Calculating
CAPM Beta
Inthispaper,wewilllookatthecapitalassetpricingmodel(CAPM),asimplebutwidelyusedfactor modelinfinance.CAPMsmainstrengthanditsprimaryweaknessisthatitassumesonesingle sourceofrisk(i.e.marketrisk)andthenbucketseverythingelseasidiosyncratic(i.e.nonsystematic). Thispaperwillpavethewaytomoreadvancedfactormodelingtechniquesincomingissues.
Wewillbeginbydiscussingtheunderlyingassumptions,definesystematicandidiosyncraticrisk,and outlinetheirinfluenceonthecovarianceamongassets.Next,usingasimpleregressionmodel,wewill attempttocomputetheCAPMsensitivityfactor(Beta)fortwodifferenttechstocks:MicrosoftandIBM. OurcoalinapplyingCAPMtothesetechstocksistocomputeeachassetssensitivity(i.e.Beta)tonon diversifiablemarketrisk.Todothat,wewilluseasimplelinearregressionmodel,thenanormalprocess tovalidatethemodelsassumptionsandensureitsstabilityoverthedatasample. Forsampledata,weusedthemonthlyreturnsbetweenJuly2001andMay2013(140observations).For themarketrisk,weselectedmonthlyreturnsoftheRussell3000Index,andforriskfree,weoptedfor the4weektreasurybills(TBILL)returns.
Background
Infinance,thecapitalassetpricingmodel(CAPM)isusedtodeterminetheappropriaterequiredrateof returnofanasset(oraportfolio).TheCAPMtakesintoaccounttheassetssensitivitytothenon diversifiablerisk(akasystematicormarketrisk).
T T E[ RiT ] RT f i ( E[ RM ] R f )
E[ RiT ] RT f
T E[ RM ] RT f
Where
E[ RiT ] istheexpectedreturnofanassetIoveraholdingperiodT.
RT f istheriskfreereturnovertheperiodT.
i isthesensitivityoftheassetsexcessreturnovertheexpectedexcessmarketreturn.
T E[RM ] istheexpectedmarketreturnoveraholdingperiodT.
T E[RM ] RT f isthemarketpremium(expectedexcessmarketreturn).
E[RiT ] RT f isreferredtoastheriskpremium(expectedexcessassetsreturn).Inotherwords,
theassetsriskpremiumequalsthemarketpremiummultipliedbyitsbeta.
CalculatingCAPMBetaTutorial
SpiderFinancialCorp,2013
Theequationabovedescribesasimplelinearregressionmodel(withzerointercept),betweenthe assetsexcessreturnsandtheexcessmarketreturn.
T T RiT RT f i ( RM R f ) i
i ~ i.i.d ~ N (0, 2 )
2 isoftenreferredtoastheidiosyncraticrisk(i.e.riskthatisspecifictotheassetitself,ratherthanthe
overallmarket). Finally,the i istheslope(sensitivity)andcanbeexpressedasfollows:
T Cov ( RiT , RM ) i T Var ( RM )
Furthermore,fortwoassets,thecovariancecanbecomputedusingCAPMasfollows:
Cov ( Ri , R j ) E[ Ri R j ] E[( Ri R f )( R j R f )]
Cov ( Ri , R j ) E[( i ( RM R f ) i )( j ( RM R f ) j )] Cov ( Ri , R j ) E[ i j ( RM R f ) 2 i ( RM R f ) j j ( RM R f ) i i j ] Cov ( Ri , R j ) i j E[( RM R f ) 2 ] i j Var ( RM )
BasedontheCAPM,thevariance(orrisk)ofeachassetconsistsoftwocomponents:systematicand idiosyncraticrisk.
T Var(RiT ) i2 Var(RM ) 2
Why do we care?
BasedontheCAPMtheory,wecancomputenotonlytheexpectedreturns,butalsoconstructa covariancematrixofthedifferentassets.Notethatthevarianceofeachassetconsistsoftwo components.
Case 1: Microsoft
MicrosoftCorporationdevelops,licenses,andsupportssoftwareproductsandservices,aswellas designingandsellinghardwareworldwide.Microsoftisapubliclytradedcompany,listedonNASDAQ withamarketcapitalof290B. LetsplotthemonthlyexcessreturnsofMicrosoftandRussell3000(marketproxy):
CalculatingCAPMBetaTutorial
SpiderFinancialCorp,2013
Next,weplotthescatterplotforthetwodatasetsanddrawalineartrendlinetooutlinethecorrelation betweenthetwo:
UsingthelinearregressionwizardinNumXL,designatethemonthlyexcessreturnsofMicrosoftasthe dependentvariable(Y)andthoseofRussell3000astheindependentvariable(i.e.X).
CalculatingCAPMBetaTutorial
SpiderFinancialCorp,2013
FromtheOptionstabintheregressiondialogbox,settheintercept/constantvaluetozero.
Note:Youmayleavetheintercept/constantfloating(i.e.unset)andtheregressionwillfindit insignificant.Tryit. Whenwearefinished,clickOK.Theregressionwizardwillgenerateseveraloutputtables.
CalculatingCAPMBetaTutorial
SpiderFinancialCorp,2013
Theregressionmodel(i.e.CAPM)isstatisticallysignificant(ANOVAtable)andcapturesabout40%of MSFTmonthlyexcessreturnvariance.TheBeta(i.e.Russell3000coefficient)hasanaveragevalueof 0.98withanerrorof0.10. Thisisgoodsofar,soletsexaminethestandardizedresidualsoftheregression(rightmosttable).The residualsexhibitapositiveskewandfattails,andthusitfailsthenormalitytest. Togetabetterideaabouttheresidualsdistribution,wecreatetheQQplotwithaGaussiantheoretical distribution:
CalculatingCAPMBetaTutorial 5 SpiderFinancialCorp,2013
TheQQPlotshowsasmalldeviationfromnormalityatpositivevalues(i.e.skew)andafatlefttail (negative). BeforewestartusingtheCAPMandourregressionbetatodeterminetheappropriaterequiredreturn ofMicrosoft,weshouldaskourselvesakeyfewquestionsfirst:
Q: Is the regression model stable? Does the Betas value significantly differ throughout the sample data?
A:Toanswerthisquestion,letsdividethesampledataintotwosubsets:dataset1includesall observationspriorto2008(~70observations)anddataset2coversobservationsstartingfromJanuary 2008toMay2013(~70observations). UsingtheRegressionStabilityTestWizardinNumXL,weconductthisimperativetest.Similartowhatwe didwiththeregressionwizard,theRussellsexcessreturnsaretheindependent(X)variable,andthe MSFTreturnsarethedependentvariable(Y).
IntheOptionstab,settheintercept/constanttozero.
CalculatingCAPMBetaTutorial
SpiderFinancialCorp,2013
Now,ClickOK.TheWizardgeneratesthestatisticalstabilitytestoutputtable.
TheBetavalueisstablethroughoutoursampledataset(2001to2013).Letscomputeandplotthe betavaluethroughoutthedataset.Theshadedareaisour95%confidenceinterval.
Q: Are the regressions standardized residuals serially (aka auto) correlated?
A:Thewhitenoisetestanswersthisspecificquestion,andisavailableintheNumXLstatisticalteststab.
CalculatingCAPMBetaTutorial
SpiderFinancialCorp,2013
IntheOptionstab,setthemaximumlagorderto12(1year).ClickOK.
Theresidualstimeseriesexhibitsnosignificantserialcorrelation. Sofar,wefoundthefollowing: ThemonthlyreturnsofMicrosoftstockhaveanaveragesensitivityof0.98withtheoverall market. Theresidual(akaidiosyncratic)risk(i.e. )isaround5.54%.
Q: Do we have observation(s) that significantly affect the regression more than others (i.e. Influential data)?
CalculatingCAPMBetaTutorial
SpiderFinancialCorp,2013
Toanswerthequestionabove,wecomputetheCooksdistanceforeachobservationinthesample 4 data.Furthermore,weusetheheuristicthresholdof toidentifythoseinfluentialpoints.Nisthe N numberofnonmissingvaluesinthedataset.
Tohandleinfluentialandatapoint,wedecidedtoremoveitbysettingtheMSFTreturnsto#N/A,thus removingtheobservationfromanyanalysis.Weremoveoneobservation(theonewiththehighest Cooksdistance)atatime,thenrecalculatetheCooksdistancefortheremainingdatapointsusingthe reduceddataset.Notethatthethresholdslightlyincreasesaswedropobservations.Wecontinuewith theprocessuntilnoapparentinfluentialdataisinsight.
CalculatingCAPMBetaTutorial
SpiderFinancialCorp,2013
4 Notethatthe thresholdisaheuristic,soweaccepteddatapointswhoseCooksdistanceisslightly N higherthanthethreshold.Recalculatingtheregression(SHIFT+F9),weobservethenewBetavalue (1.21)andregressionerror(5.07%).
PlottingtheCAPMBetavaluethroughoutthesampledata,weobservethattheBetaslightlychanges overtimeandistrendinglowerovertime.OnemayconcludethatMSFTssensitivitytomarketriskis goingdown,duetoitsmarketcaporthenatureofinvestmentthatthecompanyitselfisundertaking.
CalculatingCAPMBetaTutorial
10
SpiderFinancialCorp,2013
Case 2: IBM
InternationalBusinessMachines(IBM)Corporationprovidesinformationtechnology(IT)productsand servicesworldwide.Thecompanyoperatesinfivesegments:GlobalTechnologyServices,Global BusinessServices,Software,SystemsandTechnology,andGlobalFinancing.IBMispubliclytraded,listed onNYSEwithamarketcapof233B. LetsplottheIBMmonthlyexcessreturnsalongwiththeRussell3000(marketproxy)excessreturns.
Next,weplotthescatterplotforthetwodatasetsanddrawalineartrendlinetooutlinethecorrelation betweenthetwo.
CalculatingCAPMBetaTutorial
11
SpiderFinancialCorp,2013
Thetwoseriesdemonstrateastrongcorrelationbetweenthem.Again,usingtheRegressionWizard, designateIBMexcessreturnsasthedependentvariableandtheRussell3000astheindependent, settingtheintercept/constanttozero.
TheoutputtablesshowsimilarresultstowhatwesawwiththeMicrosoftcase.Letsexaminethe residualsdistributioncloserusingtheQQPlot.
CalculatingCAPMBetaTutorial
12
SpiderFinancialCorp,2013
TheQQplotexhibitspositiveskew,withaheavyfattailontheleft(negative)side. BeforewestartusingtheCAPMandourregressionbetatodeterminetheappropriaterequiredreturn ofMicrosoft,weoughttoaskourselvesakeyfewquestions:
Q: Is the regression model stable? Does the Betas value significantly differ throughout the sample data?
Again,welldividethedatasetinto2separatesubsets:dataset1includesallobservationspriorto 2008,anddataset2includesallobservationsstartingfromJanuary2008todate.UsingtheNumXL regressionstabilitytest,wespecifytheindependent(X)anddependentvariable(Y)valuesforeachdata set,settheintercepttozero,andclickOK.
Thetestfailed!Wehaveastructuralbreakinthedataset.ThiscanbeinterpretedastheBetavalue changedsignificantly.
CalculatingCAPMBetaTutorial
13
SpiderFinancialCorp,2013
Whatcanwedonow?LetsfirstplottheBetavalueovertimeinanattempttoidentifythepoint(s) wherestructuralchangecommenced.
TheIBMstockhasundergoneaBetastartingin2008.Thiscanbeduetointernalcompanypolicy change:typeofinvestment,particularmarketexposure,etc.Theimportantfacthereisthattheidentity oftheIBMstockmorphed(withrespecttoCAPM). Insum,weneedtotossawaytheobservationspriorto2008andusethelaterobservations(i.e.2008to May2013)toestimatetheCAPMBeta.
Examiningtheregressionoutputs(usingpost2008observations),theBetahasameanvalueof0.66. Furthermore,theresidualdiagnosistestsallpassed.Additionally,thenonsystematicrisk(i.e.regression standarderror)isaround4%. Inshort,theIBMstockmorphedfrombeingahighbetavalueabove1toavaluelowerthanone.
CalculatingCAPMBetaTutorial
14
SpiderFinancialCorp,2013
Q: Are the regressions standardized residuals serially (aka auto) correlated?
A:Thewhitenoisetestanswersthisspecificquestion,andisavailableinNumXLsstatisticalteststab.
Theresidualstimeseriesexhibitsnosignificantserialcorrelation.
Q: Do we have observation(s) that significantly affect the regression more than others (i.e. influential data)?
Toanswerthequestionabove,wecomputetheCooksdistanceforeachobservationinthesampledata post2008.
SimilartowhatwedidintheMicrosoftcase,weremovedinfluentialdatabysettingtheMSFTreturnsto #N/A,thusremovingtheobservationfromanyanalysis.Weremoveoneobservation(onewiththe highestcooksdistance)atatime,thenrecalculatetheCooksdistancefortheremainingdatapoints CalculatingCAPMBetaTutorial 15 SpiderFinancialCorp,2013
usingthereduceddataset.Notethatthresholdslightlyincreasesaswedropobservations.Wecontinue withtheprocess,untilnoapparentinfluentialdataisinsight.
Recalculatingtheregressionmodel:
Thenonsystematicerrordroppedto3.42%(from4.27%earlier),andalltheresidualsdiagnosistestsare passed.
CalculatingCAPMBetaTutorial
16
SpiderFinancialCorp,2013
PlottingtheCAPMbetavaluethroughoutthesampledata,weobservethattheBetaslightlychanges overtimeandistrendingupwardovertime.OnemayconcludethatMSFTssensitivitytomarketriskis goingup,duetothenatureofnewinvestmentthatthecompanyisundertaking.
CalculatingCAPMBetaTutorial
17
SpiderFinancialCorp,2013
Conclusion
Inthispaper,wedemonstratedtheprocessforcomputingtheCAPMBetafortwotechstock:IBMand MSFT. Inbothcases,weproposedasimplelinearregressionmodelforthestocksmonthlyexcessreturns versusthemonthlyexcessreturnsoftheRussell3000Index(marketproxy).Theregressionslopeisthe empiricalCAPMBetaandtheregressionstandarderrorisviewedasthestocksnonsystematic (idiosyncratic)error. Afterward,wecarriedonaplainregressionanalysisprocess:ANOVA,coefficientsvaluetest,residuals diagnosis,regressionstabilitytest,andinfluentialdataanalysis. ThecomputedCAPMBetasignificantlyimprovedaswecarriedourthoroughanalysistotheregression results. AlltoolsyouneedtocarryonthisexercisearepartofNumXL1.60Pro. TheCAPMisarelativelysimpleonefactormodel.Inlaterissues,welltacklemultifactors(e.g.Fama Frenchthree(3)factormodel(FFM),etc.),whichmayaddsomenumericalcomplexitywhilethebasic stepsandintuitionremainthesame.
CalculatingCAPMBetaTutorial
18
SpiderFinancialCorp,2013
