the rise in firm-level volatility - causes & consequences

8/8/2019 The Rise in Firm-Level Volatility - Causes & Consequences

1/63

This PDF is a selection from a published volume fromthe National Bureau of Economic Research

Volume Title: NBER Macroeconomics Annual 2005,Volume 20

Volume Author/Editor: Mark Gertler and KennethRogoff, editors

Volume Publisher: MIT Press

Volume ISBN: 0-262-07272-6

Volume URL: http://www.nber.org/books/gert06-1

Conference Date: April 8-9, 2005

Publication Date: April 2006

Title: The Rise in Firm-Level Volatility: Causes andConsequences

Author: Diego A. Comin, Thomas Philippon

URL: http://www.nber.org/chapters/c0072


2/63

The Rise in Firm-Level Volatility: Causes and

ConsequencesDieg o Comin and Thomas Philippon, New YorkUniversity

1 Introduction

Over the past thirty years, there has been a decline in aggregate volatil-ity (McConnell and Perez-Quiros 2000, Stock and W atson 2002). At thesame time, there has been a large increase in the volatility of firms(Comin 2000; Campbell, Lettau, Malkiel, and Xu 2001; Comin and

Mulani 2003; and Chaney , Gabaix, and Philippon 2002).Our paper has five parts. We first document the upward trend in

various measures of firm volatility. Second, we present a decomposi-tion of aggregate volatility between the average volatility of sectorsand the correlation of growth across sectors. This decomposition sug-gests that the decline in aggregate volatility is mostly due to a declinein the correlation growth rates across sectors.

Third, we explore whether the firm-level trend toward more vola-tility and the aggregate trend toward more stability are related, orwhether the two have moved in opposite directions by coincidence.The two trends appear to be related. We find that TFP growth inindustries where firms have become more volatile tends to be lesscorrelated with aggregate TFP growth. Across countries, there alsoseems to be a negative relationship between aggregate and firm-levelvolatility.

Fourth, we explore the potential explanations for the increase infirm-level volatility. We find support for the idea that firm volatilityhas increased because of higher competition in the goods market. Wefind that firm volatility increases after deregulation. We also find thatthe increase in firm-level volatility is correlated w ith h igh research anddevelopment (R&D) activity as well as more access to debt and equitymarkets. However, we find no evidence that sectors with more accessto external finance have become less correlated with the rest of the


3/63

168 Comin and Philippon

economy, while we do f ind evidence that sectors with larger increasesin R&D investment have become less correlated with the rest of the

economy.

2 Th e Facts

The decline in aggregate volatility has been docum ented by McConnelland Perez-Quiros (2000), Blanchard and Simon (2001), and Stock andWatson (2002). On the other hand, firm-level volatility has increased.Firm-level volatility can be measured using financial data or real data.Using financial data for the United States, Com in (2000) and Cam pbell,Lettau, Malkiel, and Xu (2001) document an increase in the volatility ofidiosyncratic stock returns. Using accounting data, Chaney, Gabaix,and Philippon (2002) and Comin and Mulani (2003) show an increasein the idiosyncratic vo latility of employm ent, sales, earnings, and capi-tal expenditures.

Throughout the paper, we will use aggregate data from the National

Income and Product Accounts and firm-level data from COMPUSTATand CRSP. We will also use the sectoral data set developed by Jorgen-son and Stiroh (from now on, KLEM data).1

2.1 Volatility: GD PVersus Firm SalesIn this section, we document the increase in firm volatility using realmeasures, like sales, employment, or capital expenditures. Our sampleincludes all the firms in COMPUSTAT w ith at least eleven consecutive

observations of the relevant variable. Table 3.1 contains the basic de-scriptive statistics for our sample.

Figure 3.1 shows the evolution of idiosyncratic and aggregate vola-tility. Aggregate volatility (of) is defined as the standard deviation ofthe annual growth rate(yt) of real GDP:

+5 i V 2

where yt is the average growth rate betweent 4 and t + 5. For eachfirm i, we compute the volatility of the growth rate of sales(yt J as:

1/2

- i n / > v / t + T , i / : , ! / Ky-^-l

+ 5X^ / \ 2

T=-4


4/63

The Rise in Firm-Level Volatility 169

Table 3.1Firm-Level Summary Statistics

Year

1955

1956

1957

1958

1959

1960

19611962

1963

1964

1965

1966

1967

1968

1969

1970

1971

1972

1973

1974

1975

1976

1977

1978

1979

1980

1981

1982

1983

1984

1985

19861987

1988

1989

1990

1991

N u m b e rof Firms

810

829

849

927

982

1,589

1,7271,952

2,171

2,351

2,506

2,680

2,861

3,450

3,633

3,705

3,898

4,073

4,502

6,110

6,1756,224

6,262

6,187

6,081

6,187

6,226

6,530

6,771

6,827

7,135

7,3947,448

7,295

7,202

7,239

7,375

Average RealSales

1.30

1.35

1.38

1.22

1.28

0.87

0.840.82

0.81

0.83

0.86

0.89

0.89

0.82

0.92

0.91

0.92

0.96

1.02

0.88

0.84

0.91

0.97

1.04

1.15

1.18

1.17

1.09

1.05

1.09

1.06

1.031.11

1.20

1.27

1.33

1.27

Median SalesVolatility

0.096

0.093

0.090

0.084

0.081

0.095

0.0980.099

0.099

0.098

0.100

0.108

0.114

0.120

0.122

0.128

0.141

0.139

0.134

0.139

0.1380.139

0.142

0.146

0.149

0.151

0.157

0.167

0.174

0.179

0.184

0.1880.190

0.192

0.187

0.181

0.175


5/63


Table 3.1(continued)

Year

1992

1993

1994

1995

1996

1997

19981999

2000

N u m b e rof Firms

7,786

8,907

9,288

10,101

10,282

10,020

10,28610,294

9,819

Average RealSales

1.22

1.11

1.17

1.18

1.23

1.33

1.391.51

1.76

Median SalesVolatility

0.171

0.160

0.163

0.172

0.180

0.197

0.2120.211

0.207

Av erag e sales in 2000 in billions of d ollars.

.2-

= - 1 5 -

.05-

Median Firm Volatility - GDP Volatility

r

-.04

.03OOTJ

.02 o

-.01

-0

1960 1970 1980Year

1990 2000

Figure 3.1GDP Versus Individual Firm Sales Volatili ty: 10-Year Centered Rolling Standard Devia-tion of Growth Rates


6/63


4 -

3 -

0 -

Median Firm Volatility75th Pctile25th Pctile

n^z---1960 1970 1980

Year1990 2000

Figure 3.2Distribution of Firm Volatility: 10-Year Centered Rolling Standard Deviation of SalesGrowth

We then take the median across all firms presen t in the sample at timetas our measure of typical firm volatility:

o{ = median; {


7/63


.2-

Frm V

a

ty

Excluding M&A Controlling for age and size

r1960 1970 1980 1990 2000

Year

Figure 3.3Firm. Volatility, Alternative Measures: Ten-Year Centered Rolling Standard Deviation ofSales Growth

problems, in which boundaries of organizations do not matter. Plantscould move among firms without any real consequences, yet firmswould appear to be volatile. Firms would simply no t be the right unitsof observation. One could perhaps argue that mergers and acquisitions(M&As) fall partly into the category of irrelevant ownership changes.Thus, as a robustness check, we are going to show that our results arenot d riven by M&As.

Figure 3.3 shows that the trend increase in firm volatility is notdriven by the entry of young and small firms, or by an upsurge inM&A activity.5 Another way to show that our results are economicallymeaningful is to show that they relate to results obtained in other datasets. Guvenen and Philippon (2005) show that firm volatility measuredacross industries in COMPUSTAT is a good predictor of both un-employment risk and wage inequality measured across the sameindustries in PSID. Comin, Groshen, and Rabin (2005) relate firm-level

volatility to wage volatility at the occupation level by taking advantageof a unique da ta set that contains firm-level and w orker-level informa-tion for a sample of firms in Ohio. They document a positive relation-ship between firm-level volatility and the volatility and dispersion ofwages at the occupation level. We will no t discuss these results further,bu t we note that they show that our m easures of volatility capture realeconomic risks, not just measurement error or sample composition bias.


8/63

The Rise in Firm-Level V olatility 173

.3 -

E

tPo

ty

Market value turnoverOperating income turnover

1

1960 1970 1980 1990 2000Year

Figure 3.4Turnover of Industry Leaders: 5-Year Ahead Exit Rate from Top20% of Industry

2.2 Turnover of Leaders Within Industries

The distribution of firm sizes is famously skewed, and a few firms ac-count for most of the sales in each industry. Thus, one might arguethat firm volatility is relevant only if it affects the industry leaders. Wedefine turnover in industryI at time t as the probability of leaving thetop quintile of the industry over a five-year period:

TopTurnu = P(Z it+5 < Z%*> \ Zit > Z^ m)

where Z!t is either opera ting incom e or m arket value of firmi at time t,and ZtPl w is the 80th percentile of the distribution ofZn at time t in in-dustry I(i). This measure is robust to the entry of small firms in theparticular industry. We then define average turnover as the median ofturnover across all industries.

Figure 3.4 shows the increase in turnover among leaders for bothoperating income and market value. There are too few firms in thesample in the 1950s to obtain a reasonable estimate of the probability,

so we also computed the correlation of ranking over time, using all thefirms and not only the top 20 percent. For a particular measure Z, wedefine:

RkCorr = CorrieI(rankljt(Zit),rankij(Zit+T))

where ranki^ t{Zit) is the rank of firm i in industry I at time t accordingto Z. The picture using market value or operating income is similar to


9/63


0 . 9 5

g 0.9oI 0.85

g 0-8 -Iin

cp 0.75'I

0.7-

0.65-

0.6

Correlations after 5 yearsCorrelations after 10 years

1950 1953 1956 1959 1962 1965 1968 1971 1974 1977 1980 1983 1986 1989 1992 1995

Year

Figure 3.5Correlation of Labor Productivity RankingsNote: Five and ten years ahead correlation of within sector ranking, based on sales peremployee.

the one in figure 3.4 and, for the sake of completeness, we present theresults based on labor productivity rankings.

Figure 3.5 shows the evolution of the ranking correlation of firms,

over five and ten years, based on labor produ ctivity. There has been aclear decline in the ranking correlations over time. We will return tothe interpretation of these findings when we discuss product marketcompetition.

2.3 Equity Return VolatilityReal data are probably more directly relevant for macroeconomics.However, there are at least two good reasons to explore financial data

as well. The first is that financial data will allow us to look at firm vol-atility before World War II. The second is that financial data can helpus disentangle risk from predictable variations in firm dynamics.

We start by looking at equity return s. Let r!; t; m be the return to share-holders of firm i in m onth m of year t, and let r be the monthly re-turn on the Value Weighted Index. All the returns come from CRSP.


10/63


0 -1930 1940 1950 1960 1970

Year1980 1990 2000

Figure 3.6The Declining Explanatory Power of CAPM: Mean R2 from CAPM on M onthly ReturnsNote: For each firm/year, the CAPM-beta is estimated using 12 monthly returns.

For each firm, we estimate the CAPM model over rolling windows ofthirty-six months:

rVW

4-i, V t,m ' for m = 1,...,12

We therefore allow /3t t to vary (smoothly) over time, as seems plausi-ble since we use data from 1926 to 2004. We take the m edian across allfirms /m on ths observations in yeart as our measure of idiosyncratic fi-nancial volatility:

a[m = median,,m(|e;-,t,m|)

The nice thing abou t monthly data is that it allows us to construct non-overlapping annual measures of firm volatility. We define the explana-tory power of the CAPM model as the share of total firm returnvolatility that one can explain with the m arket return, i.e., theR2 of the

CAPM regression.Figure 3.6 shows the historical decline in the explanatory power ofCAPM. CAPM used to explain 40 percent of firm returns before the1950s, but its explanatory power is now around 10 percent.R2 is theratio of two volatilities, however, and we also want to know what hashappened to the level of idiosyncratic volatility. Figure 3.7 show s a U-shaped pattern forafn. Firm volatility was high in the late 1920s, and it


11/63


.02-

1930 1940 1950 1960 1970 1980 1990 2000Year

Figure 3.7The Evolution of Idiosyncratic Return Volatility: Median Absolute Deviation of MonthlyResidual Firm ReturnsNote: Firm returns are CAPM-adjusted using betas estimated on 12 monthly returns.

increased dramatically during the market crash and the early years ofthe great depression. It then declined steadily from the mid-1930s tothe mid-1950s. At that point in time, we can make the link with thereal data presented in the previous section. Since the mid-1950s, bothreal and financial volatility have increased steadily, with large spikesaround the first oil shock and the rise and fall of the Internet bubble.For a discussion of the link between financial and real volatility at thefirm level, see Veronesi and Pastor (2003).

Finally, note that our measure of firm volatility falls from 2001 to2003. First, many firms have delisted from the stock exchanges, anddelisting is more common for small, risky firms. Second, holding con-stant the composition of the sample, there has been a decrease in firmvolatility. This is not unprecedented. The same happened in the early1990s, and we expect firm volatility to start increasing again in thenear future.

2.4 Credit Ratings and Credit SpreadsIf firms have really become more risky, then this should also bereflected in corporate bond spreads and corporate bond ratings. Forthe spread, we use Moody's seasoned Aaa corporate bond yield minusthe ten-year treasury rate. For bon d ratings, we use S&P long-term do-


12/63

The Rise in Firm-Level V olatility 177

2-

3 1-5(00)

O

(00)

1-

.5-

o-

Aaa spread over treasury Residual rating

A

-.5

-0

309wa.c01

,5 &to

- 1

-1 .5

1950 1960 1970 1980 1990 2000Year

Figure 3.8Average Credit Ratings and Credit SpreadsNote: Rating ranges from 2 (AAA) to 20 (CCC). Index adjusted for age, size, and in-dustry.

mestic issuer credit rating from COMPUSTAT, coded from 2 for AAAto 27 for D (default). We first regress the rating on firm-level character-istics (age, assets, sales, SIC code), and we then average the residualsacross firms. Figure 3.8 shows that the Aaa spread over treasury hasincreased overtime, and also that the average credit rating of firms in

COMPUSTAT has deteriorated. Both trends suggest an increase inrisk, consistent with the increase in cash flow volatility. For more onthis topic, see Campbell and Taksler (2003).

Historical default rates on corporate bonds have also varied a lotover time. The average default rate from 1900 to 1943 was 1.7 percent.It dropp ed to a mere 0.1 percent from 1945 to 1965 (Sylla 2002). It thenincrease again, to 0.64 percent between 1970 and 1985, and to 1.85percent between 1986 and 2001 (Moody's 2002). These evolutions are

also consistent with the importance of rating agencies. These agenciesplayed an important role before World War II, became largely irrele-vant in the 1950s and 1960s, and have regained their previous impor-tance in the past thirty years (Sylla 2002).

Conclusion1: Firm-level risk hasincreased over the past fifty years.Conclusion 2: Firm-level risk was higher in the 1920s and 1930s than in

the 1950s and 1960s.


13/63


3 Sectoral Evidence

We have established that the aggregate stabilization of the U.S. econ-omy has coincided w ith a large increase in firm-level risk. How ever, ina statistical sense, this is only one observation. Our goal in this sectionis to explore sectoral dynam ics and see how they relate to firm volatil-ity. We are first going to show that the decline in aggregate volatility isaccounted for by a decrease in the comovement of the different sectorsand not by a decrease in the average volatility of each sector. Second,we are going to show that sectors in which firms have become morevolatile have typically become less correlated with the aggregate. Sec-toral data comes from Jorgenson and Stiroh's 35 KLEM data set.6

3.1 Decomposition of Aggregate VolatilityWe now perform a decomposition of the aggregate variance of thegrow th rate of real value added , TFP, and real value added per workerinto sector variances and correlations. Letys t be the growth rate of the

particular variable in sectors at time t, and let coss

t

ec

be the share of salesfor sector s in the aggregate sales in the economy. Also, let V([ZT]^4)denote the variance of {Zf_4,Zf_3 ,... Z( , . . . Z t+4,Z f+5} for any ge-neric variable Z and Cou([ZT]^4, [YT]| 4) be the covariance between{Zt_4, Z f_3,. . . Z f , . . . Zt+4, Zt+5J and {Yt-4, Yt-3,... Yt,.By definition, the aggregate growth rate is:

Then, using the definition of the variance:

v(w - ~ E fo-X?-^ El u

T = f - 4 \ iLV r=

For simplicity, suppose that (off = cwssec for all the sectors i and all

years t. Then V([yT]^4) can be written as follows:7

Variance Component Covariance Component

Hence, the variance of the growth rate of aggregate sales is de-composed into two terms: the first is related to the sector level variance


14/63


of sales (variance component), and the second reflects the covariancesbetween the growth rates of sales at different sectors (covariance

component).The first two rows in figure 3.9 show the evolution of the variance

and covariance components of the variance of the growth rate of ag-gregate value added , aggregate value added per worker, and TFP. Thevariance com ponent of all three variables displays a hum p-shaped pat-tern over time, with no obvious decline over our sample period, 1959to 1996. On the other hand, for all three variables, we can observe thatthere has been a decline since the 1970s in the covariance of growthacross sectors. For value added per w orker and TFP, there has been animportant decline in the covariance of growth over our sample period,while for value added grow th there has been no trend.

For the three variables, the covariance component is substantiallylarger than the variance component. The difference in magnituderanges from twice larger (TFP growth) to an order of magnitude larger(value added growth). As a result, the relevant component for under-

standing the dynamics of aggregate volatility is the covariance ofgrowth across sectors.The covariance component is affected by the sectoral variance and by

the correlation of a sector with the others. To increase further our un-derstanding, we also compute the correlation component. Specifically,we define first the correlation of each sector with the other sectors:

(3.3)

Then we define aggregate correlation as a weighted average of the sec-toral correlations:

Corrf =

The third row in figure 3.9 shows a clear decline in aggregate correla-tion for value added, TFP, and value added per worker growth overtime. Hence, we conclude that, in order to understand the decline inaggregate volatility, we should try to understand what drives this de-cline in the correlation between sectors. The results presented in thissection are based on the KLEM sectoral data set. We have obtainedsimilar results for the decomposition of aggregate volatility us ing m an-ufacturing data from the BLS.


15/63

{V\ \

HNr O

o o o o o o o o o

the rise in firm-level volatility - causes & consequences

Documents