R Shock
No Time to Produce, R Shock
Time to Produce, TFP shock
No Time to Produce, TFP Shock

−0.12

−0.14

Figure 11: Coefficient of Output on Time to Produce
Output in each sector is calculated as the value added over the course of a year. Since not all crises started
in the first quarter, the points correspond to an average over crises starting in each of the four quarters.

0.01

−0.01

−0.02

−0.03

−0.04

−0.05

−0.06

Time to Produce, R Shock
No Time to Produce, R Shock
Time to Produce, TFP shock
No Time to Produce, TFP Shock

−0.07

−0.08

Figure 12: Coefficient of Output on Time to Produce with Labor Adjustment Costs
Output in each sector is calculated as the value added over the course of a year. Since not all crises started
in the first quarter, the points correspond to an average over crises starting in each of the four quarters.

5.5

Sensitivity

The quantitative model has many constituent parts and many parameters. Table 5 deconstructs the
role of these different parts and parameters in generating the aggregate and sectoral patterns in the
data. The initial model is one which extends the model in section 4 by adding multiple sectors, some
of which are tradables and some of which are not, inter-sectoral linkages and heterogeneity in the
shares. Each subsequent row adds an additional feature to the model: time to build in construction,
demand for durable consumer goods, capacity utilization margin and adjustment costs to labor. The
last rows show the sensitivity of the benchmark model to different parameterizations.
The key results change very little with the degree of complexity of the model or the precise
parameterizations. Under all parameterizations, a combination of time to produce and the interest
rate shock is necessary to generate a large predictive effect of the inventory/cost ratio on sectoral
output. Furthermore, time to produce enhances the impact of the interest rate shock on impact.
Lastly, under all models and parameterizations, the T.F.P. shock generates a medium-run drop in the
investment to G.D.P. which falls short of the data. In the closest entry (ζ = 8) the ratio drops by
only about half of what it drops in the data, even as it overshoots the output drop.
The addition to the model which has the greatest impact on the quantitative results is the capacity
utilization margin. Adding capacity utilization more than doubles the impact of the interest rate shock
on G.D.P. as well as the average coefficient of sectoral output on the inventory/cost ratio. In terms of
long-run ratios between G.D.P. and other aggregates it changes relatively little. Adding adjustment
cost to labor reduces the initial impact of the interest rate shock to 1.2% and reduces the average
coefficient of output on inventory/cost ratio to about 5.5%. Otherwise results are very similar.
Finally, in terms of individual parameters, there is an interesting split between the adjustment cost
to capital and the elasticity of labor supply, with the first having its major impact on the coefficient
of sectoral output on the inventory/cost ratio and the latter having a large impact on the response of
aggregate G.D.P. to the shock. The directions are as one would expect, with higher adjustment costs
leading to a reduction in cross-sectoral differences and higher elasticity of labor supply leading to a
larger drop in output.

5.6

Alternative Explanations

Here I discuss other mechanisms that do not rely on increased opportunity cost of holding capital to
explain the data in section 3 and argue why they are unlikely to account for the findings.
One plausible mechanism to explain the cross-section relies on inventory adjustment dynamics as
the one explored by Alessandria et al. (2010). Firms typically target some inventory-to-sales ratio. If,
in response to a recessionary shock, firms desire to reduce their sales, they will also want to reduce
their inventories. This requires a reduction in production above and beyond the reduction in sales.
The effect will be larger if firms hold more inventories. The reason why this mechanism cannot explain
the pattern in the data is that, since inventory/sales ratios are less than a quarter, firms can easily

T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P
T2P

I - G.D.P.
(years 1-6)
-17.0%
-10.3%
-9.6%
-2.7%
-10.3%
-9.6%
-2.7%
-10.4%
-9.7%
-2.7%
-9.2%
-8.9%
-2.4%
-9.2%
-8.9%
-2.4%
-15.8%
-15.3%
-8.3%
-7.0%
-6.9%
-0.2%
-10.4%
-10.1%
-2.5%
-8.7%
-8.3%
-2.2%

C - G.D.P.
(years 1-6)
3.1%
-0.3%
-0.2%
1.3%
-0.3%
-0.2%
1.4%
-0.4%
-0.3%
1.4%
-0.7%
-0.5%
2.1%
-0.7%
-0.6%
2.1%
0.3%
0.4%
2.6%
-0.9%
-0.8%
2.0%
-0.2%
-0.1%
2.5%
-1.5%
-1.3%
2.0%

G.D.P. Trad. - G.D.P.
Coef. on Inv/Cost
(years 1-6)
(years 1-6)
2.0%
-3.79% (O.L.S), -6.98 % (I.V.)
0.7%
-2.0%
1.1%
0.1%
0.1%
-0.9%
0.8%
-1.9%
1.2%
0.1%
0.1%
-0.9%
0.7%
-2.0%
1.0%
0.1%
0.3%
-0.8%
0.8%
-7.6%
1.9%
0.3%
-0.7%
-1.3%
0.8%
-5.5%
1.7%
0.2%
-0.5%
-0.9%
1.6%
-13.6%
2.8%
1.0%
-0.6%
-1.5%
0.6%
-6.6%
1.6%
0.2%
-0.8%
-1.3%
0.0%
-7.9%
1.3%
0.2%
0.0%
-1.0%
1.7%
-7.4%
2.6%
0.3%
-1.3%
-1.6%

Table 5: Sensitity Analysis
R, T 2P corresponds to the results given an interest rate shock and time to produce in manufacturing, R, noT 2P corresponds to
the results given an interest rate shock and no time to produce in manufacturing, and T.F.P., T 2P corresponds to the results
given a total factor productivity shock and time to produce in manufacturing.

R,
R, No
T.F.P.,
+ Time to Build in Const.
R,
R, No
T.F.P.,
+Durable Consumption
R,
R, No
T.F.P.,
+ Capacity Utilization*
R,
R, No
T.F.P.,
Adjustment Cost to Labor
R,
R, No
T.F.P.,
ζ=8
R,
R, No
T.F.P.,
ζ = 40
R,
R, No
T.F.P.,
ψ=5
R,
R, No
T.F.P.,
ψ = 1/3
R,
R, No
T.F.P.,

Data
Model 1

G.D.P.
G.D.P.
(Impact) (years 1-6)
-10.4%
-1.2%
-2.9%
-0.3%
-2.6%
-5.7%
-7.6%
-1.1%
-3.0%
-0.3%
-2.7%
-5.7%
-7.7%
-0.8%
-3.0%
-0.1%
-2.7%
-5.5%
-7.5%
-1.9%
-2.3%
0.4%
-1.7%
-8.4%
-10.0%
-1.2%
-2.2%
-0.4%
-1.7%
-10.8%
-10.2%
-1.6%
-4.2%
0.5%
-3.6%
-8.3%
-11.2%
-2.0%
-1.8%
0.3%
-1.3%
-8.4%
-9.7%
-3.5%
-4.0%
0.3%
-3.0%
-5.6%
-7.5%
-0.7%
-1.2%
0.2%
-1.0%
-10.8%
-11.9%

adjust their inventories to their new target level over the course of a year. However, production in
inventory intensive sectors is relatively depressed several years after the crisis, by which time all these
adjustment dynamics should have exhausted themselves.
A second explanation is that more cyclical sectors are naturally more likely to hold large inventories
as insurance. Since inventory/cost ratios are measured using data from American firms, this is only
plausible to the extent that the episodes in question were large versions of common US business cycles
in terms of the cross-sectional impact. Even if this were the case, most of the uncertainty that US
firms face is idiosyncratic. Insurance against aggregate shock is thus unlikely to figure as a primary
reason to hold inventories.

Conclusion

The 2008-2009 recession had a salient financial aspect to it, leading researchers to the look for links
between the financial sector and the macro-economy. The working capital channel has featured prominently as one such link. Arguably, emerging market crises were important precursors, with the difference that the role of the financial sector in mobilizing savings from foreign households to domestic
investment and production played a central role. This paper shows that financial shocks have a potentially large short-run impact on output because of their effect on production costs. This helps explain
how interruptions in capital inflows to a given economy can generate a drop in output even if openness
to foreign trade means that production in this economy is not constrained by domestic demand.
In the specific context of emerging market crises, this paper is a contribution to the debate of
whether productivity or the cost of foreign borrowing are the main drivers of business cycles in
emerging economies. The results in the paper show that a productivity shock alone cannot account
for the persistent pattern of cross-sectoral dispersion observed in the data, but an interest rate shock
does. The quantitative model also shows that, given tightly calibrated time to produce, it is possible
for such an interest rate shock to generate a sizeable immediate drop in output. In fact, the results in
the paper may understate the role of time to produce given work by Pratap and Urrutia (2011) and
Meza and Benjamin (2009) that shows that the interaction of working capital constraints with other
frictions can lead to significant fluctuations in measured productivity and in output.
The insight that changes in the cost of capital can have a quick and sizeable impact on output
is more general and applies to other setups. For example this insight might be important for the
quantitative predictions of models with financial frictions. Kocherlakota (2000) argues that, since
capital is a small share of output and depreciates slowly, financial frictions may not be quantitatively
important, and Cordoba and Ripoll (2004) verify that this is indeed the case in a calibrated business
cycle model where capital depreciates slowly. Allowing for time to produce increases the capital share
and reduces average capital depreciation. It would be interesting to see how allowing for time to
produce would change the results of this kind of model.
On a methodological note, this paper is an example of how cross-sectional evidence can be used

to discipline macroeconomic models. If different sectors react differently to a macroeconomic event,
those differences ought to provide useful information about the type of shock that is affecting the
economy and the mechanism through which that shock affects the economy.

References
Aguiar, Mark and Guita Gopinath, “Emerging Market Business Cycles: The Cycle Is the Trend,”
Journal of Political Economy, 2007, 115 (1), 69–102.
Alessandria, George, Joseph Kaboski, and Virgiliu Midrigan, “Inventories, Lumpy Trade,
and Large Devaluations,” American Economic Review, 2010, 100, 2304–2339.
, Sangeeta Pratap, and Vivian Yue, “Export Dynamics in Large Devaluations,” Unpublished
mimeo, 2012.
Barth, Marvin J. and Valerie A. Ramey, “The Cost Channel of Monetary Transmission,” NBER
Macroeconomics Annual, 2001, 16, 199–240.
Bates, Thomas W., Kathleen M. Kahle, and Rene M. Stulz, “Why do U.S. firms hold so
much more cash than they used to?,” Journal of Finance, 2009, 64, 1985–2021.
Benjamin, David and Felipe Meza, “Total Factor Productivity and Labor Reallocation: The
Case of the Korean 1997 Crisis,” The B.E. Journal of Macroeconomics, 2009, 9 (31), 1166–1187.
Bils, Mark and James A. Kahn, “What Inventory Behavior Tells Us about Business Cycles,” The
American Economic Review, 2000, 90 (3), 458–481.
Blinder, A S and L Maccini, “Taking Stock: A Critical Assessment of Recent Research on Inventories,” Journal of Economic Perspectives, 1991, (5), 73–96.
Burnside, Craig and Martin Eichenbaum, “Factor-Hoarding and the Propagation of BusinessCycle Shocks,” The American Economic Review, 1996, 86 (5), 1154–1174.
Calvo, Guillermo, “Capital Flows and Capital-Market Crises: The Simple Economics of Sudden
Stops,” Journal of Applied Economics, 1998, 1 (1), 3554.
, Alejandro Izquierdo, and Ernesto Talvi, “Phoenix Miracles in Emerging Markets: Recovering without Credit from Systemic Financial Crises,” National Bureau of Economic Research
Working Paper Series, 2006, (12101).
,

, and Luis-Fernando Mejia, “Systemic Sudden Stops: The Relevance of Balance-Sheet

Effects and Financial Integration,” National Bureau of Economic Research Working Paper Series,
2008, (14026).
Cameron, A. Colin, Jonah B. Gelbach, and Douglas L. Miller, “Robust Inference with MultiWay Clustering,” National Bureau of Economic Research Working Paper Series, 2006, (T0327).

Cerra, Valerie and Sweta C. Saxena, “Growth Dynamics: The Myth of Economic Recovery,”
American Economic Review, 2008, 98 (1), 439–57.
Chang, Roberto and Andres Fernandez, “On the Sources of Fluctuations in Emerging
Economies,” Unpublished Manuscript, 2012.
Chari, V. V., Patrick Kehoe, and Ellen McGrattan, “Sudden Stops and Output Drops,”
American Economic Review, 2005, 95 (2), 381–387.
Christiano, Lawrence J., Martin Eichenbaum, and Charles L. Evans, “Nominal Rigidities
and the Dynamic Effects of a Shock to Monetary Policy,” American Economic Review, 1992,
82 (2), 346–353. Papers and Proceedings of the Hundred and Fourth Annual Meeting of the
American Economic Association.
, Roberto Motto, and Massimo Rostagno, “Financial Factors in Business Cycles,” Mimeo,
2010.
Cordoba, Juan-Carlos and Marla Ripoll, “Credit Cycles Redux,” International Economic Review, 2004, 45, 1011–1046.
Correia, Isabel, Joao Neves, and Sergio Rebelo, “Business Cycles in a Small Open Economy,”
European Economic Review, 1991, 81 (4), 797–818.
Dell’Ariccia, Giovanni, Enrica Detragiache, and Raghuram Rajan, “The real effect of banking crises,” Journal of Financial Intermediation, 2007, 17 (1), 89–112.
Dupor, Bill, “Aggregation and irrelevance in multi-sector models,” Journal of Monetary Economics,
1999, 43 (2), 391–409.
Fazzari, Steven, Glenn Hubbard, and Bruce Petersen, “Financing Constraints and Corporate
Investment,” Brookings Papers on Economic Activity, 1988, 1988 (1), 141–206.
Foerster, Andrew, Mark W. Watson, and Pierre-Daniel G. Sarte, “Sectoral vs. Aggregate
Shocks: A Structural Factor Analysis of Industrial Production,” Journal of Political Economy,
2011, 119 (1), 1–38.
Foley, C. Fritz, Jay C. Hartzell, Sheridan Titman, and Garry Twite, “Why do firms hold so
much cash? A tax-based explanation,” Journal of Financial Economics, 2007, 86 (3), 579–607.
Gaiotti, Eugenio and Alessandro Secchi, “Is There a Cost Channel of Monetary Policy Transmission? An Investigation into the Pricing Behavior of 2,000 Firms,” Journal of Money, Credit
and Banking, 2006, 38 (8), 2013–2038.
Garcia-Cicco, Javier, Roberto Pancrazi, and Martin Uribe, “Real Business Cycles in Emerging Countries?,” American Economic Review, 2010, 100, 2510–2531.
Gertler, Mark and Simon Gilchrist, “Monetary Policy, Busines Cycles, and the Behavior of Small
Manufacturing Firms,” Quarterly Journal of Economics, 1994, 109 (2), 309–340.

, and Fabio Massimo Natalucci, “External Constraints on Monetary Policy and the

Financial Accelerator,” Journal of Money, Credit and Banking, 2007, 39 (2), 295–330.
Giannetti, Mariassunta, Mike Burkart, and Tore Ellingsen, “What you sell is what you lend?
Explaining trade credit contracts,” Review of Financial Studies, 24.
Gollin, Douglas, “Getting Income Shares Right,” Journal of Political Economy, 2002, 110 (2).
Greenwood, Jeremy, Zvi Hercowitz, and Gregory Huffman, “Investment, Capacity Utilization
and the Real Business Cycle,” American Economic Review, 1988, 78 (3), 402–417.
Horvath, Michael, “Cyclicality and sectoral linkages: Aggregate fluctuations from independent
sectoral shocks,” Review of Economic Dynamics, 1998, 1 (4), 781–808.
, “Sectoral shocks and aggregate fluctuations,” Journal of Monetary Economics, 2000, 45 (1),
69–106.
Humphreys, Brad, Louis Maccini, and Scott Schuh, “Input and output inventories,” Journal
of Monetary Economics, 2001, 47 (2), 347–375.
Iacoviello, Matteo, Fabio Schiantarelli, and Scott Schuh, “Input and output inventories in
general equilibrium,” International Economic Review, 2011, 52 (4), 1179–1213.
International Monetary Fund, “Crisis and Recovery,” World Economic Outlook, 2009, (3).
Jermann, Urban and Vicenzo Quadrini, American Economic Review, 2012, 102 (1), 238–271.
Jones, Christopher S. and Selale Tuzel, “Inventory Investment and the Cost of Capital,” Mimeo,
2010.
Kahn, Aubhik and Julia Thomas, “Inventories and the Business Cycle: An Equilibrium Analysis
of (S,s) Policies,” The American Economic Review, 2007, 97 (4), 1165–1188.
Kaplan, Steven and Luigi Zingales, “Do Investment-Cash Flow Sensitivities Provide Useful Measures of Financing Constraints?,” Quarterly Journal of Economics, 1997, 112 (1), 169–215.
Kehoe, Timothy and Kim Ruhl, “Sudden Stops, Sectoral Reallocations and the Real Exchange
Rate,” National Bureau of Economic Research Working Paper Series, 2008, w14395.
Kim, Se-Jik and Hyun Song Shin, “Industrial Structure and Corporate Finance,” Unpublished
manuscript, 2007.
King, Robert G. and Sergio T. Rebelo, “Resuscitating real business cycles,” in John Taylor and
Michael Woodford, eds., Handbook of Macroeconomics, Vol. 1, Elsevier, 1999, chapter 14.
Kiyotaki, Nobuhiro and John Moore, “Liquidity, business cycles and monetary policy,” Unpublished manuscript, 2008.
Kocherlakota, Narayana, “Creating Businss Cycles Through Credit Constraints,” Federal Reserve
Bank of Minneapolis Quarterly Review, 2000, 24 (3).

Krishnamurthy, Arvind, “Collateral Constraints and the Amplification Mechanism,” Journal of
Economic Theory, 2003, 111 (2), 277–292.
Kroszner, Randall, Luc Laeven, and Daniela Klingebiel, “Financial Crises, Financial Dependence, and Industry Growth,” Center for Economic Policy Research, 2006, 5623.
Kydland, Finn E. and Edward C. Prescott, “Time to Build and Aggregate Fluctuations,”
Econometrica, 1982, 50 (6), 1345–1370.
Long, John B. and Charles I. Plosser, “Real Business Cycles,” The Journal of Political Economy,
1982, 91 (1), 39–69.
Maccini, Louis, Bartholomew Moore, and Huntley Schaller, “The Interest Rate, Learning,
and Inventory Investment,” American Economic Review, 2004, 94 (5), 1303–1327.
Melitz, Marc J., “The Impact of Trade on Intra-Industry Reallocations and Aggregate Industry
Productivity,” Econometrica, 2003, 71 (6), 1695–1725.
Mendoza, Enrique G., “Real Business Cycles in a Small Open Economy,” The American Economic
Review, 1991, 81 (4), 797–818.
, “Sudden Stops, Financial Crises and Leverage,” American Economic Review, 100 (5), 1941–
1966.
Meza, Felipe and David Benjamin, “Total Factor Productivity and Labor Reallocation: The
Case of the Korean 1997 Crisis,” The B.E. Journal of Macroeconomics, 2009, 9 (1).
Neumeyer, Pablo and Fabrizio Perri, “Business cycles in emerging economies: the role of interest
rates,” Journal of Monetary Economics, 2005, 52 (2), 345–380.
Otsu, Keisuke, “A neoclassical analysis of the Korean crisis,” Review of Economic Dynamics, 2007,
11 (2), 449–471.
Petersen, Mitchell and Raghuram Rajan, “Trade Credit: Theories and Evidence,” Review of
Economic Studies, 1997, 10 (3), 661–691.
Pratap, Sangeeta and Carlos Urrutia, “Financial Frictions and Total Factor Productivity: Accounting for the Real Effects of Financial Crises,” Review of Economic Dynamics, 2011, 15 (3),
336–358.
Quadrini, Vicenzo and Fabrizio Perri, “International Recessions,” National Bureau of Economic
Research, 2011, (17201).
Raddatz, Claudio, “Liquidity needs and vulnerability to financial underdevelopment,” Journal of
Financial Economics, 2006, 80, 677–722.
Rajan, Raghuram and Luigi Zingales, “Financial Dependence and Growth,” American Economic
Review, 1998, 88 (3), 559–586.
Ramey, Valerie, “Inventories as Factors of Production and Economic Fluctuations,” The American
Economic Review, 1989, 79 (3), 338–354.

and Kenneth West, “Inventories,” in John Taylor and Michael Woodford, eds., Handbook of
Macroeconomics, Vol. 1, Elsevier, 1999, chapter 13.
Reinhart, Carmen and Kenneth Rogoff, This time is different: eight centuries of financial folly
2009.
Tong, Hui and Shang-Jin Wei, “The Composition Matters: Capital Inflows and Liquidity Crunch
During a Global Economic Crisis,” National Bureau of Economic Research, 2009, (15207).
Uribe, Martin and Vivian Yue, “Country spreads and emerging countries: Who drives whom?,”
Journal of International Economics, 2006, W12564.
Wen, Yi, “Input and output inventory dynamics,” American Economic Journal - Macroeconomics,
3.
Young, Alwyn, “The Tyranny of Numbers: Confronting the Statistical Realities of the East Asian
Growth Experience,” The Quarterly Journal of Economics, 1995, 119 (3), 641–680.

Appendix A: Details of Data Construction
Events
The empirical analysis centers on events where capital inflows to an emerging market drop by an
unusually large amount and there is a drop in G.D.P. either in the same year or in the following year.
The episodes satisfy the following criteria:
1) A large drop in capital inflows. Capital inflows are proxied on a monthly basis by the sum
between the trade deficit and declines in international reserves, normalized by the linear trend of
G.D.P.18 The episodes of interest feature a pronounced change in capital flows over one year. 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.

19 , 20

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 G.D.P. 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 criteria are similar to Calvo et al. (2008) and Calvo et al. (2006), except that there is no
attempt to identify exogenous or unpredictable events that may serve as natural experiments. Rather,
the methodological approach is to identify episodes which are clear cases of counter-cyclical movements
in the current account, document regularities and then compare them to a model. Because there is no
attempt to claim exogeneity or unpredictability, there is no filter for international financial conditions.
Also, for the same reason, 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. The episodes are listed in table 1, and they
include most episodes identified by Calvo et al. (2006).

Empirical Specification
I test whether time to produce is associated with lower performance at different horizons after the
crisis. For horizons spanning from the year of the crisis to 8 years afterwards, I run the following
regression:
yi,k,t∗ +h − yi,k,t∗ −1 = βτk + γXi,k,t∗ + αi + εik
18

The proxy follows from the identity capital inflows + current account = 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.
(2008).
19
Pre 1975 data were excluded as capital flows were much less volatile before the end of the Bretton Woods regime.
20
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.

where yi,k,t∗ +h is log value added in industry k, episode i, h years after the start of the event, τk is
inventory/cost ratio for industry k and Xi,k,t∗ is a vector of controls. I also allow for episode fixed
effects αi . The hypothesis is that β < 0, so that firms with longer production time do relatively worse.
Value added data are available in INDSTAT3, a data set compiled by UNIDO. The data originate
from official sources in UN member countries. The industries are classified according to three digit
ISIC Rev. 2. I do not include non-crisis data in the regressions, because the data-set includes a
reasonable amount of methodological breaks in the time series that are not properly documented by
UNIDO. This problem is especially severe among developing countries. By restricting the number of
years of data used by each country to those around the crisis, I restrict the possibility of having these
breaks contaminate the results. The use of the three digit level ISIC Rev. 2 classification allows me
to have data for the early ’80s.
Errors may be correlated both within episodes or within industries. To account for these correlations, the standard errors are robust to overlapping clusters at both the industry and episode level
(Cameron et al. 2006).21 For example, they allow the error term for shoes in Thailand in 1997 to
be correlated both with machinery in Thailand in 1997 and shoes in Mexico in 1994. Because the
standard error calculation allows for a widespread correlation between error terms, they are more
conservative than usual “one-way” clustering procedures.

Inventory/Cost Ratio
For American firms, I aggregate firm level data from COMPUSTAT into multiple sectors. COMPUSTAT includes decades’ worth of financial reports from listed firms that operate in the US. The
long time series allows one to smooth business cycle fluctuations when calculating the inventory/cost
ratios, obtaining something close to a steady state measure. Importantly, because COMPUSTAT is
based on balance sheet data, the accounting identity between inventories and costs is exact.
In the benchmark regressions, the inventory/cost ratio is calculated using firm level balance sheet
data from COMPUSTAT. The numerator is total inventories and the denominator is the cost of
goods sold. An accounting identity ensures that the two encompass exactly the same goods, making
it a precise measure. To calculate the inventory/cost ratio for a given industry, I take the median
of the ratio across firms and years. Because COMPUSTAT includes observations spanning over 20
years of data, business cycle fluctuations are smoothed out. 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.
21

I implement these in STATA using the cgmreg.ado file produced by the authors and ivreg2 for the I.V. estimates.

Controls
Sectoral reallocation can occur because of a wide range of alternative forces. These are important to
the extent that they are correlated both with the inventory ratios and the change in value added in
the event period. This section describes the controls. Descriptive statistics are summarized in table 6
and correlations are shown in table 7.

variable
mean
Inventory/Cost (U.S. firms)
2.457
Inventory/Sales (Korean firms) 1.605
Investment Share
0.066
0.014
Durable Consumption Share
Exported Share
0.132
Labor Share
0.151
Capital Share
0.177
Imported Inputs Share
0.126
0.244
External Dependence
Size
1.365

median min
2.451
0.509
1.606
0.614
0.058
0.001
0.007
0.000
0.120
0.030
0.141
0.022
0.175
0.046
0.116
0.018
0.220 -0.450
1.438 -2.057

max
sd
3.940 0.908
2.609 0.442
0.523 0.056
0.503 0.036
0.756 0.070
0.254 0.027
0.379 0.046
0.775 0.053
1.140 0.330
2.144 0.371

Table 6: Descriptive Statistics
Inventory/Cost (U.S. firms) is the median ratio of inventories-to-cost across firms and years in the COMPUSTAT database using data since 1980. Inventory/Sales (Korean firms) is the average ratio of inventoriesto-sales in the Financial Statement Analysis collected by the Korean government. Export, Investment and
Durable Consumption represent the fraction of sectoral output that eventually finds its way to either of
these final uses. Durable Consumption includes all consumption from industries whose ISIC numbers start
with 33, 36 and 38 (Wood Products, Non-Metallic Minerals and Machinery and Equipment). External Dependence is the dependence on external financing as defined by Rajan and Zingales (1998). The numbers
are taken from their paper. Size is the average over time of the ratio between employees and establishments
for each country/industry observation calculated with data from the UNIDO database.

Demand Side
Demand for different goods reacts differently in a downturn. In particular, the more durable a good,
the more pro-cyclical its demand. On the other hand, the more tradable a good, the less it should be
affected by conditions that affect domestic demand.
A measure of the share of output that is ultimately destined to the production of exports, investment goods and durable consumer goods can be calculated from input-output matrices made available
by OECD. The measure uses the Leontief inverse to allow for indirect effects through the supply chain.
For example, this clarifies that basic metals are sold largely as investment goods.

Inventory/Cost, U.S. firms
Inventory/Sales, Korean firms
Investment Share
Durable Consumption Share
Exported Share
Labor Share
Capital Share
Imported Inputs Share
External Dependence
Size

Inventory/Cost, U.S. firms
1
0.6135
0.3584
0.3977
0.3839
0.2817
-0.015
0.0677
0.1404
-0.1071

Inventory/Sales, Korean firms
1
0.2556
0.4895
0.5702
0.3606
-0.0612
0.1912
0.3606
-0.3339

Table 7: Industry Level Correlations
Correlations only refer to cross industry variation. Variables that vary across countries are first averaged
at the industry level. See notes in table 6 for details on how the variables are constructed.
Cost
During a sudden stop, there are massive realignments in relative prices of inputs. The real foreign
exchange rate depreciates, disproportionately affecting industries that use imported goods heavily.
Also the real wage rate declines, significantly affecting firms that use labor heavily. To capture these
effects I include the share of labor and imported inputs used in production as captured by the inputoutput matrices. For completeness, I also include a control for the share of fixed capital, defined as
the difference between value added, the share of materials and the share of labor. Since the interest
rate affects the user cost of fixed capital, firms that rely disproportionately on fixed capital are likely
to be more heavily affected.

External Dependence
Rajan and Zingales (1998) define External Dependence as
External Dependence =

Capital Expenditure - Cash Flow
Capital Expenditure

They calculate averages of this ratio for listed firms in the US. Rajan and Zingales argue that
the ratio captures technological aspects of production, in particular, the extent to which investment
is front-loaded when compared to production. Everything else constant, if access to finance is more
expensive, industries that require more up-front investment should do relatively worse. Dell’Ariccia
et al. (2007) and Kroszner et al. (2006) have applied this insight to banking crises, and they find
an important role for external dependence. For comparability, I use the same numbers for external
dependence as Rajan and Zingales (1998).

Size
The empirical corporate finance literature has often focused on size as a determinant of differences in
the supply of finance available to different firms.22 Firm size can also matter because large firms are
more likely to be exporters.23 The UNIDO data include data on employment and number of establishments by industries. I calculate for each country/industry pair the average employment/establishment
ratio.24

Trend Behavior
Different industries may have different long-run growth trends. I control for these in two ways. First,
for each country/industry observation I add the average growth in the six years before the crisis as a
control. This should pick up both inertia and mean reversal tendencies.
Second, I include the average growth in the industry among non-crisis countries over the same
period. These are countries that have not experienced a crisis over the previous 10 years.

Normally Expected Cyclical Behavior
I construct this control as follows: First, for each country/industry observation I run the following
regression:
yi,k,t+h − yi,k,t−1 = λi,k,h + ηi,k,h (gdpi,t+h − gdpi,t ) +
where again yi,k,t is log value added in industry k, country i, time t and gdpi,t is log G.D.P. in country
i time t. If there is a crisis in the country, I include in the regression for all years prior to the crisis and
all years more than ten years after. For countries that do not face a crisis, I include all observations
available. Since there is typically not a large number of observations for each industry, the estimates
are quite noisy. I then calculate industry specific components:

P
λ̄k,h =
η̄k,h

i λi,k,h

P I
ηi,k,h
= i
.
I

Lastly, I calculate the expected industry behavior within a given window following the crisis given
the drop in output:
22

See Fazzari et al. (1988) for the seminal paper and Gertler and Gilchrist (1994) for an application specific to interest rate
shocks. However, see also Kaplan and Zingales (1997) for a critique.
23
See Melitz (2003).
24
I also experiment with the employment/establishment ratio in the year before the crisis. The results are virtually
identical.

Normally expected cyclical behavior = λ̄i,h + η̄i,h (gdpi,t∗ +h − gdpi,t∗ −1 )

Controls for Demand
To measure the demand for durable consumer goods consumption, I assume that all the consumption
from certain sectors is durable.25
I allow for indirect demand effects through the supply chain. As a consequence, the control variable
takes into account that the demand for basic metals is heavily affected by the demand for investment
goods. Let ai,j be the i, j th entry in the matrix of technical coefficients A. If sector j produces Yj
then it uses Yi,j = ai,j Yj inputs from sector i. Let Vi denote the demand for final use of the products
of sector i. Then it must be the case that
Yi =

ai,j Yj + Vi .

In terms of percentage changes,
Yj Yj
Vi dVi
dYi X
=
ai,j d +
.
Yi
Yi Yj
Yi V i
j

Let B be a matrix with entries bi,j = ai,j Yj /Yi , D a vector with entries Zi /Yi . Also (with abuse
of notation), let dY /Y be a vector with entries dYi /Yi . Likewise, let dVi /Vi = dV /V . Then
dY
dY
dV
=B
+
Y
Y
V
so that,
dY
dV
= (I − B)−1 D
Y
V
where I is a conformable identity matrix and (I-B) is invertible.
Thus, the elasticity of output in each sector to a given change in final demand is given by the
corresponding entry in the (I − B)−1 D vector.
It is natural to extend the framework to look at the effect of changes in different components of
demand. All that is needed is to substitute D for a matrix with entries Ii,H /Yi , Ii /Yi and EXPi /Yi .
In the calibration, I use country specific input-output matrices whenever possible; otherwise I use
an average for the region.26

I define as durable consumption all the consumption of wood products, furniture, non-metallic mineral, machinery and
equipment and half of Manufacturing n.e.c. I likewise use the Leontief inverse to account for indirect effects.
26
Thus, for example, the input-output structure for Malaysia is the average between South Korea and Indonesia.

Appendix B: Setup for Quantitative Model
Model Setup
The model economy has multiple sectors. It has NT tradable sectors and NN T non-tradable sectors.
The tradables correspond to the manufacturing sectors in the data, and the non-tradables are included
so as to fully account for all general equilibrium effects.

Household
There is a representative household, who is able to borrow and lend abroad at an exogenous risk-less
gross interest rate Rt . The household supplies labor and consumes both durable and non-durable
goods. The utility function of this household is time-separable with discount rate β < 1 and the
period utility is
1
u (Ct , KH,t , Lt ) =
1−σ

d ρ1

ρ−1
ρ

+ 1 − γd

ρ−1
ρ

ρ
ρ−1

KH,t

−χ∗

1+ 1
Lt ψ

!1−σ

where Ct is non-durable consumption, KH,t is the stock of durable goods held by the household and
Lt is labor supply.

Production Functions
The production function of each sector is

Yi,t =

τi
Y

Zi,t−v (v)ωi (v) ,

ωi (v) = 1

v=0

where Yi,t is the production of the input i in t. Zi,t−v (v) denotes a composite of goods acquired in
period t − v to produce the finished good which is sold v periods ahead. It is defined as:
Zi,t−v (v) = γi At−v (ui,t−v (v) Ki,t−v (v))αK,i (Li,t−v (v))αL,i (Mi,t−v (v))1−αL,i −αK,i
where ui,t−v (v) is the rate of utilization of fixed capital at t, Ki,t−v (v) is capital stock, Li,t−v (v) is
labor and Mi,t−v (v) is a composite of materials produced in the different sectors. The index v implies
that the input is meant for production of the output that will be available v periods in the future.
Once inputs are combined into Zi,t−v (v), they cannot be reassigned to a different sector or horizon.
This will be important for the short-run dynamics of the model, as it will be a source of inertia both
in aggregate output and in cross-sectoral reallocation.
As an accounting matter, the composite Zi,t−v (v) is part of inventories. In particular, in the
beginning of each period, the inventories held by sector i are the sum of all Zi,t−v (v) with v > 0,

properly inflated by the accumulated interest rates over the period in question.
Lastly, as a national accounting matter, the production function in this section corresponds to
total output, not to the value added by the sector. To get to value added, we need to subtract from
output the cost of all materials used in production and the change in inventories.

Capital Accumulation, Utilization and Maintenance Costs
Capital is sector specific and is produced by combining old capital and investment goods.
Ki,t = (1 − δ) Ki,t−1 + Ii,t −

ζ (Ki,t − (1 + g)Ki,t−1 )2
.
2
Ki,t−1

Durable consumer goods are produced in a similar fashion:
KH,t+1 = (1 − δ) KH,t + Ii,t −

ζ (KH,t − (1 + g)KH,t )2
.
2
KH,t

Whenever a firm uses capital, it has to use up spare parts in order to make up for wear and tear
of the capital stock. This maintenance requirement is increasing in capacity utilization:27
Ωi,t =

i
µi h 1+ξ
ui,t − 1 Ki,t .
1+ξ

Composite Goods
Non-durable consumption Ct , fixed investment Ii,t , investment in durable consumption goods IH,t ,
materials Mi,t and spare parts Ωi,t are composites of the goods produced in the NT + NN T sectors
and imported goods. These composites are represented by CES aggregates:
ρ
 ρ−1
ρ−1
1
ρ (X
ρ
)
(γ
)
k,X,t
k,X
k∈{1,...,NT +NN T }

1
Xt = 
ρ−1
P
ρ
+ 1 − k∈{1,...,NT +NN T } γk,X (X∗,X,t ) ρ



where X stands in for C, Ii , Ωi or Mi . X∗,t is the amount of imported goods used for the production
P
of X and
γi,X ≤ 1.

Resource Constraints
While capital is sector specific, there is no limit in reshuffling it for the production at different horizons.
This implies the following resource constraint for capital stock for sector i:
X

ui,t Ki (v) ≤ ui Ki,t .

v
27

In the calibration, I choose units so that in steady state u = 1 and the firm does not have any maintenance costs. This
implies that spare parts needs can be negative if u < 1. This assumption is made for tractability, but does not have any
important consequences for the results.

Raw materials respect an analogous resource constraint:
X

Mi,t ≤ Mi,t .

In contrast, labor is perfectly mobile between sectors. This implies the following constraint:
n
X

Li,t (v) ≤ Lt .

k∈{1,...,NT +NN T } v

Finally, there is a resource constraint in the non-tradable sectors. This is:
X

Ci,t + Ii,H,t + Gi +

(Mi,j,t + Ii,j,t + Ωi,j ) ≤ Yi,t

j∈{1,...,NT +NN T }

where Gi are government purchases, assumed to be constant over time.

Foreign Trade and the Current Account
Exports are identical to the total output from sector i minus the sum of the domestic demands for
this good:
X

EXPi,t = Yi,t − Ci,t − Ii,H,t −

(Mi,j,t + Ii,j,t + Ωi,j,t ) .

j∈{1,...,NT +NN T }

For non-tradables, exports are equal to zero:
EXPi,t = 0 if i is non-tradable.
Total imports are the sum of all foreign produced goods demanded for different uses:
IM Pt = C∗,t + I∗,H,t +

(M∗,j,t + I∗,j,t + Ω∗,j,t ) .

j∈{1,...,NT +NN T }

The prices of the foreign goods are constant over time and are normalized to be equal to 1. The
current account identity is

Bt − Rt−1 Bt−1 = IM Pt
X
−

EXPi,t + κBt2

i∈{1,...,NT }

where Bt is the amount of net foreign debt that domestic households hold.

Extension Adjustment Cost to Labor
In the text I consider an extension of the model where there is an adjustment cost to labor. To model
this adjustment cost, I assume that in order to increase employment in sector i from Li,t−1 to Li,t it
is necessary to expend f racζL 2 (Li,t − Li,t−1 )2 units of the imported good. I calibrate ζL = ζ.

Allocation
The allocation is the solution to a planner’s problem where the planner maximizes the utility function
of the household subject to all the resource constraints.

Appendix C: Details of Calibration
To calibrate the shares I use information from the same input-output matrices used to construct
indicators of the demand shares. I assign to each of the manufacturing sectors in the input-output
table an ISIC Rev. 2 code in order to match the data in the empirical section. The correspondence is
not perfect. Some sectors that are differentiated in the empirical section, such as textiles and apparel,
only appear as a single sector in the input-output matrix. On the other hand, the input-output
matrix includes more machinery producing sectors than in the UNIDO data set. In this latter case, I
consolidate the sectors in order to build equivalents to the ones in section 3.
The service sector is really a residual sector that includes services, retail, utilities, public administration and agriculture. I treat mining as a foreign sector since it is highly tradable and employs
relatively few workers. I also experimented with treating agriculture the same way as mining. The
results are not sensitive to this assumption.
The sectors that I treat as non-tradable do, in fact, trade. I treat all their sales abroad as domestic
sales and all the purchases from their counterpart abroad as domestic purchases. I am still left with
a (small) imbalance. Also, at the aggregate level there is a trade imbalance that does not come up in
the steady state of the model. I remove these imbalances by re-scaling the size of the non-tradable
sector and of domestic absorption.
I re-scale the labor shares up by a factor of 2, conforming to the findings in Young (1995) and
Gollin (2002) that labor shares in developing countries are frequently underestimated because of the
large number of self-employed workers. This brings the aggregate labor share close to 0.6, which is the
norm for developed countries. Also, I do not allow explicitly for indirect taxes in the model, so that I
split them between labor and capital income.28 Given the input-output table, it is a straightforward
matter to calculate the factor shares for the different sectors as well as the weight of each sector in
each of the different composite goods. The values in the input-output matrices do not account for the
−1
ωi
i
. An
opportunity cost of capital. Hence, for a manufacturing sector i, αi,M = M
Yi × R + 1 − ω
analogous calculation applies to labor share and to the construction sector.

28
For an interesting account of how heterogeneity in indirect taxes provides a foundation for fluctuations in T.F.P., see
Benjamin and Meza (2009).

Full text of Working Papers (Federal Reserve Bank of Richmond) : Time to Produce and Emerging Market Crises, Working Paper 10-15R

FRASER