1 Michigan State
290% University

This is to certify that the
dissertation entitled

THREE EMPIRICAL ESSAYS IN FINANCIAL ECONOMICS
AND INTERNATIONAL FINANCE

presented by
MAREK KOLAR

has been accepted towards fulﬁllment
of the requirements for the

PhD degree in Economics

M W

Major Professor’s Signature
BILOJZoDZ

Date

MSU is an affinnative-action, equal-opportunity employer

 

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

5108 KrlProilecaPres/CIRCIDateDue,indd

THREE EMPIRICAL ESSAYS IN FINANCIAL ECONOMICS AND
INTERNATIONAL FINANCE

By

Marek Kolar

A DISSERTATION

Submitted to
Michigan State University
in partial fulﬁllment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY

Department of Economics

2008

ABSTRACT

THREE EMPIRICAL ESSAYS IN FINANCIAL ECONOMICS AND
INTERNATIONAL FINANCE

By
Marek Kolar

In the ﬁrst chapter, "Credit Reallocation", we inquire into the existence of a process of
inter-ﬁrm credit reallocation. Employing the methodology developed by Davis and
Haltiwanger (1992) for the measurement of job reallocation, we ﬁnd that at any phase of
the business cycle a signiﬁcant amount of credit ﬂows across US. ﬁrms. Credit
reallocation well exceeds net credit changes and is highly volatile. We also ﬁnd that, as a
result of a procyclical credit creation and a more mildly countercyclical credit
destruction, credit reallocation is moderately procyclical. Finally, most of the magnitude
and dynamics of credit reallocation reﬂect credit ﬂows across ﬁrms relatively
homogeneous for size, industry or location. The results imply that recent macroeconomic
models with search ﬁictions and lock-in effects in the credit market can usefully
complement established models of the ﬂight to quality and the ﬁnancial accelerator.

In the second chapter, "The Effect of Comovement Between Firms on their Liquidation
Values: Do Firms Attract More Trade Credit when their Sales Are Less Correlated?", I
analyze whether and in what ways does the degree of comovement of a ﬁrm with other
ﬁrms in its industry, measured by the correlation of their growth rates of sales, affect the
amount of trade credit extended to the ﬁrm. I ﬁnd evidence that more trade credit is

extended to ﬁrms that exhibit a lower degree of comovement with its industry peers. The

results suggest a signiﬁcant role of the liquidation value of a ﬁrm in determining its
ﬁnancial structure as proposed in Shleifer and Vishny (1992).

The third chapter is titled "How Should We Control For the Clustering in Central Bank
Intervention Data?" Central bank intervention data is clustered, with successive days of
intervention followed by successive days without intervention. We test motivations for
intervention by estimating both the autoregressive conditional hazard model and the
autoregressive conditional binomial model for interventions by the Federal Reserve,
Bundesbank, and Bank of Japan. Utilizing a variety of measures, we ﬁnd that the former
is outperformed by the latter. We ﬁnd evidence that both the Federal Reserve and
Bundesbank preferred to intervene when the market was calmer. The Bundesbank
intervened in response to exchange rates being out-of-line with long-run fundamentals,
while the spread between the 6-month Treasury Bill rate contains predictive power for
Federal Reserve interventions. We ﬁnd evidence that Japan intervened in response to
changes in the nominal exchange rate, and intervention differed before and after Eisuke
Sakakibara became Director General of the International Finance Bureau of the Ministry

of Finance in Japan.

ACKNOWLEDGMENTS

First, I would like to apologize to all the people that made this moment possible for
me who will not ﬁnd themselves mentioned here. There Were so many of you that if I
tried to list everyone’s name I would surely omit someone. Please know that I am deeply
thankful to all of you that helped me on this long journey; my professors and friends at
schools in both the Czech Republic and the United States; my tennis friends both in the
Czech Republic and Germany; and many others.

I would especially like to thank professors Ana Maria Herrera and Raoul Minetti for
serving as my main advisors. I thank professor Herrera for her valuable and speedy
feedback on my various drafts and her many helpful comments. I thank professor Minetti
for never turning me away the many times I stopped by his office to discuss my work. I
am especially grateﬁil to both of them for being most patient with me when needed.

I am also thankful to professors Richard T. Baillie and Charles Hadlock for their
valuable comments.

From my fellow classmates at MSU, I have especially enjoyed the friendship of
Nathan Cook and Chris Douglas. In addition to being a good friend, Chris also proved to
be of tremendous help as a co-author of the third chapter of this dissertation.

I could not possibly thank my family enough for their support and for putting up with
me at times when the dissertation work was affecting my personal life. I am indebted to
my wife Erin for the highest degree of understanding. It was her who gave me strength
and hope when I felt like I would never be able to ﬁnish. My parents, Zdenka Kolarova

and Vaclav Kolar, also played a big part in keeping me motivated. They have always

iv

encouraged me to pursue more education; I may not even have gone to college if it was
not for my parents. Lastly, I would like to thank my brother Vaclav Kolar for always
being a great brother to me. With him lending support to my parents in the Czech

Republic, I was free to focus on completing this dissertation.

TABLE OF CONTENTS

LIST OF TABLES ................................................................................... viii

LIST OF FIGURES ........................................ x

INTRODUCTION ....................................................................................... 1

Notes ............................................................................................. 5
CHAPTER 1

CREDIT REALLOCATION ........................................................................... 6

Introduction .................................................................................... 6

Related Literature ............................................................................ ll

Methodology ................................................................................ 13

Overview of the Data .............................................................. 13

Measurement of Credit Flows .................................................... 14

Deﬁnitions .................................................................. 14

Measurement Issues ....................................................... 15

Aggregation ................................................................ 17

Summary Statistics and Cross-Sectional Properties ..................................... l9

Magnitude ............................................................................ 19

Cross-Sectional Properties ........................................................ 22

Theoretical Background .................................................................... 26

Preliminary Observations .......................................................... 26

Second Generation Models ....................................................... 27

First Generation Models ........................................................... 29

Time Series Properties ....................................................................... 30

Long-run Overview ................................................................ 30

Volatility ............................................................................. 31

Cyclical Properties ................................................................. 35

VAR Analysis ....................................................................... 37

Response to Output Shocks: Evidence from a Structural VAR. ...38

Short and Long Run: Evidence from a Reduced-form VAR ....... 41

Conclusion .................................................................................... 43

Notes ........................................................................................... 45
CHAPTER 2

THE EFFECT OF COMOVEMENT BETWEEN FIRMS ON THEIR
LIQUIDATION VALUES: DO FIRMS ATTRACT MORE TRADE

CREDIT WHEN THEIR SALES ARE LESS CORRELATED? ....................................... 61
Introduction .................................................................................. 61
Measurement ................................................................................. 67

Measuring Trade Credit ............................................................ 68
Comovement Measure .............................................................. 68

vi

Main Independent Variables ..................................................... 70

Other Controls ..................................................................... 73
Data ................................................................................. 74
The Empirical Model ...................................................................... 75
Tariff as Instrumental Variable .................................................. 76
Results ....................................................................................... 78
Summary Statistics ................................................................ 78
Results .............................................................................. 79
Pooled OLS and Fixed Effects .......................................... 79
Instrumented Pooled OLS and Fixed Effects ......................... 82
Conclusion .................................................................................. 82
Notes ......................................................................................... 84
CHAPTER 3
HOW SHOULD WE CONTROL FOR THE CLUSTERING IN CENTRAL
BANK INTERVENTION DATA? .................................................................................. 100
Introduction ................................................................................ 100
Intervention Data ......................................................................... 104
US. and Germany ............................................................... 104
Japan .............................................................................. 106
Estimating the Probability of Intervention ............................................. 108
ACH ............................................................................... 108
ACB ............................................................................... 1 10
Explanatory Variables ........................................................... 112
Results ...................................................................................... 1 15
US. and Germany ............................................................... 115
Japan .............................................................................. 119
Comparison of Econometric Methods .......................... ' ....................... 122
Goodness-of-Fit .................................................................. 122
Model Selection .................................................................. 124
Summary .......................................................................... 127
Conclusion ................................................................................. 128
Notes ....................................................................................... 130
APPENDICES
Appendix 1: Data .......................................................................... 150
Data Sources ..................................................................... 150
Variables .......................................................................... 151
Classiﬁcation ..................................................................... 151
Remarks ........................................................................... 152
Appendix 2: Results for All Lags ....................................................... 154
REFERENCES ..................................................................................... 162

vii

1.1

1.2

1.3

1.4

1.5

1.6

1.7

1.8

2.1

2.2

2.3

2.4

2.5

3.1

3.2

3.3

3.4

3.5

3.6

3.7

LIST OF TABLES

Aggregate Credit Flows ................................................................... 49
Comparison of Credit, Job, and Capital Flows ......................................... 50
Credit Flows in Manufacturing Industries .............................................. 51
Credit Flows in Sales Quartiles (Aggregate Data) ..................................... 52
Credit Flows in Census Regions (Aggregate Data) .................................... 53
Properties of Idiosyncratic Credit Flows ................................................ 54
Correlation with Unemployment (Aggregate Data) ................................... 55
Variance Decomposition .................................................................. 56
Description of the Variables .............................................................. 90
Summary Statistics ......................................................................... 92
Comovement by 2-digit SIC (Full Sample) ............................................. 94
Baseline Results ........................ 96
Instrumental Variables ..................................................................... 98
Summary Statistics for Federal Reserve and Bundesbank Intervention ........... 134
Summary Statistics for Bank of Japan Intervention ................................. 135

ACH(1,1) Estimation Results for Federal Reserve and Bundesbank

Intervention ................................................................................ 136
ACB Estimation Results for Federal Reserve Intervention ......................... 137
ACB Estimation Results for Bundesbank Intervention .............................. 138
ACH(1,1) Estimation Results for Bank of Japan Intervention ..................... 139
ACB Estimation Results for Bank of Japan (Full Sample) .......................... 140

viii

3.8 ACB Estimation Results for Bank of Japan Intervention (1 st Subsample) ....... 141
3.9 ACB Estimation Results for the Bank of Japan (2nd Subsample) .................. 142

4.1 Results for All Lags ...................................................................... 154

ix

1.1

1.2

1.3

3.1

3.2

3.3

3.4

3.5

LIST OF FIGURES

Credit Change, Credit Reallocation, and Unemployment ............................ 57
Response to a 1% Increase in GDP Growth ....... A ...................................... 58
Correlation with GDP ..................................................................... 60
Number of Interventions Per Week for the Federal Reserve ........................ 143
Number of Interventions Per Week for the Bank of Japan .......................... 144
ACH and Probit Estimated Probability of Intervention

for the Federal Reserve ................................................................... 145
ACB(0,1,2) Estimated Probability of Intervention for the Federal Reserve ...... 146
Fluctuation Test ........................................................................... 147

Introduction

This dissertation deals with the relationship between ﬁnancial markets and the real
economy. The ﬁrst chapter is broad in scope; it is devoted to developing basic un-
derstanding of the nature and behavior of credit reallocation. The second chapter
complements the ﬁrst one by focusing on one of the aspects of cross-sectional credit
reallocation, namely the comovement of ﬁrms within a single sector of the economy,
and its relationship to the amount of credit extended to ﬁrms within the sector. The
third chapter is a stand alone chapter that presents, estimates, and compares two
empirical models of central bank intervention in the foreign exchange market.

The goal of the ﬁrst chapter1 , " Credit Reallocation", is to describe properties of
credit reallocation among ﬁrms. The need for taking a closer look at the process of
credit reallocation comes from the importance of reallocation of resources in general.
Due to continuous development of new products and technologies on the one hand,
and frequent occurrences of various macroeconomic shocks on the other, economic
conditions are constantly changing. This presence of never ending change in the
circumstances faced by ﬁrms and whole sectors of the economy gives rise to reallo-
cation of resources across ﬁrms. Davis and Haltiwanger (1992) developed measures
of gross ﬂows and reallocation to be used in their study of the reallocation of labor.
Ramey and Shapiro (1998) used the same methodology to address the reallocation
of physical capital. Both studies found a signiﬁcant degree of reallocation of these
physical resources. Our aim in this ﬁrst chapter is to improve upon the common

understanding of the process of resource reallocation by studying the reallocation of

credit across ﬁrms. We ﬁnd that most of the credit reallocation occurs within groups
of homogeneous ﬁrms, regardless of whether grouped by size, location, or product
characteristics. In other words, even though we ﬁnd that the idiosyncratic nature of
credit reallocation is weaker than in the case of job reallocation as reported by Davis
and Haltiwanger (1992), we still ﬁnd that most of the credit reallocation is driven by
the idiosyncratic nature of ﬁrm credit ﬂows. Since ﬁnancial credit, when compared to
workers as a resource, exhibits virtually no industry speciﬁcity, one would expect that
ﬁnancial capital would migrate freely across industry groups. Why then do we not
observe more ﬁnancial credit being reallocated across groups of homogeneous ﬁrms in
response to changing economic conditions? The second chapter takes a step toward
answering this question.

The second chapter, " The Effect of Comovement Between Firms on their Liq-
uidation Values: Do Firms Attract More TYade Credit when their Sales Are Less
Correlated?" , much narrower in scope, complements the ﬁrst one by focusing on one
aspect of the intra-sectorial properties of credit reallocation. In this chapter, I identify
intra-sectorial comovement between ﬁrms, measured by correlation of sales, as one of
the potential determinants of the amount of credit received. The motivation comes
from Shleifer and Vishny’s (1992) theory of liquidation value. Shleifer and Vishny
(1992) suggest a link between the comovement of ﬁrms within a sector and the col-
lateral value of their assets. The objective of this chapter is to empirically test this
link, and look for evidence of a relationship between the intra-sectorial comovement
and the ability of ﬁrms to obtain credit. Shleifer and Vishny (1992) argue that ﬁrm’s

liquidation value is negatively correlated with the degree of comovement of ﬁrms in

its industry. In turn, there is ample evidence that liquidation value of the borrower
affects the willingness of creditors to provide ﬁnancing. The evidence on the direction
of this effect, however, is mixed. On the one hand, there is evidence, for instance in
Benmelech et al. (2005), that higher liquidation values lead to higher amounts of
credit extended, due to a higher collateral. On the other hand, agency problem can
lead to a higher risk associated with higher liquidation values, which, as suggested for
instance in Jensen and Meckling (1976) and evidenced in MacKay (2003), may lead to
less credit extended. This second chapter provides empirical support for the positive
relationship between liquidation values and the ability of ﬁrms to obtain ﬁnancing. In
other words, I ﬁnd that ﬁrms in industry sectors with a lower degree of comovement
enjoy access to more credit.

The third chapterz, "How Should We Control For the Clustering in Central Bank
Intervention Data? " , estimates and compares two models of central bank intervention
in the foreign exchange market. Traditional econometric techniques do not always
take into account the nature of central bank intervention data, which exhibits a large
measure of clustering, corresponding to periods of no intervention and periods of
frequent intervention. One of the two estimated models is the autoregressive con-
ditional hazard (ACH) model developed by Hamilton and Jorda (2002), while the
second model is the autoregressive conditional binomial (ACB) model developed by
Herrera (2004). These two models take advantage of the special nature of clustered
data in a time series setting. We ﬁnd that while the ACH speciﬁcation is only a
minor improvement on a simple probit model, the ACB speciﬁcation is substantial

improvement and turns out to be the preferred estimation technique for the kind of

clustered data used in the analysis.

Notes

1This ﬁrst chapter is co-authored with Ana Maria Herrera and Raoul Minetti.

2This third chapter is co—authored with Christopher Douglas.

Chapter 1: Credit Reallocation

1 Introduction

The Schumpeterian view of creative destruction has recently gained momentum (Caballero
and Hammour, 2005). According to this view, aggregate and allocative shocks continually
alter the distribution of production opportunities, generating an intense inter-ﬁrm realloca-
tion of resources. Indeed, there is evidence that an intense reallocation of physical inputs
occurs at any phase of the business cycle. This evidence is well established for workers
(see, e.g., Davis and Haltiwanger, 1992, and Davis, Haltiwanger and Schuh, 1996) and re-
cent studies (Ramey and Shapiro, 1998; Eisfeldt and Rampini, 2005) document a process
of reallocation of physical capital.

In contrast with the attention dedicated to the reallocation of physical inputs, the role of
ﬁnance for aggregate restructuring is mostly neglected. Yet, just like employment changes
and investment disguise large inter-ﬁrm ﬂows of workers and physical capital, the aggre-
gate net change of ﬁrms’ external ﬁnancing may mask large inter-ﬁrm ﬂows of ﬁnancial
resources. In particular, two questions arise naturally: are ﬁnancial resources subject to an
intense process of inter-ﬁrm reallocation analogous to that observed for physical inputs? If
so, what are the properties of this “ﬁnancial reallocation”? Macroeconomic models with
ﬁnancial imperfections imply that addressing these issues is essential for understanding
aggregate restructuring. Consider ﬁrst the models that emphasize asymmetric informa-
tion and agency problems in the “vertical”1 relationships between ﬁrms and creditors (e.g.,

Bemanke and Gertler, 1989; Bemanke, Gertler and Gilchrist, 2000; Holmstrom and Tirole,

1997). These models imply that creditors can monitor ﬁrms and thereby enhance their ef-
ﬁcient use of workers and capital. Therefore, they imply that credit reallocation is crucial
for the reallocation of workers and capital to the most productive ﬁrms. Consider next
the models that emphasize “horizontal” heterogeneity across ﬁrms and the search frictions
that the inter-ﬁrm reallocation of ﬁnancial resources can entail. These models imply that,
by hindering the inter-ﬁrm reallocation of ﬁnancial resources, search frictions in the credit
market can hinder the reallocation of workers and capital to the most productive ﬁrms.
Thus, in both perspectives investigating ﬁnancial reallocation yields crucial insights into
the (obstacles to the) reallocation of physical inputs.

The goal of this paper is to take a step towards understanding ﬁnancial reallocation.
We focus on credit (debt from ﬁrms’ perspective),2 which constitutes the main form of ex-
ternal ﬁnance of US. corporations. Rajan and Zingales (1995) calculates that in 1991 the
aggregate debt to equity ratio of US. corporations equalled 1.13.3 Furthermore, in many
years net equity issues are negative (Myers, 2001): for example, in'1999 U.S. non-farm,
non-ﬁnancial corporations increased their external ﬁnancing by 139 billion dollars; the net
additional borrowing was 283 billion dollars while net equity issues were negative. Debt
is not only the main form of external ﬁnance of US. corporations but, according to the
literature on credit imperfections, it also has distinct costs and beneﬁts that sharply differ-
entiate it from equity. For example, unlike equity, debt is a “hard claim” that can prevent
ﬁrms from wasting cash flow and provide creditors with strong incentives to monitor; at
the same time, being a hard claim, debt can also exacerbate ﬁrms’ incentive to add risk to
their projects (for more on these and other distinct properties of debt, see Myers, 2001, and

Section 5 in this paper).

To measure credit reallocation we adopt the statistical methodology developed by Davis
and Haltiwanger (1992) and Davis, Haltiwanger and Schuh (1996) for the measurement
of job flows. Employing ﬁrm balance sheet data from US. Compustat tapes, we compute
inter-ﬁrm flows of total credit and long-term and short-term credit separately. We do so both
for the aggregate economy (excluding ﬁrms classiﬁed in the “Finance, Insurance, and Real
Estate” group) and for the manufacturing sector alone. We ﬁrst investigate the magnitude
and cross-sectional properties of the ﬂows; then, we explore the time series properties of
the ﬂows, namely their volatility and cyclical pattern. We document the following stylized

facts:

Fact 1. At any phase of the business cycle inter-ﬁrm credit flows are important. These
ﬂows well exceed those needed to accommodate net credit changes and are of the same
order of magnitude as job and capital ﬂows. Thus, an intense process of credit reallocation

is continually at work;

Fact 2. Credit reallocation is somewhat larger in manufacturing than in non-manufacturing
industries and it varies substantially across manufacturing industries. Furthermore, reallo-
cation does not vary substantially across census regions while it is negatively correlated

with ﬁrm size;

Fact 3. Inter-ﬁrm credit flows exhibit a large volatility (coefﬁcient of variation), above
that of job ﬂows. Moreover, while for total credit the volatilities of credit creation and
credit destruction are roughly equal, for long-term credit the volatility of credit destruction

exceeds that of credit creation;

Fact 4. Credit reallocation follows a moderately procyclical pattern. This is espe-
cially evident for the eighties and for the nineties when credit creation was signiﬁcantly
procyclical while credit destruction was only moderately countercyclical. This ﬁnding is
corroborated by a VAR analysis, which reveals that negative shocks to output growth de-
press credit reallocation by exerting a downward pressure on credit creation and a milder

upward pressure on credit destruction;

Fact 5. Credit ﬂows within groups of ﬁrms roughly homogeneous for size, industry
or location are substantially larger than credit ﬂows across these groups. In addition, the
volatility and procyclical dynamics of credit reallocation mostly reﬂects the idiosyncratic
behavior of debt changes within these narrowly deﬁned groups rather than mean transla-

tions of the aggregate or sectorial distributions of the debt changes.

Drawing on these facts, we inquire into the determinants of credit reallocation and its
role in aggregate restructuring. The following questions guide us throughout the analysis:
do the stylized facts we document support macroeconomic models with credit imperfec-
tions? If so, do they allow to discriminate between the second generation of these models
(e.g., Caballero and Hammour, 2005; Den Haan, Ramey, and Watson, 2003; Wasmer and
Weil, 2004), which emphasize “horizontal” ﬁrm heterogeneity and lock-in eﬂ‘ects in credit
relationships due to speciﬁcity or search frictions, and the ﬁrst generation of these models.
(e.g., Bemanke and Gertler, 1989; Bemanke, Gertler and Gilchrist, 2000), which empha-
size asymmetric information and agency problems in the “vertical” relationships between
ﬁrms and creditors? Our conclusion is clear-cut: second generation models can explain at

least three of the facts we uncover and thus appear to usefully complement established ﬁrst

generation models. Models with search frictions and lock-in effects in the credit market
predict that forming credit relationships is more difﬁcult than breaking them and therefore
can account for the larger volatility of long-term credit destruction relative to long-term
creation (Fact 3). These models can also rationalize the procyclical behavior of credit real-
location (Fact 4). In Caballero and Hammour (2005), reallocation across production units
declines during recessions because, as a result of a hold-up problem within credit matches,
the creation of credit matches drops, while cumulatively their destruction rises only moder-
ately. Finally, these models can explain the larger importance of idiosyncratic credit ﬂows
relative to credit ﬂows across groups of ﬁrms with different characteristics (Fact 5). In
Wasmer and Wei] (2004) and Den Haan, Ramey, and Watson (2003), for example, credit
reallocation is governed by a matching function and credit ﬂows to ﬁrms randomly, i.e.
independently of ﬁrms’ intrinsic characteristics (such as size or productivity).

The remainder of this paper is structured as follows. In Section 2, we review the related
literature. In Section 3, we describe the methodology used to measure credit ﬂows. Section
4 presents summary statistics for the ﬂows and investigates their cross-sectional properties.
In Section 5, we examine the predictions that different classes of macroeconomic models
with credit imperfections yield about inter-ﬁrm credit reallocation and its role in aggregate
restructuring. In Section 6, we turn to the time series properties of the ﬂows. In this
section, we also study the dynamic behavior of credit reallocation using a VAR. Section 7

concludes. We relegate details on the data and their sources to Appendix 1.

10

2 Related Literature

There is limited knowledge of the reallocation of ﬁnancial resources. Using data from US.
banks’ Call Report Files, Dell’Ariccia and Garibaldi (2005) investigates the inter-bank
reallocation of loans and ﬁnds evidence of a signiﬁcant heterogeneity in bank-level loan
dynamics. This work provides valuable information on the dynamics of liquidity in the
ﬁnancial interrnediation sector but limited information on the inter-ﬁrm credit reallocation.
A bank can contract its loans to a ﬁrm and expand its loans to another: this will induce
credit reallocation across ﬁrms but not across banks. Analogously, a ﬁrm can contract its
borrowing from a bank and expand its borrowing from another: this will induce credit
reallocation across banks but not across ﬁrms. In fact, recent studies document that a large
number of ﬁrms borrow from multiple banks and compensate for the contraction of credit
by one bank by increasing their borrowing from another (see, e.g., Petersen and Rajan,
1994, and Detragiaghe, Garella and Guiso, 2000). In conjunction with the fact that many
ﬁrms can also replace bank loans with non-bank credit, these arguments imply that periods
of intense inter-bank loan reallocation can be periods of weak inter-ﬁrm credit reallocation
and viceversa. Indeed, our results support the idea that the reallocations on the supply side
and on the demand side of the credit market are governed by different processes.4

The paper also relates to the empirical literature on the “ﬂight to quality” (see, e.g.,
Bemanke, Gertler, and Gilchrist, 1996), which supports established ﬁrst generation macro-
economic models with credit imperfections (e.g., Bemanke and Gertler, 1989; Bemanke,
Gertler and Gilchrist, 2000). One version of the ﬂight to quality argument is that follow-

ing negative aggregate shocks ﬁnanciers contract credit to information opaque borrowers,

ll

such as small ﬁrms, while they accommodate the increasing credit demand of information
transparent borrowers, such as big ﬁrms.5 In turn, this ﬂight to quality works as a ﬁnancial
accelerator of recessionary shocks. The facts we uncover imply that inter-ﬁrm credit reallo-
cation is a process with distinct properties that extends beyond the ﬂights to quality during
recessions documented by this literature. In particular, we ﬁnd that credit reallocation is
a continuous, major process driven not only by the reshuﬂiing of credit across groups of
ﬁrms, as possibly implied by the ﬂight to quality argument, but also, and to a larger extent,
by the reallocation of credit within relatively homogenous groups of ﬁrms (e.g., ﬁrms of
similar size). Consistent with the predictions of second generation macroeconomics mod-
els with credit imperfections (and with the results of the job reallocation literature), we ﬁnd
that idiosyncratic credit ﬂows are larger than credit ﬂows across groups of ﬁrms and that
they explain a disproportionate share of the time variation of total credit reallocation. Fur-
thermore, credit reallocation accelerates during booms and slows down during recessions.
This ﬁnding does not contradict the argument that ﬂights to quality oCcur during recessions
but are virtually absent during booms (Bemanke, Gertler, and Gilchrist, 1996): in fact, in
our data the procyclical behavior of credit reallocation primarily reﬂects the procyclical
behavior of idiosyncratic credit ﬂows within narrowly deﬁned groups of ﬁrms. Yet, this
ﬁnding ﬁts the predictions of second generation models, conﬁrming that these models can
useﬁrlly complement previous ones. Finally, as anticipated in the Introduction, this paper
relates to the empirical literature on the reallocation of labor and physical capital. For ex-
ample, the procyclical behavior of credit reallocation brings new arguments to the debate
on the cyclical pattern of job and capital reallocation. We elaborate on the relationship with

this literature throughout the analysis.

12

3 Methodology

3.1 Overview of the Data

Our main data source is the Standard and Poor’s Full-Coverage Compustat tapes, which
provide information on the balance sheets and income statements of all publicly traded
US. ﬁrms.6 There are advantages and drawbacks in using Compustat. On the minus side,
Compustat excludes small ﬁrms: the average ﬁrm in our sample has sales of 96 1 .26 million
dollars. On the plus side, Compustat covers a long period, which allows us to investigate
the long-run behavior of credit ﬂows: the original sample comprises annual data from 1950
to 2003 as well as quarterly from 1962Q1 to 2004Q3. Furthermore, the comprehensive
scope of Compustat enables us to analyze credit reallocation in both manufacturing and
non-manufacturing sectors. This is important because, as shown by the job reallocation
literature (see, e.g., Foote, 1998), manufacturing and non-manufacturing sectors can feature
different restructuring processes. Because we are interested in ﬁrrns that demand rather
than supply credit, we remove all ﬁrms in the industry group “Finance, Insurance, and Real
Estate”.

For the ﬁrst years of both samples and for the last years of the quarterly sample the
number of observations is disproportionately lower than for other years. Thus, we drop
these years and work with annual data from 1956 to 2003 and quarterly from 1971Q1 to
2003Q4. In what follows, we employ annual data to construct descriptive statistics of inter-
ﬁrrn credit ﬂows, thus exploiting the length of the 1956-2003 period. We employ instead

quarterly data in the VAR analysis.

13

3.2 Measurement of Credit Flows
3.2.1 Deﬁnitions

The ﬁnance literature endorses different notions of debt: perhaps the majority of studies
concentrate on long-term debt only (see, e.g., Bradley, Jarrell and Kim, 1984), while others
include short-term debt (see, e.g., Ferri and Jones, 1979). Short-term and long-term debt
typically serve different purposes. Short-term debt mostly covers the time lag between the
ﬁnancing of current business operations (e.g., payment of inventories or wages) and the
accrual of returns. Long-term debt typically ﬁnances long-term plans, such as the purchase
of equipment or structure. We are primarily interested in debt changes that reﬂect long-term
investment and production plans but we do not want to neglect the possibility that ﬁrms use
short-term debt to ﬁnance these plans. Thus, we adopt a ﬂexible approach, presenting
results for total debt and separately for long-term and short-term debt.7

Following an established practice in the literature, our notion of short-term debt in-
cludes all forms of ﬁnancial debt, such as loans from ﬁnancial institutions (e.g., revolving
credit lines) and dispersed debt (e.g., commercial bills), but it excludes accounts payable
to suppliers. There are strong reasons for this exclusion. First, trade credit is ﬁ'equently
used for purposes unrelated to ﬁnancing (Rajan and Zingales, 1995). Trade credit is often
used for transaction purposes: for example, it is used to reduce the frequency of payments
to suppliers and, hence, transaction costs or to ensure recourse in case of inferior product
quality. Trade credit is also used like advertising to differentiate products. The transac-
tion purposes of trade credit tie its dynamics to ﬁrms’ commercial policy rather than to

their ﬁnancing policy. Second, even when used for ﬁnancing purposes, trade credit and

14

other forms of short-term ﬁnance differ along important dimensions and are allegedly poor
substitutes. For example, trade credit is extended by ﬁrms; its contracts are governed by
speciﬁc factors; moreover, trade credit is very costly and ﬁrms use it only when they can-
not obtain cheaper ﬁnancing (Petersen and Rajan, 1994). Indeed, because of its cost, it is
unlikely that ﬁrms use trade credit to ﬁnance long-term investment and production plans,
which we are especially interested in.

In 1956-2003 the average ratio of long-term debt to total assets is 35.77%, while for
short-term debt net of accounts payable the average ratio is 17.34%. Both ratios increased
during the period: for example, the ﬁrst rose from an average of 17.74% in 1959-69 to an

average of 36% in 1990-99.

3.2.2 Measurement Issues

We have to tackle a number of methodological issues in measuring credit ﬂows. The ﬁrst
stems from the fact that a ﬁrm generally engages in different projects. Ideally, we would
want to measure the reallocation of credit across projects. However, the ﬁrm constitutes
our unit of observation and we cannot measure credit ﬂows within ﬁrms. Thus, we will
tend to underestimate credit reallocation.8

The second issue regards ﬁrm entry. Some ﬁrms that enter the data-set are newly
founded while others are existing ﬁrms that ﬁle with the Securities and Exchange Com-
mission, become incorporated, or result from the divestiture of bigger ﬁrms. We do not
want to count the debt of existing ﬁrms as additions to the aggregate stock of credit. To
address this problem, we adopt the criterion put forward by Ramey and Shapiro (1998).

Typically, the gross book value of physical capital of a new ﬁrm is similar to its net book

15

value. Thus, we drop ﬁrms that appear in the data-set for the ﬁrst time and have a ratio
between the end-of-period gross capital and the end-of-period net capital above 120% (see
the remarks in Appendix 1 for further details).

The third issue regards ﬁrm exit. This issue is less thorny than ﬁrm entry because
Compustat speciﬁes the reason that a ﬁrm exits from the data-set. Possible reasons are:
merger and acquisition, bankruptcy, liquidation, conversion to a private company, leveraged
buyout, or unspeciﬁed. We consider exits due to merger or acquisition, bankruptcy or
liquidation as credit subtractions, while we do not count exits for other reasons (see Ramey
and Shapiro, 1998, for an analogous approach and the remarks in Appendix 1 for further
details). There is a strong reason to consider the exit of a merged or acquired ﬁrm as
a credit subtraction. When two ﬁrms merge, the management and workforce of either
acquire control over the ﬁnancial resources of the other. Thus, from the perspective of the
ﬁnanciers of either ﬁrm, this is at least partly equivalent to reallocating credit between two
ﬁrms. Indeed, a large body of literature (e.g., Servaes, 1991) ﬁnds that the announcement of
mergers signiﬁcantly affects the stock market valuation of target and acquirer, suggesting
that mergers have signiﬁcant real effects.

The fourth issue regards the use of ﬁscal or calendar year. In Compustat the data for a
whole ﬁscal year are attributed to the calendar year in which the ﬁscal year ends if the ﬁscal
year ends after May 31st and to the previous one otherwise. We recalculated credit ﬂows
partitioning ﬁscal year data proportionally into calendar years. The results were virtually
identical and therefore we chose to use the original data.

The ﬁnal issue regards the treatment of inﬂation. Because we want to capture changes

in ﬁrms’ real exposure to ﬁnanciers and relate the dynamics of credit ﬂows to that of real

16

variables, we deﬂate the original annual data using the implicit GDP deﬂator.

3.2.3 Aggregation

The methodology used to measure credit ﬂows replicates that developed by Davis and
Haltiwanger (1992) and Davis, Haltiwanger and Schuh (1996) for the measurement of job
ﬂows. We deﬁne cf, as the average of the debt of a ﬁrm f at time t - 1 and at time t. For a
set of ﬁrms, we deﬁne Cs, in an analogous manner. We then deﬁne the time t debt growth
rate of the ﬁrm (g f,) as the ﬁrst difference of its debt divided by c f,. This measure of the
growth rate, which is a monotonic transformation of the canonical one, takes values in the
interval [-2,2]; for newborn ﬁrms g f, equals 2 while for dying ﬁrms g f, equals -2.9 The
empirical distribution of the growth rates of total debt (not shown to save space) has a peak
close to zero and two spikes of approximately 5% in correspondence of the entry (+2) and
exit (-2) of ﬁrms. About 68% of the observations are clustered in the interval [-0.5,0.5],
while 81% fall in the interval [-l,+l].

We calculate credit creation (P US) as the weighted sum of the debt growth rates of
ﬁrms with rising debt or newborn ﬁrms. Analogously, we calculate credit destruction
(NEG) as the weighted sum of (the absolute values of) the debt growth rates of ﬁrms with
shrinking debt or dying ﬁrms. Both these sums are computed weighting the debt growth

rate of a ﬁrm f by the ratio cf, /Cs,. Formally,

l7

C
pos, = Z ”(54) (1)
fest St
8ft>0

NEG, = Z |gﬂ| (FCL) (2)

fest
8ft<0

where S, is the set of ﬁrms in year t. As in Davis and Haltiwanger (1992), we also deﬁne
credit reallocation (SUM) as the sum of credit creation and credit destruction, and excess
credit reallocation (E X C) as the reallocation in excess of the net credit change (NE T)

expressed in absolute value. Formally,

SUM, = POS,+NEG,, (3)
EXC, = SUM, — |NET,|. (5)

A net increase in credit can be attained through a positive value of credit creation and a
zero value of credit destruction; similarly, a net decrease in credit can be attained through
a positive value of credit destruction and a zero value of credit creation. Thus, EX C
measures credit reallocation in excess of the minimum required to accommodate net credit

changes.

18

4 Summary Statistics and Cross-Sectional Properties

In this section, we present summary statistics for the credit creation, destruction and real-
location rates. First, we assess the magnitude of the credit. ﬂows. Then, we investigate their

cross-sectional properties.

4.1 Magnitude

In Table 1.1, Panel A, we report the average credit creation, destruction and reallocation
for the 1956-2003 period and for the 1959-69, 1970-79, 1980-89, and 1990-99 sub-periods.
We also report the average credit change and excess reallocation. Starting with total credit,
the average credit creation over the sample period equals 11.3%, while the average credit
destruction equals 5.97%. Hence, the average credit change is 5.33% and the average
reallocation is 17.27% (see Figure 1.1 for the plot of credit change and credit reallocation).
Finally, the average excess reallocation is 11.29%. These ﬁgures entail the existence of
ﬂows of credit in and out of ﬁrms which well exceed net ﬂows: on average more than 11%
of credit is reallocated across ﬁrms each year. Furthermore, the yearly data (not reported)
reveal the simultaneous presence of large positive and negative ﬂows at any phase of the
business cycle. For example, in each year since 1980 both credit creation and destruction
have exceeded 5.6%; in 1975, when net credit shrank by 5.47%, the creation rate was 4.5%;
in 1998, when net credit rose by 11.4%, the destruction rate exceeded 8%.

The reader may wonder to what extent the magnitude of the credit ﬂows is attributable
to temporary short-term ﬁnancing shortages.lo To address this issue, we report summary

statistics for long-term and short-term credit. The average reallocation of long-term credit

19

is 17.34%, which stems from a creation of 11.17% and a destruction of 6.16%. The net
average change is 5.01%, while the average excess reallocation is 12.01%. The average
creation, destruction and reallocation rates of short-term credit are larger than those of
long-term credit: the creation rate is 25.37%, the destruction rate is 18.37%, so that gross
reallocation is 43.73%; excess reallocation is 31.54%.

The ﬁgures for long-term credit suggest that the intense credit reallocation we uncov-
ered is not only the outcome of temporary short-term ﬁnancing needs. To verify this hy-
pothesis, and in line with Davis and Haltiwanger (1992), we compute a measure of the

persistence of the debt changes underlying the ﬂows as

 

debt owth rate between t and t + 2
p, = I g’ ] (6)

debt growth rate between t and t + 1’

The maximum persistence occurs when P, = 1: in this case, all the debt change between
t and t + 1 will last one additional year; the minimum occurs when P, = 0. For total debt
the unweighted average value of P across positive and negative growth rates is 0.402. The
changes of long-term debt are more persistent that those of short-term debt: the mean value
of P is 0.42 for long-term versus 0.32 for short-term debt. This conﬁrms that a sizable
portion of the credit ﬂows, especially in long-term credit, reﬂects persistent ﬁrm-level debt

changes.

Comparison with Job, Capital and Inter-Bank Loan Flows. How large is inter-ﬁrm
credit reallocation compared with the reallocation of jobs and physical capital and with

the inter-bank loan reallocation? To ease comparisons, in Table 1.2, Panel A, we re-

20

port average credit ﬂows for the 1973-88 period, together with the average ﬂows of jobs
and physical capital documented for the same period by Davis and Haltiwanger (1992)
and Ramey and Shapiro (1998), respectively. To be consistent with these studies, we re-
strict our attention to the manufacturing sector. In Panel B, we report credit ﬂows from
1979Q2 to 1999Q2 together with the inter-bank loan ﬂows obtained for the same period
by Dell’Ariccia and Garibaldi (2005). To be consistent with their study, we report average
quarterly non-deﬂated ﬂows. The reader should bear in mind two issues when interpreting
these comparisons. First, Davis and Haltiwanger (1992) uses plant-level data and Ramey
and Shapiro (1998) has information on the reallocation of capital within ﬁrms: therefore,
these studies tend to capture more reallocation than ours. Second, Ramey and Shapiro
(1998) uses Compustat data like us while Davis and Haltiwanger (1992) uses data from the
Census Longitudinal Research Dataﬁle and Dell’Ariccia and Garibaldi (2005) uses data
from banks’ Call Report Files. Thus, the comparison with Ramey and Shapiro (1998) is
particularly tight whereas the comparison with the other two works should be performed
with some caution.

During the years 1973-1988, the average credit reallocation in manufacturing is 23.01%,
the net credit change is 4.43% and the excess reallocation is 15.87%. By comparison, the
average job reallocation is 19.4%, the employment growth rate is -l.1% and the excess
job reallocation is 15.4%. For physical capital, the average reallocation is 21.2%, net in-
vestment is 1.2% and the excess reallocation is 18.9%. Hence, credit reallocation is of the
same order of magnitude as the reallocation of jobs and capital.11 The ﬁgures in Panel B
are also insightful, as they reveal that the size of inter-bank loan ﬂows is somewhat smaller

than that of inter-ﬁrm credit ﬂows. For instance, gross (excess) inter-ﬁrm credit realloca-

21

tion amounts to 8.38% (6.29%) while gross (excess) inter-bank loan reallocation amounts

to 4.6% (2.69%).

4.2 Cross-Sectional Properties

The observed credit reallocation may stem from two processes: the reallocation of credit
within relatively homogenous groups of ﬁrms and the reshufﬂing of credit across groups of
ﬁrms with different characteristics (e.g., ﬁrms of different size or in diﬁ'erent industries).
The former process reﬂects ﬁrm-level heterogeneity in debt dynamics; the latter may reﬂect
sectorial shocks or different sectorial responses to aggregate shocks. As we better argue in
the remainder of the analysis, ﬁrst generation macroeconomic models with credit imper-
fections stress the different effects that aggregate shocks may have on the debt capacity
of different groups of ﬁrms. Bemanke, Gertler, and Gilchrist (1996), for instance, treats
size as a proxy for information transparency12 and studies the incentive of ﬁnanciers to
reallocate credit from small, information opaque ﬁrms to big, information transparent ones
during recessions.13 By contrary, second generation models view the allocation of credit
as mostly independent of ﬁrms’ intrinsic characteristics. In Den Haan, Ramey and Watson
(2003) and Wasmer and Well (2004), the reallocation of credit across ﬁrms is governed by a
stochastic process summarized by a random matching function. As a result of this random
process, credit ﬂows with equal probability, say, to a ﬁrm with high and low productivity
or to a small and big ﬁrm.

In what follows, we try to disentangle the contribution of the within-group and the

cross-group credit reallocation. We ﬁrst partition our sample in industries: focussing on the

22

sub-sample of manufacturing ﬁrms, we sub-divide it into two-digit SIC industries. We then
partition the aggregate sample using two other classiﬁcation schemes, size and location.

In order to gauge the magnitude of the within-group credit reallocation, we construct the

W-index as14

S
ZGNEEtI)
s=1

_ S ,
Z SUMS,
s=1

with W, taking on the value of 1 if in year t excess credit reallocation occurs entirely

Wt=l

 

(7)

within groups and 0 if it occurs entirely across groups. The rationale is the following. If in
group 3 there is only credit creation or destruction, then SUM,» = IN E El . If this occurs
for every group, then W = 0, signalling the absence of reallocation within groups. Yet,
there could still be reallocation across groups. If, instead, IN E El = 0 for each group and
SUM, > 0 for some group, W = 1 and all the reallocation will occur within groups. In
reviewing the evidence below, it is perhaps useﬁrl to bear in mind that for jobs Davis and
Haltiwanger (1992) obtains strikingly low values of the W-index, less than 5% for each
classiﬁcation scheme (industry, size, or location). We are going to show that in the case of
credit a signiﬁcant amount of reallocation is generated by heterogeneity in ﬁrm-level debt

dynamics, although the cross-group reshuﬂ’ling is not as trivial as in the case of jobs.

23

Industry. In Table 1.3, we report the average credit creation and destruction in each man-
ufacturing industry and in the manufacturing sector over the sample period. For manufac-
turing as a whole, the magnitude of credit ﬂows is slightly larger than for the aggregate.
The average credit creation is 12.97%, while the average. destruction is 7.32%. Thus, the
average net credit change is 5.65% and the average reallocation is 20.29%. Turning to
the single manufacturing industries, we ﬁnd substantial variability in the magnitude of the
ﬂows. For example, credit creation ranges from 21% in Printing, Publishing, and Allied
Industries to 10% in Petroleum, Reﬁning and Related Industries. Leather and Leather Prod-
ucts features the highest credit destruction rate (12.82%), while transportation features the
lowest (6.33%). The cross-industry range of variation for the net change is 9.98%-1.43%;
for reallocation it is 31.64%-l7.08%. The substantial variability of the reallocation rate
reﬂects the high positive correlation (approximately 0.5) between credit creation and de-
struction across industries. The ﬁgures for long-term credit are very close to those for total
credit and we do not dwell on them (see Panel B).

The most important piece of information we learn from the table is that signiﬁcant ﬂows
of credit creation and destruction coexist within manufacturing industries. This suggests
that an important amount of reallocation may not reﬂect the reshuﬂling of credit across
industries but within-industry heterogeneity in the dynamics of ﬁrms’ debt. To verify this
hypothesis, we compute the W—index, where now the generic group s is a manufacturing
industry. The value of W ranges ﬁ'om 0.235 in 1956 to 0.746 in 1994 and its mean is 0.546.
Thus, although the reshuﬂiing of credit across industries is important, a signiﬁcant fraction

of reallocation is generated by heterogeneity in ﬁrm-level debt dynamics.15

24

Size and Location. We decompose the aggregate sample using two classiﬁcation schemes:
size and location. In Table 1.4, Panels A and B, we partition ﬁrms in quartiles according
to their sales. The data show that a signiﬁcant amount of reallocation occurs within all the
quartiles, both when we consider total credit (Panel A) and its long-term component (Panel
B). In fact, the values of the W-index are 0.632 for total credit and 0.627 for long-term
credit. Furthermore, even if we use a ﬁner partition of the sample in deciles, we obtain
values of W of 0.608 for total credit and 0.605 for long-term credit.

Interestingly, there appears to be a clear negative relationship between the intensity of
credit reallocation and ﬁrm size. Consider total credit: the average excess reallocation rate
drops from 34.02% for the quartile of smallest ﬁrms to 9.6% for the quartile of largest
ﬁrms. This ﬁnding somehow resembles that in Davis and Haltiwanger (1992): for the
1975-86 period, they ﬁnd that excess job reallocation drops from 28.1% for plants with
1-99 employees to 11.9% for plants with more than 1000 employees. Indeed, the result
we obtain here is even more noteworthy because the variation occurs across medium-large
ﬁrms.

In Table 1.5, we partition ﬁrms according to their census region. The average credit
reallocation ranges from 20.24% in West North Central to 16.51% in East South Central.
Thus, in all the census regions an intense process of reallocation is at work suggesting
that the reshuffling of credit across regions can only explain a fraction of the observed
reallocation. This impression is conﬁrmed when we compute the W-index, which takes on
the value of 0.602 (0.625 for long-term credit). Finally, note that the credit reallocation
rate varies across regions less than across industries or sales quartiles. This resembles the

ﬁndings of Davis and Haltiwanger (1992) for job ﬂows, which indeed appear to have a

25

somewhat lower variance across census regions than across industries or size classes.

5 Theoretical Background

One of the most relevant contributions of the literature on gross input (labor and capital)
ﬂows has been to discriminate among different classes of theoretical models of the aggre-
gate restructuring process. For example, the studies on gross job ﬂows have shown that
embedding shocks to demand, input costs, or technology in macroeconomic models with
search frictions in the labor market allows to explain several properties of the ﬂows (for
a discussion see, e.g., Davis and Haltiwanger, 1992). This has revealed that we can learn
important insights from search models of the labor market. In this paper, we pursue an
analogous objective. The main question that guides us throughout the analysis is whether
the facts we document support macroeconomic models with credit imperfections. More-
over, we are interested in understanding whether these facts help to discriminate between
the ﬁrst and the second generation of these models, thus shedding light on the main obsta-
cles to the process of aggregate restructuring. For this purpose, in this section we compare
and contrast the predictions of these two classes of models with respect to the properties of

credit reallocation (ﬁrrther details are in the remainder of the analysis).

5.1 Preliminary Observations

First and second generation macroeconomic models with credit imperfections differ espe-
cially in the ﬁ'ictions that shape credit reallocation rather than in the underlying shocks

that drive reallocation. In particular, in both these classes of models credit reallocation can

26

be generated by real shocks that continually redistribute production opportunities across
ﬁrms and by shocks that redistribute liquid funds. Moreover, particularly in ﬁrst generation
models, credit reallocation can be generated by continuous shocks to ﬁrms’ net worth. In
fact, it is well established in the literature that in the presence of credit imperfections the
level of debt of a ﬁrm is determined not only by its production opportunities and by the
current amount of its internal funds but also by the debt capacity of the ﬁrm. In turn, since
a ﬁrm’s net worth inﬂuences its incentives to behave and declare its characteristics to ﬁ-
nanciers truthfully, debt capacity is affected by the net worth of the ﬁrm and, hence, by its
proﬁtability and asset value. These links constitute a standard feature of models with credit
imperfections and are extensively investigated in previous papers (see, e.g., Myers, 2001,
and Harris and Raviv, 1991 and 1990, for in-depth analyses). For this reason, and because
ﬁrst and second generation models do not differ sharply in this respect, we do not dwell on
the nature of the shocks that underly credit reallocation but, for given shocks, we focus on

the ﬁ'ictions that shape the reallocation process.

5.2 Second Generation Models

Second generation models emphasize “horizontal” heterogeneity across ﬁrms and lock-in
effects in the credit relationships between borrowers and lenders due to search ﬁ'ictions
and/or speciﬁcity. Wasmer and Weil (2004) studies a model economy where ﬁrms face
search frictions both in the credit market and in the labor market. A ﬁnancier (banker) can
ﬁmd the ﬁxed cost that a ﬁrm has to sustain to post a vacancy in the labor market. The

meeting technology between creditors and borrowers has a stochastic random nature which

27

is described by a matching function. Caballero and Hammour (2005) develops a model
where ﬁrms carry out indivisible projects within matches that also involve workers and ex-
ternal ﬁnanciers. The relationship between a ﬁrm and the external ﬁnanciers of its project
features partial speciﬁcity: if the ﬁnanciers relocate capital to another match, they will face
costs. This speciﬁcity generates a hold-up problem: ex post the ﬁrm can threaten to break
the match and thereby extract rents. Finally, the hold-up implies that too many productive
matches are destroyed because ﬁnanciers are unwilling to inject ﬁrnds and offset negative
cash ﬂow shocks. Furthermore, too few matches are created because, expecting the hold-
up, ﬁnanciers are reluctant to inject funds. Den Haan, Ramey and Watson (2003) develops
a model in which the lock-in of ﬁnanciers stems from search frictions, as in Wasmer and
Weil (2004). In their context, projects are divisible but feature ﬁxed costs, so that entrepre-
neurs who obtain too little credit have the incentive to shirk. Finally, an entrepreneur who
obtains little credit from her ﬁnancier cannot immediately form a match with an alternative
ﬁnancier because of search frictions in the credit market.

All in all, second generation models deliver a number of distinct predictions that can
be contrasted with those of ﬁrst generation models (discussed below): (i) idiosyncratic
(within-group) inter-ﬁrm credit ﬂows are large and account for an important share of credit
reallocation. In Wasmer and Weil (2004) and Den Haan, Ramey, and Watson (2003) credit
reallocation is governed by a matching function and credit ﬂows to ﬁrms randomly, i.e.
independently of ﬁrms’ intrinsic characteristics (size, productivity, etc.); (ii) search fric-
tions and lock-in effects in the credit marketihinder the creation of new credit relationships,
dampening the volatility of the creation rate over the cycle. Therefore, these models im-

ply that credit creation is less volatile than credit destruction; (iii) the credit reallocation

28

rate tends to surge during booms and drop during recessions. In Caballero and Hammour
(2005), for example, this happens because the creation of credit relationships falls during

recessions, while cumulatively their destruction rises only moderately.

5.3 First Generation Models

First generation macroeconomic models with credit imperfections emphasize asymmetric
information and agency problems in the “vertical” relationships between entrepreneurs or
managers and external ﬁnanciers. In macroeconomic models with informational asym-
metries, ﬁnanciers have imperfect information on ﬁrms’ assets, proﬁtability or investment
opportunities and may charge high quality ﬁrms an “external ﬁnance premium”. This group
of models build on established corporate ﬁnance theories such as the pecking order theory
(Myers and Mijluf, 1984).16 In macroeconomic models with agency issues, debt may in-
duce managers to engage in risky projects (risk-shifting) or under-invest (debt overhang) to
reduce the expected debt repayment. This group of models, which comprise Bemanke and
Gertler (1989) and Bemanke, Gertler, and Gilchrist (2000), build on the corporate ﬁnance
analysis of Jensen and Meckling (1976).

The key empirical predictions of ﬁrst generation models differ from those of second
generation models: (i) credit reallocation is mostly a cross-group phenomenon. For exam-
ple, Bemanke and Gertler (1989) predicts that during recessions credit ﬂows from small
to large ﬁrms. In fact, the ﬂight to quality literature has investigated empirically the oc-
currence of this cross-group reallocation; (ii) credit creation is highly volatile because it

reﬂects the dynamics of ﬁrms’ net worth besides that of cash ﬂows. In particular, as dis-

29

cussed by Caballero and Hammour (2005), the high volatility of credit creation is jointly
driven by ﬂuctuations in ﬁrms’ proﬁts and in the value of their collateral assets; (iii) credit
reallocation tends to follow a countercyclical pattern. Flights to quality occur during reces-
sions because agency costs increase when ﬁrrns’ net worth drops (Bemanke, Gertler, and

Gilchrist, 1996).

6 Time Series Properties

Because of the different predictions of ﬁrst and second generation models, understanding
the dynamics of credit reallocation can help us to discriminate between these two classes
of models. We investigate the time series properties of credit ﬂows in three steps: long-run

pattern, volatility, and interaction with the business cycle.

6.1 Long-run Overview

Table 1.1 shows that credit reallocation surged signiﬁcantly in the eighties and then dropped
slightly in the nineties. In the sixties and seventies, the total reallocation was close to 14%
while in the eighties it exceeded 21%. This pattern is also evident for long-term credit.
Second generation models offer an explanation for this pattern. According to Caballero
and Hammour (2001), in the eighties and nineties the large amount of liquidity generated
by the rise in stock market prices fostered sellers’ incentives to sustain transaction and
search costs and sell assets. This increased the reallocation of physical capital. In this
view, credit reallocation may have been crucial to ﬁnance this process of reallocation of

physical capital. However, this is not the only possible explanation and ﬁrst generation

30

models can also rationalize the surge in credit reallocation. Holmstrom and Kaplan (2001)
argues that the deregulation and advances in information technology occurred in the United
States at the end of the seventies uncovered severe agency problems (of the type studied es-
pecially by ﬁrst generation models): inefﬁcient ﬁrms wasted cash ﬂow, while eﬂicient ones
were under-scaled. In the eighties and nineties, US. corporations underwent a massive re-
structuring, which entailed stock buybacks and takeovers and was largely ﬁnanced by debt
(especially bank loans and junk bonds). If creditors are indeed a source of discipline and
monitoring for ﬁrms, credit reallocation may have been critical for ﬁrrrrs’ restructuring. In
fact, credit would have migrated from eﬂicient to ineﬂicient ﬁrms that needed to discipline

their managers.

6.2 Volatility

Credit creation and destruction exhibit signiﬁcant volatility (see Table 1.1, Panel B). Ex-
cluding 1988, when a dramatic rise in short-term credit occurred andthe creation rate was
45%, credit creation ranges ﬁ'om 19.01% in 2000 to 4.45% in 1975; credit destruction
ranges from 11.01% in 1984 to 1.34% in 1964. The coefﬁcient of variation (100*standard
deviation/mean) of credit creation exceeds 39%, while that of credit destruction exceeds
42%; for gross and excess reallocation, the ﬁgures are approximately 30% and 41%.17 To
better grasp the magnitudes at stake, it is useful to compare these ﬁgures with those forjob
ﬂows. Using the data from 1973 to 1986 in Davis and Haltiwanger (1992), we calculated
that the coefficient of variation of job creation is 24.71%, that of job destruction is 28.83%,

while that of job reallocation is 8.82%. Thus, at least for this period, credit ﬂows are more

31

volatile than job ﬂows.18

The literature on job ﬂows generally ﬁnds that the volatility of job destruction exceeds
that of job creation19 and explains the sluggishness of job creation with the frictions that
ﬁrms face in hiring workers, such as search frictions (for a theoretical model, see Mortensen
and Pissarides, 1994). Seemingly in contrast to this ﬁnding, the coeﬂicient of variation of
credit creation and that of credit destruction are roughly equal. Yet, if we focus on long-
terrn credit, the coefﬁcient of variation of the credit creation rate (36%) is signiﬁcantly
lower than that of the destruction rate (46%).20 The ﬁnding for total credit is thus due to the
dynamic behavior of short-term credit (see Table 1.1, Panel B). Second generation models
can rationalize the higher volatility of long-term credit destruction relative to long-term
credit creation. In Caballero and Hammour (2005) the destruction of productive matches
and credit relationships ﬂuctuates more than the creation of new ones. This occurs because
a hold-up problem within credit relationships discourages the ﬁnancing of production units
ex ante and thereby depresses the creation of relationships. The analyses of Wasmer and
Weil (2004) and Den Haan Ramey and Watson (2003) yield an analogous implication:
just like search ﬁ'ictions in the labor market hinder job creation, search frictions in the
credit market render the formation of credit relationships a sluggish process. The corporate
ﬁnance literature allows to integrate these implications of second generation models and
perhaps explain the asymmetry between long-term and short-term credit. Long-term credit
entails a long-term commitment by ﬁnanciers: this exposes ﬁnanciers to hold-up because
they cannot recall funds at will (Diamond, 2004), possibly reinforcing the effect pointed

out by Caballero and Hammour (2005).

32

Sectorial and Idiosyncratic Effects. We have established that a signiﬁcant amount of
credit reallocation occurs within homogeneous groups of ﬁrms. Although within groups the
cross-sectional variance of the debt growth rates is large in any year, this variance could be
constant over time and most of the time variation of credit ﬂows stem from other sources.
For instance, the ﬁrst generation models of Bemanke and Gertler (1989) and Bemanke,
Gertler and Gilchrist (1996) predict that the cross-group credit reshufﬂing associated with
ﬂights to quality is volatile, accelerating during recessions and being virtually absent dur-
ing booms. To discriminate among classes of models it is then critical to assess how much
of the time variation of the ﬂows is accounted for by changes in the cross-sectional vari-
ance of the idiosyncratic debt growth rates (idiosyncratic effects) and how much by mean
translations of their aggregate or sectorial distributions (sectorial/aggregate effects).

We follow Davis and Haltiwanger (1992) and decompose the debt growth rate of each
ﬁrm in two parts: the sector growth rate (gi), which bundles together aggregate and sec-

torial effects, and an idiosyncratic component (g5?) Formally,

8ft = 3?. + 5:}.- (8)

We then recompute credit ﬂows using only the idiosyncratic component of debt growth
and denote them by superscript 1. Using simple algebra, the variance of a ﬂow equals the
sum of the variance of the idiosyncratic ﬂow, the variance of the sectorial-mean component

and twice the covariance between the idiosyncratic ﬂow and the sectorial-mean component.

33

For instance,

varSUM = varSUM’ + var (SUM — SUM’) + 2cov(SUM — SUM’ , SUM’). (9)

If the time variation of credit reallocation is entirely driven by mean translations of
the aggregate or sectorial distributions of debt changes, varSU MI /varSU M = 0; if it
is mostly driven by changes in their cross-sectional variance, varSU MI /varSU M will
take a high value. The covariance term captures the share of variance that cannot be at-
tributed directly to either effect and signals whether the sectorial/aggregate effects and the
idiosyncratic effects work in the same or in opposite directions.

In Table 1.6, we report the variance decomposition using our three preferred classiﬁ-
cation schemes: size classes, census regions, and two-digit manufacturing industries. The
results indicate that a disproportionate share of the total variance of gross credit ﬂows is
attributable to the cross-sectional variance of idiosyncratic growth rates. Consider manu-
facturing industries: about 83% of the variance of the reallocation rate in manufacturing
can be attributed directly to idiosyncratic effects while only about 4% can be attributed to
mean/sectorial effects. The results for long-term credit are even more remarkable, with
almost 86% of the total variance being accounted for by idiosyncratic effects. We obtain
similar ﬁgures when, focussing on the aggregate, we decompose the variance using size and
location as classiﬁcation schemes: for quartiles of sales, the share of variance explained di-

rectly by idiosyncratic effects is 95%.

34

6.3 Cyclical Properties

We now investigate the cyclical properties of credit reallocation. In Table 1.7, we report
pairwise correlation coeﬂicients of the credit ﬂows with unemployment for the 1956-2003
period and for the 1959-69, 1970-79, 1980-89, and 1990-99 sub-periods. The net change
of credit is procyclical: the contemporaneous correlation with the unemployment rate is
-0.7411, while the correlation with unemployment lead (lagged) one year is -0.6123 (-
0.5616). This may indicate that the contraction in credit supply that occurs during reces-
sions overwhelms the increase in credit demand by ﬁrms due to the reduced availability of
internal funds. The procyclical pattern of the net credit change stems ﬁ'om a procyclical
credit creation and a countercyclical credit destruction (see the table).

Turning to credit reallocation, we ﬁnd that it is also procyclical. The contemporaneous
correlation with unemployment is -0.2157; reallocation is also negatively correlated with
unemployment lead or lagged one year. The procyclical pattern of credit reallocation is
especially evident in the eighties and in the nineties when the correlation coeﬂicients with
unemployment equal -0.6773 and -0.8911, respectively. In these two decades, the lack
of action on the destruction margin in response to cyclical ﬂuctuations is particularly re-
markable: the credit destruction rate is countercyclical but its correlation coeﬂ‘icient with
unemployment equals only 0.1654 in the eighties and 0.3196 in the nineties. To save space,
we do not report the corresponding tables for long-term aggregate credit and for total and
long-term credit in the manufacturing sector. However, the key ﬁndings carry over: in all
these cases we ﬁnd that credit reallocation exhibits a procyclical pattem.”

The cyclical behavior of credit reallocation ﬁts the predictions of second generation

35

models (Caballero and Hammour, 2005; Den Haan, Ramey and Watson, 2003). In Ca-
ballero and Hammour (2005) reallocation across production units declines during reces-
sions. This occurs because, as a result of the hold-up problem, creation falls, while cu-
mulatively destruction rises only moderately. Indeed, we ﬁnd that in absolute value the
correlation of credit creation with unemployment is somewhat larger than that of credit
destruction. The procyclical pattern of credit reallocation is also interesting in light of the
debate on the cyclical behavior of input reallocation. Davis and Haltiwanger (1992) and
Davis, Haltiwanger, and Schuh (1996) ﬁnd that job reallocation is countercyclical; Ramey
and Shapiro (1998) ﬁnds that capital reallocation is also countercyclical. Moreover, on the
ﬁnancial side, Dell’Ariccia and Garibaldi (2005) ﬁnds that inter-bank loan reallocation is
countercyclical, although this result is not easily comparable to the previous ones because
the authors focus on the dynamics of excess reallocation.22 All these ﬁndings suggest that
the restructuring process accelerates during recessions. However, recent studies are increas-
ingly challenging this view. For example, Foote (1998) ﬁnds a procyclical pattern of job
reallocation in US. non-manufacturing industries; Eisfeldt and Rampini (2005) uses data
on acquisitions and sales of property, plant, and equipment ﬁom U.S. Compustat Tapes and
documents that capital reallocation is procyclical. Our ﬁndings for credit reallocation go in

the direction of this recent group of studies.

Sectorial and Idiosyncratic Effects. How do the idiosyncratic effects and the sector-
ial/aggregate effects contribute to the cyclical pattern of credit reallocation? In Table 1.6,
bottom row of Panels A and B, we report the correlation coeﬂicient of the idiosyncratic

credit ﬂows with unemployment. Consistently across classiﬁcation schemes, we ﬁnd that

36

the idiosyncratic ﬂows are negatively correlated with unemployment. Consider manufac-
turing industries: the pairwise correlation coefficient between unemployment and idiosyn-
cratic reallocation is -0.1617, which slightly drops to -0.1225 for long-term credit. It thus
appears that not only idiosyncratic ﬂows drive most of the time variation in credit realloca-

tion, but they also exhibit a procyclical pattern.

6.4 VAR Analysis

In Section 6.3, we analyzed the cyclical properties of credit reallocation using the uncon-
ditional correlation between the credit ﬂows and the unemployment rate. Insofar as these
correlations can help to discriminate among different theories, it is important to explore
whether the moderately procyclical behavior of credit reallocation is conﬁrmed when we
employ alternative measures of correlation. Indeed, Den Haan (2000) argues that not only
the magnitude but also the sign of the empirical correlations are sensitive to the method
employed for their calculation. Furthermore, by focusing on a single correlation coeﬂicient
(the unconditional correlation), one looses information regarding the dynamic comovement
between economic activity and credit ﬂows.

In this section, we explore the cyclical properties of credit reallocation using a VAR ap-
proach. We proceed in two steps. We ﬁrst examine the dynamic response of credit creation
and destruction to exogenous output shocks using a structural VAR. Our focus on output
shocks is motivated by the observation that second generation models with credit imper-
fections focus on the impact of real shocks on credit creation and destruction whereas they

are mostly silent about the impact of nominal (monetary) shocks. Next, we provide ﬁrrther

37

evidence on the comovement between output and credit ﬂows using a statistic proposed
by Den Haan (2000) which only requires estimation of a reduced-form VAR and, hence,
does not depend on particular identifying assumptions. Note that, while in Section 6.3 we
focused on the unconditional correlation with the unemployment rate to ease comparisons
with the labor and capital reallocation literature, we now focus on the conditional correla-

tion with output growth. In fact, the use of GDP is the norm in VAR studies.

6.4.1 Response to Output Shocks: Evidence from a Structural VAR

To study the effect of output shocks on credit reallocation we consider a quarterly VAR
describing the behavior of y,, a vector that comprises a macro block and a credit ﬂows
block. Following the parsimonious approach of Rudebush and Svensson (1999), the macro
block includes the log grth of real GDP (yyy), the log grth of the CPI (pr), and,
as a policy instrument, the federal ﬁrnds rate (y f, 0.23 The credit ﬂows block includes the
non-deﬂated credit creation rate (ya) and the non-deﬂated credit destruction rate (yam) of
the sector and maturity of interest. We use seasonally adjusted quarterly data spanning the
1971 :Q1-2003zQ4 period.24

We assume the data generating process for y, to be given by the following structural

VAR

Boy; = B (L)y:—1 + u: (10)

I o o . o o
where u, = [14“, u sz u f, ,, um, add] rs a vector of whrte norse structural mnovatrons
uncorrelated with all the variables dated t — 1 and earlier, with variance-covariance matrix

E [muﬂ = D, and B (L) is a matrix lag polynomial of order 4.

38

Our identiﬁcation strategy is as follows. Let y, = [ Yin yf ]I, where yj" is the 3 x 1

vector of macro variables and yf is the 2 x 1 vector of credit ﬂows. We rewrite (10) as

133"" o B’"’" (L) B’“ (L)

Yr = Yt-l + “t (11)

B3“ Bff B“ (L) B“ (L)
where 33"" is a 3 x 3 lower triangular matrix with ones along the main diagonal, the
elements on the main diagonal of the 2 x 2 33" matrix equal one, and there are no zero-
restrictions on the off-diagonal elements of 33". As it is common in the VAR literature,
the ordering of the variables in the macro block assumes that output is Wold-causally prior
to the CPI and the federal funds rate and that the CPI is Wold-causally prior to the federal
ﬁmds rate. This ordering reﬂects the view that monetary policy -proxied by the federal
ﬁrnds rate equation- responds contemporaneously to innovations in the real part of the
economy, as does the price level, whereas real production responds to nominal and ﬁnancial
innovations with a lag. Furthermore, by ordering the credit ﬂows last, we allow credit
creation and destruction to respond contemporaneously to output shocks. Note that we do
not pursue identiﬁcation within the credit ﬂows block, thus allowing credit creation and
destruction to be contemporaneously correlated.

In Figure 1.2, we portray the estimated impulse responses of credit creation (solid line),
destruction (dotted line), and net credit growth (dashed line) to a one unit innovation in the
GDP, that is an increase in annual output grth by 1%. The ﬁgure depicts the responses for
short-term, long-term, and total credit for the aggregate and for manufacturing. Response

coeﬂicients that are signiﬁcant at the 1% and 5% levels are respectively denoted by (0)

39

and (0) marks on the response function. Signiﬁcance levels are based on nonparametric
bootstrap intervals obtained following the methodology in Runkle (1987).”

In all cases, in the aﬁermath of the shock, the creation rate of total credit exhibits a
statistically signiﬁcant increase. For the aggregate, the peak response occurs ﬁve quarters
after the shock and entails a total credit creation rate 1.4% above the baseline value, whereas
for manufacturing the peak occurs after ﬁve quarters and entails an increase of the total
creation rate by roughly 1.9%. For long-term credit, the dynamic response of the creation
rate is similar to that for total credit: for example, for the aggregate the peak response is
realized ﬁve quarters after the shock and amounts to 0.9%. The long-run creation effect is
positive: the cumulative response of total credit creation at a 16 quarters-ahead horizon is
6.8% for the aggregate and 13.1% for manufacturing.

The shock generally triggers a drop in credit destruction, but the response is weaker
than that of credit creation and seldom statistically signiﬁcant. For the aggregate, the de-
struction rate of total credit reaches its bottom after two quarters at about 0.3% below the
baseline. Cumulating the ﬁrst 16 impulse response coefficients reveals that the shock in-
duces a slight fall of 0.5% in total credit destruction. For manufacturing, the maximal
decline of total credit occurs after four quarters but is negligible (less that 0.2%) and not
statistically signiﬁcant.

Turning to the response of the net growth of total credit granted to the aggregate econ-
omy and to manufacturing, the peak occurs after ﬁve quarters at 2.9% and 3.4%, respec-
tively. The longer term (16 quarters-ahead horizon) effect on net credit growth is positive,
taking on a value of 13.3% for the aggregate and 21.1% for manufacturing. As for total

credit reallocation, the results imply that the positive output shock generates a long-run

4o

reallocation of 4.0% for credit to the aggregate economy and 9.2% for credit to manufac-
turing; once we net out the eﬁ‘ect of the shock on inﬂation, the corresponding real ﬁgures
are approximately 2.8% and 8.0%. Thus, the shock fosters credit reallocation in an eco-
nomically signiﬁcant way.

All in all, the shape of the impulse response functions suggests that in the aftermath of a
positive shock to GDP growth credit creation rises while credit destruction does not change
or slightly drops. Thus, consistent with what found in Section 6.3 using unconditional
correlations, credit reallocation appears to have a moderately procyclical behavior.

To better understand the contribution of output shocks to the mean square error of the
k-period ahead forecast error, we also compute the variance decomposition. For concise-
ness, we only report the contribution of each identiﬁed innovation, or block of shocks,
to the 8-step and 16-step ahead forecast error variance (see Table 1.8). Regardless of the
forecast horizon and the type of credit, output shocks account for a larger percentage of
the variance of credit creation than of that of credit destruction. This is especially evident
for manufacturing where at an 8-step (16-step) ahead forecast horizon the contribution of
output innovations is 25.76% (29.36%) for total credit creation and only 3.63% (7.87%)

for total credit destruction.

6.4.2 Short and Long Run: Evidence from a Reduced-form VAR

We now explore the cyclical properties of credit reallocation using a statistic proposed by
Den Haan (2000) which requires only estimation of a reduced-form VAR and, hence, does
not rely on identifying assumptions. Another advantage of this statistic is that it captures

dynamic aspects of the comovement between two variables that are neglected by the un-

41

conditional correlation commonly used in the business cycle literature. For instance, if the
variables are stationary, the correlation statistic at horizon k converges to the unconditional
correlation as It goes to inﬁnity. On the other hand, for small k the statistic reﬂects the
comovement between the variables in the short-run, an aspect that is not captured by the
unconditional correlation.26

The correlation statistic between variables 1' and j, C 0Ri '1' (k), is computed as follows.

Consider the reduced-form VAR

Yr = ¢+t+A(L)Yr—1+ez (12)

where y, is the vector of endogenous variables described in the previous section, c is a
vector of constants, t is a time trend,27 A (L) is a matrix lag polynomial of autoregressive
coefficients, and e, is a vector of serially uncorrelated but possibly contemporaneously
correlated innovations. The k-period ahead forecast of the vector y,, given the information

at time t, is

371+“: = Et (c + (t + k) + A (L)YI+k—1 + et+k)- (13)

and the series of k-step ahead forecast errors can be generated as the difference between

42

the realization y,+k and the forecast 33+ka To compute the correlation statistic between
output, y”, and credit creation (destruction), ye“ (you), at horizon k we then estimate the
5 variable reduced-form VA R(4) in (12). Next, we compute the forecast at horizons k =
l, 2, ..., 40 and generate the series of forecast errors for the level28 of output, yﬁH, and
credit creation (destruction) ygewk (ygjﬂ). Finally, we compute the correlation coefficient
between y fft+k and ygj+k 0’25”):

Figure 1.3 plots the computed C ORW (k) coeﬂicients between output and total credit
creation and destruction. The results can be summarized as follows. For total credit to the
aggregate economy and to manufacturing, the correlation of output and creation is positive
across forecast horizons. The forecast errors for aggregate credit destruction and output are
mildly positively correlated both in the short and in the long-run. As for manufacturing,
while the forecast errors for credit destruction and output appear to be negatively correlated
in the short-run, they are mildly positively correlated in the long-run. All in all, for long
forecast horizons, credit creation appears to be procyclical whereas credit destruction is
generally acyclical. Although this result is consistent with our ﬁnding of a procyclical
credit reallocation, driven mainly by the procyclical behavior of credit creation, we should
note here that this statistic cannot be computed with a high degree of precision. In fact, we

cannot reject the null of a zero correlation at a 5% signiﬁcance level.”

7 Conclusion

In this paper, we have argued that investigating the inter-ﬁrm reallocation of credit is im-

portant for understanding aggregate restructuring and its obstacles. We have then inquired

43

into the existence of a process of inter-ﬁrm credit reallocation. The facts we have docu-
mented support recent macroeconomic models with search frictions and lock-in effects in
the credit market, implying that these models can complement more established macro-
economic models with credit imperfections. These facts can be wrapped up as follows: i)
Inter-ﬁrm credit ﬂows are large, at least as large as those of physical inputs. An impressive
credit reallocation occurs within homogeneous groups of ﬁrms, although the reshuﬂiing of
credit across the groups is non-trivial. This reallocation is signiﬁcant throughout the post-
war period and peaks in the 1980-1989 decade; ii) Credit ﬂows exhibit signiﬁcant volatility,
which primarily reﬂects time variation in the magnitude of idiosyncratic debt changes. For
long-term credit, the volatility of credit destruction is larger than that of credit creation;
iii) Credit reallocation is moderately procyclical and the results from a VAR analysis re-
veal that it rises in the aftermath of positive output shocks. In particularly, it appears that
the credit creation rate signiﬁcantly responds to the cycle while the credit destruction rate
is more mildly countercyclical. Finally, idiosyncratic ﬂows signiﬁcantly contribute to the
procyclical dynamics of credit reallocation.

We identify two directions for future research. On the theoretical side, the literature
on search frictions in the credit market is relatively recent and we still lack a model with
precise quantitative predictions on inter-ﬁrm credit ﬂows that can be compared with the
data. Building such a model is an objective of our current research. On the empirical side,
although debt and equity constitute well distinct forms of external ﬁnance, we can probably

learn important insights from comparing inter-ﬁrm credit ﬂows with inter-ﬁrm equity ﬂows.

44

Notes

1We borrow the distinction between “vertical” heterogeneity between ﬁrms and credi-
tors and “horizontal” heterogeneity across ﬁrms from Wasmer and Weil (2004).

2Following a vast literature, we consider ﬁnancial debt and exclude trade debt (see 3.2.1
for more on this).

3The ratio is constructed using book values of debt and equity. The aggregate debt/equity
ratio is deﬁned as the sum of the debt of all corporations over the sum of their equity.

4Clearly, another important difference with Dell’Ariccia and Garibaldi (2005) is that
we do not restrict our attention to bank credit.

5Firms’ liquidity needs tend to rise during recessions because of a shortage of internal
funds. A word of caution is due here. The ﬂight to quality argument does not gener-
ate strong predictions on the absolute amount of credit extended to different groups of
ﬁrms. For example, some studies ﬁnd that during recessions credit extended to small ﬁrms
declines while others ﬁnd that it is ﬂat, presumably because of the interaction between a
high credit demand and a tight credit supply.

6See Appendix 1 for details on Compustat data.

7Considering total debt also allows us to control for a measurement issue. In fact, in
Compustat long-term debt that will mature in less than one year is included in short-term
debt.

8Note that a problem of underestimation also arises from the point-in-time nature of the
data.

9For more details on the properties of the growth rate g f, see Davis and Haltiwanger

45

(1992).

10A similar issue is present in the literature on job ﬂows, where job creation and de-
struction can reﬂect short-lived establishment-level employment changes (Davis and Halti-
wanger, 1992).

11This impression is conﬁrmed when we turn to other sub-periods (values not reported).
For example, in manufacturing during the 1970-79 period the average rate of capital real-
location is 16.2%, while the average credit reallocation is 16.35%.

12Obviously, size can also reﬂect technological characteristics of a ﬁrm, besides its in-
formation transparency. Sometimes the empirical literature on credit imperfections uses
proxies of the quality of a ﬁrm such as its debt rating. However, this variable is not with-
out limits. For example, the debt rating of a ﬁrm is likely to capture especially its current
riskiness rather than more broadly its information transparency.

l3Eisfeldt and Rampini (2007) analyzes the dynamics of liquidity ﬂows in and out of the
corporate sector and relates the value of aggregate liquidity to ﬁrms’ ﬁnancing shortfalls.

14Davis and Haltiwanger (1992) uses a similar methodology.

15For short-term credit the mean value of the index is 0.578, while for long-term credit
the mean value is 0.57.

“In pecking order models (e.g., Myers and Mijluf, 1984), ﬁrms use all their cash ﬂow,
which is inexpensive because it does not entail informational asymmetries. After exhaust-
ing internal funds, they turn to risky debt, which is the cheapest form of external ﬁnance;
ﬁnally, they use more costly ﬁnance (e.g., equity).

l'7The coeﬂ'icient of variation decreases from the quartile of smallest ﬁrms (52.07%)

to the third quartile (41.75%), while it rises slightly from the third to the fourth quartile

46

(43.01%). Thus, volatility tends to drop as ﬁrm size grows larger.

”This is also true if, like Davis and Haltiwanger ( 1992), we restrict our attention to the
manufacturing sector.

19Dell’Ariccia and Garibaldi (2005) also ﬁnds that the volatility of the ﬂow of bank loan
destruction is larger than that of the ﬂow of bank loan creation.

2"In particular, the variances of the creation and the destruction rate are similar while the
average creation rate is larger than the average destruction rate.

21Note that, because credit reallocation is deﬁned as the sum of credit creation and credit
destruction (P US + NE G), the correlation between unemployment (U n) and credit real-

location can be written as

sd(P OS)
sd(SUM)

sd(NEG)

corr(SUM, U"): m

corr(P 0S, Un) + corr(N E G, Un).

In other words, this correlation equals a weighted average of the correlations of un-
employment with credit creation and destruction, where the weights are given by the ratios
of the standard deviation of creation and destruction relative to the standard deviation of
reallocation.

22If we restrict our attention to the period considered by Dell’Ariccia and Garibaldi (i.e.
1979-1999), the contemporaneous correlation of the gross reallocation of total debt with
unemployment is -0.4894.

23The results are robust to using policy instruments alternative to the federal funds rate.
In particular, we experimented with the term spread (10 year treasury constant maturity rate

- federal funds rate), the quality spread (3-month commercial paper rate - 3-month t-bill),

47

and the mix (short term debt /( short term debt+commercial paper».

24The credit ﬂows data have been seasonally adjusted with the Census X12 procedure.
See the data appendix for a detailed description of the series used in the VAR analysis.

25'We ﬁrst estimate the VAR in order to obtain B = {B3, B/(\I.)} and 1?). We then take
random draws with replacement from if, and use the bootstraped series 1711) together with
the parameter estimates E, and equation (10) to generate bootstraped data yfl). Then, the
VAR is ﬁtted to the bootstraped data to obtain BIT), and the impulse response function
‘1’ (W). We repeat this procedure 5000 times and then compute 99% (95%) conﬁdence
intervals based on the 0.5 (2.5) and the 99.5 (97.5) percentiles of the empirical distribution.

26See Den Haan (2000) for more on the properties of the statistic.

27A VAR speciﬁcation without the linear trend produces essentially identical results.

28Note that, because the VAR is estimated in growth rates, we accumulate the forecasted
changes in the series over the k-period.

2"’The standard errors are computed using the bootstrap method suggested in Den Haan

(2000).

48

Table 1.1: Aggregate Credit Flows

 

 

Panel A: Average

 

 

Type of Credit (years) POS NEG NET SUM EXC

 

Total (56-03) 1 1.30 5.97 5.33. 17.27 11.29
Long-Term (56-03) 11.17 6.16 5.01 17.34 12.01
Short-Term (56-03) 25.37 18.37 7.00 43.74 31.54

 

Total (59-69) I 1.26 2.72 8.54 13.98 5.44
Long-Term (59-69) 9.90 2.54 7.37 12.44 5.07
Short-Term (59-69) 32.95 16.69 16.26 49.64 33.38

 

Total (70-79) 8.00 5.25 2.75 13.25 8.99
Long-Term (70-79) 8.15 5.19 2.96 13.33 10.31
Short-Term (70-79) 20.70 19.11 1.59 39.81 26.49

 

Total (80-89) 13.03 8.10 4.94 21.13 15.31
Long-Term (80-89) 12.49 8.42 4.07 20.91 15.92
Short-Term (80-89) 27.88 19.68 8.20 47.57 35.60

 

Total (90-99) 1 1.59 7.55 4.04 19.15 14.37
Long-Term (90-99) 12.62 8.48 41.34 21.10 16.45
Short-Term (90-99) 25.57 19.75 5.81 45.32 31.77

 

 

Panel B: Coefficient of Variation

 

Type of Credit (years) POS NEG NET SUM EXC

 

Total (56-03) 39.70 42.48 95.88 30.10 41.25
Long-Term (56-03) 36.38 46.42 79.23 33.45 . 45.14
Short-Term (56-03) 37.87 32.04 187.59 20.66 21.74

 

 

 

 

 

Notes:

The table reports: in panel A, average annual credit ﬂows for the
sample period and for the sub-periods; in panel B, the coefficient
of variation of the credit ﬂows for the sample period.

49

Table 1.2: Comparison of Credit, Job, and Capital Flows

 

 

 

Panel A: Manufacturing in 1973-88

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Variable POS NEG NET SUM EXC
Total Credit 13.72 9.29 4.43 23.01 15.87
Long Term Credit 13.58 9.70 3.88 23.29 18.11
Short Term Credit 24.34 18.45 5.89 42.79 28.18
Jobs 9.1 10.3 -1.1 19.4 15.4
Physical Capital 11.2 10.0 1.2 21.2 18.9
Panel B: Aggregate in 1979-99

Variable POS NEG NET SUM EXC
Total Credit 5.08 3.30 1.78 8.38 6.29
Long Term Credit 5.59 3.71 1.88 9.29 6.06
Short Term Credit 15.03 12.79 2.23 27.82 22.78
Inter-Bank Loan Flows 3.18 1.42 1.76 4.6 2.69
Notes:

The table reports: in panel A, average annual credit ﬂows, average
job flows from Davis and Haltiwanger (1992) and average capital
ﬂows from Ramey and Shapiro (1998) in the 1973-88 period; in
panel B, average quarterly credit ﬂows and average interbank loan
ﬂows from Dell’Ariccia and Garibaldi (2005).

50

Table 1.3: Credit Flows in Manufacturing Industries

 

 

 

 

 

 

 

 

 

 

 

Panel A: Total Panel B: Long-Term
Industry POS NEG NET SUM EXC POS NEG NET SUM EXC
Food 14.79 9.15 5.64 23.93 15.01 14.57 8.66 5.92 23.23 15.63
Tobacco 1 1.52 7.47 4.05 19.00 5.99 12.05 7.48 4.57 19.54 6.76
Textiles 12.21 10.27 1.94 22.49 1 1.98 13.25 1 1.52 1.73 24.77 13.52
Apparel 16.69 10.34 6.34 27.03 14.42 16.96 1 1.00 5.96 27.96 16.36
Lumber 17.63 7.66 9.98 25.29 10.05 19.20 8.43 10.77 27.63 10.07
Fumiture 16.48 1 1.75 4.73 28.24 14.79 17.13 1 1.43 5.70 28.56 16.65
Paper 14.05 6.52 7.54 20.57 1 1.38 14.10 6.92 7.19 21.02 11.87
Printing 20.64 11.00 9.64 31.64 18.52 21.93 12.13 9.79 34.06 20.29
Chemicals 13.10 7.40 5.70 20.51 13.14 13.53 8.06 5.47 21.58 14.23
Petroleum 10.33 6.75 3.58 17.08 8.91 10.55 7.37 3.18 17.92 10.51
Rubber and 1 1.51 8.70 2.81 20.21 10.83 12.69 9.82 2.87 22.50 12.56
Plastics
Leather 16.40 12.82 3.58 29.22 1 1.71 16.02 12.93 3.09 28.96 I 1.49
Stone, Clay, 13.77 8.22 5.55 21.99 11.98 14.58 9.15 5.43 23.73 13.68
and Glass
Primary 10.67 9.24 1.43 19.90 12.95 11.35 10.16 1.20 21.51 14.57
Metal
Fabricated 13.55 10.00 3.55 23.56 13.45 14.46 10.44 4.03 24.90 15.17
Metal
Machinery 12.30 6.78 5.52 19.07 10.82 13.29 7.85 > 5.44 21.14 13.82
Electronic 15.27 9.14 6.13 24.40 14.60 15.51 9.64 5.87 25.15 16.45
Transportation 14.41 6.33 8.08 20.74 9.58 14.28 6.30 7.99 20.58 9.88
Instruments 18.15 9.61 8.54 27.75 15.59 19.20 1 1.06 8.14 30.25 17.95
Miscellaneous 17.52 13.37 4.15 30.90 17.34 18.60 14.26 4.34 32.86 18.39
Manufacturing 12.97 7.32 5.65 20.29 13.38 13.41 7.91 5.51 21.32 15.22
Notes:

The table reports average annual credit ﬂows for each two-digit manufacturing industry and for

the whole manufacturing sector. Panel A refers to total credit; Panel B to long-term credit.

51

Table 1.4: Credit Flows in Sales Quartiles (Aggregate Data)

 

 

 

Panel A: Total Credit

Panel B: Long-term Credit

 

 

 

 

 

 

 

 

Quartile POS NEG NET SUM EXC POS NEG NET SUM EXC
0-25% 33.39 21.02 12.36 54.41 34.02 36.29 24.05 12.45 60.34 36.92
25%-50% 24.02 13.15 10.88 37.17 24.54 25.24 14.63 10.61 39.86 27.29
50%-75% 17.16 8.22 8.94 25.38 15.44 17.14 8.70 8.44 25.84 16.74
75%-100% 10.62 5.05 5.57 15.67 9.60 10.46 5.19 5.27 15.65 10.10
Notes:

The table reports: in panel A (B), average annual total (long-term) credit ﬂows over the sample

period for quartiles of sales.

52

Table 1.5: Credit Flows in Census Regions (Aggregate Data)

 

 

Panel A: Total Credit

Panel B: Long-Term Credit

 

 

 

 

 

 

 

 

 

Region POS NEG NET SUM EXC POS NEG NET SUM EXC
New England 13.01 6.12 6.89 19.13 9.86 15.19 5.10 10.09 20.29 9.88
Middle Atlantic 10.87 6.77 4.10 17.64 11.21 12.94 5.29 7.64 18.23 9.87
South Atlantic 11.62 5.92 5.70 17.54 10.65 13.49 4.57 8.91 18.06 8.86
E. South Central 11.53 4.99 6.54 16.51 8.16 14.90 4.13 10.77 19.03 7.55
W. South Central 10.79 5.89 4.90 16.68 10.77 12.93 4.55 8.37 17.48 8.94
E. North Central 10.97 5.67 5.30 16.64 9.93 12.51 4.21 8.30 16.72 8.04
W. North Central 13.14 7.10 5.55 20.24 12.37 14.63 4.82 9.81 19.44 9.50
Mountain 13.13 6.20 6.93 19.33 11.57 15.52 5.46 10.07 20.98 10.24
Paciﬁc 1 1.94 6.75 5.20 18.69 12.36 14.32 5.29 9.02 19.61 10.39
Notes:

The table reports average annual credit ﬂows over the sample period for each census region.

Panel A refers to total credit; Panel B to long-term credit.

53

Table 1.6: Preperties of Idiosyncratic Credit Flows

 

 

Panel A: Total Credit

 

 

 

 

 

 

 

 

 

 

 

 

Share of reallocation variance due to Manufacturing Size (Quartiles) Region
1) Sectorial/aggregate effects 0.0398 0.0514 0.0275
ii) Idiosyncratic effects 0.8293 1.2521 1.1281
iii) Covariance term 0.1310 -0.3036 -0.1556
Correlation with Unemployment -0. 1617 -0.1889 -0. 1278
Panel B: Long-term Credit
Share of reallocation variance due to Manufacturing Size (Quartiles) Region
1) Sectorial/aggregate effects 0.0248 0.0225 0.0153
ii) Idiosyncratic effects 0.8559 1.1213 1.0854
iii) Covariance term 0.1 192 -0.1437 -0.1007
Correlation with Unemployment -0. 1225 -0.1578 -0.0988

 

 

 

 

Notes:

The table reports the variance decomposition of the ﬂows and the correlation of the

idiosyncratic reallocation with unemployment.

54

Table 1.7: Correlation with Unemployment (Aggregate Data)

 

 

Panel A: Full Sample

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Un.(-2) Un.(-1) Un. Un.(+1) Un.(+2)
POS 0.0131 -0.3570 -0.5470 -O.4746 -0.3560
NEG 0.3568 0.4906 0.5251 0.3879 1 0.2689
NET -0. 1689 -0.5616 -0.741 1 -0.6123 -0.4485
SUM 0.1832 -0.068 -0.2157 -0.2196 -0.1783 ‘
Panel B: 1959-69
POS -0.333 -0.7073 -0.7532 -0.6279 -0.37 17
NEG -0.0433 0.5423 0.8197 0.7122 0.8576
NET -0.2525 -0.7017 -0.7983 -0.6918 -0.5477
SUM -0.4455 -0.6940 -0.6544 -0.4858 -0.1088
Panel C: 1970-79
POS 0.3755 -0.3229 -0.8075 -0.2655 -0.1584
NEG -0.3410 0.1943 0.8996 0.5221 0.2303
NET 0.3848 -0.2729 -0.8833 -0.3942 -0.1944
SUM 0.0419 -0.2349 0.0368 0.3705 0.0172
Panel D: 1980-89
POS 0.1 152 -0.4676 -0.7394 -0.731 1 -0.7336
NEG 0.6762 0.8307 0.1654 -0.1853 -0.3555
NET -0.0349 -0.6318 -0.7651 -0.6913 -0.5472
SUM 0.2607 -0.2755 -0.6773 -0.7271 -0.77 17 .
Panel B: 1990-99
POS -0.5484 -0.8218 -0.9370 -0.7910 -0.1544
NEG -0.2209 0.1 158 0.3196 -0.3025 -0.4847
NET -0.5013 -0.8402 -0.9503 -0.7024 -0.0869
SUM -0.5823 -0.7871 -0.8911 -0.8499 -0.2162
Notes:

The table reports the correlation of credit ﬂows with unem-
ployment in the sample period (Panel A) and in the four sub-
samples (Panels BE).

55

Table 1.8: Variance Decomposition

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

8-step ahead 16-step ahead
Output ' Price Monetary Credit Output Price Monetary Credit
Shocks Shocks Shocks ﬂows Shocks Shocks Shocks ﬂows
Shocks ‘ Shocks

Panel A: Contribution to aggregate credit creation
Total 0.097 0.077 0.057 0.769 0.084 0.093 0.056 0.768
Long-term 0.066 0.057 0.051 0.826 0.062 0.106 0.044 0.788
Short-term 0.048 0.033 0.025 0.894 0.051 0.047 0.023 0.878
Panel B: Contribution to manufacturing credit creation
Total 0.258 0.025 0.086 0.632 0.294 0.032 0.076 0.599
Long-term 0.087 0.092 0.089 0.732 0.079 0.128 0.077 0.716
Short-term 0.221 0.016 0.066 0.697 0.221 0.059 0.060 0.660
Panel C: Contribution to aggregate credit destruction
Total 0.063 0.060 0.102 0.775 0.065 0.058 0.100 0.778
Long-term 0.007 0.106 0.017 0.869 0.014 0.180 0.033 0.774
Short-term 0.036 0.055 0.034 0.874 0.061 0.094 0.036 0.810
Panel D: Contribution to manufacturing credit destruction
Total 0.036 0. 144 0.049 0.771 0.079 0. 146 0.062 0.713
Long-term 0.070 0.089 0.092 0.749 0.068 0.096 0.088 0.749
Short-term 0.041 0.167 0.008 0.884 0.038 0.199 0.025 0.739
Notes:

The table reports the contribution of shocks to 8-step and 16-step ahead forecast error variance.

56

Figure 1.1: Credit Change, Credit Reallocation, and Unemployment

 

 

 

 

 

 

 

 

SE40 12
=
.835 r
a ,I 10
830. n. "
= I
N I
.225- :-
I

gm.
H
3,15.
2
£10.
0
ga-
° 0

6'

-10 0

 

—‘Na""&m—Unmmmm

Unemployment (‘5)

 

 

Figure 1.2: Response to a 1% increase in GDP growth

 

 

Aggregate Long-term Credit
2.5

 

Percentage

   

'2 4 6789101112

 

  

 

 

 

——Creation ------- Destruction —-——Net

   

 

9 indicates signiﬁcance at a 1% level 0 indicates signiﬁcance at a 5% level

 

 

  

 

 

 

 

 

 

 

 

 

 

   

 

 

 

 

 

 

Aggregate Short-term Credit
5
4 e l”“\
3 _ I” “'I1-’ \‘
o _ ’ ‘s
g 2 “'"X
C ‘11
§ 1 '
1’ ..............................
a o . . . . ___ -
_ 1‘-._ 2 ,.-'--.__4 5 8 9 1o 11 13 14 15 16
-1 “e".
-2 [— Creation ------- Destmaron - — — - Net |
0 indicates signiﬁcance at a 1% level 0 indicates signiﬁcance at a 5% level
Aggregate Total Credit
3.5
3 ‘ ,
25 - ,r‘t‘
/' i
c 2 4
n 1.
g 1.5 " ”\\ '3' ‘V”\
s 1 -
“- 0.5 -
o 1"‘~--"'3" 4"‘i--.--"'"é"" ---- ' "'é""9' 1o 11'12 '13- 14 15 16
-0.5

 

 

 

 

Figure 1.2 (continued)

 

Manufacturing Long-term Credit

 

Percentage

 

 

 

 

 

 

— Creation ------- Destruction — — — - Net

 

 

0 indicates signiﬁcance at a 1% level 0 indicates signiﬁcance at a 5% level

 

Manufacturing Short-term Credit

 

 

 

 

 

o
u
5
r:
o
2 .
a I. ..-
_2‘1 2 31x43} 6 7 8 9 10 11 12 13 14 1516
-3. ‘3"
.4

 

— Creation ------- Destmction - - — - Net

 

 

0 indicates signiﬁcance at a 1% level 0 indicates signiﬁcance at a 5% level

 

 

Manufacturing Total Credit

 

Percentage
52
,u"
r"
r
/‘

   

 

 

 

 

 

 

 

1-

0.5- __ _

o 1 2'3"“'---""5"6' 7'8 '9 '10'11'12'13'14'15‘1‘6“
-0.5

 

—Creation ------- Destruction — - - - Net

   

 

59

 

Figure 1.3: Correlation with GDP

 

 

Aggregate Total Credit

 

 

\
e O s
o D 4.. ‘
I

 

1‘3“; ? '7' I9' '11‘ '13' '15‘ '17' :«5'21 25 27 m 31 33
'02-1 .‘e .' '.. ..0

Percentage

 

 

 

 

|____Ctealion ...... Destruction I

 

 

 

 

Manufacturing Total Credit

 

 

 

 

0.7
0.6
0.5.1
0.4
36.3
8
{:02
a 0.1 l.. .- .0.
o j—TIZﬁI I I I I I I T—I—J‘g.I—1—I_I I I I I I TIITI I I I I I I II II
‘01135‘..791113;‘17 21235 $313537

‘0

8

 

.5
co

 

 

 

1— Creation """ Destructio]

 

 

Chapter 2: The Effect of Comovement
Between Firms on their Liquidation Val-
ues: Do Firms Attract More Trade Credit
when their Sales Are Less Correlated?

1 Introduction

Why do ﬁrms in the same industry get ﬁnanced differently? Is there an optimal capital
structure that ﬁrms strive to achieve? Answers to these key questions have eluded re-
searchers for decades.1 Despite vast research efforts, there is no commonly accepted uni-
fying theory to explain ﬁrm capital structure. A recent attempt to reconcile the available
theories and empirical evidence was made by Frank and Goyal (2007). As Frank and Goyal
(2007) illustrates, the level of disagreement and confusion over the precise role of the de-
terminants of capital structure can be seen in the contradictory implications of two of the
most cited studies on the determinants of capital structure, Titrnan and Wessels (1988) and
Harris and Raviv (1991). According to the empirical results obtained by Titrnan and Wes-
sels (1988), leverage decreases with ﬁrm size and leverage is not affected by ﬁrm’s growth
opportunities, the size of tangible assets and non-debt tax shields, while Harris and Raviv
(1991) report in their survey of the available empirical evidence the apparent consensus to
be that leverage increases with all of these determinants. In other words, despite decades
of research since Modigliani and Miller (1958), no consensus has been reached in this area
of ﬁnance.

Numerous determinants of capital structure have been suggested over time. More re-

61

cent empirical inquiries include for instance the role of taxes (see for instance Graham
(2000)), technology (see for instance MacKay (2003)), stock returns (see for instance
Welch (2004)), and testing the static trade-off theories versus the pecking-order theories
(see for instance Fama and French (2002)).

In this chapter I focus on a new promising line of research that takes a more detailed
look at the impact of industry characteristics on the ﬁnancial decisions of a single ﬁrm.
I identify comovement among ﬁrms measured by the correlation of sales growth rates as
a new determinant of ﬁrm ﬁnance structure. I provide empirical evidence for this newly
identiﬁed link. The results suggest a signiﬁcant, even though somewhat small, impact of
comovement on leverage.

Traditionally, the industry effects have been dealt with by simply including industry
dummies (See for instance MacKay (2003) for an example). However, recently, two pa-
pers have uncovered evidence suggesting a much more complicated relationship between
industry and leverage. MacKay and Phillips (2005) ﬁnd that in addition to overall industry
effects the position of a ﬁrm within its industry also plays an important role in determining
its ﬁnancial structure.2 Using simultaneous-equation regressions to mitigate endogeneity
problems, they ﬁnd that leverage responds to how close the ﬁrm’s capital-labor ratio is to
the industry median, whether the ﬁrm is new, established, or exiting, and to the actions
of other ﬁrms in the industry. Their evidence is consistent with the competitive industry
equilibrium models of Maksimovic and Zechner (1991) and Fries, Miller, and Perraudin
(1997). Both of these articles imply that ﬁrm decisions are affected by its industry posi-
tion; Maksimovic and Zechner (1991) express ﬁrm industry position in terms of how close

its technology is to the industry peers, while Fries, Miller, and Perraudin (1997) use ﬁrm

62

status as an entrant, incumbent, or exiting ﬁrm as an industry position indicator. In a study
related to MacKay and Phillips (2005), Almazan and Molina (2005) investigate the rela-
tionship between industry characteristics and the dispersion of leverage ratios among ﬁrms
in an industry. They ﬁnd more dispersed leverage ratios in. industries with greater liquid-
ity of assets, and in industries with higher proportion of older ﬁrms with suﬂicient growth
opportunities.

These two papers were followed by more recent studies. Almazan et al. (2007) conﬁrm
that ﬁrm’s position in an industry matters. They study physical location of ﬁrms and ﬁnd
that ﬁrms located in industry clusters take on less debt than ﬁrms located in rural areas in
order to preserve the ability to take advantage of possible acquisition opportunities, which
are more abundant within industry clusters. The presence of growth opportunities is the
reason for lower debt ratios. In a related study, Loughran (2008) reports the same ﬁnding:
ﬁrms located in rural areas take on more debt than ﬁrms located in clusters. However,
here the cause for higher leverage in rural areas is information asymmetry. Firms in rural
areas face greater information asymmetry problems due to a greater distance from outside
investors, which makes it more costly to issue equity, leading them instead to take on more
debt.

This chapter introduces a new way of looking at the industry position of a ﬁrm. Rather
than considering physical location as a measure of the relationship with the industry peers,
this chapter investigates the role of comovement among ﬁrms, measured by correlation of
sales growth rates, in the way ﬁrms get ﬁnanced. The channel through which comovement
affects ﬁrm ﬁnancing is via its liquidation value. The link between industry characteristics

and liquidation values was ﬁrst modeled in the context of an overall industry equilibrium

63

by Shleifer and Vishny (1992).3 Shleifer and Vishny (1992) put forth a theoretical model
where the sale value of ﬁrm assets depends on how similar its ﬁnances are to other ﬁrms
in the industry at the moment of liquidation. If, for instance, all ﬁrms in the industry
respond in a similar way to outside shocks, then it is likely that when a ﬁrm is liquidated
its assets will be acquired by an outside investor due to the lack of available cash on the
part of its industry peers that are likely to be in ﬁnancial distress at the same time. Since
outside investors lack the same information and ability to utilize the assets as compared
to the industry competitors, the liquidated assets are sold for a fraction of the value in the
second best use. Recognizing this ex ante, potential creditors will be more reluctant to
extend credit due to a lower collateral value to ﬁrms in industries with a higher degree of
comovement in terms of the likelihood of being ﬁnancially distressed at the same time.
This link has been gaining momentum in the ﬁnance literature. For instance Hege and
Hennessy (2007) present a model in which incumbent ﬁrms deter potential entrants by
taking actions that diminish the sale or liquidation value of entrants’ assets in the case
of entry. Similarly as in Shleifer and Vishny, the liquidation value of assets depends on
whether they are acquired by an industry peer or someone outside the industry. In Hege and
Hennessy (2007), incumbents issue additional debt creating debt overhang, which prevents
them from acquiring new assets. As a result, the collateral value of potential entrants’ assets
falls, which makes it more difﬁcult to ﬁnance the entry, and thus may be enough to prevent
entry of new competitors.

This chapter provides empirical evidence consistent with the model of Shleifer and
Vishny (1992). The baseline estimate suggests that an increase in the comovement among

ﬁrms in the industry corresponding to a change from the 25th to the 75th percentile in the

64

sample will lead to a decrease in the accounts payable to assets ratio by around 1.5%, even
after controlling for the inventory stock.4

Other empirical studies have found mixed evidence concerning the role of liquidation
values in corporate ﬁnance. For example Benmelech et a1. (2005) exploits commercial
zoning regulations to ﬁnd that commercial loans associated with assets of higher liquida-
tion value tend to be larger and come with a lower interest rate. Pulvino (1998) applies
the model of Shleifer and Vishny (1992) to the commercial airline industry and provides
evidence that the prices that ﬁnancially distressed airlines receive for their assets are below
their fundamental values, which he attributes to a temporary illiquidity caused by the in-
ability of other airlines to raise enough funds to buy these assets at the time.5 On the other
hand MacKay (2003) ﬁnds that, depending on whether a ﬁrm can commit not to substitute
assets, higher liquidity of assets could affect the ﬁrm’s debt capacity in both directions. He
ﬁnds on the one hand that, in production, higher redeployability of assets leads to a lower
debt capacity due to the inability to commit not to substitute these assets for riskier ones.
On the other hand, in the case of assets associated with investment opportunities, since it is
possible for the ﬁrm to commit, greater ﬂexibility leads to greater willingness of the lenders
to extend credit.

My work in this chapter is motivated by the treatment in Shleifer and Vishny (1992).
The main idea is when ﬁrms’ performances within an industry move in unison, it is likely
that when one ﬁrm goes bankrupt, its industry peers also experience ﬁnancial diﬂiculties
at the same time. Therefore, when comovement is high, for instance because ﬁrms in the
industry are regularly exposed to common shocks, assets of a fallen ﬁrm will be acquired

by an industry outsider, who will pay less than the value in the second best use. This

65

may happen, as explained in Shleifer and Vishny (1992), because the ﬁrrn’s industry peers
which would normally be in the best position to use the assets will ﬁnd themselves in
ﬁnancial distress and will be unable to borrow due to agency cost.6 The industry outsider,
due to the lack of expertise and information, will then underpay the value of the assets in
its second best use. Thus, when comovement of ﬁrms within a sector is high, using the
results of Shleifer and Vishny (1992), the liquidation value of an average ﬁrm will be low
and vice-versa. Comovement, then, will aﬂ‘ect the sale value of ﬁrms’ assets, thus aﬂ'ecting
their collateral value and debt capacity. I hypothesize that higher comovement lowers the
collateral value thus leading to a lower payment to the lenders in the case of default and
their decreased willingness to extend credit. The results provided in this chapter support
this hypothesis.

In this chapter I focus on the part of corporate debt that originates from other ﬁrms,
namely trade credit. Trade credit is a good starting point in this line of research because
the amount extended to a ﬁrm is likely to depend more heavily on the ﬁrm’s liquidation
value relative to other forms of credit. It is natural for the suppliers of trade credit to factor
in their ability to repossess and resell the assets before they provide them to the recipient.7
For example Boeing will consider its ability to resell an aircraft before it is provided to an
airline via trade credit. If airlines comove together, Boeing will recognize that its ability
to resell the aircraft at its fundamental value is severely limited because other airlines are
likely to experience ﬁnancial diﬂiculties at the same time. Indeed, Pulvino (1998) does ﬁnd
that airlines in ﬁnancial distress do sell their assets at a signiﬁcant discount. Is one of the
reasons for this kind of discount a higher comovement of airlines causing the demand for

these assets by other airlines to drop precisely when an airline is likely to attempt to sell

66

them?

In addition to the sale of assets being more natural for the suppliers of trade credit than
for banks, there is another reason why it makes sense to focus on trade credit. Whereas in
both the case of bank loans and the case of corporate bonds the issuers frequently account
for the difference in the default risk by adjusting the rate of interest associated with the loan,
the issuers of trade credit commonly keep the price terms the same for all ﬁrms within the
same industry, while instead adjusting the amount of trade credit.8 Therefore, with trade
credit I can simply focus on the amount of credit extended without having to track the terms
also. Indeed, I ﬁnd evidence that the amount of trade credit does depend on the degree of
comovement between ﬁrms, even after controlling for the stock of inventories which is
commonly used as a proxy for the collateral value.9 I leave it to further research to extend
this analysis beyond the area of trade credit.

The rest of this chapter is organized in the following way. Section 2 describes the
determinants of trade credit and the data. Section 3 presents the empirical model. Section

4 describes the results, and section 5 concludes.

2 Measurement

First, I discuss measuring the amount of trade credit extended to a single ﬁrm. I then pro-
ceed to discuss the explanatory variables, starting with comovement, the credit determinant
of the main interest. To this key variable I add the determinants of trade credit studied by
Petersen and Rajan (1997), who provided an extensive overview of the evidence on trade

credit, while adding evidence of their own. To this set of controls, I further add inventories

67

and market interest rate, which were recently also found to signiﬁcantly inﬂuence trade
credit (See for instance Choi and Kim (2005)). Inventories are of special interest here be-
cause higher inventories make it easier to liquidate a ﬁrm’s assets and therefore increase
the ﬁrm’s liquidation value. Thus, since I control for the effect of inventories in my regres-
sions, the coeﬂicient on comovement should indicate its true effect on the collateral value,

instead of picking up the effect of inventories.

2.1 Measuring Trade Credit

There are two balance sheet trade credit items. Accounts receivable measures the amount
of trade credit extended by the corresponding ﬁrm, while accounts payable measures the
amount of trade credit extended to the ﬁrm. In this paper, I investigate the relationship
between the level of comovement of ﬁrms in an industry and the amount of trade credit
extended to a single ﬁrm. Accounts payable is therefore the proper dependent variable to

be used in the empirical model that estimates the role of comovement in ﬁrm ﬁnancing.

2.2 Comovement Measure

I deﬁne comovement at an industry level as in Guiso and Minetti (2004), which is, in spirit,
very similar to an alternative deﬁnition of comovement at a ﬁrm level found in Comin and
Mulani (2006).10 Both measures are based on the growth rate of sales, however the industry
level of comovement is more appropriate in my analysis, because it is more consistent with
the reasoning behind the model of Shleifer and Vishny (1992). The asset illiquidity effect

delivered by the model of Shleifer and Vishny (1992) rests on the assumption that ﬁrms

68

in the industry in question are hit by common shocks.ll In fact, the effect completely
disappears if the shock that causes ﬁrm liquidation is idiosyncratic. Indeed, as Shleifer and
Vishny state in the abstract, they use the model "to explain variation in debt capacity across
industries and over the business cycle,...", and use the airline industry as an example of an
industry where asset liquidation values are well below their fundamental values because
airlines are affected in a similar way by the business cycle. In other words, I expect the
debt capacity of ﬁrms to be higher in industries where the asset illiquidity effect is present,
because ﬁrms in these industries comove together. This is the link between industry and
asset liquidity that I investigate in this chapter.

My comovement variable therefore measures the degree of comovement of all ﬁrms
within an industry sector. As stated in the introduction, 1 hypothesize that ﬁrms with a
higher industry level comovement value, ceteris paribus, should attract a lower supply of
trade credit than ﬁrms in other industries with lower levels of industry comovement.

Following Guiso and Minetti (2004) I deﬁne the sectorial comovement as the degree to
which ﬁrms’ sales comove together within a single industry. I measure this comovement in
the same way as in Guiso and Minetti (2004), by computing R-squared from a regression
of real sales growth rate on time dummy variables, using time periods from t = —4 to
t = +5. Using time period windows in constructing the comovement variable allows for
the measure to vary over time.

C omovementL, = R - Squared from the following regression

gm = a + D,6 + um, for each i in I (1)

69

where g,,, is the growth rate of real sales, a is the intercept, D, is a vector of time
dummy variables, um is the error term, I indexes an industry sector, 1' ﬁrm, t time period,
and where t = {—4, ..., +5}.

The reasoning behind this measure is that if ﬁrms within an industry exhibit a higher
level of comovement, then time dummy variables should explain most of the variation in
the growth rate of real sales. In the extreme, if all ﬁrms in an industry grow their sales by
the same percentage, then R-squared will equal one. In the extreme opposite scenario, if

ﬁrms’ sales in an industry are uncorrelated, then R-squared will equal zero.

2.3 Main Independent Variables

Following Petersen and Rajan (1997) I include measures of ﬁrm size, proﬁtability, access
to internal funds, rate of growth, ﬁrm age, and current assets. A ﬁrll description of the
measures used in the estimation is found in table 2.1.

Because of the data limitations stemming from working with Compustat data, I am
unable to control for the relationship between ﬁrms and ﬁnancial institutions. Petersen and
Rajan (1997) do not ﬁnd any evidence that relationship with ﬁnancial institutions aﬁ'ects the
amount of trade credit supplied to the ﬁrm, however they do ﬁnd some evidence suggesting
that ﬁrms with closer ties to lenders demand less trade credit.

Due to data limitations, 1 am also unable to include the price of trade credit. This
absence should not affect the results in any signiﬁcant way for two reasons. First, as stated
previously, credit terms within an industry are rarely changed. Instead, trade credit may

be reduced or withheld entirely from ﬁrms with a higher risk of default (See Petersen and

70

Rajan (1997) and Smith (1987)). Second, the inclusion of industry dummies should account
for the differences in pricing among industry sectors.12

Another potential disadvantage stemming from working with Compustat data is ab-
sence of smaller ﬁrms, because Compustat consists of data collected exclusively from pub-
licly traded ﬁrms. There is, however, some evidence suggesting that smaller ﬁrms behave in
a way similar to large ﬁrms in regard to obtaining trade credit. For instance Nilsen (2002)
evidenced that both small and large ﬁrms use more accounts payable during recessions,
even though the increase is somewhat larger for large ﬁrms. Further, Nilsen (2002) argues
that the size of a ﬁrm is not decisive for credit quality of a ﬁrm as a borrower, and shows
that even among large ﬁrms the ﬁrms with a lower credit rating use much more trade credit
during recessions, while large ﬁrms with a higher credit rating use other forms of credit,
which suggests that credit constraints play important role even among large ﬁrms. Thus, it
appears that limiting data to a pool of larger ﬁrms may still yield a mix of ﬁrms with the
borrowing behavior characteristics representative of the population of ﬁrms in the overall
economy.

The last potentially problematic omission I am aware of is the omission of the percent
of inventory in terms of ﬁnished goods. Again, this is due to data limitations. 1 do however
control for the total inventory stock, which should mitigate this problem somewhat.

Firm Size and Age. Firm size is typically used as a control variable in models where
accounts receivable is the dependent variable. While the ﬁnancial assistance view suggests
that larger ﬁrms extend more trade credit, the quality veriﬁcation theory (see Long, Malitz
and Ravid (1993)) instead implies that it is smaller ﬁrms which extend more trade credit. I

am unable to offer evidence on this point since I view the ﬁrms in my data set as the users

71

of trade credit rather than suppliers.

The effect of ﬁrm size and age on accounts payable is also unclear. On the one hand,
it is commonly thought that investment opportunities are negatively correlated with both
size and age, which means that demand for trade credit should decline with size and age
as investment opportunities shrink. On the other hand, larger and older ﬁrms are usually
associated with lower risk of default (see for instance Diamond (1989)), implying that the
supply of trade credit to these ﬁrms may be higher. I use the book value of assets to measure
ﬁrm size, which is a standard approach, used for instance by MacKay (2003). I approximate
ﬁrm age by the number of quarters it has been present in the Compustat database.l3

Proﬁtability. It is commonly believed that proﬁtable ﬁrms have more investment op-
portunities, and therefore need more ﬁnancing. However, high proﬁtability also attracts
lenders, and therefore a highly proﬁtable ﬁrm should have no difficulty in obtaining credit
elsewhere, and according to the pecking order theory (see for example Graham and Harvey
(2001)) should thus demand less trade credit.

Availability of Internal Funds. Greater availability of internal funds raises the cred-
itworthiness of the ﬁrm and may increase the amount of trade credit extended to the ﬁrm.
On the other hand, according to the pecking order theory, demand for trade credit by the
ﬁrm will decrease with growing sources of internal funding. 1 follow Petersen and Rajan
(1997) in using net income not only as a measure of proﬁtability, but also as a proxy for
internal funding availability rather than the more standard measure of cash ﬂow which adds
depreciation to net income.14 For added robustness, I also use another measure of internal
sources of funding, retained earnings, which measures the cumulative earnings that have

not been paid out as dividends.

72

Rate of Growth. Growing ﬁrms tend to have more investment opportunities and there-
fore should demand more trade credit. Following Petersen and Rajan (1997) I use change
in sales standardized by assets to proxy for how fast the ﬁrm is growing.

Current Assets. 1 can partially control for the demand for short term credit by in-
cluding a measure of current assets in the model. As established in earlier studies (see for
instance Diamond (1991)), ﬁrms ﬁnance current assets almost exclusively with short term
credit. Thus a higher share of current assets in relation to the total assets should result in a

higher demand for trade credit.

2.4 Other Controls

Inventory Stock. On the one hand, inventory stock can be positively correlated with
accounts payable due to a higher liquidation value. Since it is generally easier to sell inven-
tories than other assets, the ﬁrm’s liquidation value will be higher with higher inventories
and therefore the ﬁrm as a recipient of trade credit may be able to attract more credit as it
may be perceived to have a higher collateral value and thus represent a lower risk to the
lender.15

On the other hand, as mentioned in the introduction, assets with higher liquidation
values are more liquid and therefore make it easier for the recipient to substitute other,
more riskier assets.16 Therefore, a higher collateral value could represent a greater, not
lower, risk to the lender.

Market Interest Rate. The transaction cost theory suggests that both the demand

for (accounts payable) and the supply of (accounts receivable) trade credit should increase

73

when the market interest rate rises. Ferris (1981) shows in his model that trade credit
reduces the uncertainty of money ﬂows in case of uncertain delivery dates. If delivery
dates are uncertain, trade credit has an advantage over bank loans in that it will reduce
this uncertainty jointly, not separately, for the recipient and. the supplier of trade credit. In
Ferris (1981), trade credit is optimal if it is not more than twice as expensive as credit from
a bank. Since a higher market interest rate makes bank credit more expensive, more ﬁrms
will ﬁnd it optimal to use trade credit instead when the market interest rate rises.
Alternatively, Norrbin and Reffett (1995) also provide a rationale for a positive rela-
tionship between accounts payable and nominal market interest rate. Their cash-in-advance
model suggests that an increase in the nominal interest rate leads to an increase in the ratio

of real trade credit to real balances. They also support this result with empirical evidence.

2.5 Data

I employ quarterly data retrieved from the Compustat database.17 The upside of using
Compustat is the large number of ﬁrms covered. The downside of using Compustat is the
fact that smaller ﬁrms do not typically enter the database.18

Quarterly data ranging from 1975:l to 2005 :4 are available from Compustat for the ﬁrm
balance sheet variables outlined previously.‘9 The database allows me to distinguish among
industry sectors at the 4-digit SIC level. As is common in the studies of corporate ﬁnance
(see for instance Fama and French (2002), Molina and Preve (2005), or Opler and Titman
(1994)) all observations corresponding to the banking, insurance, and real estate industries

(SIC 6000 to SIC 6999) are removed because the operations of these ﬁrms include the

74

provision of ﬁnancial services, and are therefore inappropriate for a study of corporate ﬁ-
nance.20 Also, following Petersen and Rajan (1997), services industries (SIC 7000 to 8999)
are eliminated as well. Also in accordance with the other studies”, observations with zero
or missing accounts payable, sales or assets are deleted aswell.22 There remain 374,379
ﬁrm-quarter observations that are used in the main analysis, spanning 19762 to 2004:3.23
On average there are around 3284 ﬁrm observations in each year quarter, and within each
year quarter around 10 ﬁrm observations per each of the 340 4-digit SIC industry sectors.
Table 2.1 contains description of all of the variables used in the estimation of the em-

pirical model that is outlined in the next section.

3 The Empirical Model

Following Petersen and Rajan (1997), 1 normalize accounts payable, the dependent vari-
able, by total assets.24 Assets are used to scale accounts payable because the ratio of
liabilities to assets represents a common measure of leverage. Since accounts payable is
a liability, the ratio of accounts payable to assets can be thought of as corresponding to
ﬁrms’ balance sheet management in terms of the relationship between assets and liabilities.
It then makes sense to use assets also for scaling the dependent variables.

Then, the empirical model of interest for ﬁrm i at time t is:

AP,,,

__ = X,,,,61+..+X,-,,_4ﬂ4+y lComooe,,,+..+y 4Comove,,r—4+D,6+D1w+a,-+u,~,,,
Assets“

(2)

75

where A P,,, is accounts payable of ﬁrm i at time t (Compustat data item 46), Assets,,,
is total assets (item 44), Comovem is as deﬁned in the previous section, D, is a vector
of time (year and quarter) dummy variables, D1 is a vector of industry dummy variables
at the 4-digit SIC level, and X,,, is a vector of other explanatory variables described in
the previous section. Four lags of independent variables are included as is customary in
regressions with quarterly data.25 The error term u,,, is complemented by a time invari-
ant variable unique to each ﬁrm, denoted 0,. This speciﬁcation allows for two estimation
techniques, depending on the assumptions made about a,. At ﬁrst, I use the pooled OLS
estimator under the assumption that a,- is uncorrelated with the other explanatory variables
and therefore can be safely left as a part of the error term. Since the correctness of the
assumption underlying the pooled OLS estimation can be easily questioned, I also estimate
the model by the ﬁxed effects estimator, allowing for arbitrary correlation between a, and

the regressors.

3.1 Tariff as Instrumental Variable

Next, I relax the exogeneity assumption on Comovement by introducing an instrument, the
degree of protection from foreign competition. I measure the degree of protection by aver-
age tariﬁ‘ rates available by 4-digit SIC categories.26 Firms in industries that are protected
by tariffs from foreign competition will be shielded to a certain extent from competitive
pressures. Thus, ﬁrms in such industries will be less likely, in any given period, than oth-
erwise similar ﬁrms that are unprotected, to encounter losses or exhibit a decline in sales.

In other words, ﬁrms in industries protected by tariffs should exhibit more steady perfor-

76

mances in terms of proﬁts and sales. As a result, there should also be less disparity in
performance among ﬁrms in a protected industry, with the disparity in performance declin-
ing with increasing tariff rate. Thus, I expect a positive correlation between tariﬂ' rates and
comovement. '

Any instrument must fulﬁll two conditions. First, it must be correlated with the en-
dogenous variable that it is instrumenting, which 1 conﬁrmed by regressing the comove-
ment variable on all exogenous regressors in the model, including the instrument, ﬁnding
the coeﬂicient on tariff positive and signiﬁcant.27 Second, the tariff rates must be uncorre-
lated with the error term. One cannot test the exogeneity of the instrument, however, since
tariff is a policy variable, it is reasonable to assume that it is uncorrelated with the industry-
endogenous elements left in the error term. For this reason, tariff rates as an instrument are
more appealing than other measures of the degree of industry concentration, such as the
Herﬁndahl index, which clearly violates the instrument exogeneity assumption.

As described in the next section, the magnitude and the direction of the estimated effect
of comovement on the supply of trade credit is preserved when using the tariff rates as
an instrument. However, the signiﬁcance is lost. I conjecture that this is due to a loss in
precision of the estimate caused on the one hand by less than perfect correlation between
the instrument and the instrumented variable, and on the other by the loss of more than
half of the observations when using the tariff data. Alternatively, the change in signiﬁcance
could be caused by misspeciﬁcation of the earlier models, if the comovement variable is
endogenous. If this is the case then the true effect of comovement is much less signiﬁcant

than suggested by the results from the non-instrumented regressions.

77

4 Results

4.1 Summary Statistics

Summary statistics for the main variables are provided in Table 2.2 for the full sample,
and also for ten year time periods. The median value of accounts payable standardized by
total assets, the dependent variable, is 0.083. It seems to be fairly stable over time. The
independent variable of most interest in this paper, comovement, is almost twice as high on
average in the 1976-1985 period, the ﬁrst 10 years of the sample, than in the last two ten
year periods. A similar pattern can be observed for the book value of assets, the variable
indicating the size of ﬁrms. Even though the average size stays about the same, the median
ﬁrm size appears to be about twice as large in the 1976-1985 period than in the 1985-1995
period. One may be tempted to think that the greater proportion of smaller ﬁrms caused the
median comovement to come down during 1985-1995, however the results in section 4.2
suggest otherwise. The median proﬁts divided by total assets decline from 0.014 in the ﬁrst
decade to 0.009 and 0.007 in the last two decades of the sample, with the average proﬁts
turning negative during these two decades. Change in real sales and current assets less cash
seem to be pretty stable in proportion to assets over time. The summary statistics on age
are not of much interest, because of the nature of approximating this variable. Naturally,
the average length of time since entering Compustat is growing with time.

Summary statistics for comovement are shown separately for each 2-digit SIC industry
group in Table 2.3. The median value is the lowest in sector 13, oil and gas extraction,
only 0.004. The highest median value, 0.608, is found in group 53, general merchandise

stores. If we narrow our attention to manufacturing, the median comovement value ranges

78

from 0.023 in electronic equipment (group 36) to 0.305 in stone and glass (group 32). If
we assume that the ﬁrms in the two industries are otherwise similar, then, according to the
hypothesized effect of comovement on the supply of credit, we should observe a higher
accounts payable to assets ratio in the electronics industry, because of higher average liqui-
dation values, than in stone and glass. Indeed, looking at the sample summary statistics, we
ﬁnd that the median accounts payable to assets ratio is higher for electronics (0.085) than

for stone and glass (0.078), even though the difference does not appear to be very large.

4.2 Results
4.2.1 Pooled OLS and Fixed Effects

Table 2.4 shows the results for the non-instrumented pooled OLS and ﬁxed effects.28 The
ﬁrst row provides evidence that a ﬁrm located in an industry sector with a higher degree of
comovement can expect to see a lower amount of credit supplied to it from other ﬁrms. Let
us look at the results in more detail.

First, observe that the results are very similar across the two columns. Therefore, I
will only describe in detail the results obtained by using the preferred ﬁxed effects model.
This method is superior to pooled OLS because, as mentioned before, it corrects for the
presence of an idiosyncratic unobserved ﬁrm effect, which is very likely to be present in
the panel data obtained from Compustat. Moreover, the ﬁxed effects estimation technique
is widespread in ﬁnance studies, thus my results from the ﬁxed effects estimation are more
29

comparable to other studies in this area.

Consider then the results in the ﬁxed effects column of Table 2.4. It is estimated that

79

an increase in the comovement variable by one will lower the ratio of accounts payable
to assets roughly by 10.3%. Put in a more meaningful way, a ceteris paribus increase
in the comovement among ﬁrms in an industry sector from the 25th to the 75th sample
percentile (0.023 to 0.163) will lower the accounts payable to assets ratio for all ﬁrms in
the sector by around 1.5%. This means for instance that the median sample ﬁrm, taken by
the accounts payable to assets ratio, would experience a decline in the ratio from around
0.083 to 0.08175. This effect, despite being somewhat small is nevertheless statistically
signiﬁcant at the 5% level of signiﬁcance.30

The results suggest that the role of inventories is much bigger. Increasing the inventories
to total assets ratio from the 25th to the 75th sample percentile would lead to a 28.5%
increase in the accounts payable to assets ratio. The direction of this effect is consistent
with Shleifer and Vishny’s theory of the liquidation value. The collateral value increases
with increasing inventories, which leads to an increased amount of trade credit supplied to
the ﬁrm.

The coeﬂicient on retained earnings is negative, and insigniﬁcant.31 It is also very small
relative to the other estimated coefficients. The small magnitude comes as no surprise, since
there could very well be two effects cancelling out each other. The result provides some
support for the pecking order theory, which suggests a negative relationship between a
ﬁrm’s retained earnings and its demand for trade credit. Even though this effect seems to
dominate it appears to be partly offset by the positive effect of lower risk on the supply of
trade credit, which is also associated with higher retained earnings.

There does not seem to be much inﬂuence on the amount of trade credit from the size

and age of the potential recipient. Accounts payable appear to decline with increasing size,

80

however the estimate lacks statistical signiﬁcance. Again, it is possible that there are two
effects present working in opposite directions. Lower demand for credit due to declining
investment opportunities for older and larger ﬁrms can be offset by an increased supply of
credit caused by a lower default risk commonly attributed to more established ﬁrms.

The estimate on the effect of market interest rate, represented in my regressions by
the 3-month treasury bill, is perhaps the most consistent across speciﬁcations in my analy-
sis. The coeﬂicient on the T-Bill is highly signiﬁcant and appears to be somewhat small,
however, only until one recognizes the large variation in this interest rate over the sample
period. Going from the 25th to the 75th sample percentile in this case means going from
4.41% to 7.46% interest rate, which is estimated to increase the accounts payable to assets
ratio by roughly 3.7%. This ﬁnding is consistent with the transaction cost theory of trade
credit.

The coeﬂicient on change in sales suggests that faster growing ﬁrms demand more
trade credit as their investment opportunities grow faster. This effect is smaller for ﬁrms
with negative growth rates, nevertheless still positive. Petersen and Rajan (1997) also ﬁnd
this effect positive for ﬁrms with growing sales, however for ﬁrms with declining sales they
ﬁnd this effect negative.

Increasing proﬁt causes ﬁrms to demand less trade credit due to higher cash ﬂows, again
according to the pecking order theory. This proposition is supported strongly by my results,
even though the size of the effect is not very large. Increasing the net proﬁt ratio (when
proﬁts are positive), going again from the 25th to the 75th sample percentile (from -0.005
to 0.021) decreases the payables to assets ratio by only around 0.7%.

The coefﬁcient on the fraction of total assets in the form of current assets (excluding

81

cash) conﬁrms strongly that ﬁrms with more current assets use more trade credit.

4.2.2 Instrumented Pooled OLS and Fixed Effects

Table 2.5 lists results from the instrumented pooled OLS and the instrumented ﬁxed effects
regressions in the last two columns.32 For comparison, the second and third columns of the
table present results from the non-instrumented regressions, however this time re-estimated
on the estimation sample from the instrumented versions. Note that more than half of the
observations disappear because of missing data on tariffs.

Comparing table 2.5 with the previously discussed table 2.4, the values of the estimated
coeﬂicients seem to be fairly robust across the four estimation techniques used, while their
signiﬁcance is reduced greatly.33 As discussed earlier, when describing the instrument,
there are two possible explanations for the loss in signiﬁcance. On the one hand, the miss-
ing data could have resulted in less precision in the estimation. This explanation may seem
particularly plausible, because even the non-instrumented ﬁxed effects model provides an
insigniﬁcant coeﬂicient on comovement with the reduced sample (see the ﬁrst row of the
ﬁxed effects column in table 2.5). Alternatively, if comovement is endogenous, then the re-
sults from table 2.4 are invalid and the evidence that comovement affects the credit position

of a ﬁrm is very weak.

5 Conclusion

In this chapter, I tested the effect of comovement of ﬁrms within an industry via its effect

on the liquidation value of ﬁrrns’ assets, on the amount of trade credit supplied to a ﬁrm

82

inside the industry. In the baseline speciﬁcation, 1 found evidence that higher comovement
leads to a lower amount of trade credit extended to the ﬁrm. Even though the effect of
comovement on accounts payable is much smaller than the effect of inventories, it is statis-
tically signiﬁcant and larger in magnitude than the effect of proﬁtability. This ﬁnding rests
on the assumption that comovement is exogenous. When tariff is used as an instrument for

comovement, the evidence is much weaker.

83

Notes

lSee Harris and Raviv (1991) for a survey of theoretical research matched with the
available empirical evidence. For a more recent survey of the empirical evidence see Rajan
and Zingales (1995).

2In fact, MacKay and Phillips (2005) ﬁnd, based on their sample of 315 competitive
manufacturing industries, that only 13% of variation in ﬁnancial structure can be attributed
to industry ﬁxed effects, while the rest of the variation can be attributed to ﬁrm ﬁxed effects
(54%) and within-ﬁrm variation (33%).

3According to the model presented in Shleifer and Vishny (1992), the level of optimal
leverage of a ﬁrm increases with increasing liquidation value of its assets. This effect was
previously suggested for instance in Myers (1977) and Diamond (1991). However, there
also exists an alternative view for the role of liquidation values which argues the opposite:
higher liquidity will provide a greater ﬂexibility enabling the borrower to substitute riskier
assets once ﬁrnds are obtained, and therefore exposes the lender to a greater risk of potential
default causing the lender to withdraw funds. See for instance Jensen and Meckling (1976)
for more details about this kind of risk to the lender.

4The focus on trade credit is explained later in this section.

5Pulvino (1998) however does not speciﬁcally address the issue of comovement in his
paper.

6See Jensen and Meck-ling (1976) for a detailed explanation of the agency cost problem.

7Banks can also repossess and resell assets, however they will incur much higher cost

of doing so than issuers of trade credit who already have existing means to resell the taken

84

assets. See Petersen and Rajan (1997) on this point.

8See Petersen and Rajan (1994) for evidence, or Smith (1987).

9Titrnan and Wessels (1988) identify two proxies for the ﬁrm’s liquidation value, the
ratio of inventories to total assets and the ratio of tangible assets to total assets. Also see
Song and Philippatos (2004) for a more recent example of using inventories as a proxy of

the collateral value.

loComin and Mulani (2006) deﬁne comovement as: F L com 00 em em,“ =

t 5 t 5
Z (Sj‘t/(l — SIT))Cov(giTt-:4a gjft:4)9
j #l 91' Eli '
where I indexes an industry sector, i ﬁrm, t time period, g real sales growth rate, and
8,, share of ﬁrm 1' real sales in total real sales of the corresponding industry sector in period

t, and where C 00 (3,42, 54:3) is deﬁned as the covariance between

{gm-2, gi,t—1 , git: gi,t+1, gi,r+2}30d{gj,r-2, gj,r-I , g jr, 8131+], g j,t+2}-
11There are minor differences in the impact of shocks on the cash ﬂows of different ﬁrms,
otherwise there would be nothing to distinguish the sellers of assets from the buyers in the
model.
12Industry dummies are included in the OLS regressions. However, because they are
time invariant, industry dummies cannot be included in the ﬁxed effects regressions.
13See for instance Almazan and Molina (2005) or Molina and Preve (2006) for the same

approach.

14As Petersen and Rajan (1997) points out, if the ratio of depreciation to total assets

85

does not vary too much, then excluding depreciation makes no fundamental difference to
the estimation, since net income divided by assets will be proportionate to depreciation
added to net income and divided by assets.

15See Shleifer and Vishny (1992) for the theory of the liquidation value of a ﬁrm. Alter-
natively, a positive relationship between the inventory stock and the amount of trade credit
received may also be explained by the transaction cost theory. The theory implies that sup-
pliers may extend credit when it is advantageous to separate the delivery schedule from
the payment schedule. In other words, lower transaction cost may result if ﬁrms with high
levels of inventories are not required to pay each time a delivery occurs. See Ferris (1981)
for more details about the transaction cost theory.

16See Jensen and Meckling (1976).

l7Compustat is used frequently in ﬁnancial studies. As examples, see Molina and Preve
(2006) or Opler and Titrnan (1994).

18See section 2.3. for a discussion of the implications of the absence of small ﬁrms.

19Even though Compustat states that the quarterly data on accounts payable are available
from 1976: 1, the data is actually available for a few years prior to 1976. Nevertheless, prior
to 1975, disproportionately more data are available for the ﬁrst three quarters than for the
fourth.

20Accounting treatment of these ﬁrms is also different from other industries.

21All of these studies implicitly assume that missing and deleted values occur randomly
in Compustat.

22Also, around 300 observations containing negative values for payables, assets, or sales

were deleted. The number of observations with zero accounts payble in the sample period

86

was 3,618.

23This number holds for Pooled OLS when inventories and retained earnings are ex-
cluded. If these two variables are present in the estimation additionally around 10,000
observations are lost due to missing data.

24See for instance Almazan et al. (2007) for a more recent example of scaling capital
structure determinants by assets.

25Testing for serial correlation after the static form of regression returns a t-statistic equal
to 263 on the added lagged residual. After including 4 lags, the t-statistic on the added
lagged residual declines to 74. The remaining serial correlation in the error should be
partly due to the time invariant unobserved effect, that will be removed in the ﬁxed effects
estimation. The serial correlation is not a major problem since the standard errors are
robust to it, however unless the idiosyncratic ﬁrm effect is removed the regressors may be
correlated with the error term, which would invalidate the empirical results. I also reran
the regressions including 1, 2, 3, and 5 lags. The main results, particularly the approximate
magnitude and signiﬁcance of the comovement coefﬁcient were similar across these various
speciﬁcations. In the case of ﬁrm age, the lags are omitted since the inclusion would
introduce multicollinearity in the model.

26The data on tariff rates is only available for manufacturing, and comes from two
sources. Data for the 1989-2001 period was made available by John Romalis, and is linked
for example at www.nber.org/data. Data for the 1976-1989 comes from the Center for
International Data at UC Davis and can be found at www.internationalata.org. Since the
earlier period tariffs are measured by the 1972-SIC categories, while the later period tariffs

are measured by the 1987-SIC categories, I used a ﬁle also made available by the center at

87

UC Davis to convert the 1972-SIC data to the 1987-SIC data.

2"When running the reduced form regression using comovement as the dependent vari-
able, controlling for 4 lags of tariff, the coefﬁcient on tariff was equal to 0.002, with the
p-value equal 0.008 on this coefficient. While signiﬁcant, the size is not very large. The
tariff rates are in percentages, ranging from 0 to 13.5% in the sample, with the median be-
ing 2.47%. Increasing tariff in an industry with the median comovement rate from the 25th
to the 75th percentile would thus lower the comovement from 0.061 to 0.054. The joint test
on the tariff together with all of its four lags results in the F-statistic equal to 10.12 giving
a p-value of zero to four decimal places. When repeating this test with lags of comovement
as the dependent variable, the resulting F-statistic on the instrument and its lags ranged
from 7.1 to 9.34, again giving a zero p-value to four decimal places.

28Only contemporaneous coefﬁcients are shown in the table. Results for all lags are
given in appendix 2.

29See Petersen (2007) for a summary and frequency of the estimation approaches used
in the ﬁnance literature. According to Petersen, the ﬁxed effects approach is second to the
Fama-MacBeth procedure, with OLS being used less frequently.

30The bootstrapped standard errors indicate signiﬁcance at the 5% level, while the usual
robust standard errors indicate even stronger signiﬁcance.

3‘ Insigniﬁcant once the standard error is adjusted for the presence of generated regressor.

32Again, only contemporaneous coefficients are shown in the table. Results for all lags
can be found in appendix 2.

33 Statistical signiﬁcance measured based on the bootstrapped standard errors is lost for

all regressors in the last column, while when measured based on the usual standard errors

88

signiﬁcance is lost only for the comovement variable.

89

Table 2.1: Description of the Variables

 

 

Variable

| Description

 

 

Trade Credit

log (ﬁgﬁ), where AP is balance sheet quarterly data item 46, “ac-
counts payable” (original source: Compustat), and where ASS E T S is bal-
ance sheet quarterly data item 44, "assets - total" (original source: Com-
pustat)

 

Comovement

See equation 1 in section 2.2., where g,,, is the quarterly growth rate of
real sales of ﬁrm i, where real sales is S $113535 , where SALES is quarterly
data item 2, "sales (net " (original source: Compustat), and where PP]
is the producer price index (original source: US. Department of Labor:

Bureau of Labor Statistics)

 

 

Firm Size

10g (457%), where ASSETS is balance sheet quarterly data item 44,

"assets - total" (original source: Compustat), and where PPI is the pro-
ducer price index (original source: US. Department of Labor: Bureau of
Labor Statistics)

 

Proﬁtability
and Avail-
ability of
Internal
Funds

fig—5% if PROFIT > o, 0 otherwise, where PROFIT is balance
sheet quarterly data item 69, "net income (loss)" (original source: Compu-
stat), and where ASSETS is balance sheet quarterly data item 44, "assets

- total" (original source: Compustat)

 

%;—§ if (PROFIT < 0)&(ASALES > 0), 0 otherwise, where

PROFIT and ASSETS is the same as above, and where the variable
ASAL ES is as described in the "Rate of Growth" variable entry below

 

—7—§§gg‘1; if (PROFIT < 0)&(ASALES < 0), 0 otherwise, where

PROFIT and ASSETS is the same as above, and where the variable
ASAL ES is as described in the "Rate of Growth" variable entry below

 

Availability
of Internal
Funds

MW, where RETAINED EARNINGS is balance sheet
quarterly data item 58, "retained earnings" (original source: Compustat),
and where ASSETS is quarterly data item 44, "assets - total" (original
source: Compustat)

 

Rate of
Growth

 

 

% if ASAL ES > 0, 0 otherwise, where ASAL ES is the ﬁrst dif-
ference in SAL ES,/ PPI,, where SALES, /PPI, is as deﬁned for the
Comovement variable, and where ASSETS is quarterly data item 44, "as-
sets - total" (original source: Compustat)

90

Table 2.1 (continued)

 

 

 

 

 

 

 

 

 

 

 

 

Variable H Description

Rate of %%%S if ASAL ES < o, 0 otherwise, where ASAL ES is the ﬁrst dif-

Growth ference in SALES,/PPI,, where SALVEsr/PPI, is as deﬁned for the
Comovement variable, and where ASS E T S is quarterly data item 44, "as-
sets - total" (original source: Compustat)

Age log (1 + AGE), where AGE is the number of quarters the ﬁrm has been
present in the Compustat database
[log(l + AGE)]2, where AGE is the number of quarters the ﬁrm has
been present in the Compustat database

Current As- CURRENzggﬁgs ' CASH , where CURRENT ASSETS is balance sheet

sets quarterly data item 40, "current assets - total" (original source: Compus-
tat), where C ASH is balance sheet quarterly data item 36, "cash and short-
term investments" (original source: Compustat), and where ASSETS is
deﬁned as for the next variable

Inventory W, where [N VE N T 0R1 E S is balance sheet quarterly

Stock data item 38, "inventories - total" (original source: Compustat), and where
ASSETS is balance sheet quarterly data item 44, "assets - total" (original
source: Compustat)

Market 3-month treasury bill: secondary market rate (original source: board of

Interest Rate governors of the Federal Reserve System)

91

Table 2.2: Summary Statistics

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Time Period Accounts Payable / Assets
Median Mean Std Dev Min lstP 99thP Max Obs
1976:2-2004:3 0.083 0.1 14 0.373 0 0.006 0.482 82 374379
1976:2-1984z4 0.092 0.113 0.088 0 0.009 0.417 3.04 84497
1985:1-1994:4 0.083 0.1 15 0.457 0 0.006 0.468 82 147537
1995:1-2004:3 0.078 0.113 0.380 0 0.005 0.547 56.533 142345
Time Period COMOVEMENT (R-squared from OLS of sales growth rate on time dummies)
Median Mean Std Dev Min lstP 99thP Max Obs
1976:2-2004:3 0.061 0.131 0.170 0.001 0.003 0.775 1 374379
1976:2-1984:4 0.118 0.195 0.203 0.002 0.003 0.851 1 84497
1985: 1-1994:4 0.055 0.122 0.162 0.001 0.004 0.775 1 147537
1995: 1-2004:3 0.044 0.102 0.145 0.002 0.003 0.725 1 142345
Time Period Real Book Value of Assets (using PP1, in millions of 1982 dollars)
Median Mean Std Dev Min lstP 99thP Max Obs
1976:2-2004:3 107.924 1 141 .6 4469.2 0.001 0.856 18614 188651 374379
1976:2-1984:4 142.536 949.58 3022.5 0.034 1.491 12137 153343 84497
1985:1-1994:4 74.586 967.85 3635.9 0.001 0.701 16851 107337 147537
1995: 1-2004:3 127.152 1435.65 5767.8 0.002 0.873 23167 188651 142345
Time Period Net Proﬁt / Assets —
Median Mean Std Dev Min lstP 99thP Max Obs
1976:2-2004:3 0.010 -0.006 0.187 -47.667 -0.351 0.095 24.13 374397
1976:2-1984:4 0.014 0.010 0.058 -6.101 -0. 141 0.073 5.007 84497
1985:1-1994:4 0.009 -0.004 0.130 -18 -0.321 0.102 8.069 147537
1995 : 1-2004:3 0.007 -0.018 0.268 -47.667 -0.469 0.102 24.13 142345
Notes:

When the comovement equals one it is because there was only 1 ﬁrm present in the industry
during the 10 quarter window (these observations are not used in estimation). For each variable,
median, minimum, lst percentile, 99th percentile, and maximum are listed in the table.

92

Table 2.2 (continued)

 

 

 

 

 

Time Period Change in Real Sales / Assets (Real Sales in millions of 1982 dollars, using PPI)
Median Mean Std Dev Min lstP 99thP Max Obs
1976:2-2004z3 0.004 0.004 0.226 -1 14.55 -0.308 0.295 27.33 374379
1976:2-1984:4 0.004 0.004 0.099 -1 .313 -0.316 0.288 5.192 84497
1985:1-1994:4 0.004 0.005 0.316 -1 14.55 -0.306 0.304 7.686 147537
1995:1-2004:3 0.004 0.004 0.157 -12.82 -0.303 0.288 27.32 142345

 

 

 

 

Time Period

Age (Approximated by number of quarters since ﬁrst appearance in Compustat)

 

 

 

 

 

 

 

 

 

 

 

 

 

Median Mean Std Dev Min lstP 99thP Max Obs
1976:2-2004:3 32 39.67 29.16 0 3 121 133 374379
1976:2-1984:4 27 26.40 14.71 0 2 53 60 84497
1985: 1- 1994 :4 31 37.48 25.36 0 3 92 100 147537
1995:1-200423 43 49.83 35.06 0 4 125 133 142345

Time Period (Current Assets Less Cash) / Assets

Median Mean Std Dev Min lstP 99thP Max Obs
1976:2-2004:3 0.386 0.390 0.233 0 0.022 0.889 1 374379
1976:2-1984:4 0.460 0.434 0.229 0 0.039 0.878 0.992 84497
1985:1-1994:4 0.396 0.396 0.238 0 0.023 0.896 1 147537
1995:1-2004:3 0.340 0.357 0.226 0 0.017 0.883 1 142345
Notes:

For each variable, median, minimum, lst percentile, 99th percentile,and maximum are listed in

the table.

93

Table 2.3: Comovement by 2-digit SIC (ﬁlll sample)

 

 

 

 

 

 

 

 

2-digit SIC [Median Mean Sthev Min lstP 99thP Max Obs
1 0.045 0.057 0.049 0.010 0.010 0.255 0.258 1147
2 0.195 0.191 0.082 0.051 0.051 0.383 0.383 114
7 0.181 0.258 0.176 0.068 0.068 0.817 0.817 262
8 0.228 0.247 0.063 0.145 0.145 0.361 0.361 59
10 0.028 0.078 0.121 0.007 0.009 0.532 1 3429
12 0.191 0.229 0.140 0.034 0.056 0.651 1 594
13 0.004 0.020 0.040 0.001 0.002 0.190 0.449 20511
14 0.091 0.122 0.091 0.034 0.035 0.373 0.374 832
15 0.034 0.073 0.102 0.007 0.008 0.447 1 4144
16 0.115 0.139 0.081 0.052 0.053 0.393 0.431 1796
17 0.087 0.129 0.101 0.030 0.043 0.548 0.670 1502
20 0.127 0.185 0.170 0.019 0.029 1 1 12077
21 0.284 0.352 0.180 0.133 0.136 0.879 0.908 665
22 0.148 0.164 0.088 0.025 0.034 0.441 0.718 4730
23 0.068 0.102 0.096 0.004 0.004 0.463 0.773 5506
24 0.248 0.257 0.143 0.049 0.051 0.586 1 3590
25 0.142 0.188 0.158 0.015 0.030 0.805 1 3689
26 0.065 0.109 0.130 0.014 0.016 0.738 1 6199
27 0.104 0.158 0.162 0.015 0.017 0.793 1 7930
28 0.028 0.055 0.080 0.003 0.004 0.343 1 29473
29 0.057 0.123 0.190 0.019 0.020 0.844 1 4096
30 0.085 0.119 0.126 0.008 0.012 0.617 0.950 7175
31 0.117 0.134 0.086 0.039 0.042 0.439 0.524 1943
32 0.305 0.331 0.212 0.027 0.031 1 1 4186
33 0.107 0.149 0.146 0.014 0.017 0.845 1 8555
34 0.103 0.150 0.140 0.006 0.017 0.740 1 10214
35 0.056 0.085 0.098 0.006 0.010 0.512 1 33371
36 0.023 0.056 0.090 0.003 0.005 0.469 1 37634
37 0.092 0.133 0.148 0.011 0.015 1 1 12055
38 0.024 0.051 0.087 0.005 0.005 0.444 1 29775
39 0.097 0.167 0.174 0.016 0.021 0.914 1 5668
Notes:

When the comovement equals one it is because there was only 1 ﬁrm present in
the industry during the 10 quarter window (these observations are not used in esti-
mation). For each variable, median, minimum, lst percentile, 99th percentile, and

maximum are listed in the table.

94

Table 2.3 (continued)

 

 

Std Dev Min

 

 

 

 

 

 

 

2-digit SIC 1] Median Mean lstP 99thP Max Obs
40 0.057 0.093 0.136 0.029 0.029 1 1 2276
41 0.189 0.202 0.093 0.051 0.051 0.403 0.406 230
42 0.091 0.113 0.086 0.027 0.029 0.406 0.949 3737
44 0.067 0.096 0.113 0.018 0.019 0.995 1 1588
45 0.107 0.129 0.127 0.016 0.016 0.807 1 4032
46 0.228 0.203 0.099 0.066 0.066 0.393 0.399 117
47 0.079 0.106 0.090 0.023 0.026 0.482 0.535 1555
48 0.017 0.043 0.083 0.003 0.003 0.407 1 16366
49 0.297 0.337 0.229 0.015 0.026 0.800 1 33371
50 0.075 0.139 0.167 0.009 0.014 1 1 13810
51 0.076 0.112 0.107 0.015 0.019 0.490 1 8030
52 0.322 0.312 0.165 0.050 0.050 0.631 0.997 1451
53 0.608 0.579 0.186 0.094 0.101 0.852 0.978 4476
54 0.068 0.098 0.116 0.010 0.010 0.650 0.936 4691
55 0.142 0.195 0.168 0.036 0.036 1 1 1482
56 0.348 0.348 0.143 0.030 0.033 0.694 0.723 4194
57 0.324 0.371 0.203 0.058 0.065 0.907 0.955 2817
58 0.008 0.023 0.047 0.005 0.006 0.255 0.765 8600
59 0.078 0.175 0.187 0.011 0.012 0.800 0.894 9027
All 0.061 0.131 0.170 0.001 0.003 0.775 1 374379
Notes:

When the comovement equals one it is because there was only 1 ﬁrm present in
the industry during the 10 quarter window (these observations are not used in esti-
mation). For each variable, median, minimum, lst percentile, 99th percentile, and

maximum are listed in the table.

95

Table 2.4: Baseline Results

 

 

 

 

 

 

 

 

 

 

- Pooled OLS Frixed Eﬁ'ects
1080151. 102 (1.21m)
Comovement -0.118"* 00" 0098*" W
(0.0394) [0.0455] (0.0329) [0.0384]
log (book value of assets) -0.005 -0.019
(0.0277) [0.0400] (0.0203) [0.0264]
Net Proﬁt/Assets, if proﬁt>0 0235*" 0"" -0.121"* 0"
(0.0495) [0.0821] (0.0305) [0.0538]
Net Proﬁt/Assets, if proﬁt<0 & Asales>0 -0.420"""' 0° -0.340‘”""' 0""
(0.1213) [0.1692] (0.0889) [0.1 129]
Net Proﬁt/Assets, if proﬁt<0 & Asales <0 0390*" 0" -0.284"”"* 0”
(0.1074) [0.1688] (0.0844) [0.1332]
Asales/Assets, if positive, 0 otherwise 0614"" 00" 0.419‘" 0""
(0.1738) [0.2057] (0.1207) [0.1397]
Asales/Assets, if negative, 0 otherwise 0.103“ 0202"" 0""
(0.0623) [0.1283] (0.0381) [0.0730]
log(1+ﬁrm age) -0.101"'* 0.117
(0.0505) [0.0709] (0.1014) [0.1309]
[log (1+ﬁrm age)]2 0.016" -0013
(0.0078) [0.0113] (0.0217) [0.0278]
(Current Assets-Cash)/Assets 1.741*" 0"" 1676*" 0""
(0.0705) [0.1098] (0.0538) [0.0739]
Inventories / Assets 0739“" 0"" 0762*“ “0
(0.0924) [0.1233] (0.0743) [0.1026]
Ret.Earnings / Assets -0.001“' -0.002""""'
(0.0007) [0.0051] (0.0004) [0.0043]
T-Bill 0011*" "0" 0.012‘" 00"
(0.0010) [0.0017] (0.0009) [0.0016]
4 Lags of Regressors Included Yes Yes
Year and Quarter Dummies Yes Yes
4-digit SIC Industry Dummies Yes No

 

 

 

 

 

96

Table 2.4 (continued)

 

 

 

 

 

Pooled OLS Fixed Effects
A P A F
103 (Assets ) log (Assets)
R-squared 0.47 . 0.26
Obs 280512 280512
Firms 9369 9369

 

 

 

 

Notes:

Standard errors in parentheses are robust to heteroskedasticity and serial correla-
tion. Standard errors in brackets are bootstrapped and valid despite the presence of
the generated regressor (comovement variable). Estimation assumes independence
across ﬁrms. *""", ", and "' denote respective signiﬁcance at 1%, 5%,and 10%
based on the standard errors given in parentheses. 00", 0°, and 0 denote respective
signiﬁcance at 1%, 5%,and 10% based on the standard errors

given in brackets. Firm age is approximated by the number of quarters present in
the Compustat database. 4 lags are included for all individually listed regressors
except age.

97

Table 2.5: Instrumental Variables

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Pooled OLS Fixed Effects [ 2 SLS FE-IV
10mg?) |°8(73‘35—a) 108 (13.5...) 108 (£537)
Comovement -0.138"' 0° -0.007 -0.930 -2.537
(0.0758) [0.0581] (0.0663) [0.0606] (6.5204) [30.50] (5.6860) [86.69]
log (book value -0.008 -0.021 -0.020 -0.027
of assets) (0.0331) [0.0487] (0.0254) [0.0384] (0.0350) [0.1313] (0.0273) [0.2405]
Net Proﬁt/As- -0.369*** 0"" 0186‘" 0 0366*" -0.165**
sets, if proﬁt> 0 (0.1004) [0.1383] (0.0697) [0.1081] (0.1052) [0.3302] (0.0731) [0.7521]
Net Proﬁt/As- -0.709*** 0"" 0526*" 0"" -0.726*** 0"" 0536*”
sets, if proﬁt<0 (0.0913) [0.1264] (0.0735) [0.1082] (0.0953) [0.1518] (0.0764) [0.4281]
& Asales>0
Net Proﬁt/As- 0689*“ 00" -0.563*** 0"” 0696*" 0"" 0568*"
sets, if proﬁt<0 (0.0888) [0.1393] (0.0796) [0.1148] (0.0923) [0.1353] (0.0803) [0.5873]
& Asales <0
Asales/As- 0.420 0.292 0.427 0.295
sets, if positive, (0.3108) [0.4496] (0.2103) [0.3068] (0.3192) [0.4874] (0.2146) [0.6860]
0 otherwise
Asales/As- 0.154 0285*” 0"" 0.129 0271*"
sets, if negative, (0.1054) [0.1493] (0.0672) [0.1008] (0.1117) [0.2768] (0.0752) [2.718]
0 otherwise
log (1+ﬁrm age) -0. 109 -0.150 -0.097 -0.167
(0.0781) [0.0914] (0.1487) [0.2030] (0.0818) [0.3705] (0.1615) [2.818]
[log 0.016 0.041 0.012 0.044
(1+ﬁrm age)]2 (0.0122) [0.0144] (0.0326) [0.0454] (0.0131) [0.0625] (0.0351) [0.8481]
(Current As- 1804*“ ”0" 1696“" ”0" 1802"" 00" 1.711***
sets-Cash)/As- (0. 1 192) [0.1732] (0.0813) [0.1 149] (0.1275) [0.3465] (0.0904) [1.197]
sets
Inventories/ 0678‘“ 0"" 0697*" 0"" 0645*" 0660*"
Assets (0.1526) [0.21 13] (0.1107) [0.1441] (0.1614) [0.4820] (0.1181) [2.538]
Ret.Earnings / -0.005*** -0.003" -0.005"* -0.003"“"
Assets (0.0020) [0.0163] (0.0014) [0.0151] (0.0020) [0.0163] (0.0013) [0.0181]
T-Bill 0015*" 00" 0014‘" 00° 0025*“ 0019*"
(0.0016) [0.0024] (0.0014) [0.0021] (0.0078) [0.041 1] (0.0057) [0.1325]
4 Lags of Regre- Yes Yes Yes Yes
ssors Included
Year and Qua- Yes Yes Yes Yes
rter Dummies
4-digit SIC Ind- Yes No Yes No
ustry Dummies

 

 

 

 

 

98

Table 2.5 (continued)

 

 

 

 

 

 

 

 

Pooled OLS Fixed Effects | 2 SLS FE-IV
AP A P AP AP
10g (Ass ) log ) M45311) 10g (Assets)
R-squared 0.40 0.18 0.22 0.15
Obs 107965 107965 107965 107965
Firms 3798 3798 3798 3798

 

 

 

 

 

Notes:

Tariff is used as instrument for comovement in the last 2 columns. Standard errors
in parentheses are robust to heteroskedasticity and serial correlation. Standard er-
rors in brackets are bootstrapped and valid despite the presence of the generated
regressor (comovement variable). Estimation assumes independence across ﬁrms.
*", **, and "' denote respective signiﬁcance at 1%, 5%,and 10% based on the
standard errors given in parentheses. 0°", 0°, and 0 denote respective

signiﬁcance at 1%, 5%,and 10% based on the standard errors given in brackets.
Firm age is approximated by the number of quarters present in the Compustat data-
base. 4 lags are included for all individually listed regressors except age.

99

Chapter 3: How Should We Control For
the Clustering in Central Bank Interven-
tion Data?

1 Introduction

Periods of sterilized central bank intervention in the foreign exchange market are intrigu-
ing episodes in the monetary history of countries.1 Unlike the consensus amongst macro-
economists that a monetary policy action exhibits a real effect on output at least in the short
run, there is little consensus that intervention has an effect on exchange rates or market
volatility in the short or long run.2 Looking at volume effects alone, it would be hard
to make a case that intervention could inﬂuence exchange rates by shifting the supply or
demand curve for foreign currency. For example, the largest magnitude of intervention in
the DM/$ market in the 19808 was $1.3 billion, compared with total activity in the DM/$
market of $1,300 billion. Also, the Federal Reserve has intervened at times in amounts as
little as $50 million. Thus, such interventions are only a tiny fraction of total market activ-
ity. Yet, all central bankers surveyed by Neely (2001) believe that intervention is effective
in altering the exchange rate.

Another intriguing aspect of sterilized intervention is that the data is "clustered", with
successive days of intervention followed by successive days without intervention. This
clustering has been dealt with in many ways in the intervention literature. For example,
Herrera and Ozbay (2005) estimate a dynamic Tobit model that includes lags of purchase

and sale interventions (which are the dependent variables) for intervention by the Turkish

100

Central Bank in support of the Lira. Lagging the intervention variable has been pursued
by Kim and Sheen (2002) and Frenkel, Pierdzioch and Stadtmann (2003) for the Australian
and Japanese central banks respectively. Ito (2002) includes lags of the dependent inter-
vention variable in a standard OLS regression for the Bank of Japan reaction function, or
the function that predicts whether or not an intervention will occur on a given day.

The goal of this chapter is to determine what motivations trigger a central bank inter-
vention, making use of two models that are speciﬁcally designed to exploit its data clus-
tering; the autoregressive conditional hazard model (ACH, Hamilton and Jorda, 2002) and
the autoregressive conditional binomial model (ACB, Herrera, 2004). We utilize two in-
tervention data sets from three central banks. The ﬁrst data set consists of interventions by
the United States Federal Reserve and German Bundesbank between January 5th, 1987 and
January 22nd, 1993. These dates coincide with the Plaza Agreement and the Louvre Ac-
cord. The second data set contains unilateral interventions by the Bank of Japan between
April 1st, 1991 and February 28, 2001.

The ACH model modiﬁes Engle and Russell’s (1998) autoregressive conditional dura-
tion (ACD) model by estimating a {0,1} Bernoulli process in calendar time and allowing
for the inclusion of exogenous variables. The ACH model exploits the clustering by in-
cluding a lag of the duration, or time between successive interventions, in the central bank’s
reaction function and uses the lagged duration to impose a ﬁmctional form assumption on
the conditional probability. That is, the ACH assumes the probability of intervention is
equal to the reciprocal of the expected duration. For example, if the expected duration is 3,
the probability of intervention is 1/3. The appeal is that Hamilton and Jorda, using an ACH

model, are able to outperform a standard vector autoregression in terms of mean-squared

101

error in predicting Federal ﬁmds rate target changes for the time period 1984-1989 and
1989-2001. This is a noteworthy result because, unlike a vector autoregression, the ACH
maximizes a likelihood function rather than minimizing the sum of squared errors. Since
open market Operations exhibit similar clustering, it is natural to apply the ACH model to
intervention data.

The ACB is a calendar time modiﬁcation of the autoregressive conditional multinomial
model of Engle and Russell (2005). The ACB exploits the clustering by lagging an in-
dicator variable that takes the value of unity if an intervention occurred and lagging the
response probability of intervention. In fact, the ACB without including the aforemen-
tioned lags is a standard probit model. Like the ACH model, exogenous variables can be
included in the ACB model. Since the ACH and ACB are similar in that they are calendar
time modiﬁcations of Engle and Russell models that are designed to exploit a common
characteristic of time series data, the comparison done in this paper is a natural one. By
comparing these two models, we can learn whether the time dependence in the data stems
from durations, past probabilities of intervention, and/or past events. Thus, not only are
the results relevant for this application, but they are relevant for other applications using
similarly clustered data.

Surprisingly, in the case of the Federal Reserve, the ACH does worse, according to the
Schwarz Bayesian Criterion (SBC, Schwarz, 1978) and the pseudo R2 (McFadden, 1974),
than a simple probit model that does not deal with the clustering. We show that this is
likely due to the functional form imposed by the ACH for the estimated probability. For
the Federal Reserve, the lagged duration is very insigniﬁcant in a probit model, meaning

the ACH imposes its functional form assumption on an insigniﬁcant variable. For the

102

other two central banks, the duration is signiﬁcant, or nearly so in a probit model. Yet,
the ACH only slightly improves the goodness-of-ﬁt compared to a probit model and oﬂ'ers
little additional insight as to why the central bank intervenes. And, surprisingly, the ACH
can only successfully predict less than 1% of interventions.

Contrast this with the ACB, which outperforms the ACH for all central banks in terms
of the SBC and the pseudo R2. Unlike the probit and ACH, the ACB can successfully pre-
dict between 30% and 50% of interventions. A Rivers and Vuong (2002) test of non-nested
likelihoods rejects the ACH in favor of the ACB for all central banks and a ﬂuctuation test
proposed by Giacomini and Rossi (2007) illustrates this preference is stable throughout
the sample. Additionally, when the lagged duration is signiﬁcant in a probit model, and
dynamics are introduced in the ACB model, the signiﬁcance of the lagged duration disap-
pears. This suggests that the time dynamics are better captured through past probabilities
and past events, rather than through past durations. As a result, we argue that an ACB
model should be the ﬁrst considered over an offshoot of an autoregressive conditional du-
ration model (such as the ACH) when estimating a binary model with clustered data.

In utilizing the ACB to test motivations for intervention implied by the Plaza Agree-
ment and Louvre Accord and emphasized in Neely (2001), we ﬁnd some evidence of the
Bundesbank intervening in order to move the exchange rate into line with long-run ﬁtn-
damentals, but the Federal Reserve did not. Interestingly, we ﬁnd that both the Federal
Reserve and Bundesbank preferred to intervene when foreign exchange market volatility
was low, rather than intervening when volatility was high with the goal of reducing it. We
show a novel result in that the spread between the six month Treasury Bill rate and the Fed-

eral Funds rate cOntains predictive power as to when the Federal Reserve will intervene,

103

suggesting a link between Federal Reserve monetary policy and intervention. As far as we
know, neither of these two results have been shown in the intervention literature using daily
data. We ﬁnd strong evidence that the Bank of Japan has intervened in response to day-by-
day changes in the nominal ¥/$ exchange rate and evidence that this response is different
following Eisuke Sakakibara becoming the Director General of the International Finance
Bureau of the Ministry of Finance in Japan on June lst, 1995.3 Intervention by the Bank
of Japan can be described as "leaning against the wind" prior to Sakakibara’s appointment
and can be described as "leaning with the wind" following his appointment.4

The remainder of this chapter is organized as follows. Section 2 describes the interven-
tion data. Section 3 describes modeling the probability of intervention using the ACH and
ACB. Section 4 presents the results. Section 5 offers a comparison of the two econometric

methods, and Section 6 concludes.

2 Intervention Data

2.1 U.S. and Germany

We use daily weekday intervention and exchange rate data for the Federal Reserve and
Bundesbank interventions in the DM/$ market spanning January 5th, 1987 through Janu-
ary 22nd, 1993, which corresponds to 1387 observations.5 Thus, the data follows the Plaza
Agreement (signed September 22, 1985) and the Louvre Accord (signed February 22nd,
1987). The goal of the Plaza Agreement was to use coordinated intervention to depreciate

the dollar against its main trading currencies in hopes of reducing the U.S. current account

104

deﬁcit and quelling protectionist pressure in the U.S. Congress.6 Since the dollar had sub-
stantially depreciated against both the dollar and yen following the Plaza Agreement, the
goal of the Louvre Accord was market stability.7 Nigel Lawson, British Chancellor of the
Exchequer, called the meeting "Plaza Two [...] I see this~ meeting as the lineal descendent
of the Plaza meeting [...]Then we all agreed that the dollar should fall, now we all agree we
need stability."8

Table 3.1 presents summary statistics on the Federal Reserve and German intervention
data sets. Note that the Federal Reserve only intervened 173 times during the sample
period, or in 12.5% of the observations. The Bundesbank intervened more, 210 times, or
14.3% of the observations. Dollar buys were observed on 65 days for the Federal Reserve,
accounting for 37.6% of Federal Reserve interventions, and on 55 days for the Bundesbank,
corresponding to 26.1% of Bundesbank interventions. Compare this to dollar sales, which
were observed on 108 days for the Federal Reserve, corresponding to 62.4% of Federal
Reserve interventions, and on 155 days for the Bundesbank, corresponding to 73.8% of
Bundesbank interventions. Thus, the vast majority of interventions by the Federal Reserve
and Bundesbank taking the form of dollar sales is consistent with the goal of the Plaza
Agreement to depreciate the dollar against the Deutsche mark.

Table 3.1 also gives evidence to the clustering present in the intervention data for both
the Federal Reserve and the Bundesbank. For instance, for the Federal Reserve, only
13.3% of interventions took place in weeks with only one intervention whereas 29.5% of
interventions took place in weeks with two interventions. Not only do interventions take
place for only a small percentage of observations, but the vast majority of interventions

take place in weeks with multiple interventions. Note from the table that the clustering is

105

similar for the Bundesbank. Figure 3.1 gives a graphical representation of the clustering
for the Federal Reserve.9 This clustering is what we are trying to exploit using the ACH

and ACB.

2.2 Japan

We use daily weekday intervention data for the Bank of Japan interventions in the Tokyo
¥/$ market spanning April lst, 1991 through February 28, 2001, which corresponds to
2538 observations and overlaps with the data in Ito (2002).10

Table 3.2 presents summary statistics on the Japanese intervention data set. The Bank
of Japan intervened in 200 days in time span covered by the data, which corresponds to
7.8% of the observations. Yen buys were observed on 32 days or 16% of the interventions,
while yen sales were observed on 168 days or 84% of the interventions. This is consistent
with the Japanese goal of depreciating the yen through intervention, as an overvalued yen
made Japan’s vital export sector less competitive (Ito, 2002). Note that the smallest yen
buys and sells were ¥3.2 billion and 35.1 billion respectively. This corresponds to a buy
of $25.1 million and a sale of $45.1 million, using the spot ¥l$ exchange rate for those
days. Thus, Japanese intervention data is similar to the data for the Federal Reserve and
Bundesbank in that interventions represent a small fraction of daily market activity.

Table 3.2 also breaks down the summary statistics into two subsamples, corresponding
to intervention before and during Eisuke Sakakibara’s tenure as Director General of the
International Finance Bureau of the Ministry of Japan . Both Ito (2002) and Keams and

Rigobon (2004) note that the behavior of Japanese intervention changed with Sakakibara’s

106

appointment, which began on June lst, 1995. Whereas intervention behavior prior to
Sakakibara was to resist short-term undesirable movements in the exchange rate (or to
"lean against the wind"), intervention during Sakakibara’s tenure was to actively move the
yen back to the desired range by "leaning with the wind". For example, if the goal was
to depreciate the yen, and the yen began to depreciate the foreign exchange market on its
own, Sakakibara would intervene by selling yen with the goal of further depreciating it.

Two things are noteworthy about the different subsamples. First, intervention was
much more frequent in the ﬁrst subsample than in the second. Second, intervention was
much smaller in average magnitude in the ﬁrst subsample than in the second. Notice that
for the entire sample, the smallest yen buys and sells took place during the ﬁrst subsample
while the largest buys and sells took place during the second subsample. Also, note that
the magnitude of the largest buy in the ﬁrst subsample (¥ 76.9 billion) is nearly the same as
the smallest buy in the second subsample (¥ 76.4 billion). The reason for these differences
between the two subsamples is that Sakakibara believed that the market was becoming too
accustomed to the smaller, more frequent interventions of the preceding regime. Sakak-
ibara believed that by reducing the frequency and increasing the magnitude of intervention,
he could more successfully depreciate the yen against the dollar.ll

Table 3.2 presents evidence of the clustering of interventions both in the full sample
and in the two subsamples. For the full sample, only 22% of interventions took place
in weeks with only one intervention; 18% of interventions took place in weeks with two
interventions; 22.5% of interventions took place in weeks with three interventions; 20%
of interventions took place in weeks with four interventions and 17.5% of interventions

took place in weeks with ﬁve interventions. Figure 3.2 presents a graphical representation

107

of the clustering for the full sample of the Bank of Japan. Thus, similar to the Federal
Reserve and the Bundesbank, observations with interventions represent a small fraction of
the observations and the vast majority of the interventions occur in weeks with multiple

interventions.

3 Estimating the Probability of Intervention

3.1 ACH

Following the notation of Engle and Russell (1998) and Hamilton and Jorda (2002), we
deﬁne the variable u,- as the length of time between the 1"" and (i + I)” intervention. Then

we deﬁne w, to be the expectation of u,- given past durations 14,--1, 11,--2, Then, the

ACD(r,m) model of Engle and Russell is:

m r
V’i =w+zajui-j+ZﬂjWi—j ' (1)

which is analogous to a GARCH(m,r) process. Thus, the ACD(1,1) model posits that the

expected duration is a weighted average of past durations:
01,, = aun_1 + ,Baun_2 + ﬂzaun_3 + + ,B""2au1 + 19"-]? (2)

where ‘17 is the average duration. The parameter ,6 controls how fast the past durations
decay in predicting the n” expected duration. Engle and Russell (1998) show that for the

m r
ACD(r,m) process to be stationary, Z a j + Z 191- < 1.
i=1 J'=1

108

Following Hamilton and Jorda (2002), we transform the timing of the ACD model
to calendar time. We deﬁne the function N (t) as the cumulative number of interventions
observed at time t. Thus, if we do not observe an intervention at time t, N (t) = N (t — 1).
If we observe an intervention at time t, then N (t) = N (t — 1) + 1. Using this notation,

we can rewrite equation (1) as:12

m r
WNu) = ZajuN(t)-j + 213; WN(r)—j (3)
i=1 i=1

The hazard rate, h,, is deﬁned as the conditional probability of an intervention taking
place at time t — 1 given information observed as of time t — 1. Thus, the hazard can be
written as:

h. = Pr[N(0 94 N0 —1)Iii_i] (4)

Where Y,_.1 denotes information known as of time t. We follow Hamilton and Jorda
(2002) and assume that the conditional distribution of the hazard is exponential. Under the

exponential distribution, we can write the hazard for the ACH process as:13

1
kg = ,
VIN(r-1)+ ‘0 + 1’ zt—l

 

(5)

where 60 is a constant and 21-1 is a vector of explanatory variables observed on the previous
day. Thus, the ACH imposes the functional form that the probability of intervention is

equal to the reciprocal of the expected duration, plus or minus some exogenous variables.

109

Deﬁning the binary variable x,, taking the value of unity if there is an intervention on
date t, and zero otherwise, we can obtain estimates for the parameters 0 = (y I, a’, [3,) by

maximizing the likelihood function:

T
2135110800) + (1 — x.) log (1 — h.» (6)

(=1

3.2 ACB

The ACB model is a ﬂexible speciﬁcation that captures the intervention clustering by lag-
ging the link function and binary dependent variable. First, deﬁne the probability of an

intervention as,

hr E P (x: = 1 IXr—r, ---,x1,Zr-1, "611) (7)

where, as before, 12, is the probability of intervention, x, is the binary dependent variable
taking the value of unity if we observe an intervention, and z,_1 is a. l x n vector of exoge-
nous variables that contain information about the underlying process. For our application,
z,_1 will contain the same exogenous variables for both the ACB and the ACH. Addition-
ally for the ACB, we include the lag of the duration, 14100-1) in the 2 vector.14

The ACB(q,r,s) model is then given by,

r S

q
G’1(hr) = w + Z I],- (xr—j - ht-j) + 21010—1 (ht-j) + Z5jxr-j + 721—1 (3)
j=1 ' j=l i=1

where G (-) is a strictly increasing, continuous cumulative distribution function, such as

the standard normal or the logistic. Thus, G"l (h,_1) = z, 4:) G (2,) = h,, meaning that

110

G"l (.) is a 1-1 mapping from h, to 91. Thus, the ACB exploits the clustering by allowing
the current intervention decision to depend on the lagged response probability and the past
history of interventions. Since G (-) is strictly increasing, we can obtain x, by taking the

inverse of equation (8),

q r s
h; = G [CU-1' Zﬂj (x,_j "' ht-j) +ijG-l (ht-j) +Zajxt—j + yz,-1] (9)
. i=1

J=l i=1

Since we choose the standard normal distribution for G (-), the ACB(0,0,0) is simply
a standard probit model. Thus, including the same exogenous variables in the z,_1vector
in the ACH plus the lagged duration serves two purposes. First, comparing the ACH with
the ACB(0,0,0) allows us to compare the functional form assumption imposed by the ACH
versus the functional form of the probit. Second, introducing lags into the ACB(0,0,0)
allows us to compare the lagged duration in capturing the time dynamics of the clustered
data versus lags of G "‘ (h,) and x,.

Given initial conditions for x, and In, the path of intervention probabilities can be
constructed and estimates for the parameters 6 = {w, 711, ..., ryq, p1,..., p,, 51, ..., 6,} ob-

tained. Estimation is straightforward through maximizing the likelihood function,

111

T

Z [xt108(ht)+(1—~xt)108(1-ht)] (10)

t=max{q,r,s}+l
3.3 Explanatory Variables

We appeal to the institutional background in deciding what explanatory variables to in-
clude in the 2 vector for each data set. For the U.S. and Germany, following the Plaza
Agreement’s declared goal that "exchange rates should better reﬂect ﬁmdamental economic
conditions than has been the case",‘5 we include a measure of the deviation of the log of
the exchange rate, s,, from the log of the exchange rate based on purchasing power parity,
sf, which is the exchange rate that would prevail based on economic fundamentals (Neely

2005). Absolute purchasing power parity is deﬁned as:
S, = — (11)

where S, is the level of the exchange rate, P, is the price level in the U.S. and PK is the
price level in the foreign country. Taking the natural logarithm of both sides of equation
(11).

£=m—d on

Thus, the exchange rate based on economic ﬁmdamentals can be calculated by simply
taking the difference of the log of the price level between the U.S. and Germany.16

Following the Louvre Accord’s goal to calm disorderly markets, we include a measure

112

of daily excess volatility as suggested by Baillie and Osterberg (1997). We deﬁne excess
volatility as the difference between the conditional and unconditional variance of exchange
rate returns, or (a? — 02). The conditional variance is generated through a GARCH(1,1)

process for the log-difference of exchange rate returns: I
0,2 = a) + (8,2_1 + {0,24 (13)

The unconditional variance can then be computed ﬁ'om equation (13):

a)

0’

Fischer (2000) points out that a central bank’s domestic monetary policy objective
might be in conﬂict with current foreign exchange market conditions. For example, re-
call that the goal of the Plaza Agreement was to depreciate the dollar, which could be
achieved by expansionary monetary policy.17 However, if the Federal Reserve’s domestic
monetary policy objective at the time was restrictive rather than expansionary, restrictive
monetary policy would appreciate the dollar. As a result, the Federal Reserve may postpone
a change in monetary policy so that the exchange rate is not further undermined and con-
ﬁrsing signals are not sent to the foreign exchange market (monetary policy in support of
the dollar alongside intervention to depreciate the dollar). To examine the link between the
Federal Reserve’s monetary policy and intervention objectives we include in the 2 vector
the absolute value of the spread between the 6-month Treasury Bill and the Federal funds

rate.18

113

For Japan, rather than including the difference of the nominal logged exchange rate
from fundamentals, we include the change in the natural log of the nominal exchange
rate from date t to date t - 1.19 This captures the idea that the Japanese intervened in
response to overnight changes in the ¥l$ nominal exchange rate. Ito (2002) contains
detailed accounts of many of the intervention episodes throughout the 1990s. In general,
the Japanese wanted to keep the yen within a range of 128¥l$ to' 115¥/$, buying yen when
the nominal exchange rate crossed the upper limit and selling yen when it crossed the lower
limit. However, in 1993, the yen fell out of this range and did not return until early 1997
(Ito, 2002). Thus, a motivation for Japanese intervention cannot be to defend this range,
since the yen was not in, nor even close to it, for a large part of this sample. A more
likely motivation is to lean against the wind, or resist short term movements in the nominal
exchange rate. Ito (2002) is ripe with instances of this. For example, the intervention of
February 15, 1994 was prompted by an overnight appreciation of the yen from 104¥l$ to
102¥/$.

Neither Ito (2002) nor Keams and Rigobon (2004) explicitly model whether or not
the Japanese intervened in response to excess market volatility with the goal of reducing
it. However, lto does imply that the Japanese may have intervened in response to excess
market volatility in order to smooth the appreciation of the yen against the dollar. To test
whether or not this was a motivation for Japanese intervention, we include the measure of

conditional volatility given by equation (13).

114

4 Results

4.1 U.S. and Germany

The results for estimating the probability of intervention for the Federal Reserve using the
ACH(1,1) model are reported in Table 3.3. Note that the estimated values for a and [3 are
both signiﬁcant at the 1% level. The ACH obtains an estimate of 4.701 for the average
expected duration, meaning that on average, approximately 5 days are expected to pass
between two interventions. This corresponds to an average hazard, or the average prob-
ability of an intervention occurring on a given day, of 0.0984. This is broadly consistent
with the summary statistics in Table 3.1, where interventions happened for 12.47% of the
observations

As seen from the table, there is mixed evidence on the predictive power of the two
variables based on the Plaza Agreement and Louvre Accord as to when the Federal Reserve
will intervene. The estimated coeﬂ‘icient on (a ’24 - 02), which is the variable implied by
the Louvre Accord, is insigniﬁcant, whereas the estimated coefﬁcient on the ﬁrst difference
of (s,..1 — 3,1,), which is the variable implied by the Plaza agreement, is signiﬁcant at
the 10% level. However, upon dropping (0 f4 — 02) and re-estimating the ACH model
(which we do in speciﬁcation (2)), the ﬁrst difference of (3,..1 — 3,11) becomes signiﬁcant
at the 5% level (p-value 0.0358). This is an interesting result because it suggests that the
Federal Reserve is intervening as a result of the nominal DM/$ exchange rate moving away
from the exchange rate based on economic fundamentals.

To see this, note that if the dollar is overvalued as the U.S. claimed, then (s,..1 — 31:1) >

0 and the estimated coeﬂicient on this term in the ACH model is negative. Thus,

115

y 3 A (8,..1 — 3:11) < 0. Notice ﬁom equation (5) that this term appears in the denomina-
tor of the hazard, which is the probability of intervention. Thus, this term being negative
makes the denominator of the hazard smaller, which increases the probability of interven-
tion.

The coeﬂ'icient on the spread is negative and signiﬁcant, suggesting that for a day where
the absolute value of the spread between the 6-month Treasury Bill rate and the Federal
funds rate is high, the probability of intervention by the Federal Reserve is higher than
compared to a day when the absolute value of the spread is low.

These are novel results because, as far as we know, no study has established a statistical
link between the value of the spread and the probability of intervention. Also, studies using
the same data do not ﬁnd the Federal Reserve responding to changes in the exchange rate
(see, for example, Baillie and Osterberg 1997). Column 3 of Table 3.4 presents the results
of a probit, or ACB(0,0,0) model, containing the same explanatory variables, including the
lagged duration. Note that when we do not exploit the clustering via the ACH, we do
not ﬁnd the estimated coeﬂicient on A (s,_1 - s,‘_1) to be signiﬁcant, though the estimated
coefﬁcient on spread is positive and signiﬁcant, giving the same interpretation as before.
Thus, it would appear that the ACH is offering a new result.

However, consider the estimates of the ACB model in Table 3.4.20121 Note that in this
case, the estimated coefﬁcient on A (s,_1 - s;__1) fails to be signiﬁcant at the 10% level,
even though the sign is consistent with the result for the ACH(1,1), since it also suggests
that a more overvalued dollar increases the probability of intervention. Also, notice the
difference in the estimated coefﬁcient on (012-1 — 02). Whereas this was insigniﬁcant in

the ACH and probit models, it is signiﬁcant in the ACB(0,1,2) model. However, the neg-

116

ative sign indicates that low volatility increases the probability of intervention. Thus, the
sign is the opposite of what we would expect from the Louvre Accord, meaning the Federal
Reserve prefers to intervene when the market is calmer. Although this result is the oppo-
site of the Federal Reserve’s stated policy, studies that use ultra-high frequency data, such
as Beine and Laurent (2003 and 2005), and Fatum (2002), ﬁnd that intervention increases
market volatility. As a result, the negative estimated sign on (a ‘24 — 02) suggests that the
Federal Reserve would prefer not to make an already volatile market even more volatile
through intervention. Thus, it is interesting that we are able to replicate this result with
daily data using the dynamic binary model.

Notice that the spread continues to have positive predictive power for interventions,
meaning that days when the spread is higher are days where the Federal Reserve is more
likely to intervene. Two explanations can account for this robust result. First, suppose that
the spread is large and positive, indicating that the market is expecting an increase in the
Federal funds rate. As previously described, this would appreciate the dollar, which is in
the opposite direction of the goal of the Plaza Agreement. Thus the sign and signiﬁcance
of the spread is consistent with the idea that the Federal Reserve may have to postpone this
monetary policy action and instead use sterilized intervention to achieve its exchange rate
objective.

The second explanation is consistent with the idea that the Federal Reserve’s monetary
policy objective during this time was expansionary, rather than restrictive. Consider when
the Federal funds rate is above the 6-month Treasury Bill rate, as it is for the majority
of our sample. The high value of the Federal funds rate relative to the Treasury Bill

rate suggests that the market expected the Federal Reserve to engage in an expansionary

117

monetary policy, producing a subsequent fall in the Federal funds rate, which would also
depreciate the dollar against the Deutsche mark. However, as Funabashi (1989) points
out, the Federal Reserve wanted to conduct expansionary monetary policy jointly with the
Bundesbank. The Bundesbank was reluctant to do so, given its strong anti-inﬂationary bias
and its reluctance to relinquish its independence in setting monetary policy. Thus, rather
then engaging in expansionary monetary policy to depreciate the dollar, the Federal Reserve
chose to engage in sterilized intervention instead. This is consistent with Funabashi (1989)
who found that "The Fed used monetary policy to stimulate the U.S. economy or at least
to keep it buoyant, but it did not burden domestic monetary policy with exchange rate
management" (p57).

The results of the ACH(1,1) for Bundesbank intervention are reported in the last column
of Table 3.3. The results are similar to those obtained in the case of the Federal Reserve in
that a and [3 are both signiﬁcant, albeit a is now signiﬁcant at 5%, and the estimated coeﬂi-
cient on (03.1 — 02) is insigniﬁcant. One major difference between the two central banks
is that the coefﬁcient on A (s,_1 - s,‘_1) is insigniﬁcant for the Bundesbank, although with
the expected sign.22

Column 3 of Table 3.5 presents the results of the probit model. The estimates of the
coeﬂicients on (a ,2_1 - 02) and A (s,._1 — s"_l) are both insigniﬁcant as in the ACH case.
We include 14100-1) in the probit model and ﬁnd that it is signiﬁcant at the 5% level as in
the case of the ACH, and in the same direction, also as in the ACH case.

However, like the Federal Reserve, the results are not robust to the ACB model. The last
three columns of Table 3.5 present these results. Three interesting results are important to

note. First )1 2, the estimated coeﬂicient on A (s,_, — sf_1), is estimated to be positive and

118

signiﬁcant at the 1% level. Recall that if the dollar is overvalued then A (s,_1 — 8:4) > 0.
Thus, a positive estimate of y 3 means that the probability of intervention by the Bundes-
bank increases when the dollar becomes more overvalued. Second, when we exploit the
clustering by introducing lags of the binary indicator x,.31 and the link function G ‘1 (h,_1),
the estimated coefﬁcient on the lagged duration becomes insigniﬁcant. This suggests that
including lags of the duration does not capture the time dynamics in the data. Third, like for
the Federal Reserve, the estimated coeﬂicient on (012—1 — 02) is negative and signiﬁcant,
suggesting that the Bundesbank may prefer to intervene in times of low market volatility
so that intervention does not make an already volatile market more volatile. Again, this
result has recently been found in the literature using ultra-high frequency data. As far as

we can tell, we are the ﬁrst to ﬁnd it using daily data.

4.2 Japan

Table 3.6 presents the results of the ACH(1,1) for the entire sample of Japanese interven-
tion. Note that y 2, which is the estimated coeﬂicient on (s,_1 - s,_2), is positive and
signiﬁcant, meaning that the Japanese were "leaning against the wind", or resisting short-
terrn, overnight movements in the nominal exchange rate. Recall that the goal of Japan
was to depreciate the yen against the dollar to improve the competitiveness of their ex-
ports, and that the exchange rate is in terms of yen per dollar. Thus, (s,_1 — s,_2) < 0
means that yen appreciated (s,_1 < 3,..2). The positive estimated value of y 2 thus makes
y 2 (s,..1 — s,....2) < 0 and thus the denominator of equation (5) smaller.23

As with the Federal Reserve and Bundesbank, we compare the results of the ACH with

119

those of the probit and dynamic ACB models. Column 3 of Table 3.7 presents the results
of the probit for the entire sample. Here, we ﬁnd a negative and signiﬁcant estimate of
y 2, the coefﬁcient on (5,..1 - s,_2). Thus, the leaning against the wind behavior is robust
to the probit speciﬁcation, since y 2 (s,_1 — s,_2) > 0 fOllowing a yen appreciation, which
increases the probability of intervention.

Columns 4-6 of Table 3.7 present the results of the ACB(0,1,3) model for Japan. Like
the two previous models, the ACB(0,1,3) ﬁnds that the Japanese are leaning against the
wind, with the estimated coeﬂicient on (3,-1 — s,_2) negative and signiﬁcant. Note that
(a ’24 — 02) is largely insigniﬁcant in all models (although it is signiﬁcant at the 10% level
in the ACB), suggesting that market volatility played little role in determining Japanese
intervention, and that the Japanese were more concerned with the level of the nominal ¥I$
exchange rate.

To see how Japanese intervention differed before and after Sakakibara’s tenure, we
reestimate the ACH, probit, and ACB models before and after June 1, 1995. Columns 4
and 5 of Table 3.6 present the ACH results and Table 3.8 presents the ACB(0,1,2) results
for the lst subsample, which corresponds to intervention prior to Sakakibara. These results
are very similar to the results for the entire sample. In each model, we ﬁnd evidence for
leaning against the wind.

However, contrast this with the results of the second subsample. Column 6 of Table
3.6 and Table 3.9 present the results of the second subsample. Note that these results
are strikingly different than those for the entire sample and the ﬁrst subsample. For each
model, notice the sign change for the coeﬂicient on (s,_1 — s,_2). This suggests that

after Sakakibara, the Japanese were leaning with the wind, or intervening in response to a

120

favorable overnight movement in the exchange rate in order to further that movement. Note
that if the yen depreciates, (s,_1 — s,_2) > 0. Thus, the negative estimated coeﬂicient in
the ACH model means y 2 (s,_1 - s,_.2) < 0, which makes the denominator of equation
(5) smaller and thus the hazard larger. The positive estimated coeﬂicient in the probit and
ACB(0,1,2) means that y 2 (3,-1 — s,_2) > 0, which makes the probability of intervention
larger. This evidence suggests that Sakakibara preferred to inﬂuence the exchange rate by
intervening when market conditions were already moving in the direction he favored. That
is, if the yen was depreciating, Sakakibara intervened to further depreciate it.

Compare and contrast these results of the two Japanese subsamples with what is found
by Ito (2002). With our methodology, we ﬁnd that the Japanese leaned against the wind
prior to Sakakibara and leaned with the wind during Sakakibara’s tenure. Ito, however,
ﬁnds the lean against the wind hypothesis to be statistically insigniﬁcant during the ﬁrst
subsample but statistically signiﬁcant during the second subsample. Ito also ﬁnds no evi-
dence for leaning with the wind. Thus, we argue that our results are more consistent with
Japanese intervention behavior. Ito ﬁnds that interventions in the second subsample were
less predictable than in the ﬁrst subsample given the lower R2 in his second subsample
regression than his ﬁrst. Our result is similar. We ﬁnd that interventions in the second
subsample are less predictable than interventions in the ﬁrst. Both the pseudo R2 and
percent of interventions correctly predicted are lower in the second subsample compared
with the ﬁrst. This is consistent with the idea that Sakakibara believed that Japanese inter-

ventions were too predictable, and consequently made them less so.

121

5 Comparison of Econometric Methods

5.1 Goodness-of-Fit

To begin a comparison between the ACH, probit, and ACB, we construct two measures
of goodness-of-ﬁt based on the log-likelihood of each model. The Schwarz Bayesian
Criterion (SBC, Schwarz, 1978) adjusts the log-likelihood by subtracting (r /2) times the
natural log of of the number of observations, where r is the number of parameters estimated
by the model. The pseudo R2 suggested by McFadden (1974) is found by computing
1 - Lur /£0, for each model, where Luris the unrestricted log-likelihood and £0 is the
log-likelihood of the model under the restriction that all parameters except the constant are
zero.

Both the SBC and psudeo R2 are reported at the bottom of the table for each estimated
model. Comparing Tables 3.3-3.5 for the Federal Reserve and Bundesbank, we see that the
ACH is favored over the probit for the Bundesbank, which is expected, given the clustering
present in the data. Yet, given this, the pseudo R2 is relatively low, never getting above
0.10. Both measures are markedly different for the ACB. Both the log-likelihood and SBC
drastically fall, and the pseudo R2 rises from 0.04 for the Federal Reserve and 0.095 for
the Bundesbank to around 0.3 for each. Finally, note that it is curious how neither measure
favors the ACH over the probit for the Federal Reserve, since Table 3.2 and Figure 3.1
illustrate that the Federal Reserve data exhibits similar clustering to the Bundesbank.

Comparing Tables 3.6-3.7 for full-sample Japan, we see that the results are similar to

those for the Bundesbank. The ACH is preferred to probit by both measures, but like be-

122

fore, the pseudo R2 is relatively low. And, like with the Federal Reserve and Bundesbank,
both measures strongly prefer the ACB to the other two models. The SBC falls, and the
pseudo R2 rises from 0.04 to 0.36 for the full Japanese sample. Thus, there is consistent
evidence that the ACB ﬁts the data better. I

However, one may oﬂer the criticism that since the log-likehood is much lower for
the ACB, compared to the probit and ACH, goodness-of—ﬁt measures relying on it will
naturally prefer the ACB. To conduct a robustness check using a measure not relying on
the log-likelihood, we compute the percentage of interventions correctly predicted (PCP)
for each model.24

Comparing Tables 3.3-3.5 for the Federal Reserve and the Bundesbank, the results are
troubling for the probit; it cannot successfully predict a single intervention for either central
bank! Surprisingly, the results are just as troubling for the ACH. It cannot successfully
predict a single intervention for the Federal Reserve and only successfully predicts 0.7%
of them for the Bundesbank. However, like the goodness-of-ﬁt measures relying on the
log-likelihood, the ACB performs drastically better. It is able to successfully predict nearly
half of the interventions for both central banks. The PCP for Japan is largely the same (see
Tables 3.6-3.9).

Figures 3.3 and 3.4 illustrate above result for the Federal Reserve. Figure 3.3 prints
out the probit and ACH estimated probability of intervention. Notice that neither model
ever assigns a probability for intervention over 0.50.25 Since a probability of at least
0.50 is the standard criteria for predicting an event (see, for example, Wooldridge 2002),
neither model predicts that an intervention will occur. Contrast Figure 3.3 with Figure 3.4,

which prints the ACB estimated probability of intervention. Notice that the ACB assigns a

123

probability for intervention greater than 0.50 on many days. And, notice that these spikes
in probability are centered over intervention clusters. Even interventions the ACB fails
to predict correspond to spikes in probability centered over these interventions. It is just
that these spikes were not 0.50 or greater, which would then trigger a prediction. Overall,

Figures 3.3 and 3.4 offer strong evidence in favor of the ACB.

5.2 Model Selection

To formally test the ACH versus the ACB, we conduct a Rivers and Vuong (2002) test,
which modiﬁes the Vuong (1989) test of non-nested likelihoods for time series data. Ide-
ally, we would like to conduct a likelihood ratio test between the ACH and the ACB, as we
can between the ACB and probit. However, this is not possible as one model does not nest
the other. Fortunately, the Rivers and Vuong (2002) test allows us to do conduct such a

test. The null hypothesis of the test is,
H0 E0 [440” — 5,403] = 0,

which states that the two models are equally close to the true speciﬁcation. Rivers and
Vuong (2002) show that the test statistic is distributed Normal(0,1). Thus, if the statistic is
statistically less than zero, the ACH is preferred. If the statistic is statistically greater than
zero, the ACB is preferred. If the statistic is not statistically different from zero, the test
cannot distinguish between the two models, given the data.

The Rivers and Vuong statistic (RV) is reported at the bottom of the tables listing the

ACB results for each central bank. The probit is rejected in favor of the ACH at the 5%

124

level of signiﬁcance or lower for the Bundesbank and full-sample Japan, and the ACH is
rejected in favor of the ACB at a p-value of 0.0000 for all central banks. However, much
like the SBC and pseudo R2, the RV cannot reject the probit in favor of the ACH for the
Federal Reserve. In fact, the sign of the RV favors the [probit in this case.

Giacomini and Rossi (2007) point out that the Rivers and Vuong (RV, 2002) test se-
lects the model based on average performance. As a result, in unstable environments,
researchers may erroneously select a model based on average performance, ignoring the
fact that the selection statistic may be reversed over a portion of the sample. To check
if that is the case here, we compute Giacomini and Rossi’s (2007) in-sample "ﬂuctuation
tes ". The ﬂuctuation test computes the "ﬂuctuation statistic", F,,,,,, (which is based on the

Kullback-Leibler Information Criterion, or "KLIC") over a moving window of size m:

t+m/2
F,,m = ﬁ'lm-l Z (LfCH — EfCB) , t= m/2 +1,..., T — m/2
j=t-—m/2+1

(14)
and plots FM, along with critical values derived by Giacomini and Rossi for windows

ranging from 0.1 to 0.9 of the total sample size.26 If F,,,,, crosses the upper critical value,

the ACH is rejected in favor of the ACB. If Fm crosses the lower critical value, the ACB

125

is rejected in favor of the ACH. If EM is in the middle of the two critical values, neither
model is rejected. The motivation for plotting this statistic is to check whether or not the
conclusion reached by the RV statistic is reversed over any of the windows.

Figure 3.5 illustrates the result of the ﬂuctuation test over a moving window of 0.5 of
the sample along with the 5% critical values (-2.779, 2.779).27 Figure 3.5(a) plots the
ﬂuctuation statistic for the ACH versus the probit for the Federal Reserve. The result is
consistent with the overall RV statistic. The probit is slightly preferred, but the ACH is
never rejected in favor of the probit. However, the result of the ﬂuctuation test for the
ACH versus the probit for the other two central banks are startling. Even though the RV
statistic rejects the probit in favor of the ACH for both the Bundesbank and Bank of Japan,
the ﬂuctuation test never rejects the probit in favor of the ACH. One reason for this is
the RV statistic is distributed Normal (0, 1), thus making the critical values (-1 .645, 1.645)
much smaller than those for the ﬂuctuation test. More importantly, however, F,,,,, only dips
below zero at the very beginning of the sample and during the mid il990s for Japan (Figure
3.5(c)), and at the beginning of the sample for the Bundesbank. For the remainder of the
sample, F,’,,, is slightly above zero for both. Thus, Giacomini and Rossi’s (2007) concern
that researchers may mistakenly select the wrong model based on average performance
applies to the ACH.28 However, this is not a concern for the ACB. As illustrated by Figure
3.5(b), the ACH is consistently rejected in favor of the ACB until the end of the sample,
when intervention activity trails off (see Figure 3.1). The ﬂuctuation statistic for the test
of the ACH versus the ACB for the other two central banks is largely consistent with this

result.

126

5.3 Summary

The result that the probit is preferred to the ACH for the Federal Reserve deserves atten-
tion, as the ACH is designed to be used with clustered data while the probit is not. Why is
this the case? The answer is the functional form implied by the ACH and the signiﬁcance
of the lagged duration. Recall from equation (5) that the ACH places the functional form
assumption on the lagged duration in that it assumes the probability of intervention is equal
to the reciprocal of the lagged duration Also, recall from equation (1) that the expected
duration is a decaying lag of past durations. Notice in Table 3.4 that the lagged duration is
very insigniﬁcant for the Federal Reserve (p-value 0.7973) in the probit. Thus for the Fed-
eral Reserve, the ACH is using a very insigniﬁcant variable to impose the ACH ﬁmctional
form. Thus, it is no surprise that the probit, which does not make such an assumption,
does better than the ACH for the Federal Reserve. Additional evidence for this claim is
that for the Bundesbank and lst and 2nd subsamples of the Bank of Japan, the lagged dura-
tion enters into the probit model signiﬁcantly at least at the 10% level (for the full Japanese
sample, p-value on the lagged duration is 0.1645). Thus, the functional form assumption
made by the ACH for these central banks is not as severe and hence the ACH outperforms
the probit. However, it is important to stress that the ACH is only able to outperform the
probit globally, not locally.29

Finally, note from the tables that when we include the lagged duration in a dynamic
ACB speciﬁcation, the signiﬁcance of it disappears. For the Bundesbank, the p-value on
the lagged duration increases from 0.0264 in the probit to 0.2077 in the ACB(0,1,3). For

the 1st subsample for Japan, the p-value of the lagged duration increases from 0.0105 in

127

the probit to 0.7354 in the ACB(0,1,2) and for the 2nd subsample, the lagged duration
is increased from 0.101 in the probit to 0.5073 in the ACB(0,1,2). Finally, for the full
Japanese sample, the p-value of the lagged duration is increased from 0.1645 in the probit
to 0.9815 in the ACB(0,1,3). Thus, the time dynamics of the clustered data are better

captured through past response probabilities and past events, rather than past durations.

6 Conclusion

This chapter focused on two simultaneous goals. First, which binary model best captures
the time dynamics present in "clustered" data such as data for sterilized central bank in-
terventions in the foreign exchange market? Second, what motivations do central banks
follow in deciding whether or not to intervene? We found persuasive evidence that the
autoregressive conditional binomial model (ACB) outperformed the autoregressive condi-
tional hazard (ACH) model. The ACB was preferred by each goodness-of-ﬁt measure
examined; the Schwarz Bayesian Criterion, the pseudo R2, and the percentage of inter-
ventions correctly predicted. The Rivers and Vuong (2002) test of non-nested likelihoods
rejected the ACH in favor of the ACB for all central banks. And, when the Rivers and
Vuong test rejected the probit in favor of the ACH, the ﬂuctuation test of Giacomini and
Rossi (2007) found that this result was not stable over the sample. However, the rejection
of the ACH in favor of the ACB was stable over the sample. These results are not only
important for this paper’s application, but for other applications that examine binary data
with similar properties. That is, rather than estimating the ACH, or another offshoot of En-

gle and Russell’s (1998) autoregressive conditional duration model, the dynamics can be

128

captured using a dynamic binary model that nests a standard probit model. As we showed,
the ACB model does a superior job capturing the time dynamics.

Using the ACB to examine the motivations for central banks intervention, we found that
the Federal Reserve did not intervene in response to a “deviation of the exchange rate from
fundamentals. We found evidence that the Federal Reserve preferred to intervene when
the market was calmer, and found that the spread between the 6-month Treasury Bill and
Federal Funds rate contains predictive power for Federal Reserve intervention, suggesting
a link between intervention and monetary policy decisions. Both results are novel. As far
as we know, this is the ﬁrst time a negative relationship between intervention and volatility
has been shown using daily data, as well as the ﬁrst time a link between intervention and
the spread has been shown. We found that the Bundesbank intervened in response to
the exchange rate deviating from the one implied by fundamentals and, like the Federal
Reserve, that the Bundesbank intervened when the market was calmer. We found that
the Bank of Japan was leaning against the wind during the time period before Sakakibara,
whereas it leaned with the wind during Sakakibara. We also found little evidence for the
Bank of Japan intervening in response to excess market volatility, a previously unexamined
motivation.

Overall, the form of the time dynamics assumed in a binary model not only matters for
global and local model performance, but also for the sign and signiﬁcance of exogenous
variables. By assuming a different form of time dynamics, we were able to not only
increase model performance, but ﬁnd new and novel results using daily intervention data.
Thus, the form of time dynamics assumed in a binary model is important not only for this

application, but ﬁiture applications as well.

129

Notes

lAn intervention is sterilized if it is offset by a corresponding change in the domestic
monetary base, so that the intervention will have no net effect on the domestic money
supply.

2See Sarno and Taylor (2001) for a thorough overview of the diverse opinions on this
subject.

3In Japan, intervention is under the jurisdiction of the Ministry of Finance. Once
intervention orders are given, they are carried out by the Bank of Japan. See Ito (2002).

4A seperate question is whether or not intervention is successﬁil in altering the exchange
rate. Although an interesting question, given the ﬁrst paragraph of the introduction, it is
outside of the scope of this chapter. Rather, we focus on why central banks intervene and
how to control for the clustering present in the data.

5The authors are grateful to Richard Baillie for providing the data.

6See point 18 of the oﬂicial G-5 Plaza announcement included in Funabashi (1989).
Note that rather than using the term "depreciate", the G-5 uses the phrase "appreciate the
main non-dollar currencies against the dollar".

7The Deutsche mark had fallen to 1.8250 against the dollar from a pre-Plaza level of
2.85 and the Yen had fallen to 153.50 against the dollar from a pre-Plaza level of 240.
However, as Sarno and Taylor (2001) point out, the dollar was depreciating prior to the
Plaza agreement.

8See Lawson’s interview with the New York Times quoted in Funabashi (1989).

9The graph for the Bundesbank is very similar. Available upon request.

130

10This data is publicly available On the Japanese Ministry of Finance web site at
http:l/www.mof.go.jp/english/e l c02 l .htrn

“Sakakibara (2000) quoted in Ito (2002)

12We also follow Hamilton and Jorda (2002) in dropping the constant a) from equation
(1) and instead include a constant in the denominator of the hazard in (5).

13Following Hamilton and Jorda (2002), we ensure that (5) is continually differentiable

by replacing the denominator with a smoothing function deﬁned as:

1.0001 if v s 1

11(0) = * 240.1(15—1)2 . ‘
1.001 + 0.12+(v_1)2 1f 1 < v 51+0.1

 

0.0001+v if v 21+0.1

 

 

where v is the denominator in (5) and 1(0) is the denominator actually used in the
estimation.
14Note that the only difference between this and the ACH is the ACH contains an inﬁnite
sum of decaying lags of the duration.
15See point 18 of the oﬂicial G-5 Plaza Agreement announcement.

“We cannot reject the null hypothesis that (s,_1 — 3111) has a unit root (t-statistic: -

131

2.02). This result is robust to the inclusion of a time trend and to the inclusion of ten
lags of the diﬂ’erence of the series in an augmented Dickey-Fuller test. Thus, we replace
(s,_1 — s;_1) with the ﬁrst difference of it in the z vector of (5) for the Federal Reserve and
Bundesbank. This is consistent with Meese and Singleton (1982) and Meese and Rogoff
(1983) in that exchange rates appear to be 1(1).

17That is, monetary policy is equivalent to unsterilized intervention.

18Daily data on all interest rates in the paper were downloaded from the Federal Reserve
Bank of St. Louis FRED database, available at http://research.stlouisfed.org/fred2/

l9Dollar/yen exchange rate data is taken ﬁ'om the FRED data base. This exchange rate
data is the spot rate at 12:00P.M. in the New York market. This is advantageous because
when it is 12:00P.M. in New York City, it is 2:00A.M. on the next day in Japan. Assuming
that the Japanese are not intervening at 2:00A.M. we can believe that the ¥/$ are not being
affected by Japanese intervention for that particular day.

20In all cases, the order of the AC B model was selected by comparing the selected model
with a higher order model via a likelihood ratio test.

2'In section 6, we argue that the ACB results should be preferred for both the Federal
Reserve and Bundesbank.

22However, notice that a + ,B > 1, which violates the stationarity condition given in
section 4.1. This is obviously a disadvantage of the AC H model. We constrain ,6 to
ensure stationarity and reestimate the AC H model and obtain nearly identical results.

23However, like for the Bundesbank, the stationarity condition is violated for the Bank
of Japan. And, like for the Bundesbank, the estimates are not dramatically different when

[9 is constrained to ensure stationarity.

132

24 Since, at most, 15% of the observations for a central bank contain an intervention, a
model could have a PCP of 85% or greater simply by predicting 0 for each observation.
Thus, to prevent the PCP ﬁ'om being so upwardly biased, we restrict it to predicting only
interventions.

25Technically, the probit did assign a probability of intervention of 0.83 on 1/24/1991.
However, no intervention took place on that day.

263 is a HAC estimator of the asymptotic variance given by Giacomini and Rossi (2007)
and Rivers and Vuong (2002)

27To conserve space, we only plot the ﬂuctuation statistic for the ACH versus probit and
ACB models for the Federal Reserve, and for the ACH versus the probit model for the
Bank of Japan. For the remainder of the speciﬁcations, we report the percentage of times
the ﬂuctuation statistic is above, below, and in the middle of the critical values in the tables
reporting the probit and ACB results.

28In fact, setting the size of the moving window equal to the 1.0, Eff, becomes the RV
statistic, i.e., the "global AKLIC".

29One might wonder if the smoothing function in footnote 13 plays a role. To check
this, we remove the smoothing ﬁinction from the ACH. Speciﬁcation (2) for the Federal
Reserve cannot converge without it, but Speciﬁcation (3) can. The estimated coeﬂicients
are generally robust between the two. And for full-sample Japan, the ACH cannot converge
without the smoothing function. However, the sign and signiﬁcance of the coefﬁcients are
robust between the ACH, probit, and ACB. Thus, although by no means conclusive, we

ﬁnd evidence that the smoothing funtion does not present a problem.

133

Table 3.1: Summary Statistics for Federal Reserve and Bun-
desbank Intervention

 

 

 

 

 

 

 

 

Federal Reserve Bundesbankj

"“mm . 1387 ' 1461

of observations

nu’i’bﬂ . 173 210

of 1nterventrons

number of

dollar buys 65 55

number of

dollar sells 108 155
largest dollar buy $395 million $567 million
largest dollar sell $740 million $887.7 million
smallest dollar buy $15 million $11.7 million
smallest dollar sell $25 million $2 million
weeks with 1 intervention 23 28
weeks with 2 interventions 27 31
weeks with 3 interventions l8 18
weeks with 4 interventions 8 9
weeks with 5 interventions 2 6
weeks with 0 interventions 222 225

number of dollar buys 60 50

followed by dollar buys

number of dollar buys 4 4

followed by dollar sells

number of dollar sells 4 4

followed by dollar buys

number of dollar sells

followed by dollar sells 104 15]

 

 

 

134

Table 3.2: Summary Statistics for Bank of Japan Intervention

 

 

 

 

 

 

full sample 1st subsample 2nd subsample

"umbe’?f 2538 1038 1500

observations g

“m.“bﬂ . 200 165 35

of interventrons

number of 32 26 6

yen buys

number of

168 139 29

yen sells
largest yen buy ¥2620.1 billion ¥76.9 billion ¥2620.1 billion
largest yen sell ¥1405.9 billion ¥338.8 billion ¥1405.9 billion
smallest yen buy ¥3.2 billion 3153.2 billion ¥76.4 billion
smallest yen sell ¥5.l billion 35.] billion ¥43 billion
weeks with 1 intervention 44 23 16 T
weeks with 2 interventions 18 14 8
weeks with 3 interventions 15 14 1
weeks with 4 interventions 10 9 0
weeks with 5 interventions 7 7 0
weeks with 0 interventions 414 141 275

number of yen buys

followed by yen buys 30 25 5

number of yen buys 1 1 1

followed by yen sells

number of yen sells 2 1 1

followed by yen buys

number of yen sells

followed by yen sells 166 137 27

 

 

 

 

135

Table 3.3: ACH(1,1) Estimation Results for Federal Reserve and Bundes-

 

 

 

 

 

 

 

bank Intervention
Federal Reserve Bundesbank
parameter variable (1) (2)
0.171 0.173m 0.259"
0‘ “NM“ (0.0617) (0.0615) (0.115)
0.511m 0.563m 0.800‘"
'6 WW)“ (0.0949) (0.0898) (0.0745)
6.351‘" 6.11‘" 0.540m
‘0 ”mm" (0.831) (0.741) (0.438)
2 2 0.0167 -0.00326
y‘ (”t-1 ' ‘7 ) (0.0199) (0.00643)
ii‘ 0‘.
y 2 “bs(s”read)"‘ ((11331) ((1)336)
, -140.15* -99.48” -10.83
y 3 A (s "1 _ SH) (72.06) (47.48) (28.16)
171 expected duration 4.452 4.701 13.192
5 average hazard 0.0990 0.0984 0.0728
log lik -501.97 -502.31 -544.10
SBC -523.67 -520.40 -562.32
Pseudo R2 0.0381 0.0374 0.0954
PCP 0.000% 0.000% 0.7%
Notes:

Standard errors are in parentheses. " denotes statistical signiﬁcance at the
10% level, while ** and *** denote it at the 5% and 1% level respectively.
Only one speciﬁcation reported for each sample unless signiﬁcance of ex-
planatory variables changed when insigniﬁcant variables are dropped and
the model reestimated.

136

Table 3.4: ACB Estimation Results for Federal Reserve Intervention

 

 

 

 

 

 

 

 

 

 

 

probit ACB(0,1,2) _
parameter variable | l (1) (2) _ ﬂ (3_)
w constant -1.55* TIT-0.318" .0357?“ 1 038-1 ""
(0.0918) (0.0727) (0.0877) (0.0910)
p G_, (hm) 0.0820” 0.0832" 0.0811"
(0.0421) (0.0408) (0.0441)
120*" 1.16M 1.17M
5‘ x"‘ (0.123) (0.124) (0.124)
52 xr_2 -0.664‘" -0.691"' 0659‘"
(0.165) (0.159) (0.162)
y 1 (a, _ 02) 000195 000088" 000082"
H (0.0015) (0.00038) (0.00041)
0593' 0.0852" 0.0839"
7’2 abs amend)” (0.100) (0.0342) (0.0351)
y 3 A (SH _ 37“) (£0) (.5529)
-0.00051 0.00063
7 4 u” ("1’ (0.0020) (0.00040)
log lik -500.84 -377.29 -368.81 -370.87
sac -518.92 -391.75 -397.75 -392.57
Pseudo R’- 0.0403 0.288 0.293 0.289
PCP 0.000% 45.2% ‘ 49.2% 42.4%
RV 0.165 6.63 7.16 6.93
p—value 0.4345 0.0000 0.0000 0.0000
Fluctuation % above 0% 82.7% 83.0% 82.7%
% between 100% 17.3% 17.0% 17.3%
% below 0% 0% 0% 0%
Notes:

Standard errors are in parentheses. * denotes statistical signiﬁcance at the 10% level,
while ** and *** denote it at the 5% and 1% level respectively.

137

 

Table 3.5: ACB Estimation Results for Bundesbank Intervention

 

 

 

 

 

 

 

 

 

 

 

probit ACB(0,1,3)
parameter variable (1) Q _ (3L
(0 constant 4.06"“ 016*" .0.0575‘ .0.0423
(0.0405) (0.0770) (0.0350) (0.0275)
‘O. Oi. Oi.
p G_, (hr—1) 0.91 0.968 0.975
(0.0445) (0.0187) (0.0150)
1 llttt l 09!!!! l llltt
5 _ . . .

' x’ ' (0.121) (0.121) (0.121)
5 x -0.47*" 0541‘" 0550‘"
2 "2 (0.194) (0.197) (0.198)
a x .035": 0453‘“ -0.486""
3 ”3 (0.163) (0.130) (0.125)

2 2 -0.00154 -0.00031"* 000029“
in (a 1- 0 ) _—
‘- (0.00131) (0.000122) (9.23e—005)
0.504 3.99m 3.84""
A _ — * —
72 (3‘ ‘ SH) (3.20) (1.51) (1.37)
3 u -0.0046" 0.000177
7 W“) (0.00209) (0.000141)
log lik -597.77 411.80 405.09 406.11
SBC -608.70 430.02 434.23 431.61
Pseudo R2 0.00617 0.315 0.327 0.325
PCP 0.000% 45.2% 42.9% 42.4%
RV -232 5.94 6.10 6.14
p-value 0.0102 0.0000 0.0000 0.0000
Fluctuation % above 0% 66.7% 66.7% 66.7%
% between 100% 33.3% 33.3% 33.3%
% below 0% 0% 0% 0%
Notes:

Standard errors are in parentheses. * denotes statistical signiﬁcance at the 10% level,
while ** and *** denote it at the 5% and 1% level respectively.

138

Table 3.6: ACH(1,1) Estimation Results for Bank of Japan Intervention

 

 

 

 

 

 

 

 

 

 

 

full sample lst subsample 2nd subsample

parameter variable (1) (2)

0.376‘" 0.458' 0.486" 1.02e-010
a “N (1)—1 '

(0.132) (0.255) (0.212) (0.0994)

0.707m 3.98e-012 0.170
13 W N(r)—1 (100”

(0.0532) (0.137) (0.192)

2.44m 4.38'" 4.06m 49.12m

tant
’0 “ms (0.884) (0.654) (0.870) (12.96)
2 2 0.00515 -0.0285 0.362m

it (a.-. -v) —

(0.0255) (0.0227) (0.0405)

86.54'm 81.97' 112.32m -209303‘"
7' 2 (St-1 — Sr-z)

(27.80) (42.77) (25.27) (711.57)
T 21.428 4.812 6.145 2.15e-009
F 0.0419 0.0107 0.0912 0.0191
log 1i1t -647.63 437.07 437.91 -155.26
sac -667.23 454.43 451.80 473.54
Pseudo R2 0.0749 0.0385 0.0367 0.0653
PCP 0.04% 0.1% 0.1% 0.1%
Notes:

Standard errors are in parentheses. * denotes statistical signiﬁcance at the 10% level,
while ** and *** denote it at the 5% and 1% level respectively. #: ﬂ constrained to
0.00 to ensure convergence. Only one speciﬁcation reported for each sample unless
signiﬁcance of explanatory variables changed when insigniﬁcant variables are dropped

and the model reestimated.

139

Table 3.7: ACB Estimation Results for Bank of Japan (full sample)

 

 

 

 

 

 

 

 

 

 

 

 

 

probit ACB(0,1,3)
parameter variable (1) (2) (3)
co co t 438*“ -0217‘" 0304'" 0312‘"
"Sta" (0.0399) (0.0910) (0.0696) (0.0728)
. 0.89 t" 0.846m 0.843m
P G“ (hr—1) ——
(0.0466) (0.0352) (0.0366)
1.24'm 1.12m 1.14m
51 xi—l —-
(0.121) (0.125) (0.124)
-0414" -0.323‘ .0327'
62 20.2 —
(0.195) (0.196) (0.195)
0438‘" 0320" 0318"
63 xr—s —
(0.171) (0.146) (0.147)
2 2 -0.00293 -0.000801'
71 (01—1-0) —— ——
(0.00192) (0.000443)
(3 S ) -939" 4632'" 45.21""
72 “1 "2 (4.76) (3.60) (3.46)
u -0.00168 6.89e-005
Y3 W“) (0.00121) (0.00297)
log lik -694.61 463.05 448.80 451.04
sac -710.29 483.64 480.15 474.56
Pseudo R2 0.00780 0.339 0.359 0.356
PCP 0.000% 41.5% 41.0% 41.5%
RV -200 6.73 6.49 6.51
p-value 0.0228 0.0000 0.0000 0.0000
Fluctuation % above 0% 65.9% 66.9% 66.5%
% between 100% 34.1% 33.1% 33.5%
% below 0% 0% 0% 0%
Notes:

Standard errors are in parentheses. * denotes statistical signiﬁcance at the 10% level,
while ** and *** denote it at the 5% and 1% level respectively.

140

 

Table 3.8: ACB Estimation Results for Bank of Japan Intervention (1 st subsample)

 

-:

I

—_

 

 

 

 

 

 

 

 

 

 

 

 

 

 

TV Eobit ACB(0,1,2)
parameter variable (1) + (2) (3)
w constant -0.872'” 0411‘" 0398‘“ 0399‘“
(0.0554) (0.104) (0.0895) (0.0876)
_ 0.757m 0.770m 0.776m
p G ‘ (hr—1) _—
(0.0636) (0.0522) (0.0491)
1 22¢": 1 090‘. l 090.!
5] xr—l —— ' . O
(0.137) (0.143) (0.142)
-0.540" 4.98'" -0.497"'
52 x1—2
(0.213) (0.195) (0.192)
2 2 0.0177m 0.000971
71 (at-1 _ a ) _—
(0-0041 1) (0.00136)
(S S ) -24.14*" -34.59"* -3437”
72 "‘ "'2 (7.05) (6.32) (6.32)
u -0.00672'" -0.000292
7 3 W") (0.00262) (0.000865)
log lik 435.01 -308.61 -290.44 -290.77
sac 448.90 -32250 -314.75 -308.13
Pseudo R2 0.0431 0.321 0.361 0.360
PCP 0.6% 54.5% 57.6% 59.4%
RV 0.3531 6.74 6.84 6.84
p-value 0.3620 0.0000 ' 0.0000 0.0000
Notes:

Standard errors'are in parentheses. "' denotes statistical signiﬁcance at the 10% level,
~ while ** and *** denote it at the 5% and 1% level respectively.

141

Table 3.9: ACB Estimation Results for the Bank of Japan (2nd subsample)

 

 

 

 

 

 

probit ACB(0,1,2)
parameter variable (1) (2) (3)
-2.07"' -2.55‘” -3.14"' 312'“
tant >
‘0 “ms (0.0844) (0.526) (0.292) (0.280)
_ -0.188 -0.398‘" 0415‘"
,0 G l(ht—1) —
(0.242) (0.116) (0.111)
1.13m 1.30m 1.34m
(51 151—1 —
(0.265) (0.280) (0.273)
1.35m 1.49m 1.48'”
52 xt—Z _—
(0.393) (0.285) (0.285)
2 2 0.000145 0.00201
71 (01—1 — 0 ) —— *—
(0.00279) (0.00412)
(S s ) 16.60"" 26.82m 28.47”"
y 2 “1 ”2 (7.80) (7.90) (7.78)
u 0.00242‘ 0.00132
y 3 N ("1) (0.00148) (0.00199)
log lik 462.03 -145.26 -138.62 -139.08
SBC 476.65 459.89 464.22 457.36
Pseudo R2 0.0246 0.126 0.165 0.163
PCP 0.000% 14.3% 1 1.4% 1 1.4%
RV 4.48 0.814 1.60 1.56
p-value 0.0694 0.2078 0.0548 0.0594
Notes:

Standard errors are in parentheses. "‘ denotes statistical signiﬁcance at the 10% level,
while *"‘ and *** denote it at the 5% and 1% level respectively.

142

Figure 3.1: Number of Interventions Per Week for the Federal Reserve

 

 

 

wagons,
In“. «8:88
- «8:er
825mg
53on
111

111- Eaton?
Ill-“H. 826mm...
4 8263

l

l

. 0mm Eon?
t mwm SOQN _.
mwm COME
mam Eon;
wwm :oQN w
wwm :on
wmm Eon;
hwm :0me r
hum Eomxm

 

 

 

 

 

 

 

 

. . . lllllllrlll Baton?

5 4 3 2 1 0
0203—5020:: 50 500552

 

 

143

Figure 3.2: Number of Interventions Per Week for the Bank of Japan

 

 

 

". 884%
1|“. 80:36

114 82:45
,1 5335

r 0mm 5on

 

 

mmm Ev :0

3m CV Em

 

mam :va

H Now —.\.v Em

 

 

a

5

A _ _

4 3 2
2.05.3232. 2.0 Len—E32

1,

_ 5m Ev Em
0

 

 

Figure 3.3: ACH and Probit Estimated Probability of Intervention for the Federal Reserve

 

 

 

 

 

0.91
0.8]
0.7,
10.6
10.5
‘041
1
‘0.3
0.2 121‘
1'11““
0.1 “2.1. " ‘
ll
0 Illllllllll1
FNNWWCO®OO§OOOFFFNNN1
cocooooooococooooommmmoamormm,
mmmmmmmmmmmmmmmmmm
::::::::::::::::::1
l‘l‘hhl‘ﬁl‘ﬁl‘hl‘hhkkl‘hﬁ
Q‘Qgﬁﬂﬂ‘l‘ﬂﬁﬁﬂﬁﬁ‘ﬁﬁﬁﬁﬂ‘
VG VQNVQNVCONVQNVmN
‘- ‘_ ‘- F ‘— ‘—
g 1

1
1

 

1

Notes:
Black line corresponds to the ACH. Grey line corresponds to the Probit.
Black marks along x-axis indicate an intervention took place on that date.

145

 

    

4.

1|

_I

II

 

 

 

 

I

T

I

'I

 

 

 

 

i

 

 

I

 

 

 

 

0.9 —
0.8 4
0.7 ~
0.6 a
0.5 —
0.4 —
0.3 ~
0.2 4

 

Figure 3.4: ACB(0,1,2) Estimated Probability of Intervention for the Federal Reserve

 

0

«ENE _.
NmRNB
«aka?
PmRNN _.
..mRNB
5km?
SEEN _.
09nt
ome?
mmRNN _.
@233
@933
@258 _.
wwtha
wwRN?
ERNN _.
5ka
SEN?

 

146

Black marks along x-axis indicate an intervention took place on that date.

Notes:

 

Figure 3.5: Fluctuation Test

 

 

_——_———_—————

'1987.4 7988.8 1989.8
Dote

 

 

 

Notes:

Results of the ﬂuctuation test for (a)ACH vs. probit for the Federal Reserve, (b)ACH
vs. ACB (speciﬁcation 2) for the Federal Reserve, and (c)ACH vs. probit for Japan.
The dates correspond to the ﬁrst date of the window.

147

Figure 3.5 (continued)

(13)

 

 

 

 

 

'1987.4 1988.8 0989.8
Date

Notes:

Results of the ﬂuctuation test for (a)ACH vs. probit for the Federal Reserve, (b)ACH
vs. ACB (speciﬁcation 2) for the Federal Reserve, and (c)ACH vs. probit for Japan.
The dates correspond to the ﬁrst date of the window.

148

Figure 3.5 (continued)

 

 

 

 

KC)
NT) 4 .
C\l
.2-2
E
:0
B
m
‘T
C?
N') — —_ ___f _____
1 ’1992 1994— 1996
Date

Notes:

Results of the ﬂuctuation test for (a)ACH vs. probit for the Federal Reserve, (b)ACH
vs. ACB (speciﬁcation 2) for the Federal Reserve, and (c)ACH vs. probit for Japan.
The dates correspond to the ﬁrst date of the window.

149

Appendix 1: Data

1 Data Sources

Our main data source is the Standard and Poor’s Compustat North America, which provides
data items on over 24,000 companies that have ﬁled with the SEC. The Compustat data
items provide information on ﬁrms’ Balance Sheets, Statements of Cash Flows, Income
statements, and other supplemental information. We removed all ﬁrms classiﬁed under
the industry group Finance, Insurance, and Real Estate. Effectively, we deleted all ﬁrms
whose DNUM variable (explained below) was in the interval [6000, 7000). For the annual
sample, this left us with 20112 ﬁrms and 202897 credit ﬂow observations, which means
that the average ﬁrm stays in the database for about 10 years. For the quarterly sample, this
left us with 18753 ﬁrms and 545181 credit ﬂow observations implying an average stay of
slightly above 7 years. The numbers for credit ﬂow observations~ pertain to total credit; the
numbers for short-term and long-term credit were almost identical, with the discrapency
being due to some observations missing for the short-term credit but not for the long-term
and vice-versa.

The macroeconomic variables were downloaded from the FRED 11 database of the Fed-

eral Reserve Bank of St. Louis.

150

2 Variables

Short-Term Credit: Balance sheet annual (quarterly) data item 34 (45), “debt in current
liabilities” (original source: Compustat).

Long-Term Credit: Balance sheet annual (quarterly) data item 9 (51), “total long-term
debt” (original source: Compustat).

GDP: Quarterly GDP in billions of 2000 chained dollars, seasonally adjusted at annual
rates (original source: U.S. Department of Commerce, Bureau of Economic Analysis).

Unemployment Rate: Original source: U.S. Department of Labor, Bureau of Labor
Statistics, Series ID LNSl4000000. Annual series were computed by taking the average
across monthly values.

CPI: Consumer Price Index for all urban consumers (all items), index 1982-1984=100,
seasonally adjusted (original source: U.S. Department of Labor, Bureau of Labor Statis-
tics). Quarterly series were computed by taking the observation for the last month of the
quarter.

Federal Funds Rate: Effective federal ﬁmds rate available from the Federal Reserve
Board. Quarterly series were computed by taking the observation for the last month in the
quarter.

GDP Implicit Price Deﬂator: Original source: Bureau of Economic Analysis.

3 Classiﬁcation

Industry. There are two variables available in Compustat that identify an industry type.

The variable NAICS provides the North American Industry Classiﬁcation System (NAICS)

151

codes, while the variable DNUM provides 4-digit industry codes based on the former Stan-
dard Industrial Classiﬁcation (SIC) system. The NAICS variable contains a missing value
for around 10% of the observations. Fortunately, the variable DNUM contains no missing
values, and therefore was chosen to be used when separating ﬁrms into industry groups and
subgroups. The coding for DNUM is almost identical to the SIC classiﬁcation, and can be
safely treated as such for the purposes of this paper.

Size and Location. The subdivision among the census regions is based on the Compu-
stat variable STATE, which identiﬁes a state according to a ﬁrm’s principal location. The
subdivision into quartiles and deciles by size is based on the Compustat data item 12, Net

Sales.

4 Remarks

Firm Entry. To separate ﬁrms entering Compustat between newly created and already
existing we construct a ratio of the gross book value of capital to the net book value of

capital. If this ratio is not larger than 1.2 for a given ﬁrm then this ﬁrm is assumed to be

Dataltem7
Dataltem7 — Dataltem 196 ’

is total gross property, plant, and equipment, and data item 196 is accumulated depreciation,

 

newly formed. The ratio is computed as where data item 7

depletion, and amortization as provided by Compustat.

Firm Exit. When a ﬁrm exits the Compustat database its growth rate and total credit are
counted only if the footnote item AF TNT 35 takes on values between 1 and 4 (exit due to
acquisition or merger, bankruptcy, liquidation, or reverse acquisition). The ﬁrm is ignored

for the exiting period if the reason code equals 5 , 6, 9, or 10 (exit due to leveraged buyout,

152

conversion to a private company, or other reasons).

153

Appendix 2: Results for All Lags

Table 4.1: Results for All Lags

 

 

PooTEd OLs

 

 

 

 

 

 

 

 

‘ Fixed Effects Pooled OLS
Full Sample Full Sample Instrument Sample
log (Alf—PE) log (4:821 S) log (MAP 1
Comovement 0118:" 0"" 0098:" 0° -0.138* 0°
(0.0394) [0.0455] (0.0329) [0.0384] (0.0758) [0.0581]
lagl 0074*” 0° -0.065" 0° 0.124‘” "0°
(0.0289) [0.0323] (0.0268) [0.0291] (0.0470) [0.0373]
lag2 0.063" 0" 0.038 -0.096" "0"
(0.0254) [0.0268] (0.0234) [0.0259] (0.0473) [0.0311]
lag3 0.196‘“ 0189*" ""0 0.014
(0.0302) [0.0343] (0.0277) [0.0309] (0.0502) [0.0467]
lag4 -0.074** 0097"" 0"" 0.090 0
(0.0370) [0.0429] (0.0327) [0.0367] (0.0677) [0.0489]
log (book value of assets) -0.005 -0.019 -0.008
(0.0277) [0.0400] (0.0203) [0.0264] (0.0331) [0.0487]
1ag1 -0.157"'** 00" -0.128"* 0"" 0133‘" 0""
(0.0170) [0.0261] (0.0125) [0.0157] (0.0165) [0.0245]
lag2 0.001 0.002 0.006
(0.0180) [0.0295] (0.0123) [0.0169] (0.0203) [0.0358]
lag3 0.047" 0.008 0.008
(0.0188) [0.0310] (0.0123) [0.0175] (0.0188) [0.0239]
lag4 0112*" 0"" 0.032“ 0147‘" 0""
(0.0248) [0.0400] (0.0136) [0.0214] (0.0182) [0.0246]
Net Proﬁt/Assets, if proﬁt>0 0235*” "0° -0.121*" 0° 0369"" ”0"
(0.0495) [0.0821] (0.0305) [0.0538] (0.1004) [0.1383]
lag] 0303*" 0"" -0.153*" 0° -0.155*
(0.0730) [0.0841] (0.0429) [0.0621] (0.081 1) [0.1362]
lag2 -0.l46" -0.061* 0229"" 0°
(0.0606) [0.0925] (0.0325) [0.0607] (0.0716) [0.1 1 18]
lag3 -0.155""I 0059" 0290*" 0"
(0.0632) [0.0973] (0.0297) [0.0511] (0.0878) [0.1141]
lag4 -0. 145" -0.055"' -0.179"”"
(0.0645) [0.0981] (0.0303) [0.0593] (0.0763) [0.1106]

 

 

 

 

154

Table 4.1 (continued)

 

 

 

 

 

 

Pooled OLS Fixed Effects Pooled OLS
Full Sample Full Sample Instrument Sample
log( AP 2 log( AP ) logg AP 2
Net Proﬁt/Assets, -0.420*” 0" 0340"" ”0" -0.709"‘" 0""
if proﬁt<0 & Asales>0 (0.1213) [0.1692] (0.0889) [0.1129] (0.0913) [0.1264]
lagl -0.263"‘ 0245"“ 0"" 0476"" 0""
(0.1087) [0.1678] (0.0692) [0.0945] (0.0693) [0.0874]
lag2 -0.321"‘“ 0" 0207*" 0" -0.414*" 0""
(0.1 l 11) [0.1520] (0.0668) [0.1026] (0.0782) [0.1035]
lag3 0.055 -0.027 0.006
(0.1388) [0.1875] (0.0505) [0.0931] (0.0699) [0.1369]
lag4 -0.092 -0.066 -0.055
(0.0843) [0.0977] (0.0450) [0.0619] (0.0775) [0.1406]
Net Proﬁt/Assets, -0.390"”"* 0" 0284"" 0" 0689*" 0""
if proﬁt<0 & Asales <0 (0.1074) [0.1688] (0.0844) [0.1332] (0.0888) [0.1393]
lagl 0256"" 0 -0.179*"* " -0.488*** "0"
(0.0887) [0.1391] (0.0616) [0.0958] (0.0686) [0.1205]
lag2 -0.248" 0 -0.164*” 0" 0334*“ "0"
(0.0969) [0.1392] (0.0600) [0.0815] (0.0651) [0.1 173]
lag3 -0.26l“* 0" 0155"" "0" 0212"" 0°
(0.0707) [0.1072] (0.0354) [0.0560] (0.0685) [0.1004]
lag4 0397*" 0"" -0.220"* 0"" 0455*" 0""
(0.0641) [0.1133] (0.0279) [0.0527] (0.0624) [0.1137]
Asales/Assets, 0614*" 0"" 0419*" 0"” 0.420
if positive, 0 otherwise (0.1738) [0.2057] (0.1207) [0.1397] (0.3108) [0.4496]
lagl 0622*" 0"“ 0372"" 0"" 0.627"
(0.1647) [0.2242] (0.1099) [0.1420] (0.2802) [0.3867]
lag2 0440"“ 0" 0257*" 0" 1008"" 0""
(0.1518) [0.2178] (0.0697) [0.1068] (0.0832) [0.1360]
lag3 0.304" 0.188" 0893*" 0""
(0.1295) [0.2686] (0.0755) [0.1494] (0.0786) [0.1 150]
lag4 0276*" 0192*" 0 0839*” ""0

 

(0.0989) [0.1931]

155

(0.0636) [0.1090]

(0.0806) [0.1138]

Table 4.1 (continued)

 

 

 

Pooled OLS

 

 

 

 

 

rFixed Effects Pooled OLS
Full Sample Full Sample Instrument Sample
10g( AP )- logg AP 2 logg AP 2
Asales/Assets, 0.103“ 0202*" 0"" 0.154
if negative, 0 otherwise (0.0623) [0.1283] (0.0381) [0.0730] (0.1054) [0.1493]
lagl -0.024 0.090" 0.176"
(0.0759) [0.1457] (0.0364) [0.0710] (0.0893) [0.1166]
lag2 -0.009 -0.009 0.001
(0.1086) [0.1900] (0.0526) [0.0932] (0.1365) [0.1871]
lag3 -0.042 0.072 -0.039
(0.0821) [0.1106] (0.0449) [0.0565] (0.1075) [0.1558]
lag4 -0.071 0.033 -0.155"
(0.0592) [0.0858] (0.0346) [0.0470] (0.0679) [0.1093]
log (1+ﬁrm age) -0.101** 0.117 -0.109
(0.0505) [0.0709] (0.1014) [0.1309] (0.0781) [0.0914]
[log (1+ﬁrm age)]2 0.016" .0013 0.016
(0.0078) [0.0113] (0.0217) [0.0278] (0.0122) [0.0144]
(Current Assets-Cash)/Assets 1.741"* "0° 1676*" 0"" 1804“" 0""
(0.0705) [0.1098] (0.0538) [0.0739] (0.1 192) [0.1732]
lagl 0255"" "0" 0.200‘” 0"" 0.243‘" 0“
(0.0361) [0.0504] (0.0312) [0.0431] (0.0464) [0.0666]
lag2 0191*" 0"” 0162"" 00" 0.145"
(0.0531) [0.0587] (0.0355) [0.0372] (0.0699) [0.0959]
lag3 0.228‘" "0" 0145*" "0" 0142*" W
(0.0395) [0.0595] (0.0284) [0.0356] (0.0490) [0.0670]
lag4 0160*“ ° 0120*" 0" 0181*" 0°
(0.0557) [0.0907] (0.0357) [0.0501] (0.0609) [0.0908]
Inventories / Assets 0739*" °°° 0762*" 00" 0678*" 0""
(0.0924) [0.1233] (0.0743) [0.1026] (0.1526) [0.21 13]
lagl -l.021*"'* 0"" 0951*" 0"" -0.949"'" "0"
(0.0621) [0.0752] (0.0534) [0.0738] (0.0891) [0.1456]
lag2 0344*" ”0° 0266*" 0"" 0271*" 0°
(0.0774) [0.0955] (0.0509) [0.0628] (0.0817) [0.1077]
lag3 -0.344"”" 0"" -0.157"* 0° 0269*" 0""
(0.0749) [0.1095] (0.0461) [0.0621] (0.0580) [0.0921]
lag4 -0.080 -0.087"' -0.l78**
(0.0683) [0.1037] (0.0494) [0.0766] (0.0846) [0.1200]

 

 

 

 

156

Table 4.1 (continued)

 

 

 

 

 

 

 

 

Pooled OLS Fixed Effects Pooled OLS
Full Sample Full Sample Instrument Sample
108 (7343-13-1) 10882.) 1080.25,.
Ret.Earnings / Assets -0.001* 3 -0.002""" " -0.005*"
(0.0007) [0.0051] (0.0004) [0.0043] (0.0020) [0.0163]
lagl 0002*" 0.002‘“ -0.001
(0.0007) [0.0044] (0.0004) [0.0022] (0.0010) [0.0047]
lag2 0.0002 0.001" -0.002
(0.0011) [0.0048] (0.0006) [0.0027] (0.0014) [0.0106]
lag3 -0.010""""' -0.003“ -0.002"""
(0.0036) [0.0069] (0.0012) [0.0032] (0.0010) [0.0034]
lag4 -0.004"' 0.0005 -0.001
(0.0022) [0.0071] (0.0014) [0.0030] (0.0013) [0.0063]
T-Bill 0011*" 0"" 0012*" 0"" 0015"" 0""
(0.0010) [0.0017] (0.0009) [0.0016] (0.0016) [0.0024]
lagl -0.003"‘* -0.002*" -0.002
(0.0013) [0.0019] (0.0012) [0.0016] (0.0021) [0.0029]
lag2 0.002" 0.002" 0 0.002
(0.0011) [0.0015] (0.0010) [0.0012] (0.0017) [0.0025]
lag3 0.002" 0.001 0.002
(0.0012) [0.0018] (0.0011) [0.0016] (0.0018) [0.0025]
lag4 -0.002"' 0004“" ”0 -0.005*" 0°
(0.0012) [0.0019] (0.0010) [0.0014] (0.0017) [0.0022]
4 Lags of Regressors Included Yes Yes Yes
Year and Quarter Dummies Yes Yes Yes
4-digit SIC Industry Dummies Yes No Yes
R-squared 0.47 0.26 0.40
Obs 280512 280512 107965
Firms 9369 9369 3798

 

 

 

 

157

Table 4.1 (continued)

 

 

 

 

 

 

Fixed Effects 2 SLS FE-IV
Instrument Sample Instrument Sample Instrument Sample
log(7ﬁ%,;) balsa—.1 Iog(—-—-,:,’£,.1
Comovement -0.007 -0.930 -2.537
(0.0663) [0.0606] (6.5204) [30.50] (5.6860) [86.69]
lagl 0.082“ ° -0.956 -1.647
(0.0439) [0.0474] (6.6295) [28.48] (5.5713) [132.7]
lag2 -0.093‘”" 0° 15.46 12.41
(0.0440) [0.0420] (10.324) [24.27] (8.6969) [79.70]
lag3 0.017 -14.01* - 12.39“
(0.0464) [0.0530] (8.055) [43.41] (6.5438) [111.0]
lag4 -0.024 2.457 4.593
(0.0582) [0.0580] (4.5331) [20.27] (3.5672) [131.1]
log (book value of assets) -0.021 -0.020 -0.027
(0.0254) [0.0384] (0.0350) [0.1313] (0.0273) [0.2405]
lagl -0.134"'** 0"" -0.110*""" -0.1 17"”
(0.0158) [0.0218] (0.0277) [0.1095] (0.0241) [0.4529]
lag2 0.002 0.006 0.001
(0.0159) [0.0266] (0.0240) [0.0809] (0.0194) [0.6602]
lag3 0.002 —0.021 -0.019
(0.0147) [0.0206] (0.0283) [0.0611] (0.0217) [0.0978]
lag4 0.037" 0 0163*" 0"" 0048*"
(0.0153) [0.0193] (0.0215) [0.0578] (0.0177) [0.3977]
Net Proﬁt/Assets, if proﬁt>0 -0.l86”* " -0.366*** -0.165**
(0.0697) [0.1081] (0.1052) [0.3302] (0.0731) [0.7521]
lagl -0.043 -0.l77""" -0.049
(0.0607) [0.0936] (0.0893) [0.1792] (0.0629) [0.9018]
lag2 -0.099** -0.241 "* -0.095"”"
(0.0429) [0.0812] (0.0771) [0.2782] (0.0461) [1.421]
lag3 -0.148*** 0 0296*" -0.l41"
(0.0543) [0.0763] (0.0952) [0.2843] (0.0577) [1.919]
lag4 -0.095* -0.186" -0.086"‘
(0.0507) [0.0893] (0.0774) [0.2029] (0.0502) [0.8954]

 

 

 

 

 

158

Table 4.1 (continued)

 

Fixed Effects

Instrument Sample

 

2 SLS

Instrument Sample

 

F E-IV

Instrument Sample

 

10g( AP ) 10g( AP ) 10g( AP 2

 

 

 

Net Proﬁt/Assets, -0.526*" 0"" -0.726*” 0"“ -0.536*"
if proﬁt<0 & Asales>0 (0.0735) [0.1082] (0.0953) [0.1518] (0.0764) [0.4281]
lagl -0.322*** “”0 -0.482“""‘I 0"" -0.322***
(0.0486) [0.0718] (0.0712) [0.1284] (0.0491) [1.028]
lag2 -0.245*** 0"" 0402*" ”0" -0.234*"
(0.0458) [0.0756] (0.0777) [0.1 198] (0.0460) [1.818]
lag3 -0.015 0.003 -0.018
(0.0489) [0.1010] (0.0710) [0.1487] (0.0495) [1.426]
lag4 -0.026 -0.056 -0.026
(0.0493) [0.0930] (0.0799) [0.1519] (0.0503) [0.7399]
Net Proﬁt/Assets, -0.563*" "0" -0.696*** 0°" 0568*"
if proﬁt<0 & Asales <0 (0.0796) [0.1148] (0.0923) [0.1353] (0.0803) [0.5873]
lagl -0.370"* “0 -0.500*" "0" -0.381*"
(0.0530) [0.0821] (0.0711) [0.1791] (0.0545) [0.9523]
lag2 -0.217**"' ”0" -0.327*** " -0.215**"‘
(0.0526) [0.0720] (0.0683) [0.1734] (0.0547) [0.6915]
lag3 -0.113** 0 0199*" -0.106*
(0.0515) [0.0685] (0.0709) [0.1418] (0.0542) [0.5747]
lag4 -0.236*** “0" -0.451‘”“‘ 0" -0.229***
(0.0434) [0.0694] (0.0647) [0.1966] (0.0454) [0.3967]
Asales/Assets, 0.292 0.427 0.295
if positive, 0 otherwise (0.2103) [0.3068] (0.3192) [0.4874] (0.2146) [0.6860]
lagl 0.378" 0.624" 0.380"
(0.1909) [0.2629] (0.2809) [0.4590] (0. 1934) [1.574]
lag2 0544*" 0"" 0971*" 0"" 0510*"
(0.0705) [0.1137] (0.091 1) [0.2800] (0.0770) [1.176]
lag3 0474*" 0"" 0.916‘" 0” 0.491‘"
(0.0661) [0.0892] (0.0870) [0.3951] (0.0761) [1.851]
lag4 0485*" "0" 0324*” “0" 0466*"
(0.0444) [0.0694] (0.0854) [0.2777] (0.0505) [0.9167]

 

 

159

Table 4.1 (continued)

 

 

 

 

 

 

 

 

Fixed Effects 2 SLS FE-IV
Instrument Sample Instrument Sample Instrument Sample
log (MAP ) log 1 I”: 1 log 1 I”: 1
Asales/Assets, 0.285‘" 0"” i 0.129 0271*"
if negative, 0 otherwise (0.0672) [0.1008] (0.1117) [0.2768] (0.0752) [2.718]
lagl 0233“" ”0° 0.228” 0.266‘"
(0.0657) [0.0873] (0.1004) [0.2992] (0.0751) [3.032]
lag2 0.076 0.043 0.095
(0.0827) [0.1176] (0.1471) [0.3735] (0.0963) [2.188]
lag3 0.039 -0.101 -0.021
(0.0620) [0.0860] (0.1164) [0.4282] (0.0743) [1.412]
lag4 0.010 -0.188"* -0.029
(0.0407) [0.0493] (0.0880) [0.2841] (0.0624) [2.107]
log (1+ﬁrm age) -0. 150 -0.097 -0. 167
(0.1487) [0.2030] (0.0818) [0.3705] (0.1615) [2.818]
[log (1+ﬁrm age)]2 0.041 0.012 0.044
(0.0326) [0.0454] (0.0131) [0.0625] (0.0351) [0.8481]
(Current Assets-Cash)/Assets 1696"" 0"" 1802”" ”a” 1711*"
(0.0813) [0.1149] (0.1275) [0.3465] (0.0904) [1.197]
lagl 0163*" 0" 0355‘“ 0255”“
(0.0445) [0.0648] (0.1114) [0.3305] (0.0959) [0.9697]
lag2 0137*" W 0.045 0.052
(0.0519) [0.0689] (0.1012) [0.3876] (0.0760) [1.514]
lag3 0.095" 0 0.061 0.039
(0.0405) [0.0504] (0.0947) [0.3622] (0.0769) [1.662]
lag4 0.058 0253‘" 0.109“
(0.0480) [0.0653] (0.0824) [0.2684] (0.0656) [1.612]
Inventories / Assets 0697*" ”"0 0645*“ 0660*”
(0.1107) [0.1441] (0.1614) [0.4820] (0.1181) [2.538]
lagl 0882*" 0"" -1.035”* 0 -0.948***
(0.0753) [0.1182] (0.1224) [0.5669] (0.1063) [1.517]
lag2 -0.240"'** 0"" -0.128 -0.128
(0.0655) [0.0839] (0.1248) [0.3166] (0.0973) [2.226]
lag3 0144*" 0" 0264*" -0.l47*"
(0.0536) [0.0657] (0.0798) [0.6152] (0.0712) [2.816]
lag4 -0.115 -0.214*"' -0.146"
(0.0706) [0.0913] (0.0959) [0.3321] (0.0809) [1.671]

 

 

160

Table 4.1 (continued)

 

 

 

 

 

 

Fixed Effects 2 SLS FE-IV
Instrument Sample Instrument Sample Instrument Sample
log( AP ) logg AP ) log( AP 2
Ret.Earnings / Assets -0.003** -0.005““" -0.003"'*
(0.0014) [0.0151] (0.0020) [0.0163] (0.0013) [0.0181]
lagl 0.0003 -0.002"‘ 0.0001
(0.0007) [0.0029] (0.001 1) [0.0051] (0.0007) [0.0085]
lag2 -0.0000 -0.002 -0.00005
(0.0013) [0.0061] (0.0015) [0.0101] (0.0014) [0.0238]
lag3 -0.001 -0.002'”' -0.0004
(0.0008) [0.0028] (0.0010) [0.0042] (0.0008) [0.0060]
lag4 0.0004 -0.001 0.0004
(0.0013) [0.0042] (0.0013) [0.0064] (0.0013) [0.0104]
T-Bill 0014*" 0"" 0025*" 0019*"
(0.0014) [0.0021] (0.0078) [0.0411] (0.0057) [0.1325]
lagl -0.001 -0.031 -0.017
(0.0019) [0.0025] (0.0189) [0.0567] (0.0137) [0.1766]
lag2 0.002 0.006 0.005
(0.0016) [0.0021] (0.0070) [0.0427] (0.0058) [0.1433]
lag3 0.003 -0.008 -0.006
(0.0017) [0.0023] (0.0104) [0.0616] (0.0086) [0.2683]
lag4 0007*" 0"" -0.014 -0.015"‘
(0.0016) [0.0020] (0.0100) [0.0478] (0.0088) [0.2072]
4 Lags of Regressors Included Yes Yes Yes
Year and Quarter Dummies Yes Yes Yes
4-digit SIC Industry Dummies No Yes No
R-squared 0.23 0.22 0.15
Obs 107965 107965 107965
Firms 3798 3798 3798

 

 

Notes:

Tariff is used as instrument for comovement in the 2SLS and FE-IV columns. Standard errors
in parentheses are robust to heteroskedasticity and serial correlation. Standard errors in brack-
ets are bootstrapped and valid despite the presence of the generated regressor (comovement
variable). Estimation assumes independence across ﬁrms. **"‘, ”, and "' denote respective

signiﬁcance at 1%, 5%,and 10% based on the standard errors given in parentheses. 00", 0",
and " denote respective signiﬁcance at 1%, 5%,and 10% based on the standard errors given in
brackets. Firm age isapproximated by the number of quarters present in the Compustat database.

161

References

[l] Almazan, A. and C. A. Molina (2005) "Intra-Industry Capital Structure Dispersion,"
Journal of Economics and Management Strategy, 14, 2, 263-297.

[2] Almazan, A., A. de Motta, S. Titman and V. Uysal (2007) "Financial Structure, Liq-
uidity, and Firm Locations," NBER Working Paper 13660.

[3] Baillie, R. T., and W. T. Osterberg (1997). "Why Do Central Banks Intervene?," Jour-
nal of International Money and Finance 16, 909-919.

[4] Beine, M., and S. Laurent (2003). "Central Bank Interventions and Jumps in Double

Long Memory Models of Daily Exchange Rates," Journal of Empirical Finance
10, 641-660.

[5] Beine, M., S. Laurent, and F. C. Palm (2005). "Central Bank Forex Intervention As-
sessed Using Realized Moments," Unpublished Manuscript.

[6] Benmelech, E., M. J. Gannaise and T. J. Moskowitz (2005) "Do Liquidation Val-
ues Affect Financial Contracts? Evidence from Commercial Loan Contracts and
Zoning Regulation," Working Paper, Harvard University, University of California,
Los Angeles, University of Chicago, and NBER.

[7] Bemanke, B. and M. Gertler (1989) “Agency Costs, Net Worth and Business Fluctu-
ations,” American Economic Review, 79, 1, 14-31.

[8] Bemanke, B., M. Gertler and S. Gilchrist (2000) “The Financial Accelerator in a
Quantitative Business Cycle Framewor ,” in Taylor, J. and M. Woodford, eds.,
Handbook of Macroeconomics,. Amsterdam, North Holland, 1341-93.

[9] Bemanke, B., M. Gertler, and S. Gilchrist (1996) “The Financial Accelerator and the
Flight to Quality,” Review of Economics and Statistics, 78, 1, 1-15.

[10] Bradley, M., G. Jarrell, and E. Kim (1984) “On the Existence of an Optimal Capital
Structure: Theory and Evidence,” Journal of Finance, 39, 3, 857-878.

[11] Caballero, R. and M. Hammour (2005) “The Costs of Recessions Revisited: A
Reverse-Liquidationist View,” Review of Economic Studies, 72, 2, 313-341.

[12] Caballero, R. and M. Hammour (2001) “Institutions, Restructuring, and Macroeco-
nomic Performance,” in Advances in Macroeconomic Theory, Jacques Dreze ed.,
Pal grave New York.

[13] Choi, W. G. and Y. Kim (2005) "Trade Credit and the Effect of Macro-Financial
Shocks: Evidence ﬁ'om U.S. Panel Data," Journal of Financial and Quantitative
Analysis, 40, 4.

[14] Comin, D. and S. Mulani (2006) "Diverging Trends in Aggregate and Firm Volatility,"
Review of Economics and Statistics, 88, 2, 383-388.

[15] Davis, 8. and J. Haltiwanger (1992) “Gross Job Creation, Gross Job Destruction, and
Employment Reallocation,” Quarterly Journal of Economics, 107, 3, 819-63.

[16] Davis, 8., J. Haltiwanger and S. Schuh (1996) “Job Creation and Destruction,” Cam-
bridge: Mit Press.

162

[l7] Dell’Ariccia G. and P. Garibaldi (2005) “Gross Credit Flows,” Review of Economic
Studies, 72, 3, 665-685.

[18] Den Haan, W.J. (2000) “The Comovement between Output and Prices,” Journal of
Monetary Economics, 46, 1, 3-30.

[19] Den Haan, W. J ., G. Ramey and J. Watson (2003) “Liquidity Flows and Fragility of
Business Enterprises,” Journal of Monetary Economics, 50, 6, 1215-1241.

[20] Detragiache, E., P. Garella and L. Guiso (2000) “Multiple versus Single Banking
Relationships: Theory and Evidence,” Journal of Finance, 55, 3, 1133-61.

[21] Diamond, D. (1989) "Reputation Acquisition in Debt Markets," Journal of Political
Economy, 97, 828-862.

[22] Diamond, D. (1991) "Debt Maturity Structure and Liquidity Risk," Quarterly Journal
of Economics, 106, 3, 709-737.

[23] Diamond, D. (2004) “Short-Term Debt when Enforcement is Costly,” Journal of Fi-
nance, 59, 4, 1447-1479.

[24] Dominguez, K. M. E. (1992). "The Informational Role of Official Foreign Exchange
Intervention Operations: The Signalling Hypothesis," In Exchange Rate Eﬂi-
ciency and the Behavior of International Asset Markets, Dominguez K. (ed), Gar-
land Publishing Company: New York.

[25] Dominguez, K. M. E. (1993). "Does Foreign Exchange Intervention Matter? The
Portfolio Effect," American Economic Review 83, 1356-1369.

[26] Eisfeldt, A and A. Rampini (2007) “Financing Shortfalls and the Value of Aggregate
Liquidity,” Unpublished Working Paper, Kellogg School of Management, North-
western University.

[27] Eisfeldt, A and A. Rampini (2005) “Capital Reallocation and Liquidity,” Journal of
Monetary Economics, 53, 369-399.

[28] Engle, R , and J. Russell (1998). "Autoregressive Conditional Duration: A New
Model for Irregularly Spaced Transaction Data," Econometrica 66, 1127-1162.

[29] Engle, R. , and J. Russell (2005). "A Discrete-State Continuous-Time Model of F inan-
cial Transactions Prices and Times: The Autoregressive Conditional Multinomial-

Autoregressive Conditional Duration Model," Journal of Business and Economic
Statistics 23, 166-180.

[30] Fama, E. F. and K. R. French (2002) "Testing Trade-Off and Pecking Order Predic-
tions About Dividends and Debt," Review of Financial Studies, 15, 1, 1-33.

[31] Fatum, R. (2002). "Post-Plaza Intervention in the DEM/USD Exchange Rate," Cana-
dian Journal of Economics 35, 556-567.

[32] Ferri, M. and W. Jones (1979) “Determinants of Financial Structure: A New Method-
ological Approach,” Journal of Finance, 34, 3, 631-44.

[33] Ferris, J. S. (1981) "A Transactions Theory of Trade Credit Use," Quarterly Journal
of Economics, 96, 2, 243-270.

163

[34] Fischer, A. M. (2000). "Do Interventions Smooth Interest Rates?," Unpublished
Working Paper, Swiss National Bank and CERP.

[35] Foote, C. L. (1998) “Trend Employment Growth and the Bunching of Job Creation
and Destruction,” Quarterly Journal of Economics, 113, 3, 809-834.

[36] Frank, M. Z. and G. V. Goyal (2007) "Capital Structure Decisions: Which
Factors are Reliably Important?," Working Paper available at SSRN:
http://ssm.com/abstract=567650

[37] Frenkel, M. , C. Pierdzioch, and G. Stadtmann (2003). "Modeling Coordinated For-
eign Exchange Market Interventions: the Case of the Japanese and U.S. Interven-
tions in the 19908," Review of World Economics 139, 709-729.

[3 8] Fries, S., M. Miller and W. Perraudin (1997) " Debt in Industry Equilibrium," Review
of Financial Studies, 10, 1, 39-67.

[39] Funabashi, Y. (1989). Managing the Dollar: From the Plaza to the Louvre. The Insti-
tute for International Economics: Washington DC.

[40] Giacomini, R. , and B. Rossi (2007). "Model Selection and Forecast Comparison in
Unstable Environments," Unpublished Working Paper.

[41] Graham, J. (2000) "How Big Are the Tax Beneﬁts of Debt?," Journal of Finance, 55,
1901-1941.

[42] Graham, J. and C. Harvey (2001) "The Theory and Practice of Corporate Finance:
Evidence from the Field," Journal of Financial Economics, 60, 2, 187-243.

[43] Guiso, L. and R. Minetti (2004) "Multiple Creditors and Information Rights: Theory
and Evidence from US Firms," CEPR Working Paper 4278.

[44] Hamilton, J. D., and O. Jorda. (2002). "A Model of the Federal Funds Rate Target,"
The Journal of Political Economy 110, 1135-1167.

[45] Harris, M. and A. Raviv (1990) “Capital Structure and the Informational Role of
Debt,” Journal of Finance, 45, 2, 321-349.

[46] Harris, M. and A. Raviv (1991) “The Theory of Capital Structure,” Journal of Finance,
46, 1, 297-355.

[47] Hege, U. and C. Hennessy (2007) "Acquisition Values and Optimal Financial (In)
Flexibility," Working Paper available at SSRN: http://ssm.com/abstract=1082253

[48] Herrera, A. M. (2004). "Predicting U.S. Recessions Using an Autoregressive Condi-
tional Binomial Model," Unpublished Working Paper, Michigan State University.

[49] Herrera, A., M. and P. Ozbay (2005). "A Dynamic Model of Central Bank Inter-
vention: Evidence from Turkey," Unpublished Working Paper, Michigan State
University.

[50] Holmstrom, B. and S. Kaplan (2001) “Corporate Governance and Merger Activity in
the United States: Making Sense of the 1980s and 1990s,” Journal of Economic
Perspectives, 15, 2, 121-144.

164

[51] Holmstrom, B., and J. Tirole (1997) “Financial Intennediation, Loanable Funds and
the Real Sector,” Quarterly Journal of Economics, 112, 3, 663-691 .

[52] Ito, T. (2002). "Is Foreign Exchange Intervention Effective? The Japanese Experi-
ences in the 1990s," NBER Working Paper No. 8914.

[53] Jensen, M. and W. Meckling (1976) “Theory of the Firm: Managerial Behavior,
Agency Costs, and Capital Structure,” Journal of Financial Economics, 3, 305-
360.

[54] Keams, J., and R. Rigobon (2004). "Identifying the Efﬁcacy of Central Bank Inter-
ventions: Evidence from Australia and Japan," Journal of International Economics
66, 31-48.

[55] Kim, S.-J., and J. Sheen (2002). "The Determinants of Foreign Exchange Intervention
by Central Banks: Evidence from Australia," Journal of International Money and
Finance 21, 619-649.

[56] Long, M. S., I. B. Malitz and S. A. Ravid (1993) "Trade Credit, Quality Guarantees,
and Product Marketability," Financial Management, 22, 4, 117-127.

[57] Loughran, T. (2008) "The Impact of Firm Location on Equity Issuance," Financial
Management, Forthcoming.

[5 8] MacKay, P. (2003) "Real Flexibility and Financial Structure: An Empirical Analysis,"
Review of Financial Studies, 16, 4, 1131-1165.

[59] MacKay, P. and G. M. Phillips (2005) "How Does Industry Affect Firm Financial
Structure?," Review of Financial Studies, 18, 4, 1433-1466.

[60] Maksimovic, V. and J. Zechner (1991) "Debt, Agency Costs, and Industry Equilib-
rium," Journal of Finance, 46, 1619-1643.

[61] McFadden, D. L. (1974). "Conditional Logit Analysis of Qualitative Choice Analy-
sis," in Frontiers in Econometrics, Zarembka P (ed), Academic Press: New York,
105-142.

[62] Modigliani, F. and M. H. Miller (1958) "The Cost of Capital, Corporation Finance
and the Theory of Investment," American Economic Review, 48, 3, 261-297.

[63] Molina, C. A. and L. A. Preve (2006) "Trade Receivables Policy of Distressed Firms
and its Effect on the Costs of Financial Distress," Working Paper available at
SSRN: http://ssm.com/abstract=891624

[64] Mortensen, D. and C. Pissarides (1994) “Job Creation and Job Destruction in the
Theory of Unemployment,” Review of Economic Studies, 61, 3, 397-415.

[65] Myers, S. C. (1977) "Determinants of Corporate Borrowing," Journal of Financial
Economics, 5, 147-175.

[66] Myers, S. (2001) “Capital Structure,” Journal of Economic Perspectives, 15, 2, 81-
102.

[67] Myers, S. and N. Majluf(1984) “Corporate Financing and Investment Decisions when
Firms Have Information that Investors do not Have,” Journal of Financial Eco-
nomics, 13, 187-221.

165

[68] Neely, C. J. (2001). "The Practice of Central Bank Intervention: Looking Under the
Hood," Federal Reserve Bank of St. Louis Review 83,1-10.

[69] Neely, C. J. (2005). "An Analysis of Recent Studies of Foreign Exchange Interven-
tion," Federal Reserve Bank of St. Louis Review 87, 685-717.

[70] Nilsen, J. H. (2002) "Trade Credit and the Bank Lending Channel," Journal of Money,
Credit and Banking, 34, 1, 226-253. '

[71] Norrbin, S. C. and K. L. Reffett (1995) "Trade Credit in a Monetary Economy," Jour-
nal of Monetary Economics, 35, 3, 413-430.

[72] Opler, T. C. and S. Titman (1994) "Financial Distress and Corporate Performance,"
Journal of Finance, 49, 3, 1015-1040.

[73] Petersen, M. A. (2007) "Estimating Standard Errors in Finance Panel Data Sets: Com-
paring Approaches," NBER Working Paper 11280.

[74] Petersen, M. A. and R. Rajan (1994) “The Beneﬁts of Lending Relationships: Evi-
dence from Small Business Data,” Journal of Finance, 49, 1, 3-37.

[75] Petersen, M. A. and R. Rajan (1997) "Trade Credit: Theories and Evidence," Review
of Financial Studies, 10, 3, 661-691.

[76] Pulvino, T. C. (1998) "Do Asset Fire Sales Exist? An Empirical Investigation of
Commercial Aircraﬂ Transactions," Journal of Finance, 53, 3, 939-978.

[77] Rajan, R. G. and L. Zingales (1995) "What Do We Know about Capital Structure?
Some Evidence from International Data," Journal of Finance, 50, 5, 1421-1460.

[78] Ramey V. and M. Shapiro (1998) “Capital Churning,” Unpublished Working Paper,
University of California San Diego.

[79] Rivers, D., and Q. Vuong (2002). "Model Selection Tests for Nonlinear Dynamic
Models," Econometrics Journal 5, 1-39.

[80] Rudebusch, G. and L. Svensson (1999) “Policy Rules for Inﬂation Targeting” in J. B.
Taylor, ed.: Monetary Policy Rules, University of Chicago Press, 203-246.

[81] Runkle, DE. (1987) “Vector Autoregressions and Reality,” Journal of Business and
Economic Statistics, 5, 4, 437-442.

[82] Sakakibara, E. (2000) The Day Japan and the World Shuddered: Establishment of
Cyber-Capitalism, Chuo-Koron Shin Sha.

[83] Sarno, L., and M. Taylor (2001). "Official Intervention in the Foreign Exchange Mar-
ket: Is It Effective and, If So, How Does It Work?," Journal of Economic Litera-
ture 36, 839-868.

[84] Servaes, H. (1991) “Tobin’s Q and the Gains ﬁ'om Takeovers,” Journal of Finance, 46,
1, 409-419.

[85] Shleifer, A. and R. W. Vishny (1992) "Liquidation Values and Debt Capacity: A
Market Equilibrium Approach," Journal of Finance, 47, 4, 1343-1366.

[86] Smith, J. K. (1987) "Trade Credit and Informational Asymmetry," Journal of Finance,
42, 4, 863-872.

166

[87] Song, J .-Y. and G. C. Philippatos (2004) "Have We Resolved Some Critical Issues Re-
lated To International Capital Structure?: Empirical Evidence From The 30 OECD
Countries," Working Paper, University of Findlay and University of Tennessee.

[88] Titman, S. and R. Wessels (1988) "The Determinants of Capital Structure Choice,"
Journal of Finance, 43, 1, 1-19.

[89] Vuong, Q. (1989). "Likelihood Ration Tests fOr Model Selection and Non-nested Hy-
potheses," Econometrica 64, 1067-1084.

[90] Wasmer, E. and P. Weil (2004) “The Macroeconomics of Labor and Credit Market
Imperfections,” American Economic Review, 94, 4, 944-963.

[91] Welch, I. (2004) "Capital Structure and Stock Returns," Journal of Political Economy,
112,106-131.

[92] Wooldridge, J. M. (2002) Econometric Analysis of Cross Section and Panel Data,
MIT Press.

167

 

 

 

11

1111111111 1
7231

STA (E 1JNWER
1
1293 02956

1
I

111111

11111

 

 

 

. .5
x:

..
14.3.
s.

:.:..... ..