Fourth report on the consistency of risk weighted assets

11 June 2014 Fourth report on the consstency of rsk weghted assets Resdental mortgages drll-down analyss Contents Executve summary 4 1. Introducton 5 2. Defntons 7 General 7 By varable 7 Loan-to-Value at Orgnaton (LTVO) 7 Indexed Loan-to-Value at Orgnaton (ILTV) 8 Debt-to-Servce at Orgnaton (DTSO) 8 Loan-to-Income at orgnaton (LTIO) rato 9 3. PD and LGD estmatons 9 Use of drll-down varables n rsk parameter estmatons 9 Combnaton of varables 11 4. Quanttatve analyss 12 RW correlaton and varablty by drll-down varables 12 Credt rsk mtgaton 13 Top-down analyss EU sample: benchmarks, portfolo composton (mx effect), RW levels and RW senstvty (prce effect) 13 RW senstvty to ILTV buckets n the EU sample 16 RW senstvty to ILTV buckets: country specfctes 18 RW senstvty and drll-down varables used n the PD/LGD estmatons 19 RW senstvty by vntage of loans 20 Analyses at bankng-group level: drll-down varables and RW 22 1

5. Conclusons 24 Summary of fndngs from qualtatve nformaton analyss 24 Summary of fndngs from quanttatve nvestgaton 24 Annexes 26 Annex 1: Country-weghted averages by country 26 Annex 2: Top-down detals on the methodology 27 Annex 3: Top-down analyss for the LTVO, DTSO and LTIO varables 28 Annex 4: Average RW by bucket 30 Annex 5: Applcaton of the top-down approach at country level 31 Lst of fgures Fgure 1: Obseved varant to the LTVO core defnton across European bank sample 8 Fgure 2:Observed varant to the DTSO core defnton across European bank sample 9 Fgure 3: Percentage use of the varables n the PD/LGD/both estmatons 10 Fgure 4: Share of the drll-down varables used n the estmatons, by country of localsaton of exposures. 11 Fgure 5: Correlaton and varablty of RWs by drll-down varables 12 Fgure 6: EU benchmark RW by drll-down varables 14 Fgure 7:ILTV break-down of the prce (level and senstvty) and bucket mx effects 15 Fgure 8: ILTV results Indexed (100) standard devaton dynamc after controllng the prce (level and senstvty) and bucket mx effects results 16 Fgure 9: ILTV - Correlaton between RW level and RW senstvty n the EU sample (top-down results) 17 Fgure 10: RW senstvty by ILTV for selected countres 19 Fgure 11: Average rskweght by bucket for ILTV and LTVO for groups of banks bult on the use/non-use varables n the models (PD, LGD, both or none) 20 Fgure 12: RWsand EAD by vntage of loans at orgnaton 20 Fgure 13: Average LTVO across the European bank sample by vntage of loans 21 Fgure 14: RW devaton from the benchmark RW (non-defaulted exposures) and comparson wth the estmated experenced loss rate (EAD-weghted PD for non-defaulted exposure tmes the provson rate (provsons/ead) for defaulted exposures), IRB RM portfolo, by bank 22 2

Fgure 15: Devaton of the LTVO, ILTV, DTSO and LTIO varables by bankng group compared to the European average, bass 100 23 Fgure 16: Mnmum, maxmum and EAD-weghted average for the drll-down varables, by country 26 Fgure 17: LTVO Break-down of the prce and bucket mx effects 28 Fgure 18: DTSO Break-down of the prce and bucket mx effects 28 Fgure 19: LTIO Break-down of the prce and bucket mx effects 29 Fgure 20: Average rskweght by loan-to-value rato at orgnaton 30 Fgure 21: Average RW by ILTV rato 30 Fgure 22: Average RW by DTSO 30 Fgure 23: Average RW by loan-to-ncome rato at orgnaton 30 Fgure 24: ILTV Break-down of the prce (level and senstvty) and bucket mx effects 31 Abbrevatons CRMO DTSO EAD ILTV LGD LTIO LTVO PD RM RW RWA Credt rsk mtgaton at orgnaton Debt-to-servce rato at orgnaton Exposure at default Indexed loan-to-value Loss gven default Loan-to-ncome at orgnaton Loan-to-value at orgnaton Probablty of default Resdental mortgage Rsk weghts Rsk-weghted asset 3

Executve summary 1. As part of the analyss regardng the consstency of rsk-weghted assets n resdental mortgage portfolos, ths report summarses the fndngs of the second phase: the drll-down analyss 1. The objectve s to attempt to understand f, and how, dfferent varables descrbng the portfolos (beyond the by-country 2 clusters) could explan the dfferences n rsk weght across the EU banks 3 found n phase one. 2. The nvestgated varables are the loan-to-value at orgnaton (LTVO), the ndexed loan-tovalue (ILTV), the debt-to-servce at orgnaton (DTSO), the loan-to-ncome at orgnaton (LTIO) and the credt rsk mtgaton at orgnaton (CRMO). Data by year of orgnaton for current 4 exposures has also been collected. For ths purpose, predefned buckets for each drll-down varable were used. 3. Accordng to the answers provded by the banks about the use of such varables, the (ndexed) loan-to-value and credt rsk mtgatons are more commonly used n the models; the debt-to-servce coverage rato and loan-to-ncome are used less frequently and only n PD models. 4. The documentaton provded by the banks hghlghted the use of dfferent defntons for smlar concepts. Sometmes they reflect country-specfc features, but overall the defntons are usually bank-specfc. Ths s an mportant caveat to consder when readng the outcome of the quanttatve analyss. 5. The study confrmed the exstence of a postve correlaton between the value of the dfferent drll-down varables (LTVO, ILTV, DTSO and LTIO) and the RWs at the sample level. However, RW senstvty to drll-down varables was not always found to be a clear explanatory factor of the rsk-weght varaton wthn the EU sample. 6. The EAD dstrbuton across the bucket values of drll-down varables has lttle mpact on the dsparty of RWs across the EU sample. 7. On examnng the correlaton and varablty analyses, the ndexed loan-to-value (ILTV) s the varable whch most sgnfcantly nfluences RW varaton. 8. The analyss by vntage confrms the exstence of a close lnk between the level of LTVO and RW and the potental nfluence of the portfolo composton by vntage n explanng the varaton n RWs. 1 Drll-down results for SMEs have already been presented n the December nterm report for SMEs and Resdental mortgages: http://www.eba.europa.eu/documents/10180/15947/20131217+thrd+nterm+report+on+the+consstency+of+rskweghted+assets+-+sme+and+resdental+mortgages.pdf 2 In lne wth the frst report, the country of exposure s the country where the collateral s located. 3 Across the 14 EU jursdctons partcpatng n ths study, 43 banks submtted data for up to 10 countres. 4 By December 2012, as for phase one. 4

9. Some country specfcs have been dentfed. Credt rsk mtgants other than mortgages are mportant drvers whch should be consdered when assessng the varaton n some countres, and seems to explan the lower RW senstvty to the value of the fnanced real estate. 10. Further, when studyng the level of the average drll-down varables at the bankng-group level, the use of a specfc combnaton does not appear to explan the dfferences n RWs. 11. Fnally, the drect contrbuton of drll-down varables to the estmaton of PD, LGD or both models does not seem to dscrmnate across banks n terms of RWs. Ths s probably because when t s not reported that these varables have been used n the estmatons, t can be assumed that flterng credts based on those varables, when grantng a loan to a customer, wll play an ndrect but sgnfcant role. 1. Introducton 12. Ths note reports on the second phase of the resdental mortgage analyss, namely the nvestgaton of the so-called drll-down varables and ther mpact on RWs: loan-to-value at orgnaton (LTVO); ndexed loan-to-value (ILTV); debt-to-servce at orgnaton (DTSO); loan-to-ncome at orgnaton (LTIO); credt rsk mtgaton at orgnaton (CRMO). 13. For ths analyss, the 43 banks n the EBA sample were asked to report ther non-defaulted exposure-at-default (EAD) and ther average RW by pre-set buckets for each of the drll-down varables. Ths nformaton was collected for several European countres of exposures. Ths approach was necessary to nvestgate trends at the aggregate level. Banks were therefore allowed to make use of ther own defntons for each varable, whch they had to document. 14. The drll-down varables were selected from the varables that are commonly reported by the banks as beng major rsk drvers used n the banks rsk parameter estmatons n the frst phase of the Resdental mortgages study. The fnal selecton of the varables 5 was done by EBA usng expert judgement. 15. The frst secton of the report provdes an overvew of the defntons of drll-down varable used by banks. In the second secton, the use of drll down varables n the modellng of the rsk parameters s descrbed. The thrd secton covers the quanttatve analyss and fnally an overvew of the fndngs s gven. 5 The long lst ncluded: occuper owner/buy to let; type amortsaton, type contractual nterest rate, ndexed debt-toservce and loan-to-ncome. 5

16. The analyses were carred out to determne whether: a) RW varablty s drven by the drll-down varables; b) the portfolo composton by drll-down varables can explan dfferences n RW wthn the EU sample; c) any country-specfc patterns exst; d) the use of the varables n the PD/LGD estmatons are sgnfcant. 17. The study does not am to comment on the opportunty to use the dfferent drll-down varables, or dscuss the approprateness of the RW senstvty towards them. 18. The major lmtatons of the nformaton used n the study are: a) the use of drll-down varables on a standalone bass (rather than n combnaton) does not help to establsh whether dfferent RW senstvtes are caused by dfferent convergng or dstortng factors; b) the absence of a complementary approach (e.g. hypothetcal facltes for mult profles 6 ) to the real exposures data may severely lmt the understandng of modellng choces and the mpact of banks credt polces; c) the use of common and predefned buckets for all the countres nvolved may lack of granularty and lmts the ablty to understand RW senstvty; d) the absence of any nformaton on PD/LGD parameters does not allow drect nvestgaton of whether the drll-down varables manly nfluence a partcular parameter or detect possble compensaton effects; e) a separate but nterlnked ssue wth d) s that the absence of nformaton on when and how the mnmum LGD 10% 7 s appled lmts and potentally dstorts any assessments of the senstvty; f) the use by the banks of ther own defntons for the drll-down varables facltated the submssons (f avalable) of the data by the banks. However, the materalty of the use of dfferent defntons across the bank sample has not been possble to be assessed n terms of addtonal source of varaton.. 19. The results of the study must be read alongsde the results n the phase one report publshed n December 2013 1. A large part of the analyss was performed at EU sample level 8. 6 The use of hypothetcal facltes for resdental mortgages has been postvely tested n one country. Intally EBA evaluated ths opton but then decded that, consderng the lmted experence and complexty n provdng all the relevant detals to specfy the transactons, a hypothetcal faclty would not be used n the study. 7 Regulaton (EU) No 575/2013 of the European Parlament and of the Councl of 26 June 2013 on prudental requrements for credt nsttutons and nvestment frms (CRR), Art. 164(4). 8 An example of applcaton of some nvestgatons at country level s contaned n Annex 5. 6

20. The report has not factored n supervsory acton or model changes that have occurred snce the observaton date (31 December 2012). 2. Defntons General 21. The reported documentaton by banks ndcates a wde range of defntons of the dfferent varables, across banks and across countres. Wthn bankng groups, defntons also appear to accommodate country-specfc dfferences. 22. Not all banks report on all same aspects or n the same depth of detal for these defntons,.e. some banks provded only general nformaton. Ths does not allow a precse comparson of these defntons across banks nor an assessment of the full mpact of any defnton on the reported quanttatve data 23. In addton, some banks made use of proxes for reportng purposes, as some of the requested data were ether not avalable (not recorded) n a straghtforward way or not used nternally. However, the former often only occurred for part of the stock of loans; n recent years, many of the banks mentoned have been mprovng data collecton and storage for the types of varables we asked for. By varable 9 Loan-to-Value at Orgnaton (LTVO) 10 24. The most common defntons of LTVO used by banks nclude the followng. Numerator: the sum of all (orgnal) loan dsbursements. Denomnator: the market value of the property at orgnaton. If new loans (guaranteed by the same property) had been granted at later stages, these are generally taken nto consderaton n a new LTVO for ths property, whch takes nto account captal already rembursed. 25. In addton to ths core defnton, many varants were found, of whch a summary s presented n the followng table. 9 Ths secton provdes a summary of the nternal defntons used by the banks for reportng the data for each drlldown varable. Ths excludes CRMO for whch, accordng to the nstructons, the banks were requested to provde data breakdown for pre-defned credt rsk mtgants, and only n very few cases provded supplementary nformaton. 10 For LTVO, the closest concept to the rato was expected [loan amount at orgnaton/market value at orgnaton]; loan amount refers to the sum of loans granted aganst one property. 7

Fgure 1: Obseved varant to the LTVO core defnton across European bank sample Numerator Denomnator (+) Pror lens (-) Pror lens (-) Other CRM (+) Other CRM (+) Undrawn exposures (+) Further advances on property (+) Costs, fees (+) Non-housng loans but same collateral Splt over multple loans at orgnaton secured by the same collateral Dfferent value concepts : Estmated/expected Market value (wth harcut or not) Purchase prce Prce based on nternal models 26. There s no consstency n how the banks nclude the factors mentoned n the table above (some banks nclude the factors n the numerator, some n the denomnator, etc.). Ths makes any comparson at the sample level challengng. Indexed Loan-to-Value at Orgnaton (ILTV) 11 27. The ILTV defnton bulds logcally on the one at orgnaton for banks. Although, even more varaton seems to be added to the defnton due to the frequency of ndexng the values (quarterly, sem-annually or yearly) or the dfferent ndexaton methods used at natonal level. In the latter, next to the rather commonly used external ndcators (whch sometmes vary at country level), some banks also referred to the use of nternal models for ndexng the collateral valuaton or the use of a stressed value based on lowest prces or downturn adjustment. Debt-to-Servce at Orgnaton (DTSO) 28. The DTSO 12 was also nvestgated. In partcular, along wth the loan-to-ncome, the percentage of data not avalable was the hghest for ths varable, as qute often banks had only started to record and store those data very recently. 29. Agan, many varatons were found n the defntons used across banks. The followng table lsts a summary of the man ones n addton to the core defnton. No sgnfcant countryspecfc patterns were observed n these defntons. 11 By ndexed loan-to-value was expected any concept close to the rato (the current loan amount to the current market value). 12 By DTSO was expected any concept close to the rato (monthly nstalment/net monthly revenues avalable). 8

Fgure 2:Observed varant to the DTSO core defnton across European bank sample Numerator (+) Instalments of all non-housng loans (n ths bank or n other banks) (+) Charges (resdental costs, etc., chldcare, etc.) Denomnator (+) Updated for latest advances (+) X % rental ncome By contract or not (-) Instalments of all non-housng loans (n ths bank or n other banks) (-) Charges (resdental costs, etc., chldcare, etc.) Gross vs. net revenues - X% ncome guarantor (+) Non-regular professonal revenue Based on hypothetcal scenaros (standard credt) Based on stressed scenaros (+) Instalments of all non-housng loans (n ths bank or n other banks) Use of jont accounts or not Loan-to-Income at orgnaton (LTIO) rato 13 (+) Captal ncome Updated for latest advances (+) X% ncome guarantor 30. The LTIO often bulds on the monthly debt-to-servce defntons used by banks, as many banks smply scale the numerator and denomnator to one year, whle others drectly use the yearly avalable nputs. When scalng, factors range from 12 to 14 on average. 3. PD and LGD estmatons Use of drll-down varables n rsk parameter estmatons 31. Each bank was also asked f the varables were used as an nput n the estmaton of LGD, PD, both or none. In ths regard, the ndexed loan-to-value appears to be the man varable (58% 14 ) of the fve beng used n any of the estmatons. The second and thrd most commonly used varables are the LTVO and the credt rsk mtgaton. Loan-to-ncome s the least commonly used varable (see Fgure 3). 32. For some banks, the varables at orgnaton especally debt-to-servce, loan-to-ncome and loan-to-value ratos are ncluded n the credt assessment at orgnaton and are therefore an ndrect nput n the PD and LGD modellng. However, as they are not a formal varable n the modellng and they wll not appear n the percentages presented below. 33. Regardng the relatvely low percentage use of CRMO, the most smple explanaton s that f banks do not make use of multple credt rsk mtgaton technques (e.g. fnancal collateral, government guarantee, etc.), but smply use a mortgage, they may have mentoned not usng the CRMO varable n ther estmatons. 13 By LTIO rato, any concept close to the rato was expected (borrowed amount/(net) annual ncome). 14 Ether n PD, LGD or both models. In 42% of cases, the varable s therefore not used. 9

Fgure 3: Percentage 15 use of the varables n the PD/LGD/both estmatons Source: EBA data collecton (reference date: December 2012), EBA calculaton 34. The debt-to-servce rato and loan-to-ncome ratos are never used as nput n the LGD model only, whereas the credt rsk mtgaton varable s never used as nput n the PD model only. 35. The ndexed loan-to-value and credt rsk mtgaton are reported as beng manly used n LGD estmatons. The use of the loan-to-value at orgnaton for the PD modellng, as well as the sgnfcant use n both models (PD and LGD) for some varables (.e. ILTV, LTVO and CRMO) 16 was less expected. 36. Regardng the nfluence of the varables used n the RW senstvty, see Secton 5. 15 Out of a sample of 90 data ponts (banks exposures by country), across the 5 varables. No reportng for a varable s stll consdered to be a data pont. 16 For LTIO and DTSO few exceptons of applcaton are notable for both modellng, whle no bank reported to apply such varables for LGD only. 10

Fgure 4: Share of the drll-down varables used n the estmatons, by country of localsaton of exposures. Readng note: as an example, the green trangle for BE means that 60% of the banks wth exposures n Belgum reported to use DTSO wthn PD, LGD or both models. Countres wth less than four observatons are not reported. Source: EBA data collecton (reference date: December 2012), EBA calculaton 37. Fgure 4 provdes an overvew of the use of drll-down varables n the dfferent countres. 38. In some cases, the prevalent use of one set of varables for the reportng banks n the sample s more evdent. In partcular: - the ndexed loan-to-value s largely used for exposures n the Czech Republc, Portugal, Unted Kngdom, Span and Ireland; - the exstence of other credt-rsk mtgants s more sgnfcant for exposures n the Netherlands and France; - the DTSO s more sgnfcantly used for exposures n Italy and Belgum. Combnaton of varables 39. Analysng the use of drll-down varables combned wth others, the ndexed loan-to-value agan features as the most prevalent varable n any combnaton (more frequently used wth LTVO and CRMO). When the banks reported usng only one varable, the ILTV and the CRMO were agan the most common ones. 11

4. Quanttatve analyss 40. Banks reported ther EAD and RWA n the dfferent countres where they have exposures for each varable (LTVO, ILTV, DTSO, LTIO) and by bucket.. 41. For the LTVO, the banks also provded a further breakdown of the RWs by vntage of orgnaton. 42. Banks were also asked to provde quanttatve nformaton on the exposure amounts and rskweghted assets, dstngushng between the level and type of dfferent credt-rsk-mtgaton technques,.e. exposures fully and only secured at orgnaton by mortgages and others (CRMO). In partcular, more granular data has been collected for the exposures wth mandate, (government guarantees, fnancal nsttutons guarantees, personal guarantees and fnancal collateral). 43. The quanttatve data, combned wth the more qualtatve nformaton where possble, have been used n top-down, correlaton and varablty analyss, or more smply to produce descrptve charts at EU sample level. Country-level nvestgaton s lkely to be very useful n makng addtonal progress, facltatng the ntegraton wth frst-phase fndngs and understandng the dfferences between banks. RW correlaton and varablty by drll-down varables 44. For each drll-down varable, the correlaton between RW and the drll-down varables was measured, and the varablty (defned as the standard devaton RW n percentage of the smple average) of the RWs, makng use of the values reported by the banks for each bucket. Fgure 5: Correlaton and varablty of RWs by drll-down varables a. Correlaton b. Varablty (standard RW devaton n % smple average) Readng note: each colour represents the share of the observatons (bank rsk weghts for each drll-down varable for country x) havng dfferent range values for: a. correlaton between rsk weghts and drll-down buckets values; b. standard devaton (n % smple average rsk weghts) n the reported rsk weghts for dfferent drll-down buckets values. Source: EBA data collecton (reference date: December 2012), EBA calculaton 45. For all the varables, a correlaton above 80% was observed n the majorty of cases. In partcular, for the ILTV n around 65% of the observatons, the correlaton was above 80%. 12

46. Smlarly, among the dfferent varables observed, varablty was sgnfcantly hgher for ILTV where the standard devaton relatve to the mean was hgher than 25% n around 65% of the observatons. Credt rsk mtgaton 47. From the study of the LTVO defntons used by banks, and corroborated by the data analyses, t emerged that some banks make large use of the mortgage value as a credt mtgant when calculatng the LTVO rato. Some banks may even consder only the mortgage value n ths calculaton. 48. The exstence of other credt rsk mtgants s more relevant n some countres (see Fgure 16 n Annex 1). 49. Overall, the data collected suggest that when resdental mortgages exposures are secured by other credt rsk mtgants, the RWs are hgher, wth the excepton of some specfc natonal nstances such as: a. n France when exposure s supported by guarantees such as the Credt Logement, the average exposure-weghted RWs are lower (10%) than at portfolo level (13%); b. n the Netherlands when exposures are accompaned by the NHG (Natonale Hypotheek Garante) guarantee, the average exposure-weghted RWs are lower (8%) than at portfolo level (10%); c. n Belgum when there s a mandate, the average exposure-weghted RWs are lower (7%) than at portfolo level (10%). 50. In France, the Netherlands and Belgum, among the resdental mortgages exposures that are not only secured by mortgages, these natonal-specfc nstances are the most relevant. In some cases, hose features,, mght explan (.e. n France and the Netherlands) the lower senstvtes to LTV observed n the study. Top-down analyss EU sample: benchmarks, portfolo composton (mx effect), RW levels and RW senstvty (prce effect) 51. The data submtted by the banks for each drll-down varable have been used to calculate EU benchmarks (exposure-weghted average 17, see table below). Fgure 16 n Annex 1 also contans statstcs calculated for the drll-down varables n the dfferent countres and at EU sample level. 17 As suggested by the charts n the Annex 4, the use of the medan, n general, would produce benchmarks hgher for the large majorty of the buckets for the dfferent drll-down varables. 13

Fgure 6: EU benchmark RW by drll-down varables ILTV LTVO DTSO LTI Buckets RW Buckets RW Buckets RW Buckets RW [0-50%] 7% [0-50%] 7% [0-10%] 10% [0-1] 12% [50-60%] 8% [50-60%] 8% [10-20%] 11% [1-2] 11% [60-70%] 10% [60-70%] 10% [20-25%] 12% [2-3] 11% [70-75%] 11% [70-75%] 10% [25-30%] 13% [3-4] 12% [75-80%] 13% [75-80%] 14% [30-35%] 14% [4-5] 13% [80-85%] 15% [80-85%] 15% [35-40%] 15% [5-6] 14% [85-90%] 17% [85-90%] 17% [40-45%] 15% [6-7] 16% [90-95%] 20% [90-95%] 20% [45-50%] 16% [7-8] 18% [95-100%] 21% [95-100%] 19% [50-60%] 16% [100-105%] 24% [100-105%] 17% [60-70%] 17% [105-110%] 24% [105-110%] 16% [110-120%] 25% [110-120%] 20% [120-150%] 28% [120-150%] 13% 52. The EU benchmarks have been used to conduct top-down analyss 18 separately for each of the drll-down varables to see how the average RW at portfolo level (bank by country) s drven by a prce or a bucket mx effect (see paragraphs 53 to 55 for the defnton and Annex 2 for the calculatons). 53. The top-down results presented n ths secton are based on all the observatons avalable at EU sample level, and am to provde an overall pcture (although avalable, the nformaton by country s not used at all). Indeed, applyng the same approach at country 19 (cluster) level allows more specfc conclusons to be drawn, beneftng from the opportunty to control for market specfc nstances. Annex 5 llustrates the results of top-down analyss for one country based on the calculaton of a country benchmark. 54. The prce effect s calculated by applyng the benchmark share (average of the banks samples) to the dfference between a bank s RW and the benchmark RW. In ths way, the only dfference between banks stems from the dfference n RW level, and not from dfferent exposure to dfferent levels of drll-down varable (e.g. a dfferent dstrbuton of EAD by bucket of ILTV). 55. The prce effect has also been broken down n terms of the level effect and senstvty effect 20. The latter ams to measure the varablty, relatve to the benchmark, produced by applyng dfferent ncremental changes n the RW between two of the bank s contguous drll- 18 The top-down analyss has been performed replcatng the same methodology appled n the prevous studes. The general concept s to decompose the banks RW dfferences from the benchmark by the progressve neutralsaton of the potental varablty contrbuton caused by the drvers. 19 In the December SME and Resdental mortgages report (Phase 1), the country dmenson was dentfed as one of the key drvers n explanng RW varablty. 20 The senstvty effect s measured by comparng the ntal RW dfference to the one resultng from the mposton for each observaton of benchmark RW senstvtes (average for the banks sample for the ncremental changes n the RW for ncreasng drll-down bucket values). 14

down buckets. For level effect s ntended the resdual RW gap after controllng for senstvty, therefore tacklng the level of the RW ndependent to the RW changes across drll-down buckets 56. The bucket mx effect s calculated by applyng a bank s RW to the bank s devaton from the benchmark 21 EAD dstrbuton across the bucket. Therefore, here we are controllng for the dfference n RWs appled to the benchmark dstrbuton of EAD across bucket. 57. Notably, the prce effect s much more sgnfcant than the bucket mx effect across all drlldown varables 22. Ths means that the EAD dstrbuton over the dfferent buckets for each varable only has a mnor effect on the average RW at the portfolo level across the sample of banks. 58. The bucket mx effect, even f small at sample level, only seems to play a sgnfcant role for some bank portfolos. Some country-specfc patterns can also be seen, but wthout conclusve evdence, e.g. there s a possble LTVO bucket mx effect n Italy and the Netherlands, an ILTV-prce effect n Italy and Ireland, and a DTSO prce effect n Belgum. 59. In Fgure 7 23 and Fgure 8 below, the results for the top-down analyss for the ILTV sample are presented. Fgure 7:ILTV break-down of the prce (level and senstvty) and bucket mx effects Note: The banks are sorted by ther RW devaton. A bank may be represented several tmes f t has submtted data for more than one country. Source: EBA data collecton (reference date: December 2012), EBA calculaton 21 The benchmark EAD dstrbuton s the average EAD dstrbuton for the banks sample. 22 Among the dfferent drll-down varables, the ILTV has the hghest bucket mx effect (see Fgure 7, Annex 3 and Annex 4). 23 The results for other drll-down varables are presented n Annex 3. We focus here on the ILTV, as n the prevous sectons t has been shown to be the most used varable. 15

60. The level effect s more mportant than senstvty effect, but both are present at EU sample level (see Fgure 8) 24. There are sgnfcant dfferences n ther materalty and sgn (+/-) between the banks and countres 25 : n about 80% of the observatons, the senstvty effect s negatve (lower than the benchmark); around 70% of the banks (by country portfolo) have a postve level effect (RW after controllng senstvty hgher than benchmark). Fgure 8: ILTV results Indexed (100) standard devaton dynamc after controllng the prce (level and senstvty) and bucket mx effects results Source: EBA data collecton (reference date: December 2012), EBA calculaton 61. The bucket mx effect accounts for around one-thrd of the ndex standard devaton; and n 60% of the cases, the bucket mx effect s negatve (EAD concentraton towards lower (less rsky) ILTV buckets than the benchmark). RW senstvty to ILTV buckets n the EU sample 26 62. Upon fndng the prce effect to be the man contrbutor to the top-down analyss, an nvestgaton was undertaken to determne f there s a correlaton between level and senstvty effects, and to establsh the nature of the relatonshp between prce effects and average RWs for the ILTV sample 27. The correlaton was calculated (R squared = 85%) and smple graphcal analyss was used. 63. Fgure 9 plots the results of the top-down analyss for the RW level and senstvty effects. 64. As depcted n Fgure 9, the level (x-axs) and senstvty effects (y-axs) clearly correlate (postve level effects are assocated wth negatve senstvty effects and vce-versa). 24 The senstvty effect accounts for about 40% ((197.7-100)/((197.7-100)+(197.7-36.8)) of the prce effect and about one fourth of the overall varaton (100-36.8)/(100*40%). 25 The senstvty effect for ILTV appears more n PT, SE and the UK. 26 In Annex 3 are contans charts wth EAD and RW dstrbuton by buckets (1 st quartle, medan, exposure-weghted average and 3 rd quartle) for the dfferent drll-down varables. 27 The prelmnary correlaton and varablty RW analyss dentfy the ILTV as a potentally more useful varable (among the varables observed) n explanng RW varaton. 16

65. The chart n Fgure 9 was then used to analyse the extent to whch the dstance of the observatons, from the orgn of the coordnate (benchmark), was nfluenced by the average RWs: the lowest RWs (lower than the benchmark) are concentrated below the black lne n the thrd quadrant (south-west), and only a few appear n the lower area of the second quadrant (north-west). On the other hand, the majorty of the banks wth the hghest RWs are concentrated n the lower area of the fourth quadrant. Fgure 9: ILTV - Correlaton between RW level and RW senstvty n the EU sample (top-down results) Level effect (x axs) and senstvty effect (y axs) are the decomposed top-down results for prce effect. A bank may be represented several tmes f t has submtted data for more than one country. A small number of observatons wth extreme values n the south-east quadrant have not been reported. Source: EBA data collecton (reference date: December 2012), EBA calculaton 17

66. An n-depth nvestgaton 28 of selected observatons confrmed the fndngs of the top-down and correlaton analyss. The RW dfferences are nfluenced by the RW senstvty along the buckets for ILTV, but the relatve mportance may vary sgnfcantly among the other observatons. Smlar conclusons are reached for the RW-level effect. The presence of the RW senstvty (or level) does not correlate wth the RW gap from the benchmark. 67. In many cases, the RW dfferences appear to be explaned by the RW level beng drven by dfferences n the rskness of the portfolo (default and loss rates) or, for example, by conservatve margns n the estmates. RW senstvty to ILTV buckets: country specfctes 68. To detect the presence of any country-specfc patterns, the observatons were grouped by country, and the average RW were plotted for each ILTV bucket (see Fgure 10). 69. Exposures n the UK, Span, Sweden and Denmark show, on average, the hghest RW senstvty wth the level of the ILTV (the black lne represents the hypothetcal constant cumulatve varaton of 15% from the bass (100)). For exposures n the UK and Span, and to a lesser extent for Sweden and Denmark, ths s also n lne wth the answers provded by the banks regardng the wdespread use of these varables n the models. 70. Exposures n the Netherlands, France, Luxembourg and Ireland show the lowest RW senstvty (below the black dotted lne, whch represents the hypothetcal cumulatve varaton of 10% from the bass (100)). The lower senstvty observed n France and the Netherlands may be explaned n part by the potental nfluence of other credt rsk mtgants ( credt lodgement and NHG ). 28 Not ncluded n the report. 18

Fgure 10: RW senstvty by ILTV for selected 29 countres 1,000 100 [0;50%] [50;60%] ]60;70%] ]70;75%] ]75;80%] ]80-85%] ]85;90%] ]90;95%] ]95;100%] BE DK ES FR IE IT LU NL PT SE UK K_15% K_10% Note: Logarthm values calculated usng the frst bucket as a bass (100). Wthn the same country, there are notable dfferences among the reportng banks. Source: EBA data collecton (reference date: December 2012), EBA calculaton. RW senstvty and drll-down varables used n the PD/LGD estmatons 71. For each drll-down varable, we grouped the banks based on ther answers about the use of drll-down varables n the models (both, LGD, PD or none), and nvestgated the potental dfference n relatve RW senstvty. The analyss dd not provde strong evdence, but t was possble to observe that: - for ILTV between 50% and 80%, the RWs are sgnfcantly hgher when the banks use the varable n the LGD (pnk lne) or both models (green lne) than n the other cases; - for LTVO above 80%, the RWs are hgher when the banks use the varable n the LGD models (pnk lne). When the varable s used n both PD and LGD models, the RWs are lower n most of the buckets (green lne). 29 Only EU countres wth at least fve observatons 19

Fgure 11: Average rskweght by bucket for ILTV and LTVO for groups of banks bult on the use/non-use varables n the models (PD, LGD, both or none) a. ILTV b. LTVO Source: EBA data collecton (reference date: December 2012), EBA calculaton RW senstvty by vntage of loans 72. Usng the by vntage at orgnaton nformaton, t s possble to nvestgate f there s any notable relatonshp between vntage and RW levels, the stablty over tme of the LTVO, and the extent to whch the dfferent exstng portfolo mx by vntage explan the varablty n RWs. 73. As shown n Fgure 12 30, the RWs appear to be senstve to the year n whch loans orgnated. Indeed, by lookng at the medan and the nterquartle dstrbuton, a clear ncrease n RWs from 2001 to 2007, and a decrease n the followng years untl 2011 can be seen. Fgure 12: RWsand EAD by vntage of loans at orgnaton Note: Wthn the sample, notable banks dsplay a more stable or much stronger dynamc (ncrease/decrease) n the RW over tme. Source: EBA data collecton (reference date: December 2012), EBA calculaton 30 For data qualty ssues, the data from one outler bank have been dscarded. 20

74. Smlarly, by examnng the varaton of the average LTVO 31 by year of vntage, Fgure 13 shows that the evoluton s smlar to that of RWs across the vntage years, wth an ncrease from 2001 to 2007; n the followng years, untl 2012, the trend s less stable and not always consstent, but ths could be due to outlers dstortng the weghted average. Fgure 13: Average LTVO across the European bank sample by vntage of loans Note: Wthn the sample, notable banks dsplay a more stable or much stronger dynamc (ncrease/decrease) n the LTVO over tme. Source: EBA data collecton (reference date: December 2012), EBA calculaton 31 The statstcs are calculated based on an EAD-weghted LTVO for each vntage year, for each bankng group. Therefore the statstcs do not represent the varaton of LTVO for partcular bankng group across countres of exposures. For exposures belongng to the bucket above 150% LTVO, 160% have been appled. Otherwse for each bucket the upper bound of the bucket has been used. 21

Analyses at bankng-group level: drll-down varables and RW 75. The prevous analyses have taken nto account the country dmenson. Here, the purpose s to understand to what extent the combnaton of the drll-down varables may nfluence the RWs at the bankng-group level. 76. For ths purpose, t s necessary to study whether the average level of drll-down varables s related at the level of RWs by bankng group. 77. Frstly, Fgure 14 32 represents the relatonshp between the RW devaton and the estmated experenced loss rates for the European bank sample. 78. Fgure 14 shows that the experenced loss rate s a relevant explanatory factor for the varablty n RWs wthn the EU sample. Fgure 14: RW devaton from the benchmark RW (non-defaulted exposures) and comparson wth the estmated experenced loss rate (EAD-weghted PD for non-defaulted exposure tmes the provson rate (provsons/ead) for defaulted exposures), IRB RM portfolo, by bank Note: Banks are sorted by ther RW devaton. Source: EBA data collecton (reference date: December 2012), EBA calculaton 32 Ths fgure was frst presented n the report publshed n December 2013, p. 28, Fg. 12 https://www.eba.europa.eu/documents/10180/15947/20131217+thrd+nterm+report+on+the+consstency+of+rskweghted+assets+-+sme+and+resdental+mortgages.pdf 22

79. Secondly, t s useful to determne whether the level of the estmated experenced loss rate was drven by the dfferent levels of drll-down varables. 80. For ths purpose, the average level of drll-down varable for the same bankng groups can be seen below. 81. Fgure 15 represents the devaton from the EAD-weghted average drll-down varable for each bank. The banks are ordered by RW devaton (as n Fgure 14). Fgure 15: Devaton of the LTVO, ILTV, DTSO and LTIO varables by bankng group compared to the European average, bass 100 Note: The banks are sorted by ther RW devaton. Example case to assst wth readng the table: The banks are ordered around the spral chart by ncreasng RW, so the frst bank at the rght of the vertcal axs (Bank 13) s the one wth the lowest RW. For ths bank, the average LTIO ndex s around 110 (red damond). The bass 100 for the LTIO s 4.3, therefore ths bank has an average LTIO that s around 10% hgher than the sample average (so, close to 4.8). For the LTVO and ILTV, ths bank has ts ndexes close to 90, meanng that ths bank has an average LTVO and ILTV 10% below the sample LTVO and ILTV (whch means an average LTVO around 70% and an average ILTV around 66%, as the bass 100 for LTVO and ILTV are 76.3% and 73.6%). For DTSO, the bank s close to the benchmark wth an ndex of around 95 (a DTSO value around 26%). Source: EBA data collecton (reference date: December 2012), EBA calculaton 82. From those fgures (Fgure 14 and Fgure 15), t seems clear that the devaton for the level of drll-down varables does not explan the devaton n RWs (or the experenced loss rate), as no common pattern s found among banks wth low RW devaton, or among the ones wth hgh RW varaton. 23

83. Ths fndng has the same lmtatons as the former analyses, as the defntons of the drlldown varables are dfferent across and wthn the European bankng group, and only data at the bucket level has been collected 33. 84. Nevertheless, ths analyss has stll an advantage n that t shows the varablty among the EU bank sample n terms of the average level of the drll-down varable. As shown prevously, ths may be due to exposures n dfferent countres but also to bank specfctes (ncludng ther dfferent defntons). 5. Conclusons 85. Ths study ams to shed some lght upon the dfferent roles of selected rsk varables (LTVO, ILTV, DTSO, LTIO and CRMO) wthn modellng practces across European banks. The conclusons reflect the analytcal objectve of ths study and should feed the current debate about assessng and enhancng comparablty for resdental mortgages n Europe. Therefore the fndngs wll be part of the consderatons for any future polcy ntatves. Summary of fndngs from qualtatve nformaton analyss 86. The answers provded by the banks regardng the use of the varables n the models confrmed the mportance of (ndexed) loan-to-value and credt rsk mtgatons. Debt to servce rato and loan-to-ncome are used less frequently, except n PD models. 87. Overall, the documentaton provded by banks, although succnct, hghlghted the banks use of dfferent defntons for smlar concepts. Sometmes they reflect country-specfc features, but overall the defntons are bank specfc. 88. Although t s not possble to assess the materalty of such dfferences n nfluencng the varaton observed n the RWs n the EU sample, the dfference n defntons s an mportant caveat to consder when readng the fndngs of the study. 89. The use of nternal defntons and the dversty of such defntons are necessary to reflect the banks own experence, modellng choces and credt polces, as well as the country specfctes. Nevertheless, ths requres each competent authorty to make an effort to assess the materalty of such varous defntons and ther mpact. The European study provdes only an ntal overvew. Summary of fndngs from quanttatve nvestgaton 90. From the quanttatve analyss, the exstence of a postve correlaton between the value of the dfferent drll-down varables (LTVO, ILTV, DTSO and LTIO) and the RWs at EU sample level was confrmed. 33 The average drll-down varables by bankng group are calculated usng the medan of the bucket, but for the frst bucket the upper bound s used, and 160% for the EAD above 150% for LTVO and ILTV, 80% for the EAD above the 70% for DTSO and 9 for the EAD above the 8 bucket for the LTIO. Non-avalable data are not taken nto account. 24

91. Despte general postve evdence, the RW senstvty assgned by banks to the dfferent drlldown buckets across the varables was not always found to be a clear explanatory factor of the RW n the EU sample. Ths s the case even f rsk-weght senstvty across the buckets s generally observed. Country dfferences and non-trval bank specfctes complcate the study. 92. The EAD dstrbuton across the bucket values of drll-down varables has lttle mpact on the RW dsparty across the European sample. 93. The ILTV s the varable that more sgnfcantly nfluences RW varaton. Overall, the RW senstvty for ILTV accounts for about one-fourth of the total varablty at EU sample level; ts contrbuton vares sgnfcantly among the dfferent country bank observatons. 94. Some country specfctes have been dentfed. Credt rsk mtgants (other than mortgages) are very mportant drvers to be consdered when assessng the varaton n some countres, and seems to explan the lower RW senstvty to the value of the fnanced real estate. 95. The use/non-use of the varables n the models does not nfluence the RW senstvty as much as was ntally expected. One possble explanaton s that even f not used as drect nput n models, one can assume that the flterng of credts based on those varables when grantng loans to customers ndrectly play a sgnfcant role. 96. The analyss by vntage confrms the exstence of a close lnk between the level of LTVO and RWs, and the potental nfluence of the portfolo composton by vntage n explanng the varaton n RW. 97. Further, when studyng the level of the average drll-down varables at the bankng-group level, the use of a specfc combnaton dd not account for the dfferences n RW. 98. Investgatng the drll-down varables dd contrbute much to the study of experenced losses when explanng the dfference n RW. However, t dd complement the study, sheddng addtonal lght on the sources of varaton, and creatng the bass for both better knowledge and comparson of the banks at country level. 99. The fndngs show the relevance of usng a country-by-country approach when makng sample analyses, but also when analysng ndvdual results; each bank/observaton should also be benchmarked by country level when deeper understandng s requred. 100. Followng ths, fnal conclusons at bank level can only be drawn by the natonal competent authortes, takng the European benchmarks but also the specfc market structure for resdental mortgages nto account. 101. Ths s consstent wth the approach suggested n the Consultaton Paper regardng supervsory benchmarkng exercses, whch empower the competent authorty to perform the assessment of the nternal approaches and assess the dfferent rsk profles wthn the bank s portfolos. 25

Annexes Annex 1: Country-weghted averages by country Fgure 16: Mnmum, maxmum and EAD-weghted average for the drll-down varables, by country RW LTVO Mn Mean Max ILTV Mn Mean Max DTSO Mn Mean Max LTIO Mn Mean Max CRMO Mn Mean Max AT BE 10% 73% 80% 86% 62% 68% 73% 31% 37% 45% 3.9 4.1 5.4 24% 40% 58% CZ 26% 71% 76% 86% 70% 73% 76% 29% 35% 38% 4.3 4.5 5.6 0% 2% 7% DE 16% 73% 79% 85% 0% 22% 38% DK 12% 70% 71% 82% 72% 75% 90% 3.2 3.2 5.7 0% 0% 14% ES 17% 66% 73% 75% 59% 65% 70% 27% 32% 45% 4.7 5 7 0% 1% 16% FI 10% FR 16% 77% 85% 91% 71% 72% 84% 21% 30% 36% 3 3.3 4.8 4% 62% 75% IE 45% 73% 75% 78% 103% 109% 117% 16% 25% 36% 3.5 4 5.9 0% 0% 0% IT 15% 63% 67% 68% 59% 61% 65% 29% 37% 45% 4.4 5 6.2 0% 0% 7% LU 16% 54% 75% 87% 51% 68% 76% 5% 87% 100% NL 10% 86.6% 87.1% 89% 83% 87% 94% 23% 26% 29% 4.6 4.8 6 17% 41% 67% NO 9% PL 18% PT 22% 76% 77% 80% 66% 70% 77% 17% 31% 62% 4.3 4.9 6 0% 4% 55% SE 5% 69% 72% 75% 59% 68% 72% 0% 1% 4% SK 30% UK 11% 65% 71% 79% 64% 70% 89% 16% 23% 30% 3.1 3.6 4.8 0% 0% 0% Total 15% 54% 76% 95% 51% 74% 117% 13% 27% 62% 2.4 3.9 7.2 0% 18% 100% Note: Country-weghted averages are based on bucket medans, but the upper bound s used for the lowest bucket. For exposures above the latest bucket, 160% s taken for LTVO and ILTV, 80% for DTSO and 9 for LTIO. Exposures reported as non-avalable are excluded. Only countres wth at least four observatons are represented. CRMO statstcs are calculated on the percentages of exposures wth credt rsk mtgant (other than mortgages) over the total amount of resdental mortgages. Source: EBA data collecton (reference date: December 2012), EBA calculaton 26

Annex 2: Top-down detals on the methodology For the top-down approach, the same methodology was followed as n the frst nterm report (Annex I), http://www.eba.europa.eu/documents/10180/15947/interm-results-eba-revewconsstency-rwas_1.pdf. Appled to the current dataset, t means that the followng calculatons were used to analyse the devaton n terms of RW: RW = (RW bank RW benchmark ) = prce effect + bucket mx effect = k =1 Share_EAD benchmark RW bank RW benchmark + k =1 RW bank Share_EAD bank Share_EAD benchmark, wth beng the dfferent buckets for the drll-down varable. k The prce effect s =1 Share_EAD benchmark RW bank RW benchmark. The prce effect s then broken down between the senstvty effect and the level effect: k =1, The senstvty effect beng Share_EAD benchmark RW bank RW sens_adj wth RW sens_adj 1 = RW bank RW benchmark 1 + 1 k =1 RW benchmark The level effect beng Share_EAD benchmark RW sens_adj RW benchmark The bucket mx effect s k =1 [RW bank Share_EAD bank Share_EAD benchmark ] 27

Annex 3: Top-down analyss for the LTVO, DTSO and LTIO varables Fgure 17: LTVO Break-down of the prce and bucket mx effects Note: The banks are sorted by ther RW devaton. A bank may be represented several tmes f t has submtted data for more than one country. Source: EBA data collecton (reference date: December 2012), EBA calculaton Fgure 18: DTSO Break-down of the prce and bucket mx effects Note: The banks are sorted by ther RW devaton. A bank may be represented several tmes f t has submtted data for more than one country. Source: EBA data collecton (reference date: December 2012), EBA calculaton 28

Fgure 19: LTIO Break-down of the prce and bucket mx effects Note: The banks are sorted by ther RW devaton. A bank may be represented several tmes f t has submtted data for more than one country. Source: EBA data collecton (reference date: December 2012), EBA calculaton 29

Annex 4: Average RW by bucket Fgure 20: Average rskweght by LTVO rato Fgure 21: Average RW by ILTV rato Fgure 22: Average RW by DTSO Fgure 23: Average RW by LTIO 30

Annex 5: Applcaton of the top-down approach at country level Fgure 24: ILTV Break-down of the prce (level and senstvty) and bucket mx effects Note: The banks are sorted by ther RW devaton. Bank 1 has the hghest prce effect (both level and senstvty are materal). The banks apply hgh RW for the dfferent ILTV, but the RW senstvty to ILTV s the lowest one among the banks n the cluster when compared to the country benchmark. Bank 2 has the most sgnfcant bucket mx effect. Ths s due to the concentraton of ther exposures n the lowest ILTV buckets when compared to the benchmark. The RW appled n the dfferent ILTV buckets are, n general, above and more senstve to the benchmark. Bank 3 and Bank 4 appear very smlar. Those banks are closer to the benchmarks for RW at portfolo level. Nevertheless, they also show some dfferences n both the RW levels and the senstvty appled to each ILTV bucket relatve to the benchmark. Bank 5 has, on average, RWs slghtly lower than the benchmark, but t s also the bank most smlar to the country benchmark for both portfolo composton and RW appled to the dfferent ILTV buckets. 31