d5.3 technical report on the econometric assessment and

D5.3TechnicalReportontheEconometricAssessment

andResults

D5.3TechnicalReportontheEconometricAssessmentandResults

2/26

ExecutiveSummary

Thisreportpresentstheempiricalfindingsontherelationbetweenbuildings’energyefficiencyandtheassociatedmortgagedefaultrisk.TheanalysesareconductedforfourEuropeancountries:Belgium,Italy, the Netherlands, and UK. The results differ from country to country due to the size andcomposition of the underlying datasets, and due to varying variable definitions. Despite thesechallenges,however,theoverallmessageoftheseanalysesis:energyefficientbuildingsarecorrelatedwithlowercreditrisk.

InthecaseofBelgium,theanalysisexploitsthegovernmentinterventionsintroducedin2009thatwereannounced to revive the lendingmarket and to encourage energy-saving investments for housingimprovements.Usingacross-sectional loan-leveldataset fromaBelgianbank,weshowthat loansissued with the purpose to improve building’s energy efficiency are less likely to default thancomparable loans thatwere originated during the same period but did notmeet the governmentsubsidycriteria.

InthecaseofItaly,availabledatasetsallowonlyforanalysesatregionalandprovinciallevel.Attheregionallevel,ourfindingsindicatethatregionswithhighergovernmentenergyefficiencyinvestmentsare associated on average with lower individual mortgage default rates. Focusing on the regionLombardy,wedonotfindastrongrelationbetweenprovincesthathavearelativelylargershareofenergyefficientbuildingsandtheaverageindividualmortgagedefaultrisk.

InthecaseoftheNetherlands,weperformaloan-levelanalysiswhereresidentialbuildings’energyefficiencyisapproximatedusinginformationonconstructionyearandpropertytype.Resultsfromtwoempiricalmethodologies,theLogisticandtheextendedCoxregression,documentarobust,negativeand significant correlation between energy efficiency andmortgage default risk. Additionally, ourfindingsindicatethatthedegreeofenergyefficiencymatters,meaningthatmortgagepaymentsonrelativelymoreefficientbuildingsarelesslikelytofallintoarrears.

InthecaseofUK,weusehousehold,buildingandmortgageinformation,whichwascollectedfortheEnglish Housing Survey, to investigate the relationship between buildings’ energy efficiency andhouseholds’mortgagedefaultrisk.Quantitatively,ourresultsindicateanegativecorrelationbetweenthetwovariablesofinterest.Qualitatively,however,thesefindingslacksignificance.

Tosummarize,theresultsfromItalyandUKindicate,butdonotstrictlyconfirm,anegativerelationbetweenenergyefficiencyandtheprobabilityofmortgagedefault.Thiscanbeattributedtothenatureoftheunderlyingdatasets.Theloan-levelanalysesforBelgiumandtheNetherlands,ontheotherhand,indicatearobust,negativeandsignificantcorrelationbetweenthetwovariablesofinterest.

Source Activity: WP5/D5/3 Editors: M. Billio (UNIVE), L. Pelizzon (SAFE) Authors: T. Baccega, A. Bedin, M. Billio, I. Hristova, M.Riedel Status: Final Date: 30.04.2019 Contractual Delivery Date: M24


3/26

TableofContents

1. Introduction.....................................................................................................................4

2. Belgium.............................................................................................................................6

2.1 Background...............................................................................................................6

2.2 LogisticRegressionResults.......................................................................................6

3. Italy.................................................................................................................................10

3.1 RegionalAnalysis....................................................................................................10

3.1.1 Background.....................................................................................................10

3.1.2 LogisticRegressionResults.............................................................................10

3.2 ProvincialAnalysis..................................................................................................12

3.2.1 Background.....................................................................................................12

3.2.1 LogisticRegressionResults.............................................................................12

4. Netherlands....................................................................................................................14

4.1 Background.............................................................................................................14

4.2 LogisticRegressionResults.....................................................................................14

4.3 ExtendedCoxModelResults..................................................................................16

4.4 AdditionalFindings.................................................................................................19

5. UK...................................................................................................................................22

5.1 Background.............................................................................................................22

5.2 LogisticRegressionResults.....................................................................................22

6. Conclusion......................................................................................................................24

7. Bibliography...................................................................................................................25


4/26

1. IntroductionThepresentreportfulfilstherequirementunderthegrantagreementtodeliveratechnicalreportthataimstoinvestigatetherelationbetweenabuilding’senergyefficiency(EE)andtheassociatedprobabilityofmortgagedefault(PD).1Thisdocumentbuildsonthetwoprecedingreports:‘Technicalreportonthemethodologydesigntocarryoutportfolioanalysis’(D5.1)and‘Technicalreportontheportfolio analysis of banks’ loan portfolios’ (D5.2). The first report (i) discusses the theoreticalbackgroundandthedifferencebetweenacorrelationandcausalityanalysis,(ii)reviews13Europeancountriesandtheirprogrammestoincentivizeenergyefficiencyrenovationsintheresidentialbuildingsector,(iii)identifiesthosecountriesthatbearthepotentialforcarryingoutacreditriskanalysiswithrespect to building’s energy efficiency information, and (iv) presents the appropriate empiricalmethodologies.Thesecondreport(i)presentsindetailthecountry-specificdatasetsidentifiedinD5.1,(ii)discussesthevariabledefinitions,and(iii)providestherelevantdescriptivestatistics.Thepresentdocumentisacontinuationofthepreviousreportsandpresentsinthefollowingtheempiricalfindingsforthefourcountries:Belgium,Italy,theNetherlands,andUK.

Section2presentstheempiricalfindingsforBelgium.Inthiscase,westudytherelationbetweenEEandPDbyexploitingthetemporarygovernmentinterventionsthatwereintroducedfortheperiod2009to2011inordertorevivethelendingmarketandtoencourageenergy-savinginvestmentsforhousingimprovements.Usingacross-sectionalloan-leveldatasetfromaBelgianbank,weshowthatloansissuedforbuilding’senergyefficiencyimprovementsarelesslikelytodefaultthancomparableloansthatwereoriginatedduringthesameperiodbutdidnotmeetthegovernmentsubsidycriteria.

Section3reportstheresultsforItaly.Forthiscountry,availabledatasetsallowonlyforanalysesatregionalandprovinciallevel.Attheregionallevel,weusegovernmentenergyefficiencyinvestmentsasaproxyfortheaverageenergyefficiencyofbuildingswithinaregionandinvestigateitscorrelationwithmortgage default rates. Focusing on the region Lombardy,we compute the share of energyefficient buildings with each province and study its correlation with mortgage default rates. OurfindingsindicateanegativerelationbetweenEEandPDattheregionallevel,whilethefindingsareinsignificantattheprovinciallevel.

Section4 focusesonNetherlands. In this loan-levelanalysis,weapproximate residentialbuildings’energyefficiencyusinginformationonconstructionyearandpropertytype.ByapplyingtheLogisticandtheextendedCoxregression,wedocumentarobust,negativeandsignificantcorrelationbetweenenergyefficiencyandmortgagedefaultrisk.Additionally,our(lesssignificant)findingsindicatethatthedegreeofenergyefficiencymatters,meaningthatmortgagepaymentsonrelativelymoreefficientbuildingsarelesslikelytofallintoarrears.

Section 5 presents the empirical findings for UK. In this case, we use in the analysis household,building,andmortgageinformationthatwascollectedfortheEnglishHousingSurvey.Quantitatively,

1TheauthorsthankDianaBarro,RobertoCasarin,MicheleCostola,andXuLiuforresearchassistance.


5/26

our results indicate a negative correlation between the two variables of interest, EE and PD.Qualitatively,however,thesefindingsarequestionableduetolackofsignificance.


6/26

2. Belgium

2.1 Background

AsdescribedinthetechnicalreportD5.1,theBelgiangovernmentintroducedtwopolicymeasuresin2009,inordertofostertheinvestmentsintheresidentialmarketafterthe2008Financialcrisis.2

Using data from a Belgian EeMAP pilot scheme bank, descriptive statistics for the variableswerepresented in the technical report D5.2. The resulting main findings demonstrated considerablyreduced levels of delinquency and default likelihood for EE-labelled loans. On the basis of thepresenteddata,anon-uniformrepartitionofloansacrossregionscanbeobserved.ThespecificcaseofLiegeisnoteworthytobehighlightedas,eventhoughitpresentsoneofthelowestloanlevelsinthedataset,a largemajorityof them(73%)wereEEoriented.Anothermajor findingconcerns theexistingincomedifferencesbetweenEEandnon-EEborrowers.Namely,EEborrowershaveabetterfinancialstatuswhichinvolveslowerloantovalue(LTV)levels.Indeed,notonlytheirborrowingneedsarelessimportant,butalsotheirdwellings'valuesaremuchhigher.

Therefore, a correlationbetween loandefault risk and theunderlying buildings' EE characteristicsmightbesuspected.Inordertotestthestatisticalsignificanceoftheseassumptions,weconsideraLogisticregressionmodel.Theobtainedresultsarepresentedinthefollowingsection.

2.2 LogisticRegressionResults

AsmentionedinthetechnicalreportD5.1,theLogisticregressionmodelisappropriateformodellinga binary outcome disregarding time dimension. In the following, we use a cross sectional paneltruncatedattheendofNovember2018.Wehaveintotal42,055loan-levelobservationstoworkwith.InaccordancewiththedefinitionoftheBaselcommittee,wedefinealoanasdefaultedifitisinarrearsfor90daysorlonger.

Table 1 presents the estimated odds ratios (ORs). The first regression, column (1), evaluates therelation between the loans’ default status and the EE characteristic of the concerned dwellings,excludinganyothercontrolvariables.TheORestimateof0.2461suggeststhatgreenrenovationhasanegativeandhighlysignificantcorrelationwiththeriskofdefault.3Sincethefindingsmightbedrivenbyotherfactorslikeloanorhouseholdcharacteristics,weincludetheappropriatecontrolvariables.

2 For further information, refer to Hoebeeck and Inghelbrecht (2017) and e-Justice Belgium(http://www.ejustice.just.fgov.be/cgi/article_body.pl?language=nl&caller=summary&pub_date=09-07-31&numac=2009003261)3Anodds ratio isexpressedasanumber fromzero (eventwillneverhappen) to infinity (event iscertain tohappen).Thereferencepointisthevalueone:anoddsratioofonemeansthatthetwoeventsareindependentofeachother,anoddsratiogreaterthanoneindicatesthattheeventsarepositivelyassociated,whileanoddsratiobelowoneindicatesanegativeassociation.


7/26

Wecontrol for loanandborrowercharacteristics: currentLTV,debtservicecoverageratio (DSCR),loan term, interest rate, origination year, total income,borrower age, andemployment status.Atdwellinglevel,wecontrolforthepropertytypeandgeographiclocation.Column(2)presentstheORestimatesincludingthementionedcontrolvariablesandusingrobuststandarderrors.TheoddsratioestimateofEEvariableexperiencesan increasebut isstillbelowone,confirmingthenegativeandhighly significant correlationwith default risk. Addingmarket controls (i.e., end-of-month Belgianunemployment rate, 10-year German government bond yields,monthly volatility of daily 10-yearGerman government bond yields, and the end-of-month yield curve slope) and considering both,robuststandarderrorsandclusteredonesatregionallevel,doesnotaffectthefindings,asreportedin columns (3) and (4). Theestimatedodds ratioof0.3781 in column (4),means that theoddsofdefaultingonaloanisabout2.64timesgreateriftheborrowerdecidestonotimprovethebuilding’senergyefficiency.

Table1-LogisticregressionresultsThis tablepresentsLogistic regressionodds ratioestimates todetermine thepropensity todefaulton loansbackedbyenergyefficientdwellings.Thedependentvariableisadummyindicatingwhetheraloanisindefault(i.e.,inarrearsforatleastthreemonths)ornot.ThemainexplanatoryvariableisthedummyvariableEEthatequalstooneifabuilding’srenovationpurposeisconsidered“green”andzerootherwise.LoancontrolsarecurrentLTV,DSCR,loanterm,andinterestrate.Dwellingandborrowercontrolvariablesarepropertytype,borrowerageatorigination,borrowerincome,andemploymentstatus.Marketcontrolsareend-of-monthBelgianunemployment,10-yearGermangovernmentbondyield,monthlyvolatilityofdaily10-yearGermangovernmentbondyields,andtheend-of-monthyieldcurveslope(measuredas10-yearminus1-yearEURswaprates).Originationyearandregionfixedeffects(FE)areincludedwhereindicated.Standarderrors(reportedinsquarebrackets)areeitherrobustorclusteredatregionallevel.Statisticalsignificanceisdenotedby***,**,and*atthe1%,5%,and10%level,respectively.

(1) (2) (3) (4)EE 0.2461*** 0.3547*** 0.3781*** 0.3781***

[0.0640] [0.1261] [0.1249] [0.0598]CurrentLTV 0.5100 0.5220 0.5220

[0.5243] [0.5363] [0.4252]DSCR 30.1220*** 34.3433*** 34.3433***

[29.9714] [33.2219] [33.7193]Loanterm 1.0113*** 1.0109*** 1.0109*** [0.0032] [0.0033] [0.0039]Borrowercontrols No Yes Yes YesDwellingcontrols No Yes Yes YesMarketcontrols No No Yes YesLoancontrols No Yes Yes YesRegionFE No Yes Yes YesYearFE No Yes Yes YesSE Rob. Rob. Rob. RegionCl.Observations 42,055 40,263 40,263 40,263PseudoR-squared 0.0311 0.405 0.411 0.411

AcarefulobservationshowsthattheDSCRhasapositiveandstatisticallysignificanteffectoncreditrisk.Thisisacounterintuitiverelation,aswithahigherDSCRthedefaultprobabilityshouldactuallydecrease. The DSCR values are provided with the dataset and we do not have access to the


8/26

computationalmethodologyofthisvariable.Asarobustnessexercise,weexcludethevariablefromtheregression.UnreportedresultsindicatethatthenegativeEErelationwiththeprobabilityofloandefaultisnotaffectedqualitatively.

Wevalidatetheabovefindingswithadditionalrobustnesschecks.Wedefinethemodelincolumn(4),Table 1, as the baselinemodel specification and replace, redefine or add covariates as describedfurtherbelow.Thevariousmodelspecifications(Spec.)arepresentedinTable2,wherewereportforconveniencepurposesonly theORestimateassociatedwith theEEdummyvariable. ThebaselineregressionwasestimatedbyexcludingtheoriginalbalanceasitiscorrelatedwiththeLTV.InSpec.1,weaddtheoriginalbalancevariableandobservethattheresultsdonotchangequalitatively.Sinceitiscommontoestimatecreditriskmodelswithoriginalcovariates,wereplacetheexplanatoryvariablecurrent LTV with original LTV. As presented in Spec. 2, the main results are not driven by theobservationdate.InSpec.3,weaddtothebaselinemodeltheoriginalvalueofthepropertyandinSpec.4,wereplaceitwiththecurrentvalue.AsshowninTable2,alsothesetwolastspecificationsdonotaffectthemainfinding.

Table2-Logisticregression-robustnessofresultsThis table presents the odds ratio estimates and the respective standard errors (SE) for the EE variable using various Logistic modelspecifications.Thedependentvariableisadummyindicatingifaloanis indefault(i.e., inarrearsforatleastthreemonths)ornot.Thebaselinemodelspecificationisthemodelcolumn(4),Table1.Themodelspecifications1to4differfromthebaselinemodelaccordingtothefollowingchanges.Spec.1:theoriginalbalanceisaddedtotheexplanatoryvariables.Spec.2:thevariablecurrentLTVisreplacedbytheoriginalLTV.Spec.3:theoriginalvalueofthedwellingisaddedasanexplanatoryvariable.Spec.4:thecurrentfacevalueofthedwellingisaddedasexplanatoryvariable.Statisticalsignificanceisdenotedby***,**,and*atthe1%,5%,and10%level,respectively.

Model OR(EE) SESpec.1 0.3303*** [0.0401]

Spec.2 0.3675*** [0.0479]

Spec.3 0.3738*** [0.0684]

Spec.4 0.3839*** [0.0683]

Fromthepresentedresults,wecanconcludethatloansissuedwiththepurposeofenergyefficiencyrenovationarecorrelatedwitha lowerriskofdefault.This isaremarkablefindingastheanalysedsample is composedof comparable loanswith respect to several dimensions. First, all loanswereissuedwiththesamepurpose,namely,torenovateadwelling.Theonlydifferenceinthisrespectisthetypeofrenovation,whichiseitherrelatedtotheimprovementofthedwelling’senergyefficiencyornot.Andsecond,allloanswereissuedduringthesameperiod,i.e.between2009and2011.ThismeansthatbothEEandnon-EEloanswereoriginatedduringthesamemacroeconomicenvironmentandthattheywereofcomparableageatthetimeoftheanalysis.

The remarkableobservation is the fact that evenafter accounting for loan- andborrower-specificdifferencesbetweentheEEandnon-EEloans,wefindasignificantly lowerriskofdefaultfor loansthatwereissuedforenergyefficiencyrenovationpurposes.Thisresultisencouragingandtheanalysis


9/26

should be ideally repeated either (i) using a larger loan sample, i.e., by expanding the presenteddatasetwithloanportfoliosofotherBelgianbanks,or(ii)atcomepointinthefuture,i.e.,whentheloansbecomeolderandmoredefaultsoccur.


10/26

3. Italy

3.1 RegionalAnalysis

3.1.1 Background

As described in the technical report D5.1, for the analysis of the Italian market, we study therelationship between regional investments in buildings’ energy efficiency and borrower’s defaultprobabilityusingaloan-leveldatasetfromEuropeanDataWarehouse(EDW).4Theaimofthisanalysisistotestifandtowhatextentinvestmentsmadeatregionallevelcaninfluencetheaveragedefaultrates in the affected regions. In the following section,weemploy a Logistic regression analysis toinvestigatethecorrelationbetweenEEinvestmentsandborrowers’defaultriskatregionallevel.

3.1.2 LogisticRegressionResultsAsmentionedbefore,theapplicationofaLogisticregressionmodelisthemostcommonapproachusedformodellingbinaryoutcomes.Weapplythemodeltoacross-sectionofItalianloan-leveldatafromEDWwherethemortgageobservationswerereportedbetweenMarch2015andOctober2018.Foreachmortgage,wefocusonlyonthemostrecentobservationinthesampleperiod.Tobeprecise,inthosecaseswheremortgageshavenotexperiencedanydefault,weusethemostrecentvaluesavailable inthesample.Fordefaultedmortgages,ontheotherhand,weusethevaluesthatwerereportedonthefirstdeclarationdateofbeinginarrearsforatleastthreemonths.

As reported in the technical reportD5.2,weretrieveenergyefficiency investments fromthe2018reportoftheItaliannationalenergyagencyENEA.Inparticular,weusetheamountoftaxdeductionsgrantedforenergyefficiencyworksatregionallevel.5Weemploythisinformationasaproxyforthegovernmental involvement in regional EE investments during the period 2014 to 2017. The EEinvestmentsrefertothesumofinvestmentsreportedinthedocumentfor(i)theaggregateperiod2014to2016and(ii)theyear2017.

Table3presentstheoddsratioestimatesfordifferentmodelspecifications.Column(1)presentstheestimate for EE investments only, not controlling for any borrower, dwelling, or mortgagecharacteristics.TheestimatedORof0.8989indicatesthatborrowersdefaultlessofteninregionswithhigher EE investments. Since the results can be confounded by the borrowers' profile, dwellingparticularities,mortgage properties, or the general state of the economy,we include appropriatecontrolvariablesinordertotakeintoaccounttheseeffectsinthemodelspresentedincolumns(2)to(4).TherelationshipbetweenregionalEE investmentsand individualdefault riskremainsnegativeevenaftercontrollingforthesecharacteristics.Clusteringthestandarderrorsatregionalleveldoesnotaffectthefindingsqualitativelyasreportedincolumn(4).

4EDWprovidesarichdatasetwithperiodicallyupdateddynamicandstaticindividualloan-levelinformationofsecuritized Europeanmortgages. A comprehensive overview of loan-level data templates including detailedvariabledescriptionsonresidentialmortgages-backedsecurities(RMBS)datasetscanbeobtainedfromECB'swebsite:https://www.ecb.europa.eu/paym/coll/loanlevel/transmission/html/index.en.html5Dataareavailablehere:http://www.enea.it/it/seguici/pubblicazioni/pdf-volumi/2018/raee_2018.pdf


11/26

Table3-LogisticregressionresultsThis table presents Logistic regression odds ratio estimates to determine the relationship between governmental EE investments andborrower’sdefaultprobability.Thedependentvariableisadummyindicatingwhetheraborrowerisindefault(i.e.,inarrearsforatleastthreemonths)ornot.ThemainexplanatoryvariableisregionalEEinvestments,whichisdefinedasthetotalgovernmentalEEinvestmentsduringtheperiod2014to2017,normalizedbythenumberofresidentialbuildingsperregion.MortgagecontrolsarecurrentLTV,currentDSCR,mortgage term, and interest rate. Dwelling and borrower control variables are property type, borrower age at origination, andborrowerincome.Marketcontrolsareend-of-monthItalianunemploymentrate, 10-yearGermangovernmentbondyield,monthlyvolatilityofdaily10-yearGermangovernmentbondyields,andtheend-of-monthyieldcurveslope(measuredas10-yearminus1-yearEURswaprates).Yearandregionfixedeffects(FE)areincludedwhereindicated.Standarderrors(reportedinsquarebrackets)areeitherrobustorclusteredatregionallevel.Statisticalsignificanceisdenotedby***,**,and*atthe1%,5%,and10%level,respectively.

(1) (2) (3) (4)EEInvestments 0.8989*** 0.6397*** 0.5504*** 0.5504***

[0.0261] [0.0628] [0.0561] [0.0650]CurrentLTV 39.2642*** 39.8784*** 39.8784***

[3.1996] [3.2716] [14.8687]DSCR 0.6756*** 0.6872*** 0.6872***

[0.0121] [0.0124] [0.0247]Mortgageterm 10.4239*** 10.6948*** 10.6948*** [0.5516] [0.5698] [2.9752]Borrowercontrols No Yes Yes YesDwellingcontrols No Yes Yes YesMarketcontrols No No Yes YesMortgagecontrols No Yes Yes YesRegionFE No Yes Yes YesYearFE No Yes Yes YesSE Rob. Rob. Rob. RegionCl.Observations 291,363 282,486 282,486 282,486PseudoR-squared 0.000888 0.108 0.111 0.111

UnreportedresultsconfirmthatusingoriginalLTVinsteadofitscurrentvaluedoesnotaffectthemainfinding.Debttoincome(DTI)wasnotincludedinthemainregressionsasitiscorrelatedwithLTVandDSCR.Asarobustnesscheck,weaddDTIasanexplanatoryvariabletothemodelincolumn(4),Table3.WefindthatneitherusingoriginalnorcurrentDTIaffectsthemainfindingqualitatively.6

From thepresented results,we can conclude that theoverall risk ofmortgagedefault is lower inregionswheregovernmentalsupportforenergyefficiencyrenovationsisrelativelyhigher.ThisfindingisinlinewiththeresultsobtainedforBelgium.IntheBelgiancase,wefindthatloansthatwereissuedforenergyefficiencyrenovationandthatmetthecriteriaforagovernmentsubsidy,arelesslikelytodefaultthanotherwisecomparableloansfornon-energyefficiencyrenovationpurposes.

6Resultsareavailableuponrequest.


12/26

3.2 ProvincialAnalysis

3.2.1 Background

AsdescribedinthetechnicalreportD5.2,theItalianregionLombardyprovidesadetaileddatasetthatcontainsinformationonbuildings’energycertificatesissuedinoneofits11provinces.Thedatasetcontainsmorethan1.7milliondwellingsthatareaccompaniedwith47variables,includingtheexactaddressofthebuilding. Inthefollowinganalysis,weconsidertheETHindicator(EnergyThermal–hot),whichismeasuringthequantityofthermalenergyideallyrequiredbythebuildingenvelope.WeconsidertheETHindicatorasameasureofbuilding’senergyefficiency.Weexpecttofindapositiverelation between ETH and the probability of mortgage default as a higher ETH value should beassociatedwithlowerenergyefficiencyandthusahigherriskofdefault.WeusetheLombardydatasettocomputethemeanoftheETHindicatoracrossthedifferentprovincesandmergetheinformationwiththeEDWdatasetatprovincelevel.Theobjectiveofthisanalysis istoinvestigatewhetheranyrelation between average provincial energy efficiency level and themortgage default rates in therespectiveregionexists.

3.2.1 LogisticRegressionResults

Table4presentstheoddsratioestimates.Inthefirstregression,wedonottakeintoaccountforanyothervariables thanETH.Weobserveapositive relationbetweentheEEproxy (i.e.,ETH)and thedefault rateacross theprovinces.Theodds ratioof1.11 indicatesapositiveandhighly significantcorrelationofhighenergyusewiththeprobabilityofdefault.ThisisareasonableobservationsinceETHisameasureofbuildings’energyconsumption,andthehighertheETH,thelowershouldbeabuilding’s EE. In the next step, we include dwelling (building type), household (total income andborrower age), mortgage (LTV, DSCR, mortgage term, interest rates), and market (Italianunemployment,Germangovernmentbondyieldandvolatility,aswellasyieldcurveslope)controlvariables.Themodelincolumn(2)confirmsthepositiverelationquantitativelyandqualitativelyifwecontrolfordwelling,household,mortgagecharacteristics.However,marketcontrolsabsorbthiseffectasreportedincolumns(3)and(4).

Unreportedresultsindicatethatthelackofevidencedoesnotstemfrommodelmisspecificationsorvariabledefinitions.7Were-rantheregressionanalysiswithvariousproxiesfortheEEvariable.Amongtheproxies,weusedETC(EnergyThermal–cold),ETW(EnergyThermal–water),CO2,andpercentageofA-orB-ratedbuildingswithinaprovince.AsreportedinD5.2,ETCmeasuresthequantityofthermalenergyrequiredbythebuildingenvelopeduringthecoolingseason.EPWreportsthebuilding’stotalamountofthermalenergyrequiredforthegenerationofsanitaryhotwater.AndCO2measuresthetotalamountofCO2emittedbyabuilding.Finally,wecomputedthepercentageofbuildingswithanA-orB-ratingwithinaprovinceanduseditasanEEproxy.Wedidnotfindanysignificantresultsby

7Resultsareavailableuponrequest.


13/26

applyinganyoftheaboveEEmeasures.

Table4-LogisticregressionresultsThis tablepresentsLogistic regressionodds ratioestimates todetermine the relationshipbetweenaverageresidentialbuildingsenergyefficiencyatprovinciallevelandborrowers’defaultprobability.Thedependentvariableisadummyindicatingwhetheramortgageisindefault(i.e.,inarrearsforatleastthreemonths)ornot.ThemainexplanatoryvariableistheETHlevel(EnergyThermal-hot,inkWh/m2)ofanaverageresidentialbuildingwithinaLombardyprovince.ETHmeasuresthequantityofthermalenergyideallyrequiredbythebuildingenvelopeduring theheating season.Mortgagecontrolsarecurrent LTV, currentDSCR,mortgage term,and interest rate.Dwellingandborrowercontrolvariablesarepropertytype,borrowerageatorigination,andborrowerincome.Marketcontrolsareend-of-monthItalianunemploymentrate,10-yearGermangovernmentbondyield,monthlyvolatilityofdaily10-yearGermangovernmentbondyields,andtheend-of-monthyieldcurveslope(measuredas10-yearminus1-yearEURswaprates).Yearandprovincefixedeffects(FE)areincludedwhereindicated.Standarderrors(reportedinsquarebrackets)areeitherrobustorclusteredatprovinciallevel.Statisticalsignificanceisdenotedby***,**,and*atthe1%,5%,and10%level,respectively.

(1) (2) (3) (4)ETH 1.1181*** 1.1158*** 0.9923 0.9923 [0.0374] [0.0384] [0.0549] [0.0107]CurrentLTV 15.9670*** 26.5954*** 26.5954***

[4.9703] [9.3077] [12.2820]DSCR 0.7060*** 0.7695*** 0.7695***

[0.0438] [0.0416] [0.0512]Mortgageterm 13.1426*** 18.5829*** 18.5829*** [2.9331] [4.5145] [3.7184]Borrowercontrols No Yes Yes YesDwellingcontrols No Yes Yes YesMarketcontrols No No Yes YesMortgagecontrols No Yes Yes YesProvinceFE No Yes Yes YesYearFE No Yes Yes YesSE Rob. Rob. Rob. ProvinceCl.Observations 31,141 30,049 30,049 30,049PseudoR-squared 0.00257 0.127 0.161 0.161

From the presented results, we conclude that despite the fact that the Italian region Lombardyprovidesaverygranulardatasetonbuildings’energyconsumptioninformation,wecannotfullyutilizeitspotential.Thereasonforthisissuecomesfromdataprivacyregulations:itiscurrentlynotpossibletomergethebuilding-levelinformationone-to-onewiththeanonymizedloan-leveldatafromEDWasneitherhousenumbersnorstreetnamesareprovidedinthelatterdataset.However,thisobstaclecould be easily overcome if themortgage-issuing bankswould collect buildings’ energy efficiencyinformationatthedateofloanorigination.


14/26

4. Netherlands

4.1 Background

As described in the technical report D5.1, the Netherlands Enterprise Agency (Rijksdienst voorOndernemendNederland,inshortRVO)isagovernmentalagencycollectingdataonbuildings’energyperformanceandprovidingprovisionalenergylabelsforallexistingDutchresidentialbuildings.8Wecombined loan-leveldata fromEDWwiththeratings fromRVOfor theanalysisandpresentedthedescriptive statisticsof thevariables in reportD5.2.Themain insights fromthis statisticalanalysiswere the following: the frequencyofdefaultsdecreaseswithhigherEE ratings;notall regionsareequallyrepresentedinthedataset;borrowersofEEbuildingstendtodefaultlessoften.

Inthefollowing,weemployaLogisticregressionanalysisandconfirmourfindingsbyrunningasurvivalanalysisbyutilizingthepanelstructureoftheloan-leveldataset.


TheLogistic regressionmodel isappropriateformodellingabinaryoutcomedisregardingthetimedimension.Sinceourdatasetprovidesaquarterlytimeseriesofmortgageinformation,weresolvetothe following procedure in order to eliminate the time dimension. Namely, concerning thosemortgages thathavenot experiencedanydefault,weuse themost recent values available in thesample.Fordefaultedmortgages,ontheotherhand,weapplythevaluesthatcorrespondtothefirstdeclarationdateofbeinginarrearsforatleastthreemonths.

Table5presentstheoddsratioestimates.Column(1)reportstheresultswithoutcontrollingforanyothercharacteristicsinthemodel.TheORestimateof0.489fortheEEindicatorsuggeststhatenergyefficiencyhasanegativeandhighlysignificantcorrelationwiththeriskofmortgagedefault.Sincethisfindingmightbedrivenbyvariousmortgage,buildingorhouseholdcharacteristics,we includetheappropriatecontrolvariables.Oneofthemostimportantdrawbacksinthisanalysisstemsfromtheprovisional rating table. As mentioned in reports D5.1 and D5.2, buildings' rating categories areconstructedbyRVObasedonbuildingtypeandconstructionyearperiod.Thismeansthattheresultsmightbedrivennotbytheactualratingbuteitherbythebuildingtypeortheageofthebuilding.Todisentangle the energy efficiency effect from other building characteristics,we include as controlvariables both the type of the building and its current age category. Additionally, we control forhousehold(totalincomeandborrowerageatorigination)andmortgagecharacteristics(LTV,DSCR,mortgageterm,interestrate).Further,weincluderegionfixedeffectsatNUTS3levelandyearfixedeffects.9Modelspecification(2)showsthatthenegativerelationbetweenenergyefficiencyandthe 8Forfurtherinformation,refertoBoumeesteretal.(2008)andAgentschapNL(2011).9 The Nomenclature of Territorial Units for Statistics (NUTS) is a geocode standard for referencing thesubdivisionsofcountriesforstatisticalpurposes.ForeachEUmembercountry,ahierarchyofthreeNUTSlevelsisestablishedbyEurostatinagreementwitheachmemberstate.Amongthethreelevels,theNUTS3codesrefertothemostgranularregionspecification.


15/26

probabilityofdefaultremainssignificantandquantitativelysizeablewithanestimatedORof0.2582.Addingmarketcontrolsandclusteringthestandarderrorsattheregional leveldoesnotaffectthefindingsasreportedincolumns(3)and(4).Theestimatedoddsratioof0.1857incolumn(4),meansthattheoddsofdefaultingonamortgageisabout5.39timesgreateriftheborrower’spropertyhasCratingorworse.

Table5-LogisticregressionresultsThis table presents Logistic odds ratio estimates to determine the relationship between residential buildings energy efficiency andborrowers’defaultrisk.Thedependentvariableisadummyindicatingifamortgageisindefault(i.e.,inarrearsforatleastthreemonths)ornot.ThemainexplanatoryvariableisthedummyvariableEEthatequalstooneifabuilding'senergyefficiencyratingisAorB-ratedandzerootherwise.MortgagecontrolsarecurrentLTV,DSCR,averagemortgageterminmonths(weightedbyoriginalbalanceofindividualloancomponents),andtheaverageinterestrate(weightedbyoriginalbalanceofindividualloancomponents).Dwellingcontrolsarepropertytypeandbuilding’sagecategory(five-yearbins).Borrowercontrolsincludetotalincomeandborrower'sageatmortgageorigination(three-yearbins).Marketcontrolsareend-of-monthDutchunemploymentrate,10-yearGermangovernmentbondyield,monthlyvolatilityofdaily10-yearGermangovernmentbondyields,andtheend-of-monthyieldcurveslope(measuredas10-yearminus1-yearEURswaprates).Yearandregionfixedeffects(FE)areincludedwhereindicated.Standarderrors(reportedinsquarebrackets)areeitherrobustorclusteredatprovinciallevel.Statisticalsignificanceisdenotedby***,**,and*atthe1%,5%,and10%level,respectively.

(1) (2) (3) (4)EE(A/Brating) 0.4892*** 0.2582* 0.1857* 0.1857**

[0.0473] [0.2084] [0.1612] [0.1412]CurrentLTV 15.3931*** 21.9993*** 21.9993***

[6.0488] [9.9734] [9.6332]DSCR 0.9487 0.9676 0.9676

[0.0567] [0.0623] [0.0533]Mortgageterm 0.7435 0.5910 0.5910* [0.2262] [0.1994] [0.1640]Borrowercontrols No Yes Yes YesDwellingcontrols No Yes Yes YesMarketcontrols No No Yes YesMortgagecontrols No Yes Yes YesRegionFE No Yes Yes YesYearFE No Yes Yes YesSE Rob. Rob. Rob. RegionCl.Observations 126,036 125,560 125,560 125,560PseudoR-squared 0.00729 0.271 0.414 0.414

Wevalidatetheabovefindingswithanumberofrobustnesschecks.Forthispurpose,wetakethemodelpresentedincolumn(4)ofTable5asthebaselinemodelspecificationandreplace,redefineoraddcovariatesasdescribedfurtherbelow.ThevariousmodelspecificationsarepresentedinTable6,wherewereport forconveniencepurposesonlytheORestimatefor theenergyefficiencydummyvariable.Sinceitiscommontoestimateacreditriskmodelwithoriginalcovariates,wereplacetheexplanatoryvariablescurrentLTVandcurrenttotal incomewithoriginalLTVandtotal incomethatwasreportedattheearliestdateinthesample.AspresentedunderSpec.1,themainresultsarenotdrivenbythecovariates'reportingdate.Spec.2to5indicatethattheresultsarenotaffectedbythedefinitionofbuilding’sageandborrower’sagecategory.InSpec.2and4,weusetheactualbuildingandborrowerage,respectively.In3and5,weredefinetheagecategoriesfrom3-and5-yearcategory


16/26

to9-and15-yearforbuildingandborrowerage,respectively.Thebaselineregressionwasestimatedby omitting the DTI and the original balance due to multicollinearity concerns. The correlationbetweenDTIandLTV(betweenDTIandDSCR)isrelativelyhighat0.51(0.68).Similarly,totalincomeandtotaloriginalbalanceexhibitacorrelationcoefficientof0.73. InSpec.6 (Spec.7), currentDTI(originalbalance)isaddedtothebaselinespecification,whileSpec.8includesbothofthesetwolastcovariates.AspresentedinTable6,theinclusionofthetwocovariatesdoesnotaffectthemainresult.However,unreportedresultsshowthattheinclusionofeitherthetwoorbothvariablesdistortstheregressioncoefficientsofothercontrolvariables.

Table6-Logisticregression-robustnessofresultsThis table presents the odds ratio estimates and the respective standard errors (SE) for the EE variable using various Logistic modelspecifications.Thedependentvariableisadummyindicatingifamortgageisindefault(i.e.,inarrearsforatleastthreemonths)ornot.Thebaselinemodel specification refers toTable5, column (4). Themodel specifications1 to8 in this tablediffer from thebaselinemodelaccordingtothefollowingchanges.Spec.1:thetwoexplanatoryvariablescurrentLTVandcurrenttotalincomearereplacedbytheoriginalLTVandoriginaltotalincomethatwasavailableattheearliestdateinthesample.Spec.2:3-year-building’sagecategoryisreplacedbyactualbuildingage.Spec.3:3-building’sagecategoryisreplacedby9-year-building’sagecategory.Spec.4:5-year-borrower’sagecategoryisreplacedbyactualborrower’sageatoriginationofearliestloancomponent.Spec.5:5-year-borrower’sagecategoryisreplacedby15-year-borroweragecategory.Spec.6:currentDTIisaddedtothebaselinemodel.Spec.7:originalbalanceisaddedtothebaselinemodel.Spec.8:currentDTIandoriginalbalanceareaddedtothebaselinemodel.Statisticalsignificanceisdenotedby***,**,and*atthe1%,5%,and10%level,respectively.

Model OR(EE) SESpec.1 0.1670** [0.1308]

Spec.2 0.2089** [0.1577]

Spec.3 0.4931** [0.1765]

Spec.4 0.1770** [0.1373]

Spec.5 0.1861** [0.1367]

Spec.6 0.1927** [0.1451]

Spec.7 0.1852** [0.1407]

Spec.8 0.1934** [0.1456]

4.3 ExtendedCoxModelResults

TheCoxmodelistypicallyemployedtostudysurvivaldataovertime.Sincethepresentlyuseddatasetallowstoperiodicallytrackamortgage's‘health’,weapplytheextendedCoxmodelwithtime-varyingcovariatesfortheperiodJanuary2014toMay2018.

Before presenting the regression results, it is important to confirm if the proportional hazardsassumptionholdsasitmightaffecttheinterpretationoftheresults.Figure1presentstheempiricalsurvivor functions for energy efficient and non-energy efficientmortgages. On the basis of visualanalysis, it ispossibletoobservethatthetwocurvesneithercross,nordotheydivergetoomuch,suggesting that the proportionality assumption holds. The implication of this finding is that the


17/26

estimatedodds ratio for theenergyefficiency variable canbe assumed tobe constantover time,meaningthattheestimatesarenotdependentonthereportingtimeofthemostrecentobservation.Additionally, thesurvivorcurvessuggestthat,onaverage,energyefficientmortgagessurviveforalongerperiodthantheirnon-efficientcounterpartsasindicatedbytheslightlywideninggapbetweenthetwocurves.

Figure1–SurvivorFunctionsThisfigureshowsKaplan-Meiersurvivalcurvesfortwomortgagegroups:mortgageswithenergyefficient(EE=1)andnon-energyefficient(EE=0)buildings.TheLog-ranktest,whichtestsforequalityofsurvivorfunctions,yieldsap-valueof0.0001.

TofurtherexploretheobservedrelationbetweenEEandsurvivaltime,weestimatetheextendedCoxregressionwith time-varyingcovariatesandpresent the results inTable8.Column (1) reports theestimated odds ratio not controlling for any mortgage and other characteristics. The regressioncoefficient is below one and highly significant, confirming the findings obtained from the Logisticregression.Energyefficiencyseemstobeassociatedwithalowerprobabilityofmortgagedefault.Aswe can observe in columns (2) to (4) of Table 8, accounting for the time-varying nature of thecovariates(currentLTV,DSCR,totalincome,andthemacroeconomicvariables)doesnotqualitativelyaffectmuchthemainfinding.

0.95

0.95

0.96

0.96

0.97

0.97

0.98

0.98

0.99

0.99

1.00

% d

efa

ult−

free

0 50 100 150 200 250 300

Months default−free

EE = 0 EE = 1


18/26

Table7-ExtendedCoxmodelresultsThistablepresentsextendedCoxoddsratioestimatestodeterminetherelationshipbetweenresidentialbuildingsenergyefficiencyandborrowers’defaultrisk.Thedependentvariableisadummyindicatingwhetheramortgageisindefault(i.e.,inarrearsforatleastthreemonths)ornot.ThemainexplanatoryvariableisthedummyvariableEEthatequalstooneifabuilding'senergyefficiencyratingisAorB-ratedandzerootherwise.MortgagecontrolsarecurrentLTV,DSCR,averagemortgageterminmonths(weightedbyoriginalbalanceofindividualloancomponents),andtheaverageinterestrate(weightedbyoriginalbalanceofindividualloancomponents).Dwellingcontrolsareproperty typeandbuilding’sagecategory (fiveyear-bins).Borrowercontrols include total incomeandborrower'sageatmortgageorigination.Marketcontrolsareend-of-monthDutchunemploymentrate,10-yearGermangovernmentbondyield,monthlyvolatilityofdaily10-yearGermangovernmentbondyields,andtheend-of-monthyieldcurveslope(measuredas10-yearminus1-yearEURswaprates).Yearandregionfixedeffects(FE)areincludedwhereindicated.Standarderrors(reportedinsquarebrackets)areeitherrobustorclusteredatprovinciallevel.Statisticalsignificanceisdenotedby***,**,and*atthe1%,5%,and10%level,respectively.

To validate these results, similar robustness exercises are applied as in the case of the Logisticregression.Table8presentstheresults.Theestimatessuggestthatneitherredefiningborrower’sandbuilding’s age categories (Spec. 1 to 4), nor including additional covariates that might raisemulticollinearityconcerns(Spec.5to7)doesaffectthemainfinding(i.e.,energyefficiencyinvolvesalower default risk). Overall, the results are quantitatively and qualitatively similar to the baselineestimate,exceptforreplacingthebuildingsageinSpec2.

(1) (2) (3) (4)EE(A/Brating) 0.5683*** 0.4200** 0.4192** 0.4192**

[0.0555] [0.1454] [0.1450] [0.1519]CurrentLTV 50.6845*** 49.0551*** 49.0551***

[15.1644] [14.6160] [13.2560]DSCR 1.0242 1.0291 1.0291

[0.0463] [0.0465] [0.0399]Mortgageterm 0.3195*** 0.3401*** 0.3401*** [0.0854] [0.0928] [0.0876]Dwellingcontrols No Yes Yes YesHouseholdcontrols No Yes Yes YesMarketcontrols No No Yes YesMortgagecontrols No Yes Yes YesRegionFE No Yes Yes YesYearFE No Yes Yes YesSE Rob. Rob. Rob. RegionCl.Observations 1,173,551 1,114,615 1,114,615 1,114,615PseudoR-squared 0.00271 0.0465 0.0471 0.0471


19/26

Table8-ExtendedCoxmodel-robustnessofresultsThistablepresentstheoddsratioestimatesandtherespectivestandarderrors(SE)fortheEEvariableusingvariousCoxmodelspecifications.Thedependentvariableisadummyindicatingifamortgageisindefault(i.e.,inarrearsforatleastthreemonths)ornot.ThebaselinemodelspecificationreferstoTable7,column(4).Themodelspecifications1to7 inthistabledifferfromthebaselinemodelaccordingtothefollowingchanges.Spec.1:3-year-building’sagecategoryisreplacedbyactualbuilding’sage.Spec.2:3-building’sagecategoryisreplacedby9-year-building’sagecategory.Spec.3:5-year-borroweragecategoryisreplacedbyactualborrower’sageatoriginationofearliestloancomponent.Spec.4:5-year-borrower’sagecategoryisreplacedby15-year-borrower’sagecategory.Spec.5:currentDTIisaddedtothebaselinemodel.Spec.6:originalbalanceisaddedtothebaselinemodel.Spec.7:currentDTIandoriginalbalanceareaddedtothebaselinemodel.Statisticalsignificanceisdenotedby***,**,and*atthe1%,5%,and10%level,respectively.

4.4 AdditionalFindings

So far, theabovepresentedanalyses focusedon thequestionwhether thereexistsanysignificantrelation between a building's energy efficiency rating and the probability of its owners’mortgagedefault.Giventheaffirmative findings,wedecideto includeamoredetailedrepresentationofEE.Therefore,followingthefindingsofKazaetal.(2014),weassumethatthemoreefficientbuildingsareassociatedwitharelativelylowerriskofdefault.

For the purposes of the analysis, new indicator variables are created. They aggregate the energyefficiencyratingaccordingtofourefficiencyclasses.Efficiencyclass1assumesenergyratingsAandB,class2isassignedtoratingsCandD,class3isassignedtoratingsEandF,andclass4isreservedtoG-ratedbuildings.Allotherexplanatoryvariablesremainunchanged.Table9presentstheregressionresults for both regression methodologies (Logistic regression: columns (1) to (4), extended Coxmodel:columns(5)to(8)).Wecanobservethatthefindingsarelesspronouncedcomparedtothemainanalysis.Overall,theestimatedoddsratiosforratingclasses1to3exhibitanincreasingpatternwiththedegreeofenergyinefficiency:thelowertheEEratingclass,thehighertheassociatedriskofdefault.However,theexplanatorypoweroftheseresultsdiminisheswiththeinclusionofadditionalcontrolvariables.Thismightbeattributedtotheinherentimprecisionoftheratingsintheconstructeddataset.Inthemainanalysis,wecanassumethatthegeneralclassificationofbuildingsintothetwocategories“energyefficient”and“energyinefficient”ismoreorlessaccurate.AnymisspecificationsarelikelytoariseonlyattheB-andC-ratingthresholdandduetothelawoflargenumberstheyarenegligibleas longas thenumberofobservations is largeenough. In theanalysison thedegreeofefficiency,however,twoadditionalratingthresholdsareadded(attheD/EandtheF/Gthreshold).

Model OR(EE) SE.

Spec.1 0.2347** [0.1549]

Spec.2 0.6563 [0.1697]

Spec.3 0.4187** [0.1513]

Spec.4 0.4184** [0.1512]

Spec.5 0.3902*** [0.1389]

Spec.6 0.4155** [0.1505]

Spec.7 0.3839*** [0.1365]


20/26

This leavesadditional roomformisspecificationandcan, thus, lower significanceof theestimatedORs.Thatis,thepresentedfindingsareindicativeofarelationbetweenthedegreeofenergyefficiencyandcreditrisk.However,onlyanexactmatchingbetweenthemortgagedataandthebuilding'senergyratingwillprovidetrueinsightsintothisissue.Weleavethisforfutureresearch.

Table9-DegreeofEnergyEfficiencyThistablepresentsLogisticregression(columns(1)to(4))andextendedCoxregression(columns(5)to(8))oddsratioestimatestodeterminethepropensitytodefaultonmortgagesbackedbyenergyefficientbuildingswithdifferentdegreesofenergyefficiency.Thedependentvariableisadummyindicatingifamortgageisindefault(i.e.,inarrearsforatleastthreemonths)ornot.Themainexplanatoryvariablesarefourenergyefficiencycategories:(i)dummyvariableifabuilding'senergyefficiencyratingisAorB-ratedandzerootherwise,(ii)dummyiftheratingisCorD,(iv)dummyiftheratingisEorF,and(v)dummyiftheratingisG(theomittedcategoryintheregressions)andzerootherwise.AllothercontrolvariablesaredefinedasinTable6andTable8fortheLogisticandextendedCoxregression,respectively.Robuststandarderrorsarereportedinsquaredbrackets.Yearandregionfixedeffects(FE)areincludedwhereindicated.Statisticalsignificanceisdenotedby***,**,and*atthe1%,5%,and10%level,respectively.

Fromthepresentedresults,wecanconcludethatmortgagesbackedbyenergyefficientresidentialbuildingsarecorrelatedwitha lower riskofdefault.This statement seems tobevaliddespite theimprecisedefinitionoftheEEvariable.Namely,themaindrawbackoftheanalysisisthatabuilding’senergyefficiencylabeldependsonthebuilding’stypeandconstructionyearonly.Thismeansthatwedonotknowabuilding’sactualEEratingbutonlythestatisticalrepresentativeEEratingforanaverageDutchdwellingofsametypeandconstructionyear.Duetotherelativelylargesamplesize,however,thelawoflargenumbersshouldapply,whichmeansthatoursamplemeanshouldconvergetothepopulation mean. Furthermore, we account for building type and construction year effects bycontrollingforbothvariablesintheaboveanalyses.Ourfindingsarestatisticallysignificantandsurviveabatteryofrobustnesschecks.Additionally,theresultsindicatethatthedegreeofenergyefficiency

Logisticmodel ExtendedCoxmodel (1) (2) (3) (4) (5) (6) (7) (8)A/Brating 0.3891*** 0.0068*** 0.0074*** 0.4165 0.4947*** 0.8439 0.8533 0.8204

[0.0458] [0.0081] [0.0087] [0.5189] [0.0596] [0.5426] [0.5454] [0.5236]C/Drating 0.6472*** 0.2278** 0.2383** 1.3582 0.7272*** 2.0106 2.0136 1.9300

[0.0658] [0.1625] [0.1671] [1.2553] [0.0751] [1.0532] [1.0450] [1.0105]E/Frating 0.9642 0.3666 0.3936 0.8053 1.0900 1.5683 1.5736 1.5325 [0.1095] [0.2471] [0.2611] [0.6333] [0.1253] [0.7317] [0.7281] [0.7173]Dwellingcontrols No Yes Yes Yes No Yes Yes YesHouseholdcontrols No Yes Yes Yes No Yes Yes YesMarketcontrols No No No Yes No No No YesMortgagecontrols No No No Yes No Yes Yes YesRegionFE No No Yes Yes No No Yes YesYearFE No No No Yes No No No YesSE Rob. Rob. Rob. Rob. Rob. Rob. Rob. Rob.Observations 127,309 127,082 127,082 127,082 1,173,515 1,123,632 1,123,632 1,123,632PseudoR-squared 0.0102 0.138 0.145 0.517 0.00407 0.0439 0.0466 0.0479


21/26

alsomatters,i.e.moreenergyefficientbuildingsareassociatedwithrelativelylowerriskofdefault.


22/26

5. UK

5.1 Background

AsdescribedinthereportD5.1,theEnglishgovernmentconductsatayearlyfrequencyasurveyaboutpeople’shousingcircumstances(EnglishHousingSurvey-EHS).10OnthebasisoftheEHSdataset,thedescriptivestatisticsofthechosenvariableswerepresentedinthetechnicalreportD5.2.ThemainfindingswerethatthepercentageofEEbuildingsvariesconsiderablyamongregions,andtheshareofEEmortgages’defaultsisgenerallylowerrelativetotheirnon-EEcounterparts.Concerninghouseholdcharacteristics,theageofthehouseholdrepresentativediffersslightlybetweentheEEandnon-EEgroup.EEborrowerstendtobeyoungerwithanaverageageof42years.Further,theyhaveaslightlyhigherhouseholdsizeandannualincomerelativetotheirnon-EEcounterparts.Atlast,morerecentlyconstructed buildings are in general more energy efficient, while less efficient dwellings areconstructedmostlybefore1950.Inthefollowingsection,weemploytheLogisticregressioninordertoanalyzetherelationbetweenabuilding’sEElevelandthecorrespondingborrower’sriskofdefault.


Inthefollowing,weemployEHShousehold-levelinformationcollectedintheyears2009,2010,2012,and2013.AsoutlinedinreportD5.2,ourfinalsampleconsistsof4,737observationsaftercleaningthedataset.

OuranalysisfocusesonthecorrelationbetweentheDegreeofEEandtheprobabilityofmortgagedefault.The independentvariableDegreeofEE isdefinedas theenergyefficiency rating thatwascomputedundertheStandardAssessmentProcedure(SAP) intheUKforassessingtheenergyandenvironmentalperformanceofhomes.TheSAPmeasurestheenergyefficiencyofindividualdwellingson a scale of 1 to 100, where 1 indicates a very poor and 100 stands for an excellent energyperformance.Similar to theanalysespresented for theothercountries,wecontrol forhousehold,mortgage, and dwelling characteristics. As household controls, we use total income and age ofhouseholdrepresentative.MortgagecontrolscompriseoriginalLTV,DSCR,andmortgageageinyears.Dwellingcontrolsaccountforsize intermsoftotalarea insquaremeters.Additionally,we includeregionandyearfixedeffectsinordertoaccountforregion-specificdifferencesandtemporalshockscommontoallborrowers,whichmaystemfromfactorssuchasbusinesscycledynamics.

Table10presentstheoddsratioestimatesfromaLogisticregression.Column(1)reportstheresultswithout controlling for anyother variables in themodel. Theestimatedodds ratio 0.4481 for theDegreeofEEindicatesthatborrowerswithmoreenergyefficientdwellingsdefaultlessoftenontheirmortgage.However, the finding is insignificant. Thismightbe attributed to the small sample size.

10Formoredetails,refertoDCLG(2017)andMHCLG(2018).


23/26

Inclusionofadditionalcontrolvariables(seecolumns(2)and(3))doesnotaffectthisfinding:theORestimateremainsbelowoneandstaysinsignificant.Unreportedresultsindicatethatthisfindingdoesnot change neither (i) by redefining the EE variable nor (ii) by using “repayment difficulty” asdependentvariable.Intheformercase,wecategorizebuildingsintotercilesaccordingtotheDegreeofEEanddefineadummyvariableEEthatequalstooneifabuildingisinthehighestEEtercileandzerootherwise. In the lattercase, thedefaultdummyvariable is replacedby repaymentdifficulty,whichequalsone if a survey respondent reports repaymentdifficultieson themortgageand zerootherwise. Inboth cases, the results remain insignificantandareavailable from theauthorsuponrequest.

Table10-LogisticregressionresultsThistablepresentsLogisticregressionoddsratioestimatestodeterminetherelationshipbetweenresidentialbuildingsenergyefficiencyandborrowers’defaultrisk.Thedependentvariableisadummyindicatingwhetheramortgageisindefault(i.e.,inarrearsforatleastthreemonths)ornot.Themainexplanatoryvariableisbuilding’sdegreeofenergyefficiency,asmeasuredundertheSAP2009andtheRDSAP2009(whereSAP2009isnotavailable).MortgagecontrolsarecurrentLTV,currentDSCR,andmortgageageinyears.Dwellingandborrowercontrolvariablesaredwellingsizeinsquaremeters,householdrepresentative’sage,andtotalhouseholdincome.Yearandregionfixedeffects (FE)are includedwhere indicated.Standarderrors (reported insquarebrackets)areeitherrobustorclusteredatregional level.Statisticalsignificanceisdenotedby***,**,and*atthe1%,5%,and10%level,respectively.

(1) (2) (3)DegreeofEE 0.4481 0.8669 0.8669

[0.4510] [1.0108] [1.0831]OriginalLTV 2.4050*** 2.4050**

[0.7726] [0.8219]DSCR 0.8003*** 0.8003***

[0.0527] [0.0566]Mortgageage 1.0189 1.0189

[0.0208] [0.0182]Dwellingcontrols No Yes YesHouseholdcontrols No Yes YesMarketcontrols No No NoMortgagecontrols No Yes YesRegionFE No Yes YesYearFE No Yes YesObservations 4,694 4,669 4,669SE Rob. Rob. RegionCl.

The presented results are indicative of a negative correlation between EE and the probability ofmortgage default. However, the OR estimates are insignificant, andwe cannot draw any definiteconclusionfromthem.Thislackofsignificancecouldbeduetoanon-lineareffectoftheSAPratingrangingfrom1to100onmortgagedefaultandthefactthattherewerechangesintherating.Thelatter one resulted in relying on the Reduced Data SAP 2009where SAP 2009was not available.Therefore,amoredetailedanalysisofthisratingoraconsistentEEratingcouldimprovetheresults,whichwewillleaveforfurtherresearch.


24/26

6. ConclusionThegoalof this technical report is to investigatewhetherbuilding’senergyefficiency iscorrelatedwith theassociatedprobabilityofmortgagedefault. For this purpose,we focuson four Europeancountries:Belgium, Italy,Netherlands, andUK. Thecountry-specificdatasetsdiffer considerably interms of size, granularity, and energy efficiency definitions. Thus, our empirical analyses and thecorrespondingfindingsvaryacrosscountriesconsiderably.TheresultsfromItalyandUKindicate,butdonotstrictlyconfirm,anegativerelationbetweenenergyefficiencyandtheprobabilityofmortgagedefault. This can be attributed to the nature of the underlying datasets. In the case of Italy, thedatasetslackgranularityinordertoobtainconvincingfindings,whiletheUKdatasetisrelativelysmall.The loan-level analyses for Belgium and the Netherlands, on the other hand, indicate a robust,negative and significant correlation between the two variables of interest, EE and PD. This is anencouragingfinding,whichshouldmotivatefuturestudiesonestablishingacausallinkbetweenthetwovariables.

Thisdocumentcanbeconcludedwiththefollowingtakeaways.First,dataavailabilityisakeychallengeforconductingacreditriskanalysiswithrespecttobuildings’energyefficiency.Currently,onlyfewloan-level datasets exist that are accompaniedwith actual building energy efficiency information.Mostofthesedatasetsareeithertoosmallorhaveatooshorthistorytoperformareliablecreditriskanalysis.Second,acausalityanalysis iscurrentlyoutofscopeandis leftforfutureresearch.Third,among the most granular datasets (Belgium and Netherlands) a significant negative correlationbetween energy efficiency and mortgage default risk prevails, and it is robust to modelmisspecificationsandvariabledefinitions.


25/26

7. BibliographyAgentschapNL.(2011).Voorbeeldwoningen2011:bestaandebouw.6,1975-1991.An,X.andPivo,G.(2018).GreenBuildingsinCommercialMortgage-BackedSecurities:TheEffectsof

LEEDandEnergyStarCertificationonDefaultRiskandLoanTerms.RealEstateEconomics.(https://doi.org/10.1111/1540-6229.12228).

Boumeester,H.J.F.M.,H.C.C.HCoolen,C.P.Dol;R.W.Goetgeluk,S.J.T.Jansen,A.A.A.Mariën,and

E.Molin.(2008).ModuleConsumentengedragWoON2006,Hoofdrapport.Delft:OnderzoeksinstituutOTB.(https://www.rijksoverheid.nl/documenten/rapporten/2010/03/24/bijlage-1-hoofdrapport-module-consumentengedrag-woon-2006).

DCLH(DepartmentforCommunitiesandLocalGovernment).(2017)."EnglishHousingSurvey,2015:HousingStockData".UKDataService.SN:8186.(http://doi.org/10.5255/UKDA-SN-8186-1).

Hoebeeck,Annelies,andKoenInghelbrecht.(2017).TheimpactofthemortgageinterestandcapitaldeductionschemeontheBelgianmortgagemarket.(http://hdl.handle.net/10419/173780).

Kaza,Nikhil,RobertoQuercia,andChaoY.Tian.(2014).“Homeenergyefficiencyandmortgage

risks.”Cityscape:AJournalofPolicyDevelopmentandResearch16.(https://ssrn.com/abstract=2416949).

MHCLG(MinistryofHousing,CommunitiesandLocalGovernment).(2018).EnglishHousingSurvey:

privaterentedsector,2016-17. Underwood, J. and Alshawi, M. (2000). Forecasting building element maintenance within an

integrated construction environment. Automation in Construction, 9, 2, 169-184.(https://doi.org/10.1016/S0926-5805(99)00003-5).


26/26

This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 746205

EeMAP – Energy efficient Mortgages Action Plan is an initiative led by the European Mortgage Federation - European Covered Bond Council (EMF-ECBC), Ca’ Foscari University of Venice, RICS, the Europe Regional Network of the World Green Council, E.ON and SAFE Goethe University Frankfurt. For more information, visit: www.energyefficientmortgages.eu

d5.3 technical report on the econometric assessment and

Documents