2017年-EBA欧洲银行管理局_EBA_Report_on_IRB_modelling_practices_157页_3mb

EBA Report on IRB Modelling Practices Summary

Core Content Overview

This report provides an analysis of internal ratings-based (IRB) modelling practices used by European banks for estimating risk parameters, specifically Probability of Default (PD), Loss Given Default (LGD), and Expected Loss Best Estimate (ELBE), as well as the impact of the EBA's Guidelines (GLs) on these parameters. The findings are based on a survey conducted in January 2017, involving 102 institutions from 22 EU Member States, representing 64% of total EU credit risk-weighted exposures.

Main Points

1. Survey Scope and Participation

• 102 institutions participated in the IRB survey.
• The survey covers both high-default and low-default portfolios, as per the scope of the GLs.
• The COREP exposure class 'retail — secured by immovable property non-SME' is the most represented, with ~50% of PD and LGD models.
• Other exposure classes, such as central governments and central banks, institutions, and specialised lending, are less represented (7%, 11%, and 3% respectively for PD and LGD models).

Key Findings and Model Practices

2.1 PD Models

• PD Estimation:
  • Calibration is a key part of the process, and the GLs include a list of allowed calibration types under the CRR.
  • Calibration methods vary, with some institutions calibrating at the grade or pool level and others at the portfolio level.
  • The one-year default rate (DR) is used in PD estimation, with 45% of PD models already calculating it quarterly.
  • The final GLs require institutions to evaluate one-year DRs at least quarterly, which affects ~54% of PD models.
  • The long-run average DR is calculated as the average of observed one-year DRs, but adjustments are needed if the historical observation period is not representative.

2.2 LGD Models

• LGD Estimation:
  • The realised LGD is calculated based on the economic loss and recovery.
  • Discounting of additional recovery cash flows is a key issue, with only 4% of LGD models using the CP approach (no discounting), while 32% use a method that includes discounting.
  • The final GLs adopted the discounting approach after reviewing the pros and cons.
  • Unpaid late fees and capitalised interest are included in the economic loss in most models (52% and 44%), but not in the denominator in 20% and 26% of models.
  • The easiest approach was selected: unpaid late fees and capitalised interest after default are not included in the economic loss or the amount outstanding at default.
  • Additional drawings are included in the denominator of the realised LGD if they are included in the credit conversion factor (CCF) estimation.
  • The GLs retain the CP's approach for additional drawings, aligning them with CCF treatment.

Impact of the Guidelines

3.1 Model Changes

• The GLs do not quantify the impact on capital requirements, as this would require more resources and is subjective without clear supervisory guidance.
• Instead, the GLs focus on model changes necessary to align with the guidelines, including:
  • Calibration requirements.
  • Discounting of additional recovery cash flows.
  • Treatment of unpaid late fees and capitalised interest.
  • Inclusion of additional drawings in LGD estimates.
  • Calculation frequency of one-year DRs.

3.2 Calibration and Conservatism

• The margin of conservatism (MoC) is a key component in both PD and LGD estimation.
• The GLs clarify the use of MoC, specifying that it should be applied if the historical observation period lacks sufficient bad years or is over-represented with them.
• A benchmark is proposed for the long-run average DR: the maximum of the average of one-year DRs over the most recent five years and the whole observation period.
• Institutions may still estimate below this benchmark, but only if justified and an additional MoC is applied.

Data Quality and Methodology

4.1 Data Requirements

• Data quality is a critical factor in IRB model development.
• The survey results highlight heterogeneity in practices, especially regarding:
  • The definition of default.
  • The calculation of economic loss.
  • The discounting rate used for LGD estimates.
  • The inclusion of collateral and guarantees in LGD estimation.

4.2 Methodological Variations

• The discounting rate is specified at different granularities, depending on the exposure class.
• The treatment of incomplete recovery processes and cases with no loss or positive outcome is varied, with some models including them in the average LGD and others not.
• The GLs provide clarity on these methods and require institutions to assess the impact of their chosen calibration methods on the cyclical nature of capital requirements.

Conclusion and Implications

• The GLs reflect the most common approaches observed in the survey, ensuring consistency with current practices.
• The impact on capital requirements remains uncertain due to the complexity and variability of internal models.
• Monitoring the implementation of the GLs is essential to understand their final effect on capital.
• The survey results were used to justify policy choices, particularly in the frequency of DR calculation and the discounting of recovery cash flows.

Summary of Model Changes

Final Notes

• The survey results are subject to data quality issues due to the nature of the responses.
• Some questions were misunderstood, which may affect the interpretation of the findings.
• The final GLs are qualitative in nature, aiming to capture current practices and minimise institutional burden while supporting consistent implementation.

Key Figures and Tables Mentioned

• Figure 1: Number of banks by country.
• Figure 2–3: Share of PD and LGD models by country.
• Figure 4: Exposure values covered by PD and LGD models.
• Figure 5–60: Details on DR calculation, MoC, calibration, and LGD methodologies.
• Table 1–71: Summarises model practices, calibration types, and policy choices across exposure classes.