We acknowledge the proposed model component approach gives a clearer regulatory view about what “estimates that are appropriate for an economic downturn” should be understood for. The overall approach balances flexibility of application and prescriptivism in a way that is found appropriate, even though still exposed to unavoidable subjectivity.
With regard to its feasibility, we highlight that the consistent application of the model component approach is particularly demanding in terms of data requirements. For this reason its application is expected to be exposed to higher model risks for low default portfolios in particular. Instead of over-relying to (still subjective) expert-based choices, this may suggest making simplified approaches appropriate for those portfolios.
With regard to workability, we believe the reference to model components may enhance workability as it allows partition the overall problem and therefore specific data availability and representativeness issues might be better addressed.
Another significant source of unintended variability of results can be connected to the level of granularity, as, at the same time, from one hand some aggregation is deemed appropriate for having more robust results and from the other hand disaggregating clusters differently exposed to downturn economic conditions is required. We will be commenting further on this with reference to model component identification.
From a general point of view, a link to CF and LGD models is found appropriate and further disaggregation should be rather linked to outcomes variability over time instead and not necessarily to homogeneous pools delivering different levels of long-run average LGDs of CFs.
The general principles of the model components approach are applicable for both LGD and CF estimates.
However we also believe there are significant differences in what might be expected to be the impact of an economic downturn on the two parameters. In particular, we expect that CFs are likely to be less exposed to downturn impact as CFs depend also on the impact of tighter credit policies stemming from higher PD estimates within best practice credit monitoring policies.
For this reason we expect requirements for downturn CFs estimates to be to a certain extend unduly burdensome but yet applicable.
The concept of Model component is sufficiently clear and the reference to multi-stage modelling techniques as those simplified in the explanatory box inform of their appropriateness.
From the practical point of view, we suggest a list of expected model components to be disclosed as part of the GLs on risk parameters estimate amendment for better promoting harmonisation. This is particularly appropriate for CFs as, from one side, they do not find extensive treatment in the CP and, from the other side, it is less obvious to identify specific model components in spite of their bimodal distribution.
From the operational point of view, it would be found appropriate that only most relevant identified model components were individually estimated, as a better balance between estimation accuracy and implementation costs.
With reference to the way model components should be selected, we would deem appropriate that components most correlated with economic and credit cycle should equally be considered even though they do not necessarily explain a multimodal distribution shape, or even when a multimodal distribution cannot be observed at all.
We think that, from a portfolio perspective, rather than referring to components explaining the multimodal distribution, a more general reference that “drive the variability of observed LGDs and CFs across the economic cycle” would be both less demanding and more appropriate.
The identification of dependency proposed is deemed to be appropriate. Some feasibility concerns depend on the inherently lifetime definition of LGDs and on the duration of recoveries in some jurisdictions, in spite of the factorization in model components. We deem appropriate that further factorization might be required to fully identify dependencies, for instance by analyzing recovery profiles instead of the overall realization of LGD liquidation, and highlight that to fully evaluate the impact of most recent downturn on realized model components also information stemming from open workouts are relevant.
For this reason we suggest to take into account of such situations in the final draft of provisions under Point 2(f) of Article 3 especially as subpoint 2(f)(iii) is related (denying the opportunity to take into account the observable portion of incomplete model components)
We agree with the proposed approach and indeed we suggest that further factorization might be allowed and deemed appropriate as model components duration makes difficult to relate its realization to a specific point of the economic factors time series.
From an economic factor level, the provision is deemed appropriate as our joint interpretation of duration and severity is that downturn should be calibrated over the worst (severity) yearly (one-year duration) observation of economic factors relevant to each model component. What is however most relevant in the mechanics of downturn calibration is how the length of the workout process is taken into account as it is unreasonable both to assume than only one year of the overall process is affected by economic downturn and to assume that such conditions should affect the overall recovery process.
This is widely unaddressed in the CP even though the overall mechanics made clearer in the explanatory boxes imply that the worst realization of economic factors describing different model components are to be taken jointly into account in spite of the fact that they do not take place at the same time. This is generally found overprudential from one side and not as clear within the CP text from the other. This concept should be made therefore clearer if confirmed, otherwise different interpretations of the provisions may trigger significant unjustified variability of RWAs.
As far as the severity of downturn is considered, we deem appropriate the reference to the plausible variability of those factors for the future. At this regard, a regulatory benchmark is found most appropriate to promote greater homogeneity leveraging, for instance, on supervisory stress scenarios.
We would privilege an approach where an explicit model-based link between model components and economic factors is estimated, so that downturn correction turn to be the resulting of the impact on LGDs and CFs of stressed economic factors. Under this perspective, we suggest recourse to regulatory benchmarks similarly to the way stress testing exercises are performed. Reference to worst historical observations are not necessary under this perspective. Should this reference kept in the RTS, we suggest that reference to MoC as of Article 5(2)(a)(ii) to be limited to those cases where a shortening of time series imply excluding worst observations from the time series.
This is not expected to be the case in most jurisdictions as the most severe downturn from decades is generally acknowledged to have occurred in most recent years than the last 20 years. In such cases shorter time series are equally appropriate.
Benchmarking values of most relevant economic factors, consistent with stress testing exercises, as suggested in Question 7 feedback, may be beneficial for this evaluation. The reference to sufficient severe condition is subject to great subjectivity. This strengthening either to be included in the RTS or, better, in the GLs amendment.
We think RTS should leave an adequate level of flexibility. However, guidance shall be strengthen in terms of principles as explanatory boxes is clearer on supervisory expectations.
We have however relevant concerned on the expected dynamics.
First, assigning the realisation of a model component to the period where the majority of realisation occurred is a simplification as it might be possible to further factorise the problem (f.i. considering marginal recoveries) in order to achieve the same objective without neglecting the information contained in incomplete workouts.
Second, it should be acknowledged that the fact that downturn effects model components with different time lags t is a physiological characteristics of the recovery process. Instead of combining the worst model components, it is found more appropriate that either the worst yearly combination of model components or the combination of average model components over downturn was considered.
Third, the expected impact of downturn differ significantly depending on whether downturn occurs at the beginning of the recovery process or at later stages. Under this perspective not a unique calibration of LGD and LGD in-default is advisable and GLs should make clear that a dedicated calibration by reference date is expected for downturn add-on within LGD in-default.
Fourth, it is unclear how the definition of one-year downturn duration is consistent with the principles outlined in the explanatory box as over such a short horizon either downturnal cure rates or downturnal recoveries are experienced but normally not both jointly. The fact that economic factors and model components are to be observed with yearly frequency is not connected with the concept of downturn duration, which is normally a multi-year condition.
We deem that leaving such points fully at discretion of institutions and supervisors is not consistent with the objective of limiting unjustified variability or RWAs.
We deem appropriate that principles underlying the downturn approach should be incorporated in the RTS through a more detailed specification of Article 6 and that similarly should be done as far as Amended GLs are concerned.
However, we have some concerns both on the approach (deterministic approach vs. estimating relevant macroeconomic relations, as commented in the detailed feedback and also exposed briefly responding to Question 11) and on the underlying principles.
• the information included in incomplete workouts are neglected while they are deemed to be valuable;
• model components might be effected with different time lags, but notwithstanding this it is deemed overconservative to identify downturn as worst combination of downturnal model components, the latter defined as the worst yearly occurrence over the downturn period. Either considering the worst yearly combination or averaging model components outcomes over the downturn scenario timespan would find a better balance among conservatism and management of time lags;
• there is no reference to dedicated calibration for LGD and LGD in-default and by time-on-book for the latter; this is deemed critical as downturn effect with significant difference the same model component over time;
• it is implied that downturn effects the entire recovery process and different model components at the same time; it is straightforward that LGD is a “lifetime” parameter and workout processes may last for a series of years and thus it is inherently subject only for a part of it to downturn impacts; in supporting our customers we have over time developed methodologies to take into account that downturn effects for an expected duration (much higher than 1-year) the recovery expectation, either at the beginning of the process, at the end or in the middle but not the entire process.
The explanatory box implies a rather deterministic calibration on model components historical observations. From this point of view a series of model components are expected to be not available with 20 years historical depth. However, it is proposed for explanatory purposes and does not exclude using inferential techniques. We commented in the detailed feedback that we would assume that a macroeconomic overlay linking ELbe, stress testing, IFRS9 and downturn adjustment is found as most appropriate way to address the downturn calibration issue. In this context, as long as proper historical series are available to estimate such relations, the observation of model components for more limited time depth is not a critical issue and it would rather be considered appropriate to define the potential impact of hypothetical stress scenarios.
Calibration around agreed supervisory scenario, like in the stress testing exercise, would indeed promote greater homogeneity.
We deem appropriate the approach for the identification of the final downturn scenario should be consistent among LGD and CF estimates. However, as commented earlier a simplified approach for downturn CFs calibration would be found appropriate.
Namely it is found reasonable that realized overall CFs would normally be found as appropriate unique model component.
As we commented with reference to RTS and especially to the intended principles underlying explanatory box to Article 6 (see Question 10 and paragraph 2.4 of the feedback document), they are found not clearly established in the regulatory text. We commented about the fact that fully convincing and over-conservative earlies there are not considered fully convincing and over-conservative earlies, but should nevertheless explicitly stated in the GLs if confirmed as appropriate. We have also earlier commented on the fact that considering complete workouts limit the use of valuable information stemming from partial recovery observations, especially being downturn relatively recent and potentially still influencing (due to time lags) recoveries in some jurisdictions.
As we commented in the introductory part of the detailed feedback, reasons orelated tor materiality and modellability (low default portfolios) suggest regulation should make available simpler approaches as supervisory calibrated add-ons.
We deem appropriate limiting the model component approach to LGDs as CFs dynamics is less complex and mostly related to credit processes rather then to the economic cycle. As credit processes tend to be more accurate and conservative in downturn conditions, it is expected from a portfolio point of view that barely adjustment for downturnal CFs are required. At a minimum, directly evaluating the relation of the realized CFs and economic factors at an appropriate granular level should be considered sufficient.
Other simplification are required for immaterial portfolios (based on proportionality) and for low default portfolios (based on expected robustness of outcomes). In such conditions, supervisory guidance is deemed appropriate.
We do not see a significant simplification in the proposed reference value approach as it rather allows reducing prescriptiveness but still requiring dedicated analysis in a “comply of explain” approach that would necessarily be calibrated around conservative supervisory expectations. When set at EU or jurisdiction level, it would act similarly to benchmark transfer models in the stress testing exercise promoting compliance to such a value but not reducing the complexity of the issue to address compared to the model component approach. It is indeed preferred that rather thenthan acting as a challenging value or an hard floor, such a regulatory benchmark was provided as a “supervisory add-on” allowed as shortcut within supervisory review and validation based on materiality and modellability reasoning. Our opinion is that such approach may actually be a simplification if it acts as an hard shortcut making any inherently complex analysis unrequired since agreed to be unduly burdensome from a portfolio perspective or even unfeasible from the modelling point of view.
Such an approach would be different from the identified options of the “supervisory add on approach” as, like in the stress testing, reference values are expected to be fully economically grounded (on pooled data, benchmarks, structural models, etc.) and calibrated –although conservatively – on expected impacts of downturn on LGDs and CFs.
This is not the case for both the distributional approach (based on observed recovery volatility without reference to downturn conditions) and the downturn discounting approach (based on a regulatory add-on on the spread).
The economic grounding of such approaches in weak and would neither allow an appropriate risk-based calibration nor achieve effective comparability of RWAs.
They would mostly achieve only simplicity of estimation and validation in spite of the outcomes. A regulatory calibration of standard add-ons, similarly to what provided in the reference value approach, would certainly promote greater comparability and not necessarily lower risk-sensitivity if defined at appropriate level of granularity.
We supported a series of customers, mostly in Italy, in identifying downturn impact on LGDs based either on the estimate of satellite transfer model of economic factors (mostly macroeconomic) and on the empirical calibration of average LGDs to observed downturn LGDs averages. The granularity of the problem was set at a “model component-like” level, even though more often the focus was on LGD liquidation as danger rates, along with PDs, tended to be already calibrated on most recent “downturnal” years.
In the adopted deterministic approach, differently from the model component approach depicted in the text boxes of the CP, we considered rather the average model component observation over downturnal scenarios (multi-year) compared to long run averages. Due to the length of recovery processes, this was done through a overall LGD decomposition into marginal (yearly) recoveries so that as long run average recovery rate can be expressed as the sum of long run average marginal recoveries, similarly downturnal recovery rates was based on compounding the impact of downturnal on such marginal recoveries. One point often discussed with our customers was whether all impacts had to be summed up (lifetime downturn) or they should have been limited to the expected downturn duration. Usually customers priviledged the later with duration threshold set between 5 and 7 years as the most recent “double-downturn” turned to be appropriate. The length was strictly related to the time horizon identified as downturnal. In some cases the analysis was properly extended to the impact on survival rates as not all the unrecovered amount should be considered lost as long as workout time increase due to delay. Depending on the asset class the impact of this was negligible - unsecured positions whose recoveries concentrate on initial recovery stages, even limited to liquidation and therefore not as result of cure rates dynamics – or particularly significant – especially on positions secured by collaterals.
Such an approach is specialized for defaulted assets, as for LGD it might be considered that the first 5-7 years were effected while for positions already in default it was rather assumed that “the following” 5-7 years were effected. In some circumstances our customers preferred a simplified solution where a unique add-on (basically a multiplier reflecting the differences of adjusted LGDs from long-run LGDs) was defined, but taken into consideration the portfolio perspective where downturn effects differently positions with different time-on-book and therefore it was grounded on averaging “5-7 year long” impacts set at different time-on-book based on the average downturnal time-on-book distribution of the portfolio of application.
Setting aside details, the major finding of such analysis was that as recovery is a multi year process and in spite of isolating cure rates effects then overall LGD liquidation either observed as referred to liquidation entry, end or considered referred to a calendar time equal entry+observed recovery duration, show weaeker correlations to doiwnturnal scenarios than it is possible to observe on marginal recoveries by time-on-book . This also for the possibility to include in the latter only incomplete workouts. This is critical as NPL-ratios well known supervisory attention is linked to the fact that too many defaults triggered by recent downturn are still under workout.