UniCredit Group supports EBA’s efforts to harmonize IRB framework and appreciates the opportunity to comment on “Draft RTS on the specification of the nature, severity and duration of an economic downturn”. We understand the multiplicity of the exercise to reduce undue RWA variability however would like to stress the importance of choices especially with regard to downturn determination since those choices can have a significant impact on RWA.
As a general comment, we are concerned about the complexity of the model components approach and the fact it may lead to excessively conservative quantification of downturn. With that, being driven by many expert based choices, the approach could likely introduce a huge range of subjectivity in the framework. Therefore, we would welcome further amendments into suggested approach to make sure that benefits of LGD modelling are not put aside and RWA variability is not unintentionally increased. In addition, reduction of complexity is also an enabler for proper model maintenance (e.g. back testing, recalibration and regulatory approval). In our view, alternatives basing downturn adjustment on behavior of realized losses during downturn periods are the most simple and transparent ones.
In principle, we agree that requirements of Article 1 are reasonable and economically sound. Notwithstanding, we have the following concerns around the current proposal.
In the first place, the approach likely brings a huge range of subjectivity as being rather complex and relying on a lot of expert based choices (e.g. identification of economic factors and economic scenarios, addressing of MoC etc.). Downturn quantification would be even more subjective for low default portfolios or whenever historical data is not available (and consequently model components have to be estimated). In the end banks can arrive to vastly different outcomes and we have a doubt whether suggested approach will be able to reduce RWA variability, or will potentially increase it.
Secondly, estimation of economic downturn separately for exposures in different jurisdictions but treated under the same model might be hampered due to limited number of observations per each jurisdiction. Therefore, we propose to revise the requirements of Article 1 (2) b and to allow considering exposures in different jurisdictions jointly or under simplified approaches even if it cannot be proven empirically that economic factors are characterized by strong co-movements.
Thirdly, as being rather multiplex the approach requires a lot of work from institutions in terms of identification of downturn conditions for each model component etc. In order to optimize development efforts, we deem necessary to get clarified that estimation of downturn can be performed not only at level of homogenous pools used for LGD quantification (e.g. exposure buckets for SME customers) but on more aggregated level of homogeneously managed exposures (e.g. SME customers all together without separation for exposure buckets) unless the first drives multimodality of the loss distribution. Besides this, we suggest to specify whether more flexible approaches can be adopted if mixed exposures are treated under the same model. For example, unsecured LGD might be estimated on customer level in order to respect the fact that workout process is set up on customer level as well. Thereby if customer has more than one transaction in different exposure classes it can be tricky to split LGD accordingly and ensure separate quantification of downturn for each type of exposure.
We believe that the same principles of the model components approach are applicable for both LGD and CF estimates.
The difference between two parameters may appear when looking at order of downturn and default occurrence. In general, downturn affects LGD estimates starting from the moment of default or even later and the effect lasts through the workout process. On the contrary, CF estimates are typically affected by downturn before default occurs as customers get into trouble with deterioration of economic conditions. In case there is a monitoring process taking into account PD estimates (that in principle should be in place in line with the use test requirements) the following situation can happen: as the result of economic conditions worsening monitoring outcome triggers a tightening of credit line availability; consequently realized CFs get lower and analysis of dependency between model components and economic factors can lead to ambiguous outcomes. As such, if the models component approach is chosen, it would be helpful to provide implementation details separately for LGD and CF estimates.
Explanatory box brings much more clarity into description of the model components concept, however we presume necessary to include more explanations into Article 2 itself as reading of the current version may lead to confusion.
First and foremost, having too many model components defined involves high implementation and maintenance costs because economic downturn has to be defined for each model component. No doubt that performing of analysis directly on realized LGD level may come at not being able to capture the dependency between the economic factors and realized loss. However, we deem correct to determine model components as only major drivers of multimodality of loss distribution to ensure a proper balance between accuracy of estimates and implementation/ maintenance costs. In this context we challenge the requirements of Article 2 (2) (b) and 2 (2) (c):
1. Text of Article 2 (2) (b) interpretation is not straightforward. Let’s considerer an example, where components identified to calculate secured part of LGD estimates are market value of the collateral forecasted to the moment of liquidation, rates of collateral realization proceeds and costs set over market value of the collateral, time to liquidation of the collateral and discount rate; some of above components may relate to features of realized loss distribution. From one hand, these components have to be understood as model components for downturn quantification because the RTS asks, as a minimum, to use the same components already identified in the course of producing own-LGD estimates. On the other hand, the whole secured part can be understood as a model component as one driving modality of loss distribution (in a sense that collateralized exposures usually produce lower losses). In our view, a rephrasing of Article 2 (2) (b) is needed with a focus that only components as major drivers of loss distribution shape have to be considered to avoid misinterpretation.
2. With regard to Article 2 (2) (c) we propose to modify the requirements and allow using realized LGD or realized drawings as only model components respectively not just when portfolios do not show multimodal distribution anymore but whenever major drivers of loss distribution shape have been already identified. Moreover, there can be situations when multimodal distribution can’t be easily eliminated/ separated into further components anymore. For example, multimodality of loss distribution for liquidated cases can be obvious based on graphical analysis but it is not always trivial to explain the modality by underlying reasons. Since liquidation scenario is a major driver of modality of loss distribution itself we believe there is no need to look for further components for liquidated cases in such a situation.
At any rate, we have to stress, that even if only major drivers of multimodality are considered, estimation of downturn could require an important effort whenever identified model components differ from those already presented in the model itself.
In addition, it is not always clear what the difference between risk factors and model components is. So, concerning LGD in-default, time in default and workout scenario (e.g. whether liquidation phase has been initiated) can serve as risk drivers and can be the model component at the same time. In this regard explanatory box makes a contradictory statement that model components are not risk factors. Thus we expect a little bit more clarity on this issue to avoid wrong reading. Next, a further guidance how to combine different model components is needed, for example how to combine workout scenarios with time in default if both prove to be model components.
In case of CF estimates, it will be beneficial to give some examples of model components to be considered.
In general the idea behind dependency approach is clear. Again, it would be meaningful to detail the article as some important hints are only presented in explanatory box. Apparently, the dependency analysis tends to be rather complex as it has to be performed for each economic factor/ indicator and each model component.
List of economic factors
We suggest that whenever certain economic factor is deemed not relevant for exposure type performing of dependency analysis is not needed. Therefore, EBA list of economic factors should be considered as an indicative one to avoid undue burdens.
Panel of experts
Regarding a panel of experts it is not obvious, who the experts are and what their role is in assessing the link between model components and economic factors.
Even if it is true that pure quantitative analysis might be too mechanic and restricted (e.g. due to low number of observations and diminishing of representativeness), estimation of expected correlation by panel of experts also brings a lot of subjectivity. Moreover, nature of downturn defined solely according to the expert’s judgment seems to be arbitrary and will be likely questioned by Regulator. In practice, it could be difficult to find internal experts independent from the modelling unit and at the same time having enough knowledge of economics and risk management to define which the right level of correlation is. Under this approach we deem downturn can’t be modelled robustly and the burden to prove the outcome quality would be placed solely on the Bank.
Paradigm of higher losses/ worse model components during economic downturn
When performing dependency analysis, it has to be kept in mind that paradigm of higher losses during downturn is not necessarily proper and some components can show an inverse relation with economic indicators. Inverse relation can appear naturally and not as a result of changes in processes.
From operational perspective, further specifications related to the following topics will be welcomed:
1. As pursuant to Article 3 (2) (f), no less than yearly frequency of data for economic factors has to be used. As further clarified in Article 4, one year duration of economic downturn shall be applied for each economic factor and it seems that annual observation windows have to be chosen for measuring of economic factors. We think that it needs to be further specified how dependency analysis has to be performed if more than yearly frequency is prescribed. Indeed, the outcome of dependency analysis is heavily driven by a choice of reference dates and observation windows for economic factors calculation. For example, year on year GDP growth can be measured with quarterly reference dates as of 31.03, 30.06, 30.09, 31.12 or just with yearly reference date as of 31.12 (or any other). In the first case, observation windows are overlapped and whenever economic decline in quarters Q1 and Q2 changes with economic growth in quarters Q3 and Q4, outcomes of dependency analysis could be different relying on whether yearly or quarterly frequency of economic factors is chosen.
2. Dependency analysis has to be performed for each economic factor separately in line with Article 3 (3) and investigation of joint impacts of several economic factors seems to be neglected. Apparently, performing solely of univariate analysis would be beneficial in case of short time series. However, if the worst realization of economic factor happens beyond the date when realizations of model components are available, multivariate analysis can bring certain advantages into estimation of model components value that has to be performed according to Article 5. Later we argue that duration of economic downturn is not necessarily limited to one year and hence consideration of joint impacts of economic factors makes sense especially in situations when peaks/ troughs of those factors do not take place simultaneously but are nonetheless the effects of one single overall downturn.
In general, we agree with the approach to compute the time series of the realized model components.
Nevertheless, such a crucial aspect as a moment when the model component is realized has to be detailed. As prescribed in the explanatory box, time dimension for each model component is defined as a moment when the majority of the realizations of the model component are observed. However it is not straightforward to understand, whether such a time dimension refers to a period where the highest recoveries are realized or where most of the recoveries have been already realized. Furthermore, with regard to component time-in-default it is not straightforward to understand why suggested time dimension (the year of completion of workout process) should have a higher dependence with economic factors as compared to other alternative definitions of the time dimensions.
Next, a further specification is needed with respect to LGD-in default. Specifically, it is not clear how to account for downturn conditions when realizations of the model component has been realized prior reference date but workout process has not been completed yet.
Finally, the consideration of incomplete workouts only in case realization of the model component under analysis has already been observed reduces subjectivity when measuring dependency with economic factors. But disregarding of incomplete workouts for such components as a cure rate may set incorrect relation between cured and non-cured cases therefore introducing biases when defining overall nature, severity and duration of an economic downturn.
We support the choice to use a fixed horizon of one year for purpose of dependency analysis described in Article 3 and consider this solution as a pragmatic one in terms of simplicity and level of standardization. However we strongly disagree with a proposal to use only one year duration of economic downturn and especially in situation of persisting economic decline.
First of all, it is not economically sound to assume economic worsening lasting exactly for one year.
Secondly, a restriction of the period to one year may lead to overly conservative quantification of downturn. In point of fact, when assessing overall nature, severity and duration of economic downturn, it is proposed to use individual downturn period for every single model component. While doing so, downturn estimate produced is not fully realistic because worst realizations of economic factors describing different model components do not necessarily take place simultaneously but with time lags even being linked to the same economic scenario. In this way the approach could hide natural offsetting effects that may exist among model components and overly conservative downturn quantification can appear.
Consequently, we argue that consideration of longer periods may lead to more accurate and realistic quantification. In the end, when looking at model components jointly, the same period has to be considered for quantification of model components if those are affected by downturn. We propose to set one year duration as a minimum backstop and to ensure alignment with other regulatory products such as ECB Guide for the Targeted Review of Internal Models (according to Article 67 (c) of TRIM Guide “the specified downturn period should be a minimum of one year, although longer periods are acceptable in order to account for cases where the historical data show longer stress periods for some indicators, or where the peaks or troughs of different economic indicators are not reached simultaneously but are nonetheless the effect of one single overall downturn. In such cases, the downturn period should be long enough to reflect the continued stressed situation”).
The proposed approach for the identification of the severity of an economic downturn is clear. However, we express the following concerns.
First, most of the banks do not have a complete twenty years history of realized model components. Ultimately, even more than 20 year of historical data is needed as discounting of cash flows is performed to the moment of default that precedes the moment of model component realization. Accordingly, model components may have to be projected for early years. An accuracy of the projection might be low due to limited time series available and especially when choice of economic factors is heavily driven by expert decision (e.g. no dependency with model components).
Besides that, in case of longer periods it is not trivial to separate effect of changes in processes and effect of downturn itself (projection could represent a mix of both effects). Addressing of margin of conservatism when there is no data concerning the realized model components is also something problematic and it is not clear how such a margin has to be estimated. After all, downturn quantification could result to be rather subjective due to data availability issue. Next, it is not always crystal clear what has to be understood behind the worst period for certain economic factors. For example, with regard to interest and inflation rates both negative and positive (if high) values can indicate unfavorable economic conditions and it is not obvious what periods have to be picked up.
Further concern relates to a choice of reference dates used for economic factors quantification. We believe that reference dates have to be chosen coherently with the ones utilized for dependency analysis. However, whenever quarterly (or even monthly) frequency of economic factors is used, further clarifications are needed. For example, year to year GDP growth can be measured with quarterly reference dates as of 31.03, 30.06, 30.09, 31.12 or just with yearly reference date as of 31.12. Whenever economic decline in quarters Q1 and Q2 changes with economic growth in quarters Q3 and Q4, very different downturn periods can be identified depending whether quarterly of yearly frequency of reference dates is chosen.
Yes, we think that more details should be included into Article specifying a definition of “sufficiently” severe economic conditions. First of all, the idea of having a recurring cycle that always looks identical is very abstract and simplistic. Along this, an assumption that each downturn will be the same and there can be only slight variations (“plausible variability”) of economic factors in future is never fully realistic (every crisis is different and exhibit different levels of downturn severity). Secondly, measuring of representativeness with future realizations of economic factors is something puzzling and only expert judgments can be done for nearest future. Indeed, a clear distinction with stress test scenarios has to be made when looking forward. In the end, there are no incentives to consider economic conditions as not sufficiently severe since otherwise only complexity and subjectivity of downturn estimation increase (e.g. estimation of model components beyond dates where actual realizations are available etc.). We suggest that if data starts from 2007-2008 crisis, this should be considered as sufficient.
Note: we assume that the question refers to Article 5 (3) and not to Article 2 (3).
Yes, we suppose that Article 6 should pin down the steps for the joint impact analysis for proper understanding of Article 6 itself as well as the whole concept of downturn quantification.
The approach described in an explanatory box might be too penalizing since the worst realizations of economic factors do not necessarily take place simultaneously but with a time lag even being driven by one economic scenario. Moreover, the approach could hide natural mitigating effects that may exist among model components and as a consequence overall downturn quantification could be disconnected from reality. As already reported in the reply to question 6, this issue can be mitigated if the same downturn period is considered for all model components affected.
It is fuzzy how worst realizations of economic factors have to be attributed to economic scenarios and we expect a little bit more clarifications with this regard. Even through sometimes attribution can be done in a straightforward manner, in other cases it is up to discretion of experts and subjectivity could be introduced into downturn determination. On the other hand, there could be an incentive to have as many economic scenarios as possible to avoid overly conservative downturn quantifications caused by combining of worst realizations of economic factors. We think major economic scenarios could be defined centrally by Regulator on a single country level with the aim to reduce unjustifiable variability of capital requirements.
Besides the above mentioned, explanatory box does not give any clue how issues of unrepresentativeness (e.g. changes of portfolio composition, changes in process etc.) have to be taken into account when defining the final downturn scenario. In our opinion, identification of final downturn scenario based on the highest LGD is not always correct as higher losses can occur due to, e.g., diminishing of representativeness.
First of all, we would welcome EBA to provide additional guidance with respect to estimation of model component. As mentioned in the reply to question 4, we deem meaningful a consideration of joint impacts of economic factors to estimate model components. Therefore, analysis of dependency under Article 3 should not be limited to solely univariate analysis per each economic factor. Additional comments are presented in the reply to question 7 where we report that accuracy of the model components projection might be low and addressing of conservatism margin is something non trivial. In case time series of realized model components are not sufficiently long, estimation of model components might be rather complex and could bring subjectivity into downturn estimates.
Although we agree that the target approach for the identification of the final downturn scenario should be the same for LGD and CF estimates, we deem necessary to give examples separately for LGD and CF estimates.
We believe that downturn adjustment methodology has to be described in more details in RTS on downturn itself and not in GLs on PD estimation, LGD estimation and treatment of defaulted exposures.
Further, we think it is essential to specify what should be a level of downturn application, i.e. whether it is overall model itself or every single component. When reading the text of explanatory box it seems that downturn has to be applied on overall model level. According to BCBS consultation document “Reducing variation in credit risk-weighted assets – constraints on the use of internal model approaches, March 2016” application of downturn is supposed to be done on model component level, e.g. it is stated that collateral haircuts should reflect downturn conditions. Application logic has to be appropriately reflected while developing the model (e.g. allocation of secured and unsecured exposures) and consequently has a direct impact on components estimates themselves.
In addition, there are the following concerns with regard to margin of conservatism:
1) proposal on Article 160 (a): we disagree to quantify MoC in case there is no dependency between model components and economic factors; indeed, the requirement goes into contradiction with a concept of margin of conservatism introduced in EBA CP on PD estimation, LGD estimation and the treatment of defaulted exposures; absence of dependency can’t be attributed to any of deficiency categories A, B, C and D mentioned in EBA CP.
2) proposal on Article 160 (b): whenever economic conditions are not sufficiently severe, institutions have to look further back into historical data; consequently, MoC related to backward estimation of model components might be needed (in line with EBA/CP/2016/21, section 4.4) and the application of an additional MoC in light of the absence of sufficiently severe economic conditions in the 20 years timespan would be excessive, possibly leading to a double counting of the effect.
3) proposal on Article 160 (c): contrary to explanatory box, we think that MoC is not always needed if there are no realized model components and they have to be estimated for downturn adjustment purposes; for example, if it is proven that satellite model (the one used for estimation of model components) overestimates realizations of model component, application of additional margin of conservatism seems to be unnecessary.
Yes, we think that alternative and less sophisticated approaches for downturn adjustment should be considered in following situations:
1) low default and low data portfolios (e.g. banks/ financial institutions and sovereign exposures where expert judgment plays a prominent role in LGD modelling);
2) simulation based models (e.g. for specialized lending portfolios).
We welcome additional guidance on how to calculate this adjustment, and how it should be implemented. Proportionality measured in terms of size and scope of the institution represents in our opinion no trigger for differentiated approaches for downturn adjustment unless special situations mentioned above.
We agree with points presented in Table 1 of draft RTS and here report only additional comments.
Reference value approach
Generally, we do not consider reference value approach (at least in the way it is described in the consultation paper) as a valid alternative to model components approach. Despite of being more simplistic, reference value approach can turn out to be over conservative and not unconditionally being able to reflect a link with downturn conditions accurately.
First of all, we think that disregarding of any model components is not appropriate and it is quite natural to analyze major components already presented in the model separately. It is true that performing of analysis on aggregated level reduces development efforts but it also goes with a cost of losing sensitivity to economic factors.
Secondly, since there is no prescription for dependency analysis, quantification of downturn may lead to non-causal relations with economic factors and therefore can bring subjectivity into downturn quantification. Moreover, we highlight that a study of relation with economic/credit factors is required at least for ELbe estimation and, as it is further clarified in EBA/CP/2016/21 (page 88), “the analysis of the relevant economic and credit factors and their dependence with loss rates should follow the general guidance that will be provided by EBA in the context of the RTS specifying the nature, severity and duration of an economic downturn”. Thus we encourage EBA to align requirements on dependency analysis because under the current proposal counterintuitive outcome can be faced while comparing LGD in-default and ELbe (e.g. when dependency for downturn analysis is skipped).
Thirdly, with regard to reference value we do not consider a proposal to use two years with the highest realized LGDs as sufficiently grounded and explained. At this step elements having no relationship with economic downturn might be introduced and it is not always trivial to prove the opposite. More by token, reference value defined at level of EU or jurisdiction is considered even less appropriate because it demolishes risk-sensitivity of internal models. Certainly, realized loss is strongly driven by portfolio structure and internal workout process and thus it is pretty unlikely to set up appropriate value suitable for all possible types of exposure.
Fourthly, we stress a need to ensure consistency between SSM and EBA views as currently the requirements of EBA TRIM Guide are not identical to EBA proposal.
The issue of recovery rates volatility is partially addressed when determining a discount rate for calculation of economic loss. For this reason we assume that suggested approach deals with volatility not yet captured by discounting effect. Anyway, the approach can be also appropriate for quantification of margin of conservatism and consequently a clear breakdown with downturn estimation is needed. Since the link to economic conditions is not intuitive the approach can be relevant in situations where no clear downturn period is identified. However, we stress that there can be cases when volatility of recoveries in a long-term perspective is lower than in a short-term and as a result downturn effect might be diminished.
Supervisory add-on approach. Downturn discounting rate with fixed add-on
First of all, the validity of the approach depends on a decision EBA will take with regard to discounting effect to be used for calculation of economic loss. Secondly, we would welcome a bit more clarifications on what should be understood behind downturn discounting rate (i.e. base rate + spread 10%) and how it should be interpreted; currently the approach does not appear methodologically intuitive. Thirdly, we have a doubt that the same discounting rate can be applied disregarding jurisdiction or portfolio type. Although we consider the approach based on downturn discounting rate as a potential option for low data portfolios, we stress that spread of 10% seems to be inappropriate and in certain jurisdictions it will have a dramatic impact on RWA (the effect will be even worse for Banks with longer recovery processes).
Downturn periods (nature of downturn) are defined on expert based rules linked to realization of economic factors, usually focusing on consecutive periods with negative GDP growth rate. Dependency analysis between economic factors and model outputs/ components is generally performed. Duration of economic downturn is not limited to one year and severity of downturn is not attributed to the worst realizations of economic factors.
Depending on the structure of the LGD model, the level of downturn estimation can be on the overall LGD or on specific model components (e.g. unsecured LGD, LGD_liquidation). In any case downturn adjustment is calculated comparing average of realized LGDs for defaults opened/ closed during the downturn periods with long run LGD estimates.
In case no relation between macroeconomic factors and LGD is found, an expert value is applied.