Industry Classifications and Return ComovementFAJ Financial Analysts Journal Volume 63 • Number 6...

FAJ Financial Analysts JournalVolume 63 • Number 6©2007, CFA Institute

Industry Classifications andReturn Comovement

Louis K.C. Chan, Josef Lakonishok, and Bhaskaran Swaminathan

A company's industry iijfiliation is commonly used to construct homogeneous stock groupings forportfolio risk management, relative valuation, and peer-group comparisons. A variety of industryclassification systems have been adopted, however, creating disagreements as to companies' industryassignments. This analysis of the Global Industry Classification System (GICS) and Fama-Frcnchsystem indicates that common movement in returns and operating performance resulting fromindustry effects is stronger for stocks of large companies than for those of small companies. Also,increasingly fine levels of disaggregation improve discrimination up to six-digit GICS codes, afterwhich the benefits tail off. Stock groupings based on industry exhibit stronger out-of-samplehomogeneity than groups formed from statistical cluster analysis.

Financial researchers, analysts, and decisionmakers frequently grapple with the issue ofidentifying homogeneous groups of stocks.Researchers and analysts, for example, often

wish to analyze the consequences of such eventsas corporate reorganizations or changes in finan-cial and investment policies. A common procedureis to pair stocks in the sample with others that havenot experienced the event but are similar in allother respects. The behavior of the sample in ques-tion is then assessed against the reference group.In presenting financial statements to the public,corporate managers frequently assess how theircompanies are faring in terms of peer-group com-parisons. And many investors follow strategiesbased on identifying stocks that are matched alongeconomically relevant dimensions but that trade atdifferent valuations, which perhaps indicates rela-tive mispricing.

Acaden:iic researchers and investn^ent practi-tioners follow a variety of approaches to constructhomogeneous stock groupings. Purely statisticalprocedures can be applied to the problem. Earlyheuristic approaches for partitioning stocks intosimilar groups were proposed by Farrell (1974) andElton and Gruber (1970). Brown and Goetzmarm(1997) clustered mutual funds into distinct invest-

Louis K.C. Chan is Hoeft Professor of Finance ami JosefLakonishok is professor of finance at the University ofIllinois, Urtjaim-Champmgn. Bhaskaran Swaminathanis partner and director of research at LSV Asset Man-agement, Chicago, and professor of finance at CornellUniversity, Ithaca, New York.

ment style categories. Alternatively, stocks can beassigned to groups on the basis of a priori economicattributes, such as market capitalization or operat-ing performance.

Perhaps the most popular method of establish-ing sets of economically similar stocks is to followtheir industry affiliations. Numerous academicstudies provide evidence that industry influencescapture a large portion of the extra-market correla-tions in stock returns.^ Gonnor (1995), for instance,found that industry factors dominate fundamentalstock attributes in terms of ability to fit the in-samplebehavior of returns. In practice, research coverageby financial analysts is typically structured alongindustry classifications, as is the analysis carried outby such popular publications as the Value Linehtvestment Survey. Many quantitative risk modelswidely used by asset managers and consultants arebased on industry factors. Most notably, wheninvestment managers structure portfolios, they typ-ically take into consideration the industry affilia-tions of the component stocks. A portfolio that issubstantially overweight in an industry relative tothe benchmark may experience increased tracking-error volatility.

Nonetheless, deciding which companiesbelong to an industry is far from straightforward.Standard Industrial Glassification (SIG) codes,which aggregate companies selling related end-products or using similar production processes,have traditionally been used for this purpose.Ghanges in the variety of products, the growingimportance of services, and shifts in technologyand the makeup of businesses, however, have

56 www. cf apubs .org ©2007, CFA Institute

Industry Classifications and Return Comovement

called into question the usefulness of the SIC sys-tem {see Clarke 1989). The Fama and French (1997)system starts from companies' four-digit SIC codesand reorganizes them into 48 industry groupings.For instance, the Fama-French (FF) "automobilesand trucks" industry aggregates some companiesfrom SIC major (two-digit) groups—namely, 22,23,30, and 35—with some from the standard "trans-portation equipment" SIC group, 37. The FF classi-fication has been highly influential and is widelyused in academic studies of asset pricing (e.g., Bren-nan, Wang, and Xia 2004; Daniel and Titman 2006;Ferson and Harvey 1999; Hong, Torous, and Val-kanov 2007; Moskowitz and Grinblatt 1999; Pastorand Stambaugh 1999; Purnanandam and Swami-nathan 2005), corporate finance (e.g., Flannery andRangan 2006; Graham and Kumar 2006), account-ing (e.g., Chan, Frankel, and Kothari 2004; Francis,LaFond, Olsson, and Schipper 2005; Richardson2006), and economics (e.g., Bebchuk and Grinstein2005; VVulf 2002).

Fama and French (1997) did not provide anyevidence about how well their classification systemproduces groups of economically similar compa-nies. It is still an open issue, therefore, as to whetherthe FF specification of industries deserves its statusas the default choice in academic studies.

Another approach to industry classificationenjoys widespread use among investment practi-tioners. Portfolio managers and analysts have grav-itated to the Global Industry Classification System(GICS). Its categorization is based not only on acompany's operational characteristics but also oninformation about investors' perceptions of whatconstitutes the company's main line of business.

Because GICS codes consider investors' atti-tudes, the SIC and GICS schemes can disagreeabout a company's industry assignment. For exam-ple, GATX Corporation, which leases and operatesrailroad equipment and ships, is classified in thefinancial sector by GICS but is placed in the trans-portation equipment sector by SIC codes. Otherinstances of conflict arise in the case of severalcompanies that trade as real estate investmenttrusts (hence, GICS puts them in the financial sec-tor) but also manage timberlands or paper mills (sotheir SIC codes are those for the wood or paperproduct industries).

The GICS approach has received relativelyless coverage in the research literature, and itsvalidity has not been extensively documented (anotable exception is Bhojraj, Lee, and Oler 2003).Our aim is to resolve the disparity between aca-demic and practitioner approaches to industryclassification by comparing the FF industry group-ing system, which is based initially on SIC code.

with the GICS groupings. We evaluate eachmethod's ability to form groups of stocks whoseprice movements are related.

The existing literature also provides scantguidance on choosing between the different classi-fication levels available under the GICS standard.On the one hand, a coarse partitioning level (suchas two-digit GICS codes) blurs commonalities thatmay emerge more strongly in subgroups. On theother hand, finer partitions are not necessarily pref-erable: Past a point, the increasingly specific factorsthat account for homogeneity may become lessimportant than the general factors affecting themajority of companies in a broadly defined indus-try. Also, moving to a finer detail of classificationproduces groupings with fewer companies, mak-ing comparisons noisy and unreliable. The relevantquestion is how much is lost at each level of aggre-gation. A second objective of this article, therefore,is to provide evidence to aid in such a choice:Specifically, we document how each GICS aggre-gation level performs with respect to highlightingwithin-group similarity in stock price movements.

Our analysis measures homogeneity in termsof coincidence in stock price movements. Two com-panies that are economically similar may not expe-rience strong return covariation, however, overshort horizons. Their comovement may bedrowned out by the idiosyncratic portion ofreturns, including noneconomic forces, such asinvestor sentiment. As a check that our conclusionsabout industry classification are not limited by ourfocus on return correlations, we verify that the FFand GICS specifications also capture commonali-ties in underlying operating performance as prox-ied by sales growth. Finally, we use statisticalcluster analysis to form stock categories with highwithin-group return correlations. Because this pro-cedure overlooks industry affiliation entirely, itsperformance provides a yardstick for the useful-ness of the FF and GICS grouping methods.

In addition to analyzing the GICS approach,Bhojraj et al. (2003) considered various methodsfor assigning companies to industries, includingthe FF, SIC, and North American Industry Classi-fication System codes. They focused on the abilityof industry indices to capture the cross-sectionaldispersion in stock returns, valuation multiples,financial ratios, and growth rates. Unlike us, theydid notexaminehow correlations in returns and inoperating performance vary depending onwhether companies belong to the same or differentindustries. As a result, their findings are notdirectly applicable to issues in portfolio analysisand risk management. Moreover, they did not con-front industry-based grouping procedures with

November/December 2007 www.cfapubs.org 57

Financial Analysts Journal

alternative classifications that take no account ofcompanies' industry membership, nor did theyexamine different levels of detail in the GICS code.

MethodologyWe first describe our procedures for evaluating tbedegree to which each classification scheme isolatesgroups of companies that belong to an industryfrom companies that do not belong to that Industry.

Industry Classification Schemes. Manystudies follow the SIC scheme to partition compa-nies into industry groups. The scheme was estab-lished to categorize all industries in the U.S.economy and is currently administered by the U.S.Office of Management and tbe Budget. Starting from11 categories, successively finer partitions aredefined in terms of major groups (corresponding totbe first two digits of the code), industry groups(three-digit codes), and industries (four-digit codes).The SIC scheme aggregates into an industry compa-nies tbat use similar production processes or whoseproducts tend to be used or distributed together (seeEconomic Classification Policy Committee 1994).Accordingly, tbe groupings are intended to aid eco-nomic and marketing analysis and do not directlyaddress the concerns of investors.^

In contrast, the GICS scheme is expresslydesigned to cater to financial analysts and invest-ment managers. This classification, maintained byMorgan Stanley Capital International together withStandard & Poor's, identifies a company's principalbusiness activity by considering the sources of itsrevenues, its earnings, and the market perceptionof its business. The scheme comprises 10 sectors(corresponding to the leading two digits of thecode), which are broken down into 24 industrygroups (denoted by the first four digits of the code),64 industries (the first six digits of tbe code), andthen 139 subindustries (the full eight-digit code).

Fama and French (1997) drew up their own setof homogeneous stock groups. Starting from four-digit SIC codes, they categorize companies into 48industry sectors.

Comparing the Performance of IndustryClassification Systems. Ifequity market partic-ipants consider a set of companies closely related,then stocks in the group should experience coinci-dent movements in their stock returns. Thecomovement in their returns with stocks outsidethe group should be relatively weaker. Accord-ingly, we judge each classification system's abilityto produce homogeneous groupings by comparingthe magnitude of return correlations between

stocks in the same industry with the magnitude ofcorrelations between within-industry stocks andou tsi de-industry stocks.

The correlations were calculated as follows. LetK be the number of stocks in the sample. We applieda particular industry classification system (sucb asfour-digit GICS codes) to each stock / = 1, . . . , K toidentify its industry, !. Suppose stock i's industrycontains N stocks (inclusive). We averaged tbe pair-wise correlations between stock i's return and thereturn on eacb of the other members of its industry:''

(1)N-\

where p,y is the time-series correlation between thereturn on stocks / and /. Similarly, the average pair-wise correlation between stock i's return and thereturns of all otber stocks not in its industry is

(2)K-N

We then defined the average within-industrycorrelation over all stocks in the sample as

(3)K

and the average correlation between a stock andotber stocks not in its industry as

,-1 yu

K(4)

By comparing the values of p^ and ^j, we canassess the degree to which an industry classifica-tion distinguishes between similar and dissimilarstocks. Classification schemes tbat highlightgroups with strong commonalities will tend to pro-duce large positive differences between the within-industry and outside-industry correlations.

Data. To calculate correlations, we used returndata for all U.S.-listed domestic common equityissues available in the University of Chicago CRSPdatabase for the 1975-2004 sample period. Becausethe correlation structure of individual stock returnsmay not be stationary over time, we divided theperiod into nonoverlapping five-year blocks andreport results averaged over the six epochs. Foreach subperiod, we examined stocks that existedfor the full five years. To mitigate measurementerror problems in correlations arising from infre-quent trading and microstructure-related biasesresulting from bid-ask bounce, we required that astock trade at above $2 per share at tbe beginningof tbe subperiod.

58 www.cfapubs.org ©2007, CFA Institute


For eacb subperiod, based on the five years ofmonthly return data, we calculated the within-industry correlation, pj, and tbe outside-industrycorrelation, ^j. Companies were assigned to anindustry under the following classificationschemes, based on increasingly finer industrygroupings: two-digit GICS code, four-digit GICScode, six-digit GICS code, and eight-digit GICScode. In addition, we report results using tbe FFindustry groups. To improve the reliability of theresults, we required tbat any industry classifica-tion consist of at least three companies over tbesubperiod; any industries not meeting tbis crite-rion were dropped from tbe analysis.

ResultsWe present in this section the results of the proce-dures for evaluating classification schemes for large-capitalization and small-capitalization stocks, andto take tbe perspective of a delegated investmentmanager who is concerned with tracking error, wealso provide evidence at the level of returns net of amarket index.

Correlations for Large Companies. Somelines of business tend to be more narrowly definedand uniform, so the contrasts between within-industry and outside-industry correlations shouldbe more pronounced for companies engaged insuch activity. Accordingly, in presenting the results,we break out average correlations for subsets ofcompanies. To streamline the formatting across allthe tables, we organize the findings for subsets interms of two-digit GICS sectors.*' Commonality inreturn movements is likely to show up morestrongly for large companies because of their rela-tively stable behavior (see Chan, Karceski, andLakonishok 1999). Also, their return correlationsare less likely to be contaminated by infrequent ornonsyncbronous trading issues. Finally, becausetbe bulk of equity assets is concentrated in large-cap stocks, the issue of industry diversification intbis market segment is of particular concern toinvestors. Accordingly, we begin our analysis byfocusing in Table 1 on large companies, those tbatfall in the top six deciles when stocks are ranked bymarket capitalization based on NYSE breakpoints.

When industries are defined on the basis of two-digit GICS codes, companies in the same industryshare correlations averaging between 0.31 (for con-sumer staples. Sector 30) and 0.48 (for energy. Sector10). Correlations tend to be high also inside tbeutilities group (Sector 55) and financials group (Sec-tor 40), which suggests that these hnes of business

involve activities that are fairly uniform. The meanwithin-industry correlation is 0.39 when weightedby the number of companies in each industry or 0.38when each industry is equally weighted.

Correlations between companies in the sameindustry are uniformly higher than correlationsbetween tbose companies and companies outsidethe industry. Evidently, all the classificationschemes succeed in identifying clusters of homo-geneous large-cap stocks. For example, in theeight-digit GICS classification, the simple meanwitbin-industry correlation is 0.44 whereas the cor-responding mean outside-industry correlation is0.27, yielding an average difference of about 0.18across tbe sectors. A test that there is no differencebetween correlations inside and outside an indus-try under tbe eight-digit classification based on the10 sectors in Table 1 produced a f-statistic of 8.32.

In the FF classification system, the simplemean witbin-industry correlation is 0.40, comparedwith an outside-industry correlation of 0.26. Thedifference of 0.14 is of the same magnitude as thatbased on a four-digit GICS classification. Tbe FFprocedure achieves this level of discrimination byusing 48 categories, however, or twice as many asthe number of groups at the four-digit GICS level.Assigning companies to industries on the basis ofsix-digit GICS codes is more effective, in that itgenerates a higher average within-industry corre-lation (0.43) and a starker contrast (0.17) betweentbe inside- and outside-industry correlations (ver-sus 0.13 for the FF system). Put differently, basedon tbe FF definitions, about 16 percent (the squareof the within-industry correlation) of a stock'sreturn volatility, on average, is accounted for by thereturn on an industry peer. In six-digit GICSgroups, the percentage accounted for is 18.5 per-cent, representing a relative improvement inexplanatory power of 16 percent.

Up to a point, finer industry partitions generallysharpen the differences between within-industryand outside-industry correlations. Moving fromtwo-digit GICS codes to four-digit GICS codes has ameager effect on the difference (the simple meandifferences are 0.13 and 0.14, respectively), but tbemean difference rises to 0.17 under a six-digit GICSclassification and 0.18 under an eight-digit GICSclassification. Notably, for the more homogeneoussectors (energy, financials, and utilities), the parti-tioning based on six-digit GICS codes is generallyassociated with a jump in the differences withrespect to within- and outside-industry correlations.For example, in the case of companies within theutilities sector, the spread between the two correla-tions jumps from 0.26 to 0.28.



oo0)

w

•a

3

o

•o

0)

oc•DC

nc

oCC

uo(0

l §l i e

It(AC.0

'4-1

(0V

00

ise

3nQ.

Ol(0J-

' 00)•a0Nw

" — •

cks

0

55Q .

-Ca

0

u

o .a

go

D

f S ' - > o o O t - ' i — i f - o r ^ • — - r - o

o o o o d d d d d d d d i > ^

d d d d d d d d d d d d i r i

CO 1—' I— T—̂ 1—̂ 1—f \—' ^ - t— f N 1—I I—' r o

o o d d d d d d d d d d o d

d d

d d d d d d d d

d d d d d d d d

d d d d d d d d d d

r O f N r - J f M r o o J f N c ^

0 0 0 0 0 0 0 0 0 0

o o a ' t ^ i n t - i < - - i ( N f - r ^ t r j• ^ r O f O r o r O r O ' 1 ' " * r O ' i "c 5 d d c S i : i c i d d d < : 6

GO O*' " * ^r o r O r O f ^ ^ - ^ r O ' T ) -

d d d d d d d d

• : t " T f r ' ) r O r i ' ' + - : f " md d d d d d d d

•X3 OO CO

S

(X in

E Ep 3

ir>3->̂bO0

chno

l

0)

£0

10

£

afio

n

'-''c3E

udJ

i x; oc .5P cbo 0; O

i •£

OJ ( )J3

13 ;„ -=

QJ -k- ^

:2 O tl, -C

S||t

111 'be S

(U . = 1/1 Ul

CU S o

^5 O u5QJ b60.2

B ?o > cQJ oj

l^ §

Q. 2i E

C T3

E

(ft

cc

e

rau:n

?Ul

0

4 *

'.J

u

ui

u-̂OJ

1-0

a, c

.2. o J3

r .̂ ft l V qj fs - r

r* C " ^ UJ _CI

C S. o ra c^ ^ rlj > _ .

c u oj •£ aj

S •£ p •£ XC C flj C :^ '



Table 2 reports tbe mean correlations ar\d thedifferences for the subperiods. The individual sub-period results confirm the findings from Table 1. Inevery subperiod, the FF and four-digit GICSschemes are roughly equivalent. Singling out theturbulent period of 2000-2004, for instance, we findtbe simple mean witbin-industry correlations to be0.34 for FF and 0.36 for the four-digit GICS codes,representing differences of 0.17 and 0.19, respec-tively, from the outside-indusfry correlations. Animprovement occurs by going to the six-digit GICScodes, but little is gained from further detail Forthe 2000-04 period, the within-industry correlationfor six-digit GICS codes is 0.41 (a gain of 0.23 overthe outside-industry correlation) and for eight-digit GICS codes, 0.43 (a difference of 0.25 over theoutside-industry correlation).

Consistent with the findings of Campbell, Let-tau, Malkiel, and Xu (2001), Table 2 provides someevidence that tbe average stock's correlation witbother stocks has declined over time. For the 1975-79epoch, for example, the average outside-industrycorrelation under a two-digit GICS classification is0,36. For the 2000-04 epoch, the mean outside-industry correlation drops to 0.17. Fama and Frencb(2004) documented that the jump after 1980 in thenumber of newly listed stocks, which tend to exhibitlower profitability and higher growtb, on average,than more established stocks, is a contributing factorto the increase in idiosyncratic volatility.

Because the correlations reported in Tables 1and 2 are based on raw returns, they are directlyrelevant to investors whose concern is how a port-folio performs in terms of gross returns. Professionalinvestment managers are beld to a benchmark, how-ever, such as the S&P 500 Index. Consequently, theyare focused on structuring a portfolio's return andrisk versus tbe baseline index. From this standpoint,if stocks in an industry covary strongly in terms oftheir deviations from the benchmark, then any over-weighting of that industry will tend to raise themanager's exposure to the risk of straying from thereference portfolio (tracking error). Our approach iseasily modified to conform to a typical investmentmanager's perspective. Specifically, we measuredreturns in excess of the return on a general marketindex and calculated correlations based on monthlyexcess returns. We chose tbe equally weighted indexof stcxrks in the top six deciles by market capitaliza-tion based on NYSE breakpoints so as not to attachmore importance to some companies than others inthe benchmark. Table 3 reports how well eacb clas-sification system captures within-industry com-monality in return movements net of the index.

Deducting the market index from each stock'sreturn partially removes a prime source of perva-sive return movement. As a result, the averagecorrelations of excess returns in Table 3 tend to belower than the raw return correlations in Table 1.For instance, the simple mean correlation of rawreturns is 0.38 across sectors in the two-digit GICSclassification; the simple mean correlation is 0.17for excess returns.

After the effect of the market is removed, thecovariation resulting from industry is more appar-ent and the consequences of choosing differentclassification methods show up more clearly. Inthese circumstances, the GICS classification dis-plays a bigger advantage over the FF procedure.Correlations of excess returns under four-digitGICS codes are slightly higher than those in the FFgrouping method. Larger improvements in excessreturn correlations occur wben six-digit and eigbt-digit GICS classifications are used. The simplemean correlations move from 0.18 based on FFindustries to 0.20 based on four-digit codes, andthey rise to 0.24 for six-digit codes. In terms of theproportion of an average stock's extra-marketreturn variance explained by an industry peer'sexcess return, FF industries yield 3.24 percentwhereas six-digit GICS codes yield 5.26 percent, arelative improvement of about 78 percent.

To summarize, industry groups based onfour-digit GICS codes do about as well as the FFindustry aggregations in terms of capturing thecommonality in the returns of large-cap stocks. Atthe same time, the four-digit GICS scheme is moreparsimonious with respect to the number of clas-sifications (24 for four-digit GICS codes comparedwith 48 for FF). If a finer breakdown is the objec-tive, six- or eight-digit GICS categories arerougbly equivalent. They deliver similar levels ofimprovement over FF groupings on tbe basis ofhigher correlations between industry peers andlarger differentiation between tbe inside-industryand outside-industry correlations.

Correlations for Small Companies. Table 4reports correlations between an as erage small-capstock and otber small-cap stocks inside and outsideits industry and also provides the correspondingdifferences for the five industrial assignment crite-ria. (Small-cap stocks are defined as those fallingbelow the median market capitalization of NYSEissues.) In the case of small-cap stocks. Table 4shows that tbe comovement in returns associatedwith common industry affiliation is weak. Even atthe finest partitioning level of eigbt-digit GICScodes, the within-industry correlations for small-cap stocks average only 0.23, as opposed to 0.44 for



tn

0)

*O

tn3•acTC!E

0)DC

c(0

^ ^ CM

« in

I?̂6

ra .2

•OC

if

.2 o

t N

5.2o

re a>ai P2 w< - ICM

u

o

o

U

60 CL, ^

E <•?

00 00 K

5 3-S E

2 o

t/i -1

j j rS

2 ^1 I•E en

rn r- CJ ^

j> .5p "i- '̂Oi > .=I-. > en

00 00

d d

CO oc

d d

E t3

0, C

IN rj in

00 00rH rH

d d

c/i

• *

f^i x : ..=. -r̂cL .£P Q, «B ^ .B ^

I



Table 3. Average Pairwise Correlations between Individual 1Stocks' Excess Returns and ExcessReturns of Within-industry Stocks: Large-Cap Stocks (size deciles 5-10),

Description(GICS sector)

Energy (10)

M<iU'riais(15)

IiuiiistrialB(20)

Consumer discretionary (25)

Consumer staples (30)

Health care (35)

Financials (40)

Intitrmation technology (45)

Telecommunication services (50)

Utilities (55)

Weighted mean

Simple mean

GICS8

0.44

0.19

0.14

0.17

0.19

0.21

0.22

0.23

0.28

0.47

0.24

0.25

Average Correlation within Industry

GICS6

0.40

0.16

0.13

0.13

0.18

0.20

0,20

0.22

0.28

0.47

0.22

0.24

GICS4

0.35

0.11

0.07

0,11

0,17

0,16

0.19

0,20

0.24

0,41

0.19

0.20

GICS2

0,35

0.11

0.05

0.05

0.12

0.12

0.13

0,16

0,24

0,41

0.16

0.17

FF48

0,32

0.14

o.n •o.n0,15

0,18

0.19

0,15

0,13

0.39

0.18

0,18

,1975-2004

Average No,of Comp.inies

35

60

96

92

47

42

101

48

13

86

Notes: In eoch nonoverlapping five-year block of the period, each company in a sector was assigned to an industry on the basis ofone of five classification systems. The average correlation between the company's monthly excess stock return and the excess returnof every other member of its industry was calculated for the five-year period and averaged across all companies in the sector. Thedifference between the two average correlations was calculated. The results w^ere averaged over the six five-year periods. Excessreturns are measured as the stock's return net of the return on the equally weighted market index. For definitions and industryassignment, see Table I.

large-cap stocks (Table 1). Even for a fairly homo-geneous set of stocks, sucb as utilities, Ibe within-industry correlation across small-cap stocks is low(0.25) compared with the correlation for large util-ity companies (0.50). Perbaps more telling is thattbe average correlation between two small-capstocks belonging to the same industry is not asstrong as the average correlation between twolarge-cap stocks in different industries: The simplemean correlations in an eight-digit GICS classifica-tion are 0.23 for within-industry small-cap stocksbut 0.27 for outside-industry large-cap stocks.

Tbe low correlations suggest tbat the idiosyn-cratic component of returns is the main source ofvariation in small-cap returns. Another possibilityis that small-cap stocks may not be as efficientlypriced as large-cap stocks. Investors may tbus missany underlying economic commonality in small-cap stocks, which dilutes the correlation in thestocks' returns. In any event, altering tbe finenessof the industry partitions yields negligible changesin tbe witbin-industry correlations. The differencebetween the within- and outside-industry correla-tions in simple means is 0.04 under a four-digitGICS classification and 0.05 under six-digit oreight-digit GICS classifications.

Average within-industry correlations basedon excess returns for the small-cap sample arereported in Table 5. Returns are measured in excess

of tbe return on the equally-weighted portfolio ofall stocks with market capitalizations less than tbemedian NYSE breakpoint. Here as well as in Table4, successively finer partitions by GICS codes donot produce large differences in correlations forstocks within an industry, Tbe simple mean corre-lation varies from 0.08 for two-digit GICS codes to0.09 for eigbt-digit codes. FF industry groupingsyield roughly the same magnitude of excess returncorrelations (0.07) for tbe small-cap sample.

The basic message from the small-cap sampleis that, compared with the large-cap sample, thecovariation in both raw and excess returns is muchattenuated. Given the limited degree of comove-ment, the various levels of industry disaggregationall do about the same with respect to capturing thewithin-group correlations.

Industry Groupings andEconomic RelatednessThis section offers additional evidence to confirmthat the GICS and FF classification systems yieldgroups of economically related stocks. In particular,we check whether a mechanical grouping procedurethat ignores industry affiliation can do better interms of capturing the commonality in returns. Wealso show that the commonality in return move-ments witbin GlCS-based and FF-based groups mir-rors commonality in tbeir operating performance.



ta

3•aciiT3'ta

o

(A3T3C

o(0c3̂̂at

OCTlcnc

0)CC

(0

Is> o_c inc ai

> inQ) J_

li.2o

(A

8)9si<

2 t

o

LJ

o o o o o o o o oc o o o

CM CM t— <N•-; p q pd o d d 0 0 0 0 0 0

N N OO O O0 0 0

0 0 0 0

to ^ in ro op p p p rHd d d d d

0 0 0 0

0 0 0 0

fi-l •£) IT) tp p p pd d d d

o o o o o o a

t>. o o o0 0 0 0 0 0

fN TT (N f̂ '

r4 rj H (N

d d d d

fe ^

— .2

nT3

s =

_2 CO d .is

X LI: i

0 0 m

O O ro

0 0 0

0 0 0 0

CO 00 00

d d d d d d d d d d

0

CO LJ EC

d d d d d d d d d00 CO

d d

c 0

ON

0 0 0 0

CO

na>0

ocn

00

d d d d d d d d d d

0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0

o o o o o o o o o o

s .̂

" ' ^ n j • • - '

6c CL,

5 I



Table 5. Average Pairwise Correlations betweenReturns of Within-lndustry

Description(GICS sector)

Energy (10)

MatLTidls(15)

Industrials (20)

Consumer discretionary (25)

Consumer staples (30)

Health care (35)

Financials (40)

Information technology (45)

Telecommunication services (50)

Utilities (55)

Weighted mean ,

Simple mean

Note: See notes to Table .

G1CS8

0.19

0.03

0.02

0.04

0.03

0.04

0,13

0,06

0,06

0.34

0.07

0,09

ndividual Stocks'Stocks: Small-Cap Stocks (size

Average Correlation within Industi-)

GICS6

0,18

0.03

0,02

0,03

0.03

0.03

0,12

0.06

0.06

0.34

0.07

0.09

G1CS4

0,16

0.02

0.01

0,02

0.03

0.03

0.12

0.05

0.06

0.33

0.06

0,08

GICS2

0.16

0.02

0.01

0.01

0.02

0.03

0.10

0.04

0,06

0,33

0,05

0,08

Excess Returns and Excessdeciles 1-5), 1975-2004

FF48

0,14

0.03

0.01

0,03

0.02

0.03

0,12

0,03

0,04

0,29

0,tl6

0,07

of Companies

43

78

273

271

76

92

248

163

6

55

Comparisons with Statistical Clusters.Economic intuition supports the idea that compa-nies in the same industry share higher return cor-relations than companies in different industries.Criteria other than industry affiliation can be used,however, to categorize stocks into homogeneousgroups. For instance, classifications can be based onsuch company attributes as market capitalizationor valuation ratios. Whether a partition based onindustry yields strong within-group correlationsrelative to competing classifications is an openquestion. To give some assurance that it does, wechecked that the within-industry comovement ofstock returns is at least as strong as the results froma mechanically based classification scheme basedon statistical cluster analysis.

For this comparison, we divided the 1975-2004sample period into six nonoverlapping five-yearperiods. For each period, we used hierarchical clus-ter analysis to assign stocks to groups so as to min-imize the average within-group distance betweengroup members. Distance was measured as 1.00minus the correlation coefficient between the twostocks' returns. The number of clusters was set tomatch roughly the number of industries under theFF classification. Relative to the pseudo-industryclassification produced by cluster analysis, GICSindustry groupings can be determined (7 pn'or/with-out requiring a separate estimation period. To putthe comparison on an even footing, therefore, wecalculated correlations on an out-of-sample basis byusing returns for the following five years. Thisapproach let us assess whether GICS groupingssharpen the contrast between the within-group and

outside-group correlations relative to the resultsfrom cluster analysis. We repeated the process foreach nonoverlapping five-year block. Results aver-aged over ail the periods appear in Table 6.

By design, the clusters have high in-sample andwithin-group correlations of returns relative to theoutside-group correlations. For the sample of large-cap stocks (Panel A), over the "Estimation Period,"the average within-cluster correlation is0.51 and theaverage outside-cluster correlation is 0.29, for a dif-ference of 0.22. In the five-year "Testing Period" forthe large-cap sample, within-cluster correlationsfall sharply, to 0.39 on average, and exceed theoutside-cluster correlations by only 0,13. The mag-nitude of the within-group correlation, as well asthe level of discrimination relative to the outside-cluster correlation, is on a par with that provided bythe two-digit GICS groupings in Table 1. The statis-tical procedure reqviires 40 categories, however, toobtain this level of differentiation (versus 10 sectorswhen the two-digit GICS code is used).

For the small-company sample (Pane! B inTable 6), a spread of 0.15 is observable in the estima-tion period between within-cluster and outside-cluster correlations. On an out-of-sample basis,however, statistical clusters perform worse thanGICS codes. The average within-cluster correlationis 0.18, which is only marginally larger than theaverage outside-cluster correlation of 0.16. Becausethe variability in small-cap returns is predominantlyidiosyncratic in nature, the assignment of stocks tostatistical clusters is likely to be driven by the corre-lations of company-specific returns. The correlationstructure of idiosyncratic returns is unlikely to be

November/December 2007 vvww.cfapubs.org 6S


Tabie 6. Average Pairwise Correiations between individuai Stocks'Returns and Returns of Stocks within and outside StatisticalCiusters, 1975-2004

SamplePeriod

Average Correlation

Estimation Period

WithinCluster

A. Larj^e-cap stocks

1975-1984

1980-1989

1985-1994

1990-1999

1995-2004

Average

0.57

0,50

0.58

0,47

0.44

0,51

B. Small-cap stocks

1975-1984

1980-1989

1985-1994

1990-1999

1995-2004

Average

0,45

0,35

0,36

0,26

0.27

0,34

OutsideCluster

0.37

0.26

0.38

0.25

0.19

0.29

0.30

0,20

0,23

0.11

0,12

0,19

Difference

0.20

0.24

0.20

0.22

0.25

0.22

0,15

0,15

0,13

0,15

0,15

0,15

Testing Period

WithinCluster

0,39

0,52

0,37

0,32

0,34

0,39

0,24

0,26

0,14

0,13

0.12

0,18

OutsideCluster

0,26

0,39

0,25

0,19

0,21

0.26

0,22

0.25

0.12

0.12

0.10

0.16

Difference

0,13

0.13

0.12

0.13

0.13

0.13

0.02

0.01

0.02

0.01

0,02

0.02

Average No.of Companies

536

468

538

509

511

512

611

682

814

748

771

725

Nofes: In each nonoverlapping five-year block over the sampie period, hierarchical cluster analysis wasused to classify stocks on the basis of their monthly returns into one of 40 groups. Clusters were formedto maximizewithin-group Pearson correlations of returns and minimize correlations in returns betweengroups. Each cluster contains at least three companies. Given the assignment of companies to clusters,the correlation between each company's stock return and the return of every other member of its cluster,as well as its correlation with all other stocks not in its cluster, was calculated and averaged across allcompanies. This pr(x:edure was done for returns in the five-year block over which the clusteringalgorithm was applied (the estimation period) and in the subsequent five years (the testing period).The difference between the two average correlations was calculated. The results were then averagedover all estimation or testing periods.

stable over time. In comparison, within-industrycorrelations based on four-digit GICS codes average0.22 and are larger than outside-industry correla-tions by 0.04.

In summary, GICS codes compare favorablywith purely statistical classifications with respect toproducing homogeneous sets of stocks.

Commonaiity in Operating Performance.As further evidence that GICS groupings and FFindustries yield sets of economically related stocks,we look behind the commonality of return move-ments to see whether there is commonality in oper-ating performance across companies in an industry.Our measure of operating performance is thegrowth rate of sales, measured as the year-to-yearpercentage change in a trailing sum of sales overthe previous four quarters. Sales growth is animportant driver of profitability and behaves lesserratically than other fundamental indicators, suchas growth in earnings or dividends.

Specifically, suppose for company / in month t,the most recently available financial statement data

are for quarter q (assuming a publication delay ofthree months). Sales growth for four quarters,GS4Q, is then

GS4Q. = (5)

where salesia is net sales for company i in quarter q?As in previous tests, we calculated correlations ofsales growth ra tes between every pair of companiesand averaged correlations within and outside anindustry for various levels of GICS codes and forthe FF classification. Table 7 reports the results forthe large-company sample. (We are not reportingresults for small companies because of the limitedcommonality in returns of small-cap stocks withinan industry.)

Table 7 shows that companies in an industryexperience lower comovement in their sales growthrates than the covariation in their stock returns(Table 1). In every industry, the average salesgrowth correlation is smaller than the averagereturn correlation. In the case of utility stocks based

66 www.cf apubs, org ©2007, CFA Institute


w3

•D

•a

3

ocn

3•occ

"o

O(0

raCO• ocQ

sowora

uow"(53"O

OCM

c0)0)

4 - 1

(f)corao

oo.—

raQ.VO)rad̂)

(0

u0)• o0)

_N

m

a0

a.raO6

(A

uo(0

I ' r .

d d

Z 30 0 0

IN 00 h- f-.p p p r-;d d d d

O <-;d d

00 a-p pd d

in •*• (N cc -.o o

O i Dd d d d

2;pd

C^ * ^

d d d

p p p t-; rH pd d d d d d

?i 2 ?]

d d d d d d d d d

d d d d d d d d d

2 K3d d d

in in 1%.- r- 1̂d d in

o oI—I I—

d d

o od d

o o o o o o o o o

o o o o o o o o o

d d

2 M)

01•B

0)

bera

ceni

0

rate

•r o if!

i; o

3 .S ^

' iH3

C O

5;| |i -1Ji " cC VI ko o .2

il̂

•<C m « O O



on four-digit GICS codes, for example, the correla-tion of sales growth rates is 0.22 whereas the corre-lation of returns is 0.45. The simple mean correlationof sales growth for companies in an industry classi-fied by four-digit GICS codes is 0.22, compared witba simple n^ean return correlation of 0.40.

Nevertheless within-group correlations of salesgrowth rates exceed outside-group correlations. Forfour-digit GICS groups, for instance, the simplemean within-group correlation exceeds the outside-industry correlation by 12 percent. Tbese findingsboost confidence in the argument tbat GICS classi-fications create homogeneous groups of companiesthat share similar underlying economic features.

Tbe FF industries capture roughly the samedegree of commonality in sales growth as four-digitGICS groups do. For tbe FF categories, the simplemean within-group correlation of sales growth is0.23, or 0.13 bigher tban the outside-group correla-tion. Moving to six-digit GICS codes improves dis-crimination over FF groups by raising tbe spread incorrelations to 0.15; going to eigbt-digit GICS codesprovides little additional improvement.

In general, therefore, the results for correla-tions in sales growth echo the results in Table 1based on return correlations.

ConclusionWe have provided evidence to resolve tbe diver-gence in defining industry groups between aca-demic research and investment practice. Wecompared the GICS and FF grouping procedures interms of their ability to isolate common return move-ments of stocks within an industry relative tocomovements witb stocks outside tbe industry. Inaddition, our analysis documented the gains fromsuccessively finer industry partitioning. We also ver-ified that the industry groups correspond to collec-tions of economically similar companies in tworespects: (1) Industries are better at capturing out-of-sample return covariation tban are statistical clus-ters formed without regard to industry affiliationand (2) industries reflect common movement incompanies' underlying operating performance, asmeasured by sales growth.

Large-cap stocks that belong to the sameindustry classification, as defined by two-digitGICS codes, share a simple mean correlation of 0.38in returns, compared with a mean correlation of0.26 for stocks that belong to different industries.Measured net of an equally weighted market index.

return correlations average 0.17 for large-capstocks inside an industry defined by two-digitcodes. Tbe magnitudes of within-industry com-monality in movements of raw and excess returnshighlight the potential benefit of using industryaffiliation as one dimension for managing portfoliorisk and tracking-error volatility in tbe case of largecompanies. For smaller companies, however, thecomovement in returns associated with common-ality in industry membership is much less pro-nounced. In part, their common response toindustry effects may be drowned out by the highervolatility of small-cap returns.

The Fama-French categories are comparable tofour-digit GICS groups in terms of the magnitudeof average within-industry correlation. Both proce-dures deliver a simple mean within-group correla-tion oi 0.40 for large-cap stocks, corresponding toan improvement of 0.13 over the mean outside-group correlation. The FF method requires 48industries, however, whereas there are only 24four-digit GICS groups.

Finer levels of industry disaggregation tend todo better with respect to spreading out within-industry correlations versus outside-industry cor-relations. The benefits generally tail off, bowever,beyond six-digit GICS codes. For the sample oflarge-cap stocks, correlations between industrymembers differ from correlations between non-member companies by, on average, 0.13 at the two-digit level, 0.14 at the four-digit level, 0.17 at thesix-digit level, and 0.18 at the eight-digit level.

Companies in an industry share weakercomovement, on average, in their sales growthrates than in their returns. Nevertheless, affiliationin the same industry generally translates intoheightened covariation in sales growth. The meancorrelation in sales growtb between two large-capcompanies with the same four-digit GICS code is0.22, versus a correlation of 0.10 for two companiesin different groups. Tbe FF procedure, despiteusing twice the number of categories as four-digitGICS groups, yields about tbe same distinctionbetween within- and outside-industry correlations.Pseudo-industry groups formed from statisticalcluster analysis of stock returns do not match theperformance of industry classifications on an out-of-sample basis. Both GICS and FF industries havelarger within-group correlations and sharper dis-crimination over outside-group correlations.

This article qualifies for 1 PD credit.



Notes1. Studies of the influence of industry include Grinold, Rudd,

and Stefek (1989), King (1966), Lessard (1974), and Roll(1992).

2. Gm't'rnmt'nt agi-iicies nnd data vendors use their owncriteria to dcterniini.' which SIC code applies to a givencompany. As noted in Guenther and Rosman (1994) and inKahle and Walkling (1996), this practice frequently pro-duces classifications for the same company that conflictamonj; data providers.

3. Recently, the number of industries was expanded to 49 (com-puter software and computer hardware were split). Thecomposition of the industries is described in detail on Ken-neth French's website (http://mba.tuck.dartmouth.edu/pages/faculty/ken. french/Da ta_Library/changes_ind.html).

4. Other criteria for judging classification schemes are possi-ble. Fur example, Bhojraj et al. (2003) compared the abilityof different industry classification schemes to account forthe variability in the cross-section of returns, valuationmultiples, earnings growth, and financial ratios.

5. To avoid clutter, we do not write /(() to reflect the depen-dence of the industry assignment on the stock or /Vj to reflectthe variation in the number of companies among industries.

6. Note that this breakdown is solely for expository purposesand does not bias the findings for or against any of theschemes. To be specific, for each company in a two-digitGICS sector, we used one of the industry assignmentschemes to identify its pt-ers in the same industry. Allpairwise correlations between the company and its indus-try cohorts were averaged (Equation 1). The mean of thesestatistics over all companies in the two-digit GICS sectorwas then computed (Equation 3) and is reported as thewithin-industry correlation for this subset. In the same way,we applied Equation 2 and Equation 4 to calculate the meanof the average pairwise correlations between a companyand all others not in its industry; this result is reported asthe outside-industry correlation for the subset.

7. If quarterly data were unavailable, we measured the per-centage change in the most recently available annual netsales relative to a year ago, still under the assumption of athree-month delay before the release of accounting data.

ReferencesIk'bchuk, Lucian, and Yaniv Grinstein. 2005. "The Growth ofExecutive Pay." Oxford Review of Economic Policy, vol. 21, no. 2

(Summer) :283-303.

Bhojraj, Sanjeev, Charles M.C. Lee, and Derek K. Oler. 2003."What's My Line? A Comparison of Industry ClassificationSchemes for Capital Market Research." Jourmil of AccountingRi'^eiircli, vol. 41, no. 5 (December):745-774.

Brennan, Michael J., Ashley W. Wang, and Yihong Xia. 2004."Estimation and Test of a Simple Model of Intertemporal Cap-ital Asset Pricing." ]oimuil of Finance, vol. 59, no. 4 (August):1743-1776.

Brown, Stephen J., and William N. Goetzmann. 1997. "MutualFund Styles." }ourntil of Financial Economics, vol. 43, no. 3(March):373-399.

Campbell, John Y., Martin Lettau, Burton G. Malkiel, and YexiaoXu. 2001. "Ha\e Individual Stix'ks Become More Volatile? AnEmpirical Exploration of Idiosyncratic Risk." journal of Finance,vol. 56, no. 1 (February);l-43.

Chan, Liiuis K.C, Jason Karceski, and Josef Lakonishok. 1999."On Portfolio Optimization: Forecasting Covariances andChtKJsing the Risk Model." Review of Financial Sliidies, vol. 12,no. 5 (Winter):937-974.

Chan, Wesley S., Richard Frankel, and S.P. Kothari. 2004."Testing Behavioral Finance Theories Using Trends andConsistency in Financial Performance."/in(ni(i/o/'Aa"t)i/H//»^flHrfEcoinmiics. vol. 38 (December):3-50.

Clarke, Richard N. 1989. "SICs as Delineators of EconomicMarkets." journal of Business, vol. 62, no. 1 O^nuary): 17-31.

Connor, Gregor)'. 1995. "The Three Types of Factor Models: AComparison of Their Explanatory Power." Financial Analystsjournal, vol. 51, no. 3 (May/Iune):42-46.

Daniel, Kent, and Sheridan Titman. 2006. "Market Reactions toTangible and Intangible Infonnation." Journal of Finance, vol. 61,no. 4 (August); 1605-1643.

Economic Classification Policy Committee. 1994. Report No. 1:Economic Concepts Incorporated in the Stanttard IndustrialClassification Industries of tbe United States. Washington, DC:Bureau of Economic Analysis, U.S. Department of Commerce.

Elton, Edwin J., and Martin J. Gruber. 1970. "HomogeneousGroups and the Testing of Economic Hypotheses." journal ofFinancial and Quanlitative AuaU/sis, vol. 4, no. 5 (January):581-^2.

Fama, Eugene E., and Kenneth R. French. 1997. "Industry Costsof Equity." Journal of Financial Economics, vol. 43, no. 2(February): 153-193.

. 2004. "New Lists: Fundamentals and Survival Rates."journal of Financial Economics, vol. 73, no. 2 (August) :229-269.

Farrell, James, Jr. 1974. "Analyzing Covariation of Returns toDetermine Homogeneous Stock Groupings."/('KHM/ii/7J».sfK('ss,vol. 47, no. 2 (April):186-207.

Ferson, Wayne E., and Campbell R. Harvey. 1999. "Condition-ing Variables and the Cross Section of Stock Returns." journal ofFinance, vol. 54, no. 4 (August):1325-1360.

Flannery, Mark J., and Kasturi P. Rangan. 2006. "PartialAdjustment toward Target Capital Structures." journal ofFinancial Economics, vol. 79, no. 3 {March):469-506.

Francis, Jennifer, Ryan LaFond, Per Olsson, and KatherineSchipper. 2005. "The Market Pricing of Accruals Quality."Journal of Account ing ami Economics, vol. 39, no. 2 (June): 295-327.

Graham, John R., and Alok Kumar. 20t)6. "Do DividendClienteles Exist? Evidence on Dividend Preferences of RetailInvestors." Journal of Finance, vol. 61, no. 3 (June):1305-1336.

Grinold, Richard, Andrew Rudd, and Dan Stefek. 1989. "GlobalFactors: Fact or Fiction?" journal of Portfolio Management, vol. 16,no. 1 (Fall):79-88.

Guenther, David A., and Andrew J. Rosman, 1994. "Differencesbetween COMPUSTAT and CRSP SIC Codes and RelatedEffects on Research," journal of Accounting and Economics, vol. 18,no. 1 (July): 115-128,



Hong, Harrison, Walter Torous, and Rossen Valkanov. 2007."Do Industries Lead Stock Markets?" Journal of FinancialEconomics, vol. 83, no. 2 (February):367-*396.

Kahle, Kathleen M., and Ralph A. Walkling. 1996. "The Impactof Industry Classifications on Financial Research." Journal ofFinancial ami Quantitative Analysis, vol. 31, no. 3 (September):309-335.

King, Benjamin F. 1966. "Market and Indusby Factors in StockPrice Behavior." journal of Business, vol. 39, no. 1, pt. 2Ganuary):139-190.

Lessard, Donald R. 1974. "World, National, and IndustryFactors in Equity Returns." journal of Finance, vol. 29, no. 2(May):379-391.

Moskowitz, Tobias J., and Mark Grinblatt. 1999. "Do IndustriesExplain Momentum?" Journal of Finance, vol. 54, no. 4 (August):1249-1290.

Pastor, Lubos, .lnd Robert F. Stambaugh. 1999. "Costs of EquityCapital and Model Mispricing." Journal of Finance, vol. 54, no. 1(February ):67-121.

Purnanandam, Amiyatosh, and Bhaskaran Swaminathan. 2005."Do Stock Price.s React to SEO Announcements? Evidence fromSEO Valuation." Working paper, Cornell University.

Richardson, Scott. 2006. "Over-Investment of Free Cash Flow."Review of Accounting Studies, vol. 11, no. 2-3 Oune-September):159-189.

Roll, Richard. 1992. "Industrial Structure and the ComparativeBehavior of International Stock Market Indices." Joiirnal ofFinance, vol. 47, no. 1 (March):3-41.

Wulf, Julie. 2002. "Internal Capital Markets and Firm-LevelCompensation Incentives for Division Managers," Journal ofLabor Economics, vol. 20, no. 2, pt. 2 (April):S219-C262.


Industry Classifications and Return ComovementFAJ Financial Analysts Journal Volume 63 • Number 6...

Documents

Transcript of Industry Classifications and Return ComovementFAJ Financial Analysts Journal Volume 63 • Number 6...