CFA Institute

The Three Types of Factor Models: A Comparison of Their Explanatory Power Author(s): Gregory Connor Reviewed work(s): Source: Financial Analysts Journal, Vol. 51, No. 3 (May - Jun., 1995), pp. 42-46 Published by: CFA Institute Stable URL: http://www.jstor.org/stable/4479845 . Accessed: 07/02/2012 13:51
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

CFA Institute is collaborating with JSTOR to digitize, preserve and extend access to Financial Analysts Journal.

http://www.jstor.org

The

Three

Types
Their

of

Factor

Mocels:

Comparison of

Explanatory Power

Connor Gregory
Multifactor models security of market returns bedivided threetypes:macroeconomic, can into fundamental,and statisticalfactor models.Macroeconomic factor models use observable and economic series,suchas inflation interest time rates,as measures thepervasive of shocks to securityreturns.Fundamental associated with factormodelsuse the returnsto portfolios such as dividend observed securityattributes yield, the book-to-market and industry ratio, identifiers. Statistical factormodelsderivetheirpervasive factors fromfactoranalysisof the paneldataset of securityreturns.Thispapercompares explanatory the powerof thesethree approaches U.S. equityreturns. for

W ith some blurringat the boundaries,multifactormodels of asset returns can be divided into three types: macroeconomic, statistical, and fundamental. Our empirical findings confirm the conventional wisdom that statisticalfactor models and fundamental factor models outperform macroeconomic factor models in terms of explanatory power. The findings also indicate that the fundamental factor model slightly outperforms the statistical factor model. This result is at first surprising, because statisticalfactormodels are estimated by maximizingexplanatorypower. So, how can an alternativeoutperformthem by this criterion?The explanation lies in the much larger number of external data sources used in fundamental factor models, particularlythe large set of industry dummies. Another empirical finding is that the marginal explanatorypower of a macroeconomicfactor model is zero when it is added to the fundamental factor model. This result may indicate that the fundamental factors (in some unknown combination) capture the same risk characteristicsas the macroeconomicfactors.

TYPESOF FACTOR MODELS

Macroeconomic factor modelsare the simplest and most intuitive type. They use observableeconomic time series as measures of the pervasive factorsin securityreturns.1Some of the macroeconomicvariables typically used as factors are inflation, the
GregoryConnor is directorof research, Europe,for BARRA International.

percentage change in industrial production, the excess returnto long-termgovernment bonds, and the realized return premium of low-grade corporate bonds relative to high-grade bonds. The random returnof each security is assumed to respond linearly to the macroeconomic shocks. As in all factor models, each security also has an assetspecificreturnunrelatedto the factors.A security's linear sensitivities to the factorsare called the factor betas of the security. A drawback to macroeconomic factormodels is that they requireidentification and measurement of all the pervasive shocks affecting security returns. A small number of pervasive sources of risk may exist, but without knowing exactly what they are, or lacking data to measure them, they are of little use in explaining returns. Statistical factormodelsuse various maximumlikelihood and principal-components-basedfactor analysis procedures on cross-sectional/time-series samples of security returns to identify the pervasive factors in returns. Macroeconomicand statistical factor models both estimate a firm's factor beta by time-series regression. Given the nature of security returns data, this limitation is substantial. Time-series regression requires a long and stable history of returns for a security to estimate the factor betas accurately. Fundamental factormodelsdo not require timeseries regression. They rely on the empiricalfinding that company attributes such as firm size, dividend yield, book-to-marketratio, and industry

42

FinancialAnalysts Journal / May-June 1995

classification explain a substantial proportion of common return. A fundamental factormodel uses observed company attributes as factor betas. The factors in a fundamental factor model are the realized returns to a set of mimicking portfolios designed to capture the marginal returns associated with a unit of exposure to each attribute.For example, the dividend yield factor is the realized return per extra unit of dividend yield, holding other attributes constant. In the case of a fundamental factor model, the factor betas are exogenously determined, firm-specificattributes rather than estimated sensitivities to random factors, and the factor returns are empiricallydetermined random returns associated with these various attributes.2

The relationship among the three types of models, outlined in Table 1, can be described in terms of inputs, estimation technique, and outputs. In all cases, factormodel estimation involves time-series regression or cross-sectionalregression or both. Some statistical factor models rely on nonlinear maximum-likelihood theory for their derivation, but the implementation the nonlinear of maximum-likelihood procedure is usually done through a recursive series of repeated cross-sectional/time-seriesregressions.3 The three types of factor models are not necessarily inconsistent. In the absence of estimation error and with no limits on data availability, the three models are simply restatements or (to use a technical term from factor modeling) rotationsof one another. In this eclectic view of the world, the three factor models are not in conflict and all can hold simultaneously. In a macroeconomicfactor model, the factors are defined by economic theory and observed externally to the security returns data. In a statistical factor model, the factors are estimated from the sample returns data by maximizing the fit of the model. Statistical factors, because they are essentially unlabeled statistical artifacts, can be recombined linearly without altering the model. Recombining a set of statistical factors linearly

produces an alternative set of statistical factors, equally valid, or a rotation of the original set. Suppose that both factor models are true-that is, the macroeconomicfactors capture all the pervasive movements in security returns and the statistical factors and the macroeconomic factors both are measured without error. In this case, the two factorrepresentationswill differonly by a rotation and the statistical factorscan be linearlyrecombined to be made identicalto the macroeconomic factors. A fundamentalfactormodel and the other two types are not theoretically inconsistent. Suppose that a fundamentalfactormodel correctlycaptures the individual assets' sensitivities to the pervasive risks in the economy, and suppose that a macroeconomic model also correctlycaptures the pervasive sources of risk. Then, the firm-specific attributesused in a fundamental factormodel could be combined to produce the factor betas from the macroeconomicfactor model. For example, a typical macroeconomic factor is term structure risk (often measured by the realized monthly returnon a long-term government bond portfolio minus the short-termgovernment bond return).The sensitivity of a security's return to this factor is the security's term structurebeta. Two typical fundamental risk attributes are firm leverage and dividend yield. So, for example, perhaps each security's dividend yield and firm leverage attributes could be linearly recombined to equal its term structure beta. This again is a type of rotation, because the attributes in the fundamental factor model can be linearly recombined to equal the factor betas in the macroeconomicfactor model.

EXPLANATORY POWER A DESCRIPTIVE AS STATISTIC

The theoreticalconsistency among the three types of models requires more emphasis on empirical work to choose among them. In a recent paper, Connor and Korajczykproposed a test for determining the number of statisticalfactorsin security returns.4 The intuition for the test is straightforward. Suppose that one wishes to test whether

Table 1. An Overview the Empirical of ProceduresForthe ThreeTypes of FactorModels

Factor Model Type Macroeconomic Statistical Fundamental Inputs Security returns and macroeconomic variables Security returns Security returns and security characteristics Estimation Technique Time-series regression Iterated time-series/crosssectional regression Cross-sectional regression Outputs Security factor betas Statistical factors and security factor betas Fundamental factors

Financial Analysts Journal / May-June 1995

43

five factors are sufficient to explain all the pervasive movements in security returns. Connor and Korajczyksuggested estimating both a five-factor model and a six-factormodel on the same data set. Their test relies on the difference in the crosssectional average of asset-specific variances between the two models. If five factors are enough, then this difference should be near zero because the sixth factor should have negligible explanatory power. The reason is that under the hypothesis of five factors, the sixth factoris a "pseudofactor."It can explain some of the asset-specific variance for a few securities, but its average explanatorypower across all the securities must be close to zero. If six factors are necessary, then the difference in the average asset-specific variance should be strictly positive because the sixth factoris a true factorand therefore has pervasive influence in the crosssection of assets. The increase in explanatory power from adding a factoris the basis for the test of whether the additional factor is needed. Let oj denote the total return variance of
security i and
o-d

Table 2. Explanatory Powerof FiveStatistical

Factors

Factor 1 2 3 4
5

Increasein Explanatory Power from Adding EachFactor 29.0% 3.5 3.1 1.8
1.7

All factors

39.0

denote the asset-specific return

variance (the part of i's total variance that is not

explained by the factors). Let ocand o,, denote the

averages of oi and o,,, respectively, across all the securities. Define the explanatorypower of a factor model as 1 minus the average asset-specific variance divided by the average total variance (explanatory power = 1 - a, /c). The explanatorypower of each individual factor in the factor model is the change in the explanatory power of the model when the factor is added to it. Table 2 shows the results from applying this test to a statisticalfactor model estimated on U.S. returns data over the 108-monthperiod from January 1985 through December 1993 for 779 highcapitalizationU.S. equities.5This table reproduces using a one of the results of Connor and Korajczyk somewhat different sample.6 Connor and Korajczyk mentioned that their measure for explanatory power is applicable to the types of factor models

other than statisticalfactormodels. The methodology can be used to test whether a proposed macroeconomic factor or fundamental factor has a pervasive influence (and so is a true factor)or has only limited explanatorypower (and so is a pseudofactor). Table 3 shows the same statistic applied to a macroeconomicfactor model with five chosen factors (inflation, term structure, industrial production, junk-bond premium, and unemployment) estimated on the same sample.7 In the first column, as in Table2, the explanatorypower statistic is applied with each macroeconomicfactorin turn used as the fifth factor. Note that both the lowgrade bond premium and the term structure premium must be included to get a reasonable fit in the model. If either of those two variables is dropped, the fit of the model falls dramatically. None of the other variables contributes much to the explanatorypower of the model. Because macroeconomic factors are not ordered, each can be treated as the first factor, and its explanatorypower when used alone can also be measured, as in the right-handcolumn of Table3. Inflation, which has almost no marginal explanatory power, given the other four variables, does have some explanatorypower when used alone.8 Table4 applies the test to a fundamentalfactor model using a set of BARRAfundamental factors consisting of 55 industry dummies plus 12 risk o mic Factors
Increasein Explanatory Power from Adding Each Factorto All the Others 0.0% 7.7 0.3 8.1 0.1 10.9

Powerof te Mac Table 3. The Explanatory

Factor Inflation Termstructure Industrialproduction Default premium Unemployment All five factors

Power from Using Explanatory EachFactorAlone 1.3% 1.1 0.5 2.4 -0.3

44

FinancialAnalysts Journal / May-June 1995

Table 4. The Explanatory Powerof te Fundamental Factors

Explanatory Power from Using EachFactorAlone 16.3% 4.3 2.8 1.4 1.4 3.0 2.2 1.5 2.5 0.9 0.7 2.2 2.9 Increasein Explanatory Power from Adding Each Factorto All the Others 18.0% 0.9 0.8 0.6 0.5 0.4 0.6 0.6 0.4 0.5 0.4 0.5 0.4 42.6

Factor Industries in Variability markets Success Size Tradeactivity Growth Earningsto price Book to price Earningsvariability Financialleverage Foreigninvestment Laborintensity Dividend yield All factors

indexes.9 The industry dummies are added to the model together as a single set of factorsratherthan one at a time. Otherwise, the table has the same form as Table 3. Note the high explanatorypower of the industry dummies.10 The statistic from Connor and Korajczykalso has a naturalapplicationto comparisonsacross the three types of factormodels. For example, the test can be used to determine whether a five-factor statisticalfactormodel explains all of the pervasive comovements in returns against the alternative that a set of macroeconomicfactors has pervasive explanatory power not captured by the statistical factors. All of the macroeconomic factors are added simultaneously as a set of pseudofactors. The difference in cross-sectional average assetspecific variancebefore and after the set of macroeconomic factors is added to the statistical factor model gives the test statistic. This same procedure can be applied symmetrically to each of the models compared with any other one. Comparing three models, one of each type, with one another requires six tests-each of the three models tested against the other two. The matrixin Table 5 shows the results. The first row,

drawing upon Tables 2 through 4, shows the explanatory power of each factor model considered separately. The next three rows show the additionalexplanatorypower attained by combining any two types of factor models. The column headings are the first models estimated, and the row headings are the second. This table contains much of interest, but I will note only two findings. First, the fundamental factor model slightly outperformsthe statisticalfactormodel. How can this happen when the statistical factor model is estimated by maximizing explanatory power? The answer lies in the much larger number of explanatory variables that can be used in a fundamental factormodel (especiallythe industry dummies). Of course, using more explanatoryvariablesrequires a large amount of external data, whereas the statisticalfactor model only requires returns data. The much largerdata set used by the fundamental factor model compensates for the fact that the statisticalfactormodel is estimated by maximizing explanatorypower. Another interesting finding is that the macroeconomic factor model has no marginal explanatory power when added to the fundamental factor

Powerof EachTypeof FactorModel Table 5. The Marginal Explanatory

FirstModel Second Model Total explanatorypower Macroeconomic Statistical Fundamental Macroeconomic 10.9% 31.0 43.0 Statistical 39.0% 38.2 45.6 Fundamental 42.6% 42.4 44.8

Financial Analysts Journal / May-June 1995

45

model. This result implies that the risk attributesin the fundamental factor model capture all the risk characteristics capturedby the macroeconomicfactor betas. It is not clear (and is left as a problem for future research) how to rotate the fundamental risk attributesto equate some combinationof them to the macroeconomic factor betas.11This future research might provide insight on how corporate characteristicsor industry categories are related to return sensitivity to various macroeconomic shocks. The statisticalfactormodel also eliminates all of the explanatorypower of the macroeconomic factormodel, but this result is less useful than that for the fundamental factors because the statistical factors have no theoreticalcontent.

CONCLUSION
We have compared the explanatorypower of the three types of factor models typically applied to

security market returns. A particularspecification of each of the three types of models was estimated, and the total and marginal explanatory power of each was examined. The statisticaland fundamental factormodels substantiallyoutperformthe macroeconomic factor model. The fundamental factor model slightly outperforms the statistical factor model. The comparison of explanatorypower is only one criterion by which to evaluate the relative worth of the three approaches to factor modeling. By other important criteria such as theoretical consistency and intuitive appeal, using a macroeconomic factor model is probably the strongest (ratherthan the weakest) of the three approaches. Thus, our results, although interesting and useful, are not the final word in choosing among the three
approaches.
12

FOOTNOTES
factormodels is 1. The seminal referencefor macroeconomic N.F. Chen, R. Roll, and S.A. Ross, "EconomicForcesand of the Stock Market," TheJournal Business,vol. 59, no. 3 (July 1986):383-404.The Chen, Roll, and Ross model has been extended and refined by, among others, M.A. Berry, E. Burmeister, and M.B. McElroy, "Sorting Out Risks vol. Journal, Analysts Using Known APTFactors,"Financial 1988):29-42;and K.C. Chan, N.F. 44, no. 2 (March/April Investigationof the Chen, and D. Hsieh, "An Exploratory vol. Economics, 14, no. of FirmSize Effect,"Journal Financial 3 (September1985):451-71. 2. Examplesof this type of factormodel include B.A. RosenComponents of Covariance in Seberg, "Extra-Market and Analyof curity Returns,"Journal Financial Quantitative S. sis, vol. 9, no. 2 (March1974):263-73; Beckers, R. Griof nold, A. Rudd, and D. Stefek, "TheRelativeImportance Common Factors across the European Equity Markets," and vol. 16, no. 1 (March1992): of Journal Banking Finance, 75-96; and R. Grinold and R.N. Kahn, "Multiple-Factor Guide to Models of Portfolio Risk," in A Practitioners Models(Charlottesville,Va.: The ResearchFoundaFactor tion of the Instituteof CharteredFinancialAnalysts, 1994): 59-78. Founda3. See B. Lehmanand D.A. Modest, "The Empirical of tions of the ArbitragePricingTheory,"Journal Financial Economics, 21, no. 2 (September1988):213-54. vol. "A 4. G. Connorand R.A. Korajczyk, Test for the Numberof of Factors in an ApproximateFactorModel," TheJournal vol. 48, no. 4 (September1993):1263-92. Finance, 5. The cross-sectionalsample consists of all securitiesin the BARRAhigh-capitalizationuniverse of U.S. equities that had complete historiesof monthly returnsover the sample time period. See Connor and Korajczyk,"A Test for the Number of Factors,"TableI, p. 1277. The inflation factoris the change in the naturallog of the consumer price index, the term structure factor is the differencebetween the return on a long-termgovernment bond index and the one-month Treasurybill return, the industrialproductionfactoris the change in the naturallog of the industrial production index, the default premium factor (the junk-bondpremium)is the differencebetween the return on a high-yield bond index and a long-term governmentbond index, and the unemployment factoris the change in the unemployment rate. All varianceestimates are correctedfor regressiondegrees of freedom, so the marginalcontributionof a variablecan be zero or negative. Fordetaileddescriptionsof the industrycategoriesand risk
indexes, see The United States EquityModel Handbook(Berke-

6. 7.

8.

9.

ley, Ca.: BARRAInc., 1994). 10. J.K. Kale, N.H. Hakansson, and W.G. Platt, in "Industry Factorsversus Other Factorsin Risk Prediction,"working paper, Universityof California,Berkeley,1991, also found that the industry attributesare the most powerful component of a fundamentalfactormodel. 11. A recent paper by E.F. Famaand K.R. French, "Common RiskFactorsin the Returnsof Stocksand Bonds,"Journal of uses Financial 1993):3-56, vol. Economics, 33, no. 1 (February a mixed approach that combines elements of macroeconomic and fundamentalfactormodels. 12. I would like to thank Ian Buckleyand MarielleDejong for researchassistance.

46

FinancialAnalysts Journal / May-June 1995