CFA二级前导班：框架介绍_道德 数量 经济 固收 组合.docx

前导课CFA二级培训项目讲师：Vincent而宜Mi城。"BaPr中村。唯TopicWeightingsinCFALevelIISessionNO.ContentWeightingsIStudySession1-2Ethics&ProfessionalStandards10-15StudySession3QuantitativeMethods5-10StudySession4Economics5-10StudySession5-6FinancialReportingandAnalysis10-15StudySession7-8CorporateFinance5-10StudySession9-11EquityValuation10-15StudySession12-13FixedIncome10-15StudySession14Derivatives5-10StudySession15AlternativeInvestments5-10StudySession16-17PortfolioManagement10-153-106专业创新i曾值二级学习方法>一级二级各科目权重比较科目一级权重二级权重Ethics&ProfessionalStandards1510-15QuantitativeMethods105-10Economics105-10FinancialStatementAnalysis1510-15CorporateFinance105-10EquityValuation610-15FixedIncome1110-15Derivatives115-10AlternativeInvestments65-10PortfolioManagement610-15II i90TSSP-BPUelSIeuOfSSaoJdOSaffiI三Ii901，9SylPOpoospBPUBISpu3PoUs>u<Sued、u-pUoQ:SS-dpuNSIunpuou0OspBPUBpu3sJoPoDIHspBpuels-euo.ssJO,ldPUe-ravz£ISSOSES-AFramework>SSlEthicalandProfessionalStandardsR2GuidanceforStandardsI-VII/1:Professionalism/II:Integrityofcapitalmarkets/III:Dutiestoclients/IV:Dutiestoemployers/V:Investmentanalysis,recommendations,andactions/VI:Conflictsofinterest/VII:ResponsibilitiesasaCFAInstitutememberorCFAcandidateKnowledgeofthelaw1. Professionalism遵守最严格法律;/了解与工作相关的法律、法规、准则和协会CodeandStandards：/不得故意参与违法行为；/有疑问咨询上司/合规部，不需要向政府部门举报，除非法律明文规定：/发现违规行为，自己不作为，视为违规；Independenceandobjectivity/区分礼物，基金经理的客户（事后）给的是小费，broker给基金经理或者上市公司给analyst都是要影响客观独立性的不能收，除非是modestgift：/基金公司（买方）施压分析师（卖方）：/防火墙应该隔离投行部与其他所有部门，考试特别强调与研究部的隔离；/ISSUerPaidreSearCh可行，但必须收flatfee,并进行披露：/差旅费自己出，除非商业工具无法到达的情况，可接受目标公司的一般安排；Misrepresentation/没有及时改正的打字错误:/不能胡说八道（关于个人资质、关于公司服务范围、模型结果没有交代清楚、业绩度拿和归因ProCeSs,比如不可以Cherrypicking、不能担保不该担疾而U攵益、使用夕卜部基金经理要披露、业绩benchmark选择不番当）；/不能抄袭，必须恰当引用：Misconduct/欺诈盗窃等不诚信行为：/喝酒Or醉酒；/任何伤害专业诚信、信誉、工作胜任能力的行为都违反misconduct；/与工作无关的个人信彳卬、政治j顷向的事长，不违反misconduct；/非欺诈导致的个人破产可以免责，档案期内要披露。Materialnonpublicinformation/重大必须同时具备消息来源可靠，以及对股价有重大影响（举例参见课件）；2. Integrity of capital markets非公开，向SeleaiVe对象披露不属于公开;Mosaic理论:/做市商掌握了MNl不可以停止做市，应该做消极对手方；/从事无风险套利交易时，若获得了MNI,除非公司有能力证明流程和记录规范才可以继续交易，否则要停止；Marketmanipulation/两种形式（关键看动机）：散布假消息、基于交易扭曲量价；/以避税为目的的交易，不违反市场操纵；/基于特定的交易策略，不违反市场操纵；/如果为了增加流动性，期货交易所会员，书面协议，并对外事先披露，则不算违规。Guidanc6forStandardsELl1.oyaltprudenceandcareFldUCiary需要履行extracare,higherstandard;J四类客户：individual,beneficiarymandainvestingpublic;J基于组合整体进行判断；_Softdollar(softCoinmiSSicln)政策，brokerage是客户的财产必须100%让客户直接受益Directbrokerage的情况下仍然有义务寻找bestexecution和bestprice;kVotingproxy必须i行，底非其于t价比考虑，要向客户披露ingproxypolicies;13DutyToClients卜FairdealingFair并非equal;可以有PrenIiUInleVelServiC4但必须披露，同时不伤害其他客户利益；J确保客户有公允的机会对投资推荐做出反应；J采取round-lotbasis,避免Odd4otdistributions;femlymemberacunts并非beneficiary要与其他客户-视同仁；J基于客户的IPSCrrttllW,蓟每年更新，投资重大变更前必须先修改ips；J必须分散化，基于组合整体角度考虑；J客户固执己见的交易，如果对整体影响不大可以答应。影响大必须修题客户不同意修改ips,就从管理账户中剥离资金交让客户自己管理；PerformancepresentationJ过去业绩不能暗示在未来可以达成；J业绩简报可以，但必须可以后续提供详细信息；J组合组：加权平均，终止的组合，所有相似风格组合，相关支持记录保存；Preservationcfnfidentiality保密针对客户：过去客户、现在客户、合格潜在客户；是否保密首先必须考虑法律规定。芸业创新增值10-1061.oyalty/IndePendentPraCtiCe是指与雇主业务在内容上、时间上、精力上相竞争的业务，必须告知雇主性质、expectedduration以及ComPenSatiOn并且得到许可；/离开雇主前不拿雇主一针一线，正式离职前不可以先开始拉原客户；/仅仅知道几个客户是可以的，但不可以背诵客户名单；Additionalcompensationarrangements4. DUtyToemPloyerS r/各方书面同意；/告知COmPenSatiOrl的性质、大概金额、duration:Responsibilityofsupervisors/下属违规上司就违反了主管的职责，除非表明已经充分尽责;/可以将工作指派其他人负责，但最终后果自负；/在接受领导岗位之前，必须确保公司有充分的合规程序，如果不合规一定要提出改进措施，公司改正好接受岗位；/如果发现违规，必须立刻行动起来彻底调查，限制当事人工作或者加强监管。Diligenceandreasonablebasis/写分析报告或者做投资推荐，必须分析宏观经济、行业、公司基本面等全部因素；/确保第三方信息勤勉尽责（四个方面）；/选择外部顾问要勤勉尽责（四个方面）；/如果是使用量化模型进行推介，必须真懂，包括输入变量、假设前提、局限性等；/研发量化模型，需要了解模型的方方面面，需要对模型进行测试；/集体报告如果不同意结论，但过程严谨仍然可以署名；/HOtiSSUeS没有尽责；Communicationwithclients/区分事实与观点；5. Investment/投资流程中的董大变更要及时告知投资者，如模型、投资决策流程、投资范围、投资限制、投资策略、关键人员改变等；/推荐可以是CaPSUleform,但只要投资者要，必须给出详细版本；/主星风险与局限性；Recordretention/纸质版电子版保存皆可；/前公司的record未经批准不能带走。如果没有带走支持数据，在新公司不能发布旧公司业绩或研报，除非通过公开信息重建;/遵守当地法律规定，当地没有规定的，协会建议保存7年。Disclosureofconflicts/利益冲突，指潜在可能伤害客户或者投资公众的，主要不针对雇主；/必须事先平实的语言告知；/个人交易持仓；投资标的公司担任董事；与投资标的公司有业务关系(做市商、企业融资等)：分析师与投行部之间：市场部与分析师；重大个人关系；/如果奖金激励与客户利益有冲突必须披露；PriorityOftransaction/Client>employer>individual(beneficialowner),间隔时间不能太短，十天半月才合适；/区分fam订yaccount与beneficialowner的区别；/个人交易要申报获得批准才可以进行；Referralfees6. Conflicts of interest/事先披露，方便客户判断介绍是否客观，以及服务的真实成本；/最少每季度向雇主披露介绍费的性质及金额。I三Il90TSIspo三w>ne±rIUerlASS2QuantitativeMethodsR4IntroductiontoLinearRegression/SimpleLinearRegression/Calculateandinterpretregressioncoefficient/Assumptionsofthelinearregressionmodel/Hypothesistestingabouttheregressioncoefficient/Confidenceintervalforaregressioncoefficient/AnalysisOfVariance(ANOVA)/CalculateandinterpretSEEandR-squared/Calculatethepredictedvalueforthedependentvariable/DescribelimitationsofregressionanalysisI三Il90TZIsque>>4B三Bnbq-MS-po、UoqBSJPBdSS-IU-3PouJBqKUSBa、0-0>ooELLJ、SqqBEPAAELUnQSsPSPBPUB*工SUoUJnssBUo-SSaj-d三nSlUpM=JOUUo-SS636-1dIUlUO-SSaJbaI-EOs-SBqIl一、UOSSHd三口工S(XSpoq3nEWwenbZSS41OMgEeUL.人I三Il90IoIEUoUBIlIOuJI.MUo-SS6工XlPQsBpqsojsell-BuOWPUOUONABUoRBISu>ooUOABWOUOlnBONSISBUOM-1B-D-G(H<S-Pon>-ssgJOIn<、(PUaIIB三0-PUBPU一另U=)PoPUal一、S-SA-BUVs一SIU二9HSPOEanEHnbZSS>oMEelt.人SS3QuantitativeMethodsR7MachineLearningOverviewofMachineLearningSupervisedMachineLearning+PenalizedRegression+SupportVectorMachine+K-11earestNeighbor+QassificationandRegressionTree+RandomForestUnsupervisedMachineLearning+PrincipalComponentsAnalysis+ClusteringNeuralnetwordeeplearningnets,andreifrcementlearning)芸业创新增值ASS3QuantitativeMethodsR8BigDataProjects/BigDataIntroduction/StructuredDataAnalysisConceptualizationofthemodelingtaskDatacollectionDatapreparationandwranglingDataexplorationModeltraining/UnstructuredDataAnalysisTextproblemformulationData(text)curationTextpreparationandwranglingTextexplorationModeltraining>SS3QuantitativeMethodsR9Excerptfrom,ProbabilisticApproaches:ScenarioAnaIysiszDecisionTrees,andSimulations"/Simulation/ComparingtheApproachesI三IlgoTZeS-SA-BUV-BaPanEmlSUnWBa6.sSM-Bu<BQP952£s-Bo-8JoqNISBN>!二工>l-m6£Il-七3>0l6uq-etp-Jw73£uB_|zuOSdAlISydo3三>-JSUMO-dxwA0Ild'l/0IdAl7Bn-BAd/IU/6-ls一S-SBqlOdAHTSydoll>-JJO>>>6u三o-dxUJPUeM>aHypothesisTestingabouttheRegressionCoefficientARegressioncoefficientconfidenceintervalS1±(tc×s)CInotincludethehypothesizedvalue,reject>SignificancetestforaregressioncoefficientH0:b1=0aTeststatistics:,=4-hypothesizedV出UeOf仇疗=入2Decisionrule:rejectH0if+tcritical<t,ort<-1criticalRejectionofthenullmeansthattheslopecoefficientisdifferentfromthehypothesizedvalueofb1>p-value:thesmallestsignificancelevelforwhichthenullhypothesiscanberejectedRejectH0ifp-value<FailtorejectH0ifp-value>HypothesisTestingATypeIerrorandTypeIIerrorTypeIerror:拒真，rejectthenullhypothesiswhenitsactuallytrue/Significancelevel():theprobabilityofmakingaTypeIerror/Significancelevel=P(TypeIerror)Type11error:取伪，failtorejectthenullhypothesiswhenit'sactuallyfalse/Powerofatest:theprobabilityofcorrectlyrejectingthenullHypothesiswhenitisfalse/Powerofatest=l-P(TypeIIerror)HypothesisTestingH0isactuallytrueH0isactuallyfalseDonotrejectH0CorreCtTypeIIerrorRejectH0P(TypeIerror)=thesignificancelevelCorTeCtPoweroftest=1-P(TypeIIerror)>Withotherconditionsunchanged,eithererrorprobabilityarisesatthecostoftheothererrorprobabilitydecreasing.>Howtoreducebotherrors?IncreasetheSampleSize.1.TypesofMachineLearningAMachinelearningisbroadlydividedintothreedistinctclassesoftechniques:supervisedlearning,unsupervisedlearning,anddeeplearning.SupervisedlearninguseslabeledtrainingdatatoguidetheMLprogramtowardsuperiorforecastingaccuracy./Labeleddataset:onethatcontainsmatchedsetsofobservedinputsandtheassociatedoutput.Inunsupervisedlearning,theMLprogramisnotgivenlabeledtrainingdata;instead,inputs(i.e.,features)areprovidedwithoutanyconclusionsaboutthoseinputs./Thealgorithmseekstodiscoverstructurewithinthedatathemselves.2.MLchallenge-Overfitting>OverfittingisanissuewithsupervisedMLthatresultswhenalargenumberoffeaturesareincludedinthedatasample,resultingthatthefittedalgorithmdoesfitwelltotrainingdatabutnotgeneralizewelltonewdata.Itresultsininaccuracyforecastsonoutofsampledata,randomnessismisperceivedtobeapattern/Whenamodelgeneralizeswell,itmeansthatthemodelretainsitsexplanatoryPoWerWherlitisappliedtonew(i.e.,out-of-sample)data.3.SupervisedML:K-NearestNeighborAK-nearestneighbor(KNN).Morecommonlyusedinclassification(butsometimesinregression),thistechniqueisusedtoclassifyanewobservationbyfindingsimilarities("nearness")betweenthisnewobservationandthetrainingsample.A.KNNVViHjNCzVObservation,K=IB.KNNWithNewObscruationfK三5K-NearestNeighborATwovitalconcernsTheresearcherspecifiesthevalueofkzthehyperparameter;triggeringthealgorithmtolookforthekobservationsinthesamplethatareclosesttothenewobservationthatisbeingclassified./Ifkistoosmallwillresultinahigherrorrate,/ifitistoolarge,itwilldilutetheresultbyaveragingacrosstoomanyoutcomes./Ifkiseven,theremaybeties,withnoclearwinner.Analystsneedtohaveaclearunderstandingofthedataandunderlyingbusinesstodefine"similar"(ornear).#袈.添花留SCBIPoWgUopPjOIdXEPQ寸tuouPPUPLMPUBUoI÷jPJBdaidWQEUoDoHOO3EQZ上昌MUyPPOUIq+,UoPBZyPn°UojTwd窃Aybj0u写Ou&sIp0三菖JOOSbXUlnSUO°PIdlUPX§Sn曰MM二PoUl-WCQ2署哲oalgcs包BP-sPA0AsS町4SLp.H3OjSeSAIFUVEQPaPamQOrLIQSIFQPa留寸StructuredDataAnalysisAStep3:Datapreparationandwrangling.ThisstepinvolvesCleaningthedatasetandPreParingitforthemodel.Datapreparation(Cleaning)istheprocessofexamining,identifying,andmitigatingerrorsinrawdata,includesaddressinganymissingvaluesorverificationofanyout-of-rangevalues.DataWrangling(Preprocessing)datamayperformstransformationsandcriticalprocessingstepsonthecleanseddatatomakethedatareadyforMLmodeltraining,involvingaggregating,filtering,orextractingrelevantvariables.Data Collection/ CurationModel Training* ResultsStructuredDataPreparation(Cleansing)AThepossibleerrorsinarawdatasetincludethefollowing:Incompletenesserroriswherethedataarenotpresent,resultinginmissingdata./Themostcommonimputationsaremean,median,ormodeofthevariableorsimplyassumingzero.InvalidityerroriswherethedataareOUtSideofameaningfulrange,resultingininvaliddata./Thiscanbecorrectedbyverifyingotheradministrativedatarecords.InaccuracyerroriswherethedataarenotameasureOftruevalue./Thiscanberectifiedwiththehelpofbusinessrecordsandadministrators.StructuredDataPreparation(Cleansing)AThepossibleerrorsinarawdatasetincludethefollowing:Inconsistencyerroriswherethedataconflictwiththecorrespondingdatapointsorreality./Thiscontradictionshouldbeeliminatedbyclarifyingwithanothersource.Non-uniformityerroriswherethedataarenotPreSentinUnidenticalformat./Thiscanberesolvedbyconvertingthedatapointsintoapreferablestandardformat.DuplicationerroriswhereduplicateObSetVationSarepresent./Thiscanbecorrectedbyremovingtheduplicateentries.DataPreparation(Cleansing)InvalidityerrorInconsistencyerrorInaccuracyerrorNon-uniformityerrorIDNameGenderDateofBirthSalaryIncomeState2Mr.ABCM12/5/1970$50,200$5,000VA32*Ms.XYZM115Jan,1975$60,500$0NY43EFGO1/13/1979S65,0$1,000CA54Ms.MNO/FI1119FL65Ms.XYZ/F15/1/1975S605$076Mr.GHl/M9/10/1942S55,0TX87Mr.TUV/M2/27/1956$300,000S50,0CT98Ms.DEF/F4/4/1980$55,0$0BritishColumbiaDon'tKnoYYes<NoCredItCard.CNNIncompletenesserrorDuplicationerrorDataWrangling:Transformation>Datapreprocessingprimarilyincludestransformationsandscalingofthedata.Datatransformations/Extraction:AnewvariablecanbecreatedfromthecurrentvariableforeaseofanalyzingandusingfortrainingtheMLmodel./Aggregation:TwoOrmorevariablescanbeaggregatedintoonevariabletoconsolidatesimilarvariables./Filtration:Thedatarowsthatarenotneededfortheprojectmustbeidentifiedandfiltered./Selection:Thedatacolumnsthatareintuitivelynotneededfortheprojectcanberemoved./Conversion:Thevariablescanbeofdifferenttypes:nominal,ordinal,continuous,andcategorical.ADatabeforetransformation（以1IDSalaryOther IncomeStateCredit CardNameGenderDate of Birth21Mr. ABCM12/5/1970USD 50200USD 5000VAY32Ms. XYZF1/15/1975USD 60500USDONYY43Mr. EFGM1/13/1979USD 650USD 1000CAN56Mr. GHIM9/10/1942USDOUSD 550TXN67Mr. TUVM2/27/1956USD 300000CTY78Ms. DEFF4/4/1980CAD550CADOBritishNColumbia1IDGenderTotalIncomeStateCreditCardAge21M4855200VAY32F4360500NYY43M39660CAN56M765S000TXN(1)(2)36-106DataWrangling:Scaling>Scalingisaprocessofadjustingtherangeofafeaturebyshiftingandchangingthescaleofdata.AHerearetwoofthemostcommonwaysofscaling:Normalizationistheprocessofrescalingnumericvariablesintherangeof0,1.Y_'iXmin!(normalized)=AmaXAmin/sensitivetooutliers,sotreatmentofoutliersisnecessarybeforenormalizationisperformed./usedwhenthedistributionofthedataisnotknown.Standardizationistheprocessofbothcenteringandscalingthevariables.!(standardized)。/lesssensitivetooutliersasitdependsonthemeanandstandarddeviationofthedata./Thedatamustbenormallydistributedtousestandardization.5.UnstructuredDataAnalysisAUnstructured,te×ted-baseddataismoresuitableforhumanuse.Thefivestepsinvolvedneedtobemodified(thefirstfour)inordertoanalyzeunstructured,text-baseddata:1.Textproblemformulation.Theanalystwilldeterminetheproblemandidentifytheexactinputsandoutputofthemodel.2.Datacollection(curation).This