FRM二級(jí)培訓(xùn)項(xiàng)目：市場(chǎng)風(fēng)險(xiǎn)測(cè)量與管理（閱讀版）

「市場(chǎng)風(fēng)險(xiǎn)測(cè)量V

\與管理Z

FRMPartIIProgram■基礎(chǔ)班

講師：CrystalGao

e［史由+葉間晞的|hProfQuiomGsvn

TopicWeightingsinFRMPartII

SessionNO.Content%

Session1MarketRiskMeasurementandManagement20

Session2CreditRiskMeasurementandManagement20

Session3OperationalRiskandResiliency20

LiquidityandTreasuryRiskMeasurementand

Session415

Management

Session5RiskManagementandInvestmentManagement15

Session6CurrentIssuesinFinancialMarket10

2-201

行業(yè)?創(chuàng)新?憎值

ModelingDependence:CorrelationsAnd

Copulas

⑥Framework?SomeCorrelationBasics

i?EmpiricalPropertiesofCorrelation

、MarketRiskMeasurement

\/?FinancialCorrelationModeling

andManagement/EmpiricalApproachestoRiskMetricsand

Hedges

TermStructureModelsofInterestRates

?TheScienceofTermStructureModels

rVaRandotherRiskMeasures?TheEvolutionofShortRatesandthe

?ParametricApproachesShapeoftheTermStructure

?Non-parametricApproaches?TheArtofTermStructureModels:

?Semi-parametricApproachesDrift

?Extremevalue?TheArtofTermStructureModels:

,BacktestingVaRVolatilityandDistribution

?VaRNappingVolatilitySmiles

,RiskMeasurementfortheTradingBook

3-201

VaRandotherRiskMeasures

4-201

行業(yè)?創(chuàng)新?憎值

Parametric

.

Approaches

VaRandotherRiskMeasures

5-201

?l.ProfitandLoss

>Profit/Loss

P/L=Pt+Dt-P1

>ArithmeticReturnData:

Pt+Dt—Pt-iPt+Dt

r=-----------------=----------1

tPP

t-it-i

jGeometricReturnData:

P+D

Rt=皿與t-t-)=ln(l+r)

vt

t-i

6-201

行業(yè)?創(chuàng)新?憎值

?l.ProfitandLoss

>Thedifferencebetweenthetworeturnsisnegligiblewhenbothreturnsare

small,butthedifferencegrowsasthereturnsgetbigger-whichistobe

expected,asthegeometricisalogfunctionofthearithmeticreturn.

>Sincewewouldexpectreturnstobelowovershortperiodsandhigher

overlongerperiods,thedifferencebetweenthetwotypesofreturnis

negligibleovershortperiodsbutpotentiallysubstantialoverlongerones.

7-201

行業(yè)?創(chuàng)新?憎值

?2.NormalVaR

>Approach1:NormalVaR

?Weassumethatarithmeticreturnsarenormallydistributedwithmean叩

andstandarddeviationo

VaR=-(n-zaa)VaR=-(|i-ZaO)P.i

-10

Profit(-t-Vloss(-)

8-201

行業(yè)?創(chuàng)新?憎值

?2.NormalVaR

圜>Example:

?Assumethattheprofit/lossdistributionforXYZisnormally

distributedwithanannualmeanof$16millionandastandard

deviationof$11million.CalculatetheVaRatthe95%and99%

confidencelevelsusingaparametricapproach.

VaR(5%)=-$16million+Sllmillionx1.65

=$2.15million

VaR(l%)=-$16million+Sllmillionx2.33

=$9.63million

9-201

行業(yè)?創(chuàng)新?憎值

?3.LognormalVaR

>LognormalVaR

?Assumethatgeometricreturnsarenormallydistributedwithmeanp

andstandarddeviationo.Thisassumptionimpliesthatthenatural

logarithmofPtisnormallydistributed,orthatPtitselfislognormally

distributed.NormallydistributedgeometricreturnsimplythattheVaRis

lognormallydistributed.07

VaR=1-

3

=64

3

zQW

O

VaR=(l-e^?)PJ

d3

t-iO6.

2

-08-06-04-02002040808

Loss(4>Vbrofit(-)

10-201

行業(yè)?創(chuàng)新?憎值

?3.LognormalVaR

圜,Example:

?Adiversifiedportfolioexhibitsanormallydistributedgeometric

returnwithmeanandstandarddeviationof11%and21%,

respectively.Calculatethe5%and1%lognormalVaRassumingthe

beginningperiodportfoliovalueis$100.

LognormalVaR(5%)-100x(1-e011-0-21x1-65)-$21.06

LognormalVaR(l%)=100x(1-e011-0-21x2-33)=$31.57

11-201

行業(yè)?創(chuàng)新?憎值

4.Quantile-QuantilePlots

>Weareinterestedinasking:

?Ifdatalooksrightwhenweuseparametricapproach?

?Whatwedois

JPlotourdataonahistogramandestimatetherelevantsummary

statistics.

/Considerwhatkindofdistributionmightfitourdata.

>Aplotofthequantilesoftheempiricaldistributionagainstthoseofsome

specifieddistribution.TheshapeoftheQQplottellsusalotabouthowthe

>Inparticular,iftheQQplotislinear,thenthespecifieddistributionfitsthe

data,andwehaveidentifiedthedistributiontowhichourdatabelong.

12-201

行業(yè)?創(chuàng)新?憎值

4.Quantile-QuantilePlots

4

3

2

8

=

c

1

cn

b

e-

-2O

d-

E-

山

-1

-2

13-201

行業(yè)?創(chuàng)新?憎值

?4.Quantile-QuantilePlots

-10

Normalquantiles

14-201

行業(yè)?創(chuàng)新?憎值

Non-parametric

Approaches

VaRandotherRiskMeasures

15-201

?l.HistoricalSimulation

>Allnon-parametricapproachesarebasedontheunderlyingassumptionthat

?Withnon-parametricmethods,therearenoproblemsdealingwith

va種甲nce-covarianciematrices,cursesofdimensionality；etc.~

Loss(+)/profit(-)

16-201

行業(yè)?創(chuàng)新?憎值

?l.HistoricalSimulation

>BootstrappedHistoricalSimulation

■Thebootstrapisveryintuitiveaodeasytoapply.

?Wecreatealargenumberofnewsamples,eachobservationofwhichis

obtainedbydrawingatrandomfromouroriginalsampleandreplacing

theobservationafterithasbeendrawn.

?Eachnew'resampled'samplegivesusanewVaRestimate,andwecan

takeour'best'estimatetobethemeanoftheseresample-based

estimates.Thesameapproachcanalsobeusedtoproduceresample-

basedESestimates-eachoneofwhichwouldbetheaverageofthe

lossesineachresampleexceedingtheresampleVaR—andour'best'ES

estimatewouldbethemeanoftheseestimates.

>Abootstrappedestimatewilloftenbemoreaccuratethana'raw'sample

estimate,andbootstrapsarealsousefulforgaugingtheprecisionofour

estimates.

17-201

行業(yè)?創(chuàng)新?憎值

?l.HistoricalSimulation

>DrawbacksofHS

?BasicHShasthepracticaldrawbackthatitonlyallowsustoestimate

VaRsatdiscreteconfidenceintervalsdeterminedbythesizeofourdata

set.

?Forinstance,theVaRatthe95.1%confidencelevelisaproblembecause

thereisnocorrespondinglossobservationtogowithit.

?Withnobservations,basicHSonlyallowsustoestimatetheVaRs

associatedwith,at-best,ndifferentconfidencelevels.

18-201

行業(yè)?創(chuàng)新?憎值

?l.HistoricalSimulation

>Non-parametricDensityEstimation

?Non-paQmetricdensityestimationoffersapotentialsolution.

?Drawinstraightlinesconnectingthemid-pointsatthetopofeach

histogrambar(Polygon).

?Treatingtheareaunderthelinesasapdfthenenablesustoestimate

VaRsatanyconfidencelevel.

(a)Originalhistogram(b)SurrogAfedensin*function

19-201

行業(yè)?創(chuàng)新?憎值

?2.ExpectedShortfall

>TheConditionalVaR(expectedshortfall)

?TheexpectedvalueofthelosswhenitexceedsVaR.

?Measurestheaverageofthelossconditionalonthefactthatitisgreater

thanVaR.

?CVaRindicatesthepotentiallossiftheportfoliois"hit"beyondVaR.

BecauseCVaRisanaverageofthetailloss,onecanshowthatitqualifies

asasubadditiveriskmeasure.

04

3

O.H

^

全o

z

wO.2

a

d

20-201

行業(yè)?創(chuàng)新?憎值

?2.ExpectedShortfall

圜,Example:

?Giventhefollowing30orderedpercentagereturnsofanasset:

-16,-14,-10z-7Z-7Z-5Z-4-—L-L0,0,0,L22Z4Z

6,7,8,9,11,12,12,14,18,21f23.

CalculatetheVaRandexpectedshortfallata90%confidencelevel:

?Solution:

VaR(90%)=7,ExpectedShortfall=13.3

21-201

行業(yè)?創(chuàng)新?憎值

?3.VaRvsES

>VaRcurveandEScurve:plotsofVaRorESagainsttheconfidencelevel.

22-201

行業(yè)?創(chuàng)新?憎值

?3.VaRvsES

>Thelongerthewindow,thesparsertheVaRcurve.

>TheVaRcurveisfairlyunsteady,asitdirectlyreflectstherandomnessof

individuallossobservations,buttheEScurveissmoother,becauseeach

ESisanaverageoftaillosses.

jAstheholdingperiodrises,thenumberofobservationsrapidlyfalls,

andwesoonfindthatwedon'thaveenoughdata.

>Evenifwehadaverylongrunofdata,theolderobservationsmight

haveverylittlerelevanceforcurrentmarketconditions.

23-201

行業(yè)?創(chuàng)新?憎值

?4.A/DofNon-parametricMethods

>Advantages

?Intuitiveandconceptuallysimple;

?Donotdependonparametricassumptions;

?Accommodateanytypeofposition;

?Noneedforcovariancematrices,nocursesofdimensionality;

?Usedatathatare(often)readilyavailable;

?Arecapableofconsiderablerefinementandpotentialimprovementif

wecombinethemwithparametric“add-ons“tomakethemsemi-

parametric.

24-201

行業(yè)?創(chuàng)新?憎值

?4.A/DofNon-parametricMethods

>Disadvantages

?Verydependentonthehistoricaldataset;

?Subjecttoghosteffect;

?Ifourdataperiodwasunusuallyquiet,non-parametricmethodswill

oftenproduceVaRorESestimatesthataretoolowfortheriskwe

actuallyfacing,viceversa;

?Havedifficulty(actslowly)handlingsh+fe(permanentriskchange)that

takeplaceduringoursampleperiod;

25-201

行業(yè)?創(chuàng)新?憎值

?4.A/DofNon-parametricMethods

?Havedifficultyhandlingextremevalue

/Ifourdatasetincorporatesextremelossesthatareunlikelytorecur,

theselossescandominatenon-parametricriskestimateseven

thoughwedon'texpectthemtorecur;

JMakenoallowanceforplausibleeventsthatmightoccur,butdid

notactuallyoccur,inoursampleperiod.

26-201

行業(yè)?創(chuàng)新?憎值

?4.A/DofNon-parametricMethods

>ProblemsfromLongWindow

?Thelongerthewindow:

/Thegreatertheproblemswithageddata;

?Thelongertheperiodoverwhichresultswillbedistortedby

unlikely-to-recurpastevents,andthelongerwewillhavetowaitfo『

/Themorethenewsincurrentmarketobservationsislikelytobe

drownedoutbyolderobservations;

/Thegreaterthepotentialfordata-<olleetioA-problems.

27-201

行業(yè)?創(chuàng)新?憎值

?5.CoherentRiskMeasures

>Acoherentriskmeasureisaweightedaverageofthequantilesofour

lossdistribution.

1

0=I0(P)P

0

?①(p)=weighingfunctionspecifiedbytheuser.

>ExponentialWeightingFunction

-(i-)/

J：thedegreeofourrisk-aversion

28-201

行業(yè)?創(chuàng)新?憎值

?5.CoherentRiskMeasures

jEstimatingexponentialspectralriskmeasuresasaweightedaverageof

VaRs(=0.05)

ConfidencelevelWeight

aVaR<P(a)xaVaR

(a)ct)(a)

10%-1.281600.0000

20%-0.841600.0000

30%-0.524400.0000

40%-0.25330.00010.0000

50%00.00090.0000

60%0.25330.00670.0017

70%0.52440.04960.0260

80%0.84160.36630.3083

90%1.28162.70673.4689

Riskmeasure=mean(0(a)timesaVaR)0.4226

29-201

行業(yè)?創(chuàng)新?憎值

?5.CoherentRiskMeasures

>Theestimatedoeseventuallyconvergetothetruevalueasngetslarge.

Estimatesofexponentialspectralcoherentrisk

measureasafunctionofthenumberoftailslices

Estimateofexponential

Numberoftailslices

spectralriskmeasure

100.4227

501.3739

1001.5853

5001.7896

10001.8197

50001.8461

10,0001.8498

50,0001.8529

100,0001.8533

500,0001.8536

30-201

行業(yè)?創(chuàng)新?憎值

Semi-parametric

Approaches

VaRandotherRiskMeasures

31-201

?l.Age-weightedHistoricalSimulation

>OnereturnobservationwillaffecteachoftheFieKW^-ebsewatieRS-inourP/L

series.Butafternperiodshavepassed,theobservationwillfalloutofthe

datasetusedtocalculatethecurrentHSP/Lseries,andwillthereafterhave

noeffectonP/L.

>Thisweightingstructurehasanumberofproblems.

?Oneproblemisthatit

samplepeHodthesameweight.

?Theequal-weightapproachcanalsomakeriskestimatesunresponsive

tomajorevents.

?Theequal-weightstructurealsopresumesthateachobservationinthe

sampleperiodisequallylikelyandindependentoftheothersovertime.

However,this'iid'assumptionisunrealistic.

32-201

行業(yè)?創(chuàng)新?憎值

?l.Age-weightedHistoricalSimulation

?Itisalsohardtojustifywhyanobservationshouldhaveaweightthat

suddenlygoestozerowhenitreachesagen.

?Ghosteffects

/wecanhaveaVaRthatisundulyhigh(orlow)becauseofasmall

clusterofhighlossobservations,orevenjustasinglehighloss,and

themeasuredVaRwillcontinuetobehigh(orlow)untilndaysorso

havepassedandtheobservationhasfallenoutofthesampleperiod.

33-201

行業(yè)?創(chuàng)新?憎值

?l.Age-weightedHistoricalSimulation

>Boudoukh,RichardsonandWhitelaw(BRW:1998)

?w⑴istheprobabilityweightgiventoanobservation1dayold.

?A入closeto1indicatesaslowrateofdecay,anda入farawayfrom1

indicatesahighrateofdecay.

A3(x)1A2(JO1入313]

|J4M3M21

入1（1—入|

3⑴+入3⑴+,?，+入吁1(x)(])=1T3。）=一二J

34-201

行業(yè)?創(chuàng)新?憎值

?l.Age-weightedHistoricalSimulation

>Majorattractions

?ItprovidesanicegeneralizationoftraditionalHS,becausewecan

regardtraditional屋asewithzerodecay,or入11.

?AlargelosseventwillreceiveahigherweightthanundertraditionalHSZ

andtheresultingnext-dayVaRwouldbehigherthanitwouldotherwise

havebeen.

?Helpstoreducedistortionscausedbyeventsthatareunlikelytorecur,

andhelpstoreduce

/Asanobservationages,itsprobabilityweightgraduallyfallsandits

influencediminishesgraduallyovertime.Whenitfinallyfallsoutof

thesampleperiod,itsweightwillfallfrom入MQ)tozero,insteadof

from1/ntozero.

35-201

行業(yè)?創(chuàng)新?憎值

?l.Age-weightedHistoricalSimulation

>Majorattractions

■Age-weightingallowsus

observation,soweneverthrowpotentiallyvaluableinformationaway.

Thiswouldimproveefficiencyandeliminateghosteffects,becausethere

wouldnolongerbeany“jumps"inoursampleresultingfromold

observationsbeingthrownaway.

36-201

行業(yè)?創(chuàng)新?憎值

?2.Volatility-weightedHistoricalSimulation

>HullandWhite(HW1998)

?WeadjustthehistoHcalretumstoreflecthowvolatilitytomorrowis

believedtohavechangedfromitspastvalues.

/rti=actualreturnforassetiondayt

Jat>i=volatilityforecastforassetiondayt

/aTi=currentforecastofvolatilityforasseti

37-201

行業(yè)?創(chuàng)新?憎值

?2.Volatility-weightedHistoricalSimulation

>Majorattractions

?Ittakesaccountofvolatilitychangesinanaturalanddirectway.

?Itproducesriskestimatesthatareappropriatelyseroitive4G-WTOfrt

volatilityestimates.

?ItallowsustoobtainVaRandESestimatesthatcanexceedthe

maximumlossinourhistoricaldataset.

/Inrecentperiodsofhighvolatility,historicalreturnsarescaled

upwards,andtheHSP/LseriesusedintheHWprocedurewillhave

valuesthatexceedactualhistoricallosses.

?ProducessuperiorVaRestimatestotheBRWone.

38-201

行業(yè)?創(chuàng)新?憎值

?3.Correlation-weightedhistoricalsimulation

>Correlation-weightedhistoricalsimulation

?Correlation-weightingisalittlemoreinvolvedthanvolatility-weighting.

?Toseetheprinciplesinvolved,supposeforthesakeofargumentthatwe

havealreadymadeanyvolatility-basedadjustmentstoourHSreturns

alongHull-Whitelines,butalsowishtoadjustthosereturnstoreflect

changesincorrelations.

39-201

行業(yè)?創(chuàng)新?憎值

?4?Filteredhistoricalsimulation

,Filteredhistoricalsimulation(FHS)

?CombineshistoricalsimulationmodelwithGARCHorAGARCHmodel.

>Thestepsareasfollows:

?Firstly,usethehistoricalreturntofindanysurpriseandthusreproduce

volatilitywithGARCHorAGARCHmodel.

?Secondly,thesevolatilityforecastsarethendividedintotherealized

returnstoproduceasetofstandardizedreturns,whichisLED..

?Thethirdstageinvolvesbootstrappingfromthesetofstandardized

returns.

?Finally,eachofthesesimulatedreturnsgivesusapossibleend-of-

tomorrowportfoliovalue,andacorrespondingpossibleloss,andwe

taketheVaRtobethelosscorrespondingtoourchosenconfidence

level.

40-201

行業(yè)?創(chuàng)新?憎值

?4?Filteredhistoricalsimulation

>Majorattractions

?Combinethenon-parametricattractionsofHSwithasophisticated(eg,

GARCH)treatmentofvolatility,andsotakeaccountofchangingmarket

?Itisfest,evenforlargeportfolios

estimatesthatcanexceedthemaximumhistoricallossinousdataset.

?Itmaintainsthecoirelationstructureinourreturn

?Itcanbemodifiedtotakeaccountofautocorrelationsinassetreturns

?ItcanbemodifiedtoproduceestimatesofVaRorESconfidence

intervals.

?ThereisevidencethatFHSworkswell.

41-201

行業(yè)?創(chuàng)新?憎值

Extremevalue

VaRandotherRiskMeasures

42-201

?l.Introduction

“Thefitteddistributionwilltendtoaccommodatethemorecentral

observations,ratherthantheextremeobservations,whicharemuch

sparser.

>Theestimationoftherisksassociatedwithlowfrequencyeventswithlimited

dataisinevitablyproblematic.

>Extreme-valuetheory(EVT):

?Centraltendencystatisticsaregovernedbycentrallimittheorems,but

centrallimittheoremsdonotapplytoextremes.Instead,extremesare

governedbyextreme-valuetheorems.

43-201

行業(yè)?創(chuàng)新?憎值

?2.GeneralizedExtremeValueDistribution

>SupposewehavearandomlossvariableXzandweassumetobeginwith

thatXisindependentandidenticallydistributed(iid)fromsomeunknown

distribution.ConsiderasampleofsizendrawnfromF(x)zandletthe

maximumofthissamplebeMnIfnislarge,wecanregardMnasanextreme

value.

>Underrelativelygeneralconditions,thecelebratedFisher-Tippetttheorem

thentellsusthatasngetslarge,thedistributionofextremes(i.e.zMn

convergestothefollowinggeneralizedextreme-value(GEV)distribution:

44-201

行業(yè)?創(chuàng)新?憎值

?2.GeneralizedExtremeValueDistribution

>Thisdistributionhasthreeparameters.

x-U—x

exp[-(1+m丁)刃,"0

F(x)=Ix°_

exp[-exp(-----=0

r“thelocationparameterofthelimitingdistribution,whichisameasureofthe

centraltendencyofMn.

r,thescaleparameterofthelimitingdistribution,whichisameasureofthe

dispersionofMn.

r,thetailindex,givesanindicationoftheshape(orheaviness)ofthetailofthe

limitingdistribution.

?When5>0:Frechetdistribution,heavytails,I次et-dist,Paretodist.

?When5=0:Gumbeldistribution,lighttails,likenormalorlognormaldist.

■When5<0:Weibulldistribution,verylighttails,notusefulformodelling

financialreturns.

45-201

行業(yè)?創(chuàng)新?憎值

?2.GeneralizedExtremeValueDistribution

三

S

U

E

>

三

q一

R

q

o

」

d

46-201

行業(yè)?創(chuàng)新?憎值

?2.GeneralizedExtremeValueDistribution

>HowdowechoosebetweentheGumbelandtheFrechet?

?WechoosetheEVdistributiontowhichtheextremesfromtheparent

distributionwilltend.

?Wecouldtestthesignificanceofthetailindex,andwemightchoose

theGumbelifthetailindexwasinsignificantandtheFrechetotherwise.

?Giventhedangersofmodelrisk,theestimatedriskmeasureincreases

withthetailindex,asaferoptionisalwaystochoosetheFrechet.

47-201

行業(yè)?創(chuàng)新?憎值

?2.GeneralizedExtremeValueDistribution

>EstimationofEVP