ECB’s Revised Guide to Internal Models: changes in the Credit Risk chapter

On July 28th, the European Central Bank (ECB) published its revised guide to internal models ECB publishes revised guide to internal models. On top of changes necessary for alignments with CRR3, the ECB improves their guidelines based on supervisory experience with the aim of harmonized and transparent internal modelling practices at credit institutions. 

In this article, we share our perspective on the changes in the credit risk chapter, focusing on the impact on PD, LGD and CCF modelling: 

1- PD: Institutions must use at least five years of default data and demonstrate a meaningful correlation between default rates and macroeconomic indicators for LRA DR calibration. 

2- LGD: The ECB will benchmark calibration windows against the 2008–2018 period. Downturn LGD calibration must cover all relevant components and include yearly elevated LGDs, even if they don’t perfectly match downturn periods. The LGD reference value is now an active challenge in model validation, requiring consistent calculation and action if weaknesses appear. 

3- CCF: For CCF modeling, risk drivers must use data exactly 12 months before default, 0% CCFs for non-retail exposures are no longer allowed, and negative CCFs must be floored at zero, while high CCFs remain uncapped.

The following chapters elaborate on these three proposed amendments in more detail.

1 PD

1.1 Minimum requirements historical data 

The ECB now specifies data requirements for the representativeness analysis referenced in paragraphs 82 and 83 of the EBA Guidelines on PD and LGD estimation (EBA/GL/2017/16), which assess whether the PD calibration dataset reflects a balanced mix of "good" and "bad" years. 

According to paragraph 236(A) of the revised EGIM, institutions must use a minimum of 5 years of historical data as of the calibration date. Additionally, they must have enough one-year default rates to calculate a statistically meaningful correlation between default rates and relevant (macro)economic indicators. 

If an institution lacks the required data (either the 5 years or enough data for meaningful correlation), paragraph 236(B) states the period is automatically deemed not representative, and the institution must apply adjustments and quantify a Margin of Conservatism (MoC). However, paragraph 236(C) allows that if an institution has enough data but cannot find significant correlations with macroeconomic indicators, the historical period may still be considered representative if it includes both the minimum and maximum of the institution’s internal one-year default rates. 

1.2 LRA DR reference value 

The ECB introduces a reference calibration level for LRA DR in paragraph 237, based on the period January 2008 to December 2018, which is considered to be representative of a full economic cycle. Institutions must calculate this reference LRA DR and use it as an anchor value. If an institution proposes a lower LRA DR based on a different time period, it must justify why that period is more representative. Although not a strict floor, the reference LRA DR serves as an anchor value for ECB assessment. Even without sufficient default data, institutions must estimate the reference LRA DR using the accounting definition of default (DoD) as an approximation for the prudential definition. 

2 LGD

For the loss given default (LGD) risk parameter, the ECB now sets out clearer expectations on two areas: the estimation of downturn LGD based on observed impact and the calculation and use of the LGD reference value.  

2.1 Calibration of downturn LGD based on observed impact 

The revised EGIM retains flexibility in how downturn LGD is calibrated but introduces stricter requirements for the analyses underpinning the calibration based on observed impact. Zanders expects that most institutions already conduct these analyses and may only need to refine their application. 

The ECB emphasizes that all analyses¹ required under paragraph 27(a) of the EBA guidelines must be conducted. If calibration is done at the component level (e.g., secured vs. unsecured), these analyses should be done separately for each component, with results aggregated into a total downturn LGD. At minimum, the most impactful component must be included; others may be needed if it doesn't capture downturn effects sufficiently. 

Regardless of calibration method, elevated realized LGDs (e.g. defined as the average of defaults in a given year) must be used.. Although downturns are often defined more granularly, the ECB states that yearly averages should still be used, even if they don’t align exactly with downturn periods. 

2.2 Reference value 

The LGD reference value, unlike the new LRA DR, is an existing non-binding benchmark from the EBA guidelines meant to challenge an institution’s downturn LGD estimates. While still non-binding, the ECB has raised expectations for how it should be calculated and used. Zanders supports this, seeing it as a useful tool for diagnosing weaknesses in LGD calibration. 

The reference value should follow paragraph 37 of the EBA guidelines², typically calculated as the average LGD in the two years with the highest economic loss. The ECB emphasizes that incomplete recovery processes must be included in identifying these years and in the LGD calculations, and that this should be done at least at the calibration-segment level. 

For comparing the reference value with actual downturn LGD estimates (per paragraph 19), the ECB offers guidance when the reference value is higher. If this isn’t due to a missed downturn period, institutions must reassess their calibration methodology for possible flaws. If a missed downturn may be the reason, they are expected to re-evaluate their downturn identification and consider timing lags between downturn events and losses. 

3. CCF 

The main changes to the Credit Conversion Factor (CCF) guidelines focus on calibration and aim to reduce variability across institutions' modelling practices. The ECB is aligning with the EBA’s goal of lowering RWA variability: EBA’s Revised Definition of Default - Zanders.  

Key updates include: 

• AIRB Approach: Institutions can now only use risk driver data from exactly 12 months before default (the reference date), eliminating the option to consider longer-term behavioral patterns.  

• FIRB Approach: The ECB has removed the option for institutions to justify using a 0% CCF for non-retail exposures through an annual materiality analysis. As a result, institutions not using their own CCF models must now apply the standardised (SA) CCFs under Article 168(8a). 

• Negative CCFs: In line with CRR3 (effective January 2025), negative observed CCFs must be floored at zero. However, very high CCFs (over 100%) are not capped, and the ECB provides no specific guidance on handling such outliers. Zanders recommends isolating these into separate grades or pools and reviewing the data in detail to correct any structural issues. 

Conclusion 

This post highlights the most relevant changes in the ECB guide to internal models for IRB credit risk modelling. Institutions must use at least five years of default data and demonstrate a meaningful correlation between default rates and macroeconomic indicators for LRA DR calibration. The ECB will benchmark calibration windows against the 2008–2018 period. Downturn LGD calibration must cover all relevant components and include yearly elevated LGDs, even if they don’t perfectly match downturn periods. The LGD reference value is now an active challenge in model validation, requiring consistent calculation and action if weaknesses appear. For CCF modeling, risk drivers must use data exactly 12 months before default, 0% CCFs for non-retail exposures are no longer allowed, and negative CCFs must be floored at zero, while high CCFs remain uncapped. 

Reach out to our experts John de Kroon and Dick de Heus, if you are interested in getting a better understanding of what the proposed amendments mean for your credit risk portfolio.  

Zanders actively monitors regulatory updates relevant for (credit) risk modelling. Keep a close eye on our LinkedIn and website for more information or subscribe to our newsletters here.