结构改革与经济增长：机器学习方法（英）

StructuralReformsandEconomicGrowth:AMachineLearningApproach AnilAri,GaborPulaandLiyangSunWP/22/184 IMFWorkingPapersdescriberesearchinprogressbytheauthor(s)andarepublishedtoelicitcommentsandtoencouragedebate. TheviewsexpressedinIMFWorkingPapersarethoseoftheauthor(s)anddonotnecessarilyrepresenttheviewsoftheIMF,itsExecutiveBoard,orIMFmanagement. 2022 SEP ©2022InternationalMonetaryFundWP/22/184 IMFWorkingPaper EuropeanDepartment StructuralReformsandEconomicGrowth:AMachineLearningApproachPreparedbyAnilAri,GaborPulaandLiyangSun AuthorizedfordistributionbyIvannaVladkovaHollarSeptember2022 IMFWorkingPapersdescriberesearchinprogressbytheauthor(s)andarepublishedtoelicitcommentsandtoencouragedebate.TheviewsexpressedinIMFWorkingPapersarethoseoftheauthor(s)anddonotnecessarilyrepresenttheviewsoftheIMF,itsExecutiveBoard,orIMFmanagement. ABSTRACT:Thequalitativeandgranularnatureofmoststructuralindicatorsandthevarietyindatasourcesposesdifficultiesforconsistentcross-countryassessmentsandempiricalanalysis.Weovercometheseissuesbyusingamachinelearningapproach(thepartialleastsquaresmethod)tocombineabroadsetofcross-countrystructuralindicatorsintoasmallnumberofsyntheticscoreswhichcorrespondtokeystructuralareas,andwhicharesuitableforconsistentquantitativecomparisonsacrosscountriesandtime.Withthisnewlyconstructeddatasetofsyntheticstructuralscoresin126countriesbetween2000-2019,weestablishstylizedfactsaboutstructuralgapsandreforms,andanalyzetheimpactofreformstargetingdifferentstructuralareasoneconomicgrowth.Ourfindingssuggestthatstructuralreformsintheareaofproduct,laborandfinancialmarketsaswellasthelegalsystemhaveasignificantimpactoneconomicgrowthina5-yearhorizon,withonestandarddeviationimprovementinoneofthesereformareasraisingcumulative5-yeargrowthby2to6percent.Wealsofindsynergiesbetweendifferentstructuralareas,inparticularbetweenproductandlabormarketreforms. RECOMMENDEDCITATION:Ari,A.,Pula,G.,&Sun,L.(2022).StructuralReformsandEconomicGrowth:AMachineLearningApproach.IMFWorkingPaper,WP/22/184 JELClassificationNumbers: E02,C54,C55,D58,O43,O47 Keywords: Structuralreforms,institutions,economicgrowth Author’sE-MailAddress: aari@imf.org,gpula@imf.org,lsun20@cemfi.es WORKINGPAPERS StructuralReformsandEconomicGrowth:AMachineLearningApproach PreparedbyAnilAri,GaborPulaandLiyangSun1 1TheauthorsaregratefultoIvannaVladkovaHollar,IppeiShibata,MarinaMendesTavaresandseminarsparticipantsattheIMFforhelpfulcommentsandsuggestions.ExcellentresearchassistancewasprovidedbySamuelVictorRomeroMartinez.Thedatasetofsyntheticstructuralscoresisavailableuponrequestfromtheauthors.Allerrorsareourown. Contents I.Introduction3 II.StructuralIndicators5 A.Dataoverview5 B.SyntheticstructuralscoresviaPartialLeastSquares6 C.PLSestimationprocedure7 D.SyntheticstructuralscoreasthepredictedvaluefromthePLSmodel8 III.StructuralIndicators12 A.Impactofstructuralreformsongrowth12 B.Synergiesofstructuralreformsongrowth17 C.Theroleofstructuralreformsduringcrises20 IV.Conclusions21 References21 Appendix23 A.Listofstructuralindicators23 B.Descriptionofimputationformissingindicators27 C.ComparisonbetweenthePLSstructuralscoreandsimple-averagescore27 I.Introduction Policymakersoftenpursuestructuralreformstoaidrecoveryfromcrisesandstimulateeconomicgrowth.Thisplacestheonusonpolicymakerstoidentifywhichcombinationsandsequencesofstructuralreformswouldbethemostgrowth-enhancing(IMF,2015;Rodrik,2010).However,akeychallengeisthatstructuralreformsareinherentlydifficulttomeasureastheyofteninvolvepoliciesthataregearedtowardsimprovingefficiencyofmarkets.Commonapproachesquantifystructuralreformsbasedonthestrengthofregulatorychangesthatremoveinefficiencies(seee.g.,Alesinaetal.,2020).Whiletheseapproachesprovidevaluableinsightsontheimpactofpolicyactions,theymaynotfullyreflectreformoutcomes,whichdependonthespecificsofpolicyimplementationaswellastheenvironmentinwhichreformsareimplemented.Anotherdrawbackoftheseapproachesisthattheyhavelimitedcountrycoverageduetolimiteddataavailability.Otherapproachesrelyonsurvey-basedindicatorsofstructuraloutcomestoassesstheimpactofstructuralreformsandconductcross-countryanalysis(seee.g.,EgertandGal,2016;Egert,2017).Whiletheseindicatorsareinformativeaboutstructuralperformance,empiricalanalysisiscomplicatedbythelargenumberofindicators,thecorrelationbetweenthemandbiasesthatmayarisefromtheirsubjectivenature. Weuseamachinelearningapproachtoconstructsyntheticstructuralscoresfromalargenumberofstructuralindicators.Ouranalysiscontributestotheexistingliteraturebyusingpartialleastsquares(PLS)toaggregatestructuralindicatorsforgrowthanalysis,insteadofsimpleaveragingoradhocweighingschemes.OurPLSweightingschemeassignshigherweightstoindicatorsthataremorepredictiveofhighGDPpercapita,therebyextractingusefulinformationfromavailabledatawhileremovingthenoiseandbiasesassociatedwithsubjectiveandsurvey-basedindicators.Ourapproachalsoaccountsforthecorrelationandredundancyamongstructuralindicators,thereforeavoidingtheduplicationbiasthatsimpleaveragingwouldsufferfrom.1 Oursyntheticstructuralscore