Response to consultation on Guidelines for the estimation of LGD appropriate for an economic downturn
Go back
Also, as mentioned above, in the case of global crises that have affected at least one economic factor in (almost) all jurisdictions, it should then be possible to determine the influence of this crisis on the LGD globally. This can make sense in particular for low default portfolios in order to be able to observe a downturn impact in the first place, and not to distort it by statistical fluctuations due to the required distribution of observations.
The following example illustrates the problem:
Especially with low default portfolios, determining the downturn impact at the level of the calibration segment can be difficult because there are not enough observations. For example, an institution has over 200 defaults and related observed LGDs over an observation period of 10 years. The institution has established 4 calibration segments for the LGDs with 50 cases each. This results in the following long-run average LGD, for example for calibration segments:
1: 10%
2: 15%
3: 20%
4: 25%
At the level of the overall portfolio, the average LGD is 17.5%.
In order to determine the development of the LGDs observed over time, the LGD per default must be determined. If this development is analysed at the level of the calibration segments, there are on average only 5 cases per year for which an average value is determined. These averages are statisti-cally very uncertain due to the very low number of cases. Consequently, the downturn impact at the level of the calibration segment can be very distorted.
In this case, we believe that it would make more sense to monitor the development of the realised LGDs over time at the level of the overall portfolio. This would mean an average of 20 cases per year and the statistical uncertainty would be considerably lower. If a realised LGD of 28% is now observed at the overall level for the downturn phase, this then corresponds to a relative increase of 60% com-pared with the long-run average LGD of 17.5% at the overall level. This 60% increase at the overall level must also be reflected on average at the level of the calibration segment. Consequently, the downturn LGD can be derived for each calibration segment by multiplying the corresponding long-run average LGD by 1.6 (x=0.6) per calibration segment. This results in the following downturn LGD per calibration segment:
1: 16%
2: 24%
3: 32%
4: 40%
Observing the downturn impact at the overall level and transferring it to the calibration segments should therefore be a permitted option in low-default portfolios.
• For any economic factor that has its most pronounced value in this multi-year downturn phase, the worst outcome, even if the peaks and valleys do not occur in the same year? or
• Average value over the multi-year downturn phase per economic factor? or
• Calculate each year individually with the correspondingly observed values?
Similarly, in the case of multi-year downturn phases, there is no clear definition of which values are to be applied for the economic factors that do not have their most pronounced value in this multi-year downturn period:
• worst during this downturn period or
• average value or
• calculate each year separately?
In our opinion, however, the lack of a precise definition should be avoided. The institutions should develop portfolio-specific meaningful requirements that would have to be approved by the auditors.
Article 30 requires a strictly positive MoC in category A if a haircut or extrapolation approach is applied for downturn LGD estimation. As a general rule, this MoC should be able to be equal to 0 with corre-sponding evidence, for example in the case of the haircut, that the model reacts sufficiently sensi-tively to the economic factors. Such evidence could be provided, for instance, by calculating the LGD with the haircut method comparatively for another downturn phase in which relevant loss data is available and the LGD downturn was determined in accordance with section 5. If it can be demon-strated that the LGD determined for this downturn phase using the haircut method is greater than or equal to the LGD based on observed impacts in line with section 5, then no MoC should be required for earlier downturn phases when the haircut method is applied.
The general assumption of the approach is that a dependency between observed realised LGDs and macroeconomic variables exist. This is probably true given a long observation period with different macroeconomic circumstances, but may not apply for any observation period. Nevertheless, in par-ticular given a retail portfolio with a large number of observations, a statistical model may derive a dependency between the observed realised LGDs and a macroeconomic factor although the underly-ing observation period covers a stable economic environment. In fact, the derived dependency is just a pseudo-correlation and the LGDs obtained by backward extrapolation are unsuited for calculation the downturn add-on.
Example:
Assume a macroeconomic factor ‘a’ varying between 98 and 102 during the observation period. Given the framework described above the result might be a simple linear model: LGD = 1.5 – 0.01*a. If the macroeconomic factor drop off by ~20 percent to 80 during a downturn period the LGD would in-crease by ~40 percent which is in particular not plausible for a retail portfolio.
Moreover we like to annotate that it may not even be possible to calculate a downturn add on by the proposed extrapolation approach if the dependency between the observed realised LGD and the mac-roeconomic factor does not lead to higher LGDs during the downturn period, which is in fact possible even given a stable portfolio.
Finally, paragraph 51 points out that a MoC has to be considered if the extrapolation approach is used to cover the resulting uncertainty. However, the approach may already lead to very conservative es-timates.
According to EBA/CP/2018/07 and EBA/CP/2018/08, an institution-specific default and loss history of 20 years is required in order to have a chance of avoiding the application of an MoC. We consider this to be a legally unacceptable escalation and hence a breach of the CRR requirements in Article 181(1)(j) and Paragraph 2 sentence 2 of the CRR (minimum data history). It is our understanding that, as a matter of principle, it must be possible for an institution that is new to IRBA approval and that meets the CRR requirements – and in particular the minimum data history requirements – to obtain approval without applying an MoC. In the absence of historical institution-specific loss data for an eco-nomic downturn period, a strictly positive margin of conservatism (MoC) must be applied to the inter-nal downturn LGD estimation for this downturn period in accordance with paragraph 32 of EBA/CP/2018/08. Furthermore, in accordance with paragraph 37(a) of “EBA/GL/2017/16 – Guidelines on PD estimation, LGD estimation and the treatment of defaulted exposures”, an MoC must be ap-plied if the default definitions are not consistent in the historical observation period. In accordance with Article 181(1)(j) of the CRR, for exposures to corporates, institutions and central governments and central banks, estimates of LGD must be based on data over a minimum of five years, increasing by one year each year after implementation until a minimum of seven years is reached, for at least one data source. If the available observation period spans a longer period for any source, and the data is relevant, this longer period must be used. However, according to Article 179(1) of the CRR, data collected prior to 1 January 2007 may only be used if appropriate adjustments are possible and have been made to achieve a broad equivalence with the definition of default in Article 178 or the definition of loss. Both CRR articles thus clash with a requirement of 20 years’ default and loss history, so that an MoC is inevitably being forced on the institutions because there were economic downturns in the last 20 years, including before 2007 (in Germany, in particular the 2000/01 dotcom bubble). Either the institution has to accept an MoC because it has no default and loss data (that are repre-sentative of the current definition of default) for the crisis before 2007, or it has to accept an MoC because it uses internal data before 2007 based on a different definition of default. This runs counter to the EBA’s own requirements in paragraph 50 of “EBA/GL/2017/16 – Guidelines on PD estimation, LGD estimation and the treatment of defaulted exposures” that institutions should develop a plan to rectify data and methodological deficiencies as well as any other potential sources of additional uncer-tainty, and to reduce the estimation errors within a reasonable timeframe.
According to paragraph 32 of EBA/CP/2018/08 and the preliminary remarks in section 3.6, following a previously very complex and time-consuming procedure for determining the downturn LGDs, they should be compared with a reference value calculated as the mean of the 2 years with the highest observed losses, and deviations from this value should be justified. If harmonisation of the model results is to be achieved mainly by using a reference value, we believe that the complex require-ments for deriving a downturn LGD for each downturn phase and calibration segment are unneces-sarily detailed and onerous for the institutions.
Question 1: Do you think that additional guidance around the estimation of LGD in-default, which reflect downturn conditions, is needed? If yes, could you provide examples of sound methodologies for transposing downturn LGD estimates from performing to non-performing exposures?
Depending on data availability, it should generally be left open as to how the LGD in default is derived under downturn conditions for the individual downturn phases. For example, if a relative increase of 20% over the long-term average was determined for the LGD for the living portfolio under downturn conditions, it can be expected that for the LGD in default, which analyses partial cash flows of the LGD for the living portfolio, there will also be a mean relative increase of 20% due to downturn effects. Taking into account the complexity of determining the LGD for the living portfolio under downturn conditions, such reconciliations to the LGD in default should also be permitted, and no determination of the downturn effect on the LGD in default per reference time should be required.Question 2: Do you share the concern that the proposed policy in paragraph 15 could create an undue burden if applied to every downturn period identified? If yes, in order to better balance the accuracy of the estimations and its operational complexity what evidence should be provided by institutions in order to justify the exemption of identified downturn periods from the proposed policy in paragraph 15?
From our point of view, it seems unnecessarily complex to consider all downturn phases for all cali-bration segments and all jurisdictions. This is especially the case when the informative value for downturn phases long in the past is already limited due to the availability of relevant loss data, and the most recent crisis exhibits a more pronounced impact overall on the economic factors than pre-vious crises.Also, as mentioned above, in the case of global crises that have affected at least one economic factor in (almost) all jurisdictions, it should then be possible to determine the influence of this crisis on the LGD globally. This can make sense in particular for low default portfolios in order to be able to observe a downturn impact in the first place, and not to distort it by statistical fluctuations due to the required distribution of observations.
Question 3: Do you agree with the proposed level of downturn LGD estimation set out in paragraph 14? In particular, do you support the concept that the downturn LGD estimates of different calibration segments could be based on different downturn periods? Is the policy on the level of downturn LGD estimation as well as the relation between the level of downturn LGD estimation and the relevant downturn periods sufficiently clear?
Depending on how the calibration segments are formed, it may also make sense to start from com-mon downturn phases and, especially for low default portfolios, to measure the downturn effect across all calibration segments and to compare the long-run average LGD also measured at the over-all level. For example, if a relative increase of the LGD by x% under downturn conditions compared with the long-run average LGD can be observed at the overall level, the long-run average LGD deter-mined for each calibration segment could be multiplied by the factor (1+ x%) in order to obtain a downturn LGD per calibration segment.The following example illustrates the problem:
Especially with low default portfolios, determining the downturn impact at the level of the calibration segment can be difficult because there are not enough observations. For example, an institution has over 200 defaults and related observed LGDs over an observation period of 10 years. The institution has established 4 calibration segments for the LGDs with 50 cases each. This results in the following long-run average LGD, for example for calibration segments:
1: 10%
2: 15%
3: 20%
4: 25%
At the level of the overall portfolio, the average LGD is 17.5%.
In order to determine the development of the LGDs observed over time, the LGD per default must be determined. If this development is analysed at the level of the calibration segments, there are on average only 5 cases per year for which an average value is determined. These averages are statisti-cally very uncertain due to the very low number of cases. Consequently, the downturn impact at the level of the calibration segment can be very distorted.
In this case, we believe that it would make more sense to monitor the development of the realised LGDs over time at the level of the overall portfolio. This would mean an average of 20 cases per year and the statistical uncertainty would be considerably lower. If a realised LGD of 28% is now observed at the overall level for the downturn phase, this then corresponds to a relative increase of 60% com-pared with the long-run average LGD of 17.5% at the overall level. This 60% increase at the overall level must also be reflected on average at the level of the calibration segment. Consequently, the downturn LGD can be derived for each calibration segment by multiplying the corresponding long-run average LGD by 1.6 (x=0.6) per calibration segment. This results in the following downturn LGD per calibration segment:
1: 16%
2: 24%
3: 32%
4: 40%
Observing the downturn impact at the overall level and transferring it to the calibration segments should therefore be a permitted option in low-default portfolios.
Question 4: Do you consider the description of the approaches to be sufficiently clear?
In the event that a longer downturn phase than 12 months has been identified in a downturn due to time lags in the economic factors, there is no clear definition of which corresponding values should then be used for the economic factors in the haircut or extrapolation approach:• For any economic factor that has its most pronounced value in this multi-year downturn phase, the worst outcome, even if the peaks and valleys do not occur in the same year? or
• Average value over the multi-year downturn phase per economic factor? or
• Calculate each year individually with the correspondingly observed values?
Similarly, in the case of multi-year downturn phases, there is no clear definition of which values are to be applied for the economic factors that do not have their most pronounced value in this multi-year downturn period:
• worst during this downturn period or
• average value or
• calculate each year separately?
In our opinion, however, the lack of a precise definition should be avoided. The institutions should develop portfolio-specific meaningful requirements that would have to be approved by the auditors.
Article 30 requires a strictly positive MoC in category A if a haircut or extrapolation approach is applied for downturn LGD estimation. As a general rule, this MoC should be able to be equal to 0 with corre-sponding evidence, for example in the case of the haircut, that the model reacts sufficiently sensi-tively to the economic factors. Such evidence could be provided, for instance, by calculating the LGD with the haircut method comparatively for another downturn phase in which relevant loss data is available and the LGD downturn was determined in accordance with section 5. If it can be demon-strated that the LGD determined for this downturn phase using the haircut method is greater than or equal to the LGD based on observed impacts in line with section 5, then no MoC should be required for earlier downturn phases when the haircut method is applied.
Question 5: Do you agree to the limitation of approaches for quantification of downturn LGD estimates? If not, which other approaches should be considered? Would you prefer the alternative policy considered – if yes how should a minimum MoC be established in this case?
The general idea to extrapolate an downturn add-on from a macroeconomic time series which covers a longer time period than the observed loss data is comprehensive, but from our point of view the results of the extrapolation approach described in paragraph 49 et seq. may be inappropriate and unreliable.The general assumption of the approach is that a dependency between observed realised LGDs and macroeconomic variables exist. This is probably true given a long observation period with different macroeconomic circumstances, but may not apply for any observation period. Nevertheless, in par-ticular given a retail portfolio with a large number of observations, a statistical model may derive a dependency between the observed realised LGDs and a macroeconomic factor although the underly-ing observation period covers a stable economic environment. In fact, the derived dependency is just a pseudo-correlation and the LGDs obtained by backward extrapolation are unsuited for calculation the downturn add-on.
Example:
Assume a macroeconomic factor ‘a’ varying between 98 and 102 during the observation period. Given the framework described above the result might be a simple linear model: LGD = 1.5 – 0.01*a. If the macroeconomic factor drop off by ~20 percent to 80 during a downturn period the LGD would in-crease by ~40 percent which is in particular not plausible for a retail portfolio.
Moreover we like to annotate that it may not even be possible to calculate a downturn add on by the proposed extrapolation approach if the dependency between the observed realised LGD and the mac-roeconomic factor does not lead to higher LGDs during the downturn period, which is in fact possible even given a stable portfolio.
Finally, paragraph 51 points out that a MoC has to be considered if the extrapolation approach is used to cover the resulting uncertainty. However, the approach may already lead to very conservative es-timates.
Question 6: Do you expect that the total exposure amount or share which is treated with the policy proposed in Section 7 is material?
We presume that section 7 will be of greater significance at least for some institutions.Question 7: Do you have specific examples of types of exposures which will fall under the policy proposed in Section 7?
In principle, this could particularly affect certain jurisdictions such as China or other emerging mar-kets, but also developed markets where no real crisis has been observed in the last 20 years and where prolonging the data history does not seem reasonable for reasons of representativeness. It should therefore be possible to take over downturn impacts observed globally or for similar jurisdic-tions in these cases.Question 8: Do you agree to require a minimum MoC quantified via a fixed add-on to the long-run average LGD? If not, which of the alternatives should be considered? Do you see reasons for differentiating the fixed add-on according to exposure classes?
The GBIC is critical of the fact that the MoC in the situation described in section 7 is excessively con-servative (31(b), p. 42, 20% floor). A fixed add-on of 20% is too high. It should also be considered that this floor must be regarded as cumulative with the input floors decided under Basel IV. It should be noted that lending values as inputs are already geared to the long term. In addition to real estate this also applies to other categories of collateral. Of course the 20% should also cover reduced re-covery rates and reduced contributions in a recession, but we still think 20% is too high. We think the hierarchical approach is positive and flexible.Question 9: Do you agree to the minimum MoC as the max(0,min(20%, 105% - LRAVLGD)?
The GBIC is critical of the fact that the MoC in the situation described in section 7 is excessively con-servative (31(b), p. 42, 20% floor). A fixed add-on of 20% is too high. It should also be considered that this floor must be regarded as cumulative with the input floors decided under Basel IV. It should be noted that lending values as inputs are already geared to the long term. In addition to real estate this also applies to other categories of collateral. Of course the 20% should also cover reduced re-covery rates and reduced contributions in a recession, but we still think 20% is too high. We think the hierarchical approach is positive and flexible.Question 10: Is the policy regarding the reference value sufficiently clear? Alongside with the potentially limited applicability of the reference value to the downturn LGD estimation according to paragraphs 18-19, for what reasons could the reference value feasibly be omitted? Do you agree to the proposed clarification of the role of the reference value?
Paragraph 32(a)(ii) is not clearly defined because the highest realised ratio to the EAD does not nec-essarily have to be accompanied by the highest economic loss in the numerator in the case of great-er EAD portfolio fluctuations.According to EBA/CP/2018/07 and EBA/CP/2018/08, an institution-specific default and loss history of 20 years is required in order to have a chance of avoiding the application of an MoC. We consider this to be a legally unacceptable escalation and hence a breach of the CRR requirements in Article 181(1)(j) and Paragraph 2 sentence 2 of the CRR (minimum data history). It is our understanding that, as a matter of principle, it must be possible for an institution that is new to IRBA approval and that meets the CRR requirements – and in particular the minimum data history requirements – to obtain approval without applying an MoC. In the absence of historical institution-specific loss data for an eco-nomic downturn period, a strictly positive margin of conservatism (MoC) must be applied to the inter-nal downturn LGD estimation for this downturn period in accordance with paragraph 32 of EBA/CP/2018/08. Furthermore, in accordance with paragraph 37(a) of “EBA/GL/2017/16 – Guidelines on PD estimation, LGD estimation and the treatment of defaulted exposures”, an MoC must be ap-plied if the default definitions are not consistent in the historical observation period. In accordance with Article 181(1)(j) of the CRR, for exposures to corporates, institutions and central governments and central banks, estimates of LGD must be based on data over a minimum of five years, increasing by one year each year after implementation until a minimum of seven years is reached, for at least one data source. If the available observation period spans a longer period for any source, and the data is relevant, this longer period must be used. However, according to Article 179(1) of the CRR, data collected prior to 1 January 2007 may only be used if appropriate adjustments are possible and have been made to achieve a broad equivalence with the definition of default in Article 178 or the definition of loss. Both CRR articles thus clash with a requirement of 20 years’ default and loss history, so that an MoC is inevitably being forced on the institutions because there were economic downturns in the last 20 years, including before 2007 (in Germany, in particular the 2000/01 dotcom bubble). Either the institution has to accept an MoC because it has no default and loss data (that are repre-sentative of the current definition of default) for the crisis before 2007, or it has to accept an MoC because it uses internal data before 2007 based on a different definition of default. This runs counter to the EBA’s own requirements in paragraph 50 of “EBA/GL/2017/16 – Guidelines on PD estimation, LGD estimation and the treatment of defaulted exposures” that institutions should develop a plan to rectify data and methodological deficiencies as well as any other potential sources of additional uncer-tainty, and to reduce the estimation errors within a reasonable timeframe.
According to paragraph 32 of EBA/CP/2018/08 and the preliminary remarks in section 3.6, following a previously very complex and time-consuming procedure for determining the downturn LGDs, they should be compared with a reference value calculated as the mean of the 2 years with the highest observed losses, and deviations from this value should be justified. If harmonisation of the model results is to be achieved mainly by using a reference value, we believe that the complex require-ments for deriving a downturn LGD for each downturn phase and calibration segment are unneces-sarily detailed and onerous for the institutions.