(AGENPARL) – gio 05 dicembre 2024 Mercati, infrastrutture, sistemi di pagamento
(Markets, Infrastructures, Payment Systems)
Rating the Raters. A Central Bank Perspective
Number
December 2024
by Francesco Columba, Federica Orsini, and Stefano Tranquillo
Mercati, infrastrutture, sistemi di pagamento
(Markets, Infrastructures, Payment Systems)
Rating the Raters. A Central Bank Perspective
by Francesco Columba, Federica Orsini, and Stefano Tranquillo
Number 55 – December 2024
The papers published in the ‘Markets, Infrastructures, Payment Systems’ series provide
information and analysis on aspects regarding the institutional duties of the Bank of
Italy in relation to the monitoring of financial markets and payment systems and the
development and management of the corresponding infrastructures in order to foster
a better understanding of these issues and stimulate discussion among institutions,
economic actors and citizens.
The views expressed in the papers are those of the authors and do not necessarily reflect
those of the Bank of Italy.
The series is available online at http://www.bancaditalia.it.
Printed copies can be requested from the Paolo Baffi Library:
Editorial Board: Stefano Siviero, Massimo Doria, Giuseppe Zingrillo, Guerino Ardizzi,
Paolo Libri, Paolo Bramini, Luca Filidi, Tiziana Pietraforte,
Antonio Sparacino.
Secretariat: Yi Teresa Wu.
ISSN 2724-6418 (online)
ISSN 2724-640X (print)
Banca d’Italia
Via Nazionale, 91 – 00184 Rome – Italy
Designed and printing by the Printing and Publishing Division of the Bank of Italy
Rating the Raters. A Central Bank Perspective
by Francesco Columba*, Federica Orsini*, and Stefano Tranquillo*
Abstract
We use the Bank of Italy’s credit assessment system for non-financial corporations as a benchmark
to assess the ratings assigned by commercial banks through their own internal systems, which
are also used for monetary policy purposes. We examine the distribution of ratings on bank
loans pledged as collateral in monetary policy operations in Italy and test for underreporting
of risk, which might generate unwarranted exposure for the central bank. The rating systems
of commercial banks and of the central bank both show satisfactory discriminatory power
and predictive ability, suggesting that they evaluate credit risk adequately. We find that banks’
models are, on average, slightly less conservative than the central bank model for borrowers with
loans eligible as collateral. We observe only some mild evidence of low economic significance
that banks may strategically manage the credit risk assessment for borrowers whose loans are
pledged. We find no evidence that banks using more central bank liquidity are more lenient in
assigning default probabilities to their debtors.
JEL Classification: D82, G21, G24, G32, E52.
Keywords: Model-based ratings, Credit risk, Collateral, Central bank refinancing.
Sintesi
Utilizziamo il sistema di valutazione del merito creditizio delle imprese non finanziarie della
Banca d’Italia per valutare i rating assegnati dalle banche commerciali italiane con i loro sistemi
interni, usati anche a fini di politica monetaria. Analizzando la distribuzione dei rating delle
due fonti per i prestiti stanziati come garanzia nelle operazioni di politica monetaria in Italia,
verifichiamo l’esistenza di una eventuale sottostima del rischio da parte delle banche, che
potrebbe esporre la banca centrale a rischi indesiderati. Entrambi i sistemi mostrano un potere
discriminante e una capacità predittiva soddisfacenti, indicando una valutazione adeguata del
rischio di credito. I modelli interni delle banche tendono a risultare leggermente meno prudenti
rispetto al modello della Banca d’Italia per i prestiti idonei come garanzia. Per i prestiti stanziati,
rileviamo evidenze deboli e di limitata rilevanza economica di una sottostima strategica del
rischio. Non troviamo evidenze che le banche con maggiore utilizzo della liquidità di banca
centrale siano più indulgenti verso i loro debitori.
* Bank of Italy, Financial Risk Management Directorate.
CONTENTS
1. Introduction……………………………………………………………………………………………………………… 7
2. Literature review……………………………………………………………………………………………………… 10
3. The Italian banks’ IRBs and the Bank of Italy ICAS features……………………………………………. 11
4. The performance of the credit assessment systems………………………………………………………… 15
5. Empirical analysis…………………………………………………………………………………………………….. 18
6. Robustness analysis………………………………………………………………………………………………….. 23
7. Conclusions…………………………………………………………………………………………………………….. 23
References………………………………………………………………………………………………………………….. 25
Tables and figures…………………………………………………………………………………………………………. 27
Appendix 1. Dataset……………………………………………………………………………………………………… 44
Appendix 2. Proximity statistics……………………………………………………………………………………… 45
1. Introduction1
Collateralized credit operations are a standard and widely used tool for monetary policy
implementation. Central bank funding has been crucial for credit institutions’ support to the real
economy in the years following the financial crises and in the pandemic period, especially in the
jurisdictions where the financial sector was more fragile. In 2020-21, during the Covid-19 pandemic,
the average share of Eurosystem refinancing on total Italian bank liabilities was about 12 per cent.2
In this paper we employ a central bank credit assessment system for non-financial firms, the Inhouse Credit Assessment System (ICAS) of the Bank of Italy (BoI), which is part of the Eurosystem,
as a benchmark to assess the ratings set by commercial banks with their internal ratings-based systems
(IRBs) and used for monetary policy purposes. We examine the distribution of ratings on bank loans
pledged as collateral in monetary policy in Italy and test for risk under-reporting, found by Behn et al.
(2022) and Calza et al. (2021), which might generate an unwarranted exposure for the central bank.
Our goal is to detect differences, if any, in ratings between Italian IRBs and the central bank’s system,
and investigate their determinants.
Credit claims are an essential source of collateral for banks. At the end of September 2023,
bank loans were the main type of collateral in the Eurosystem, accounting for about one third of the
total,3 and the Eurosystem and the BoI employ several layers of protection against potential losses. To
address the risk of counterparty default, banks accessing credit operations must be financially sound
(first layer of protection). Credit risk is further mitigated by lending only against adequate collateral,
which represents the second layer of protection.4
The accurate design of this system has ensured that the Eurosystem has never experienced
losses in its credit operations so far.5 To be accepted as collateral in Eurosystem monetary policy
We thank for useful comments and suggestions an anonymous referee, Francesco Calise, Giorgio Donato, Davide
Giammusso, Alessandra Iannamorelli, Riccardo Lo Conte, Francesco Monterisi, Gerardo Palazzo, Tommaso Perez, Andrea
Polinici, Antonio Scalia, Stefano Siviero, Alberto Maria Sorrentino, seminar participants at Bank of Italy and at the 2024
International Risk Management Conference.
For an in-depth analysis of the extent to which public guarantees in Italy created additional credit with respect to preexisting levels, see Cascarino et al. (2022).
The use of credit claims as collateral reached record volumes following the easing measures taken in response to the
Covid-19 pandemic in 2020. See the press releases of 7 April 2020 and 22 April 2020 for the adoption of the package and
the press release of 10 December 2020 for the extension of the collateral easing measures until June 2022. For an in-depth
description of rules of the collateral framework see Antilici et al. (2023). The other types of collateral posted by banks in
Eurosystem monetary policy operations, at the end of September 2023, are: covered bank bonds (25.4 per cent), assetbacked securities (20.4 per cent), government securities (11 per cent), unsecured bank bonds (4.5 per cent), corporate bonds
(2.9 per cent), other marketable assets (2.1 per cent).
Collateral is valued daily and is subject to haircuts. The daily valuation and the haircuts are further layers of risk protection.
For a complete overview of the financial risk management of the Eurosystem’s monetary policy operations see ECB (2015).
Few cases of counterparty default events were offset by the proceedings deriving from the posted collateral. The main
case of default concerned Lehman Brothers Bankhaus AG (LBB), the German subsidiary of Lehman Brothers Holdings
operations, bank loans must comply with some eligibility criteria, including a minimum credit quality
threshold, assessed via a rating assigned by one of the following sources: i) the IRBs operated by banks;
ii) the ICASs managed by some national central banks (NCBs); 6 iii) the external credit assessment
institutions (ECAIs), or rating agencies. The rating is a key determinant of the haircut imposed on the
collateral.7
To shed light on the consistency of the IRB ratings for non-financial firms with those produced
by the Italian central bank, while we draw inspiration from Calza et al. (2021), we use a more recent
and complete dataset and we employ a richer and more robust estimation strategy. The dataset includes
all domestic IRBs used also for monetary policy purposes over the years 2015-2023,8 providing a full
characterization of banks’ policies in this area over a full business cycle, while the ECB authors cover
the 2014-2018 period. Additionally, while Calza et al. (2021) consider only credit claims pledged in
the general collateral framework (known as the Eurosystem Credit Assessment Framework, ECAF),
we take into account also those pledged in the temporary collateral framework (known as Additional
Credit Claims, ACCs).9 Considering the whole internal rating span enables us to avoid the sample
selection bias that affects the analyses based only on the ratings accepted within the general collateral
perimeter (i.e. IRB’s PDs below 0.4 per cent), resulting in a potential overestimation of the average
difference between IRB and ICAS ratings. Finally, as for the estimation approach, we test additional
specifications to measure, if any, the degree of risk under-reporting by banks in the context of monetary
policy operations.
Inc. whose liabilities vis-à-vis the Bundesbank from monetary policy operations stood at around €8.5 billion at the time of
the insolvency of the bank. The Bundesbank eventually managed to recover all the amount of LBB exposure. For further
details see the press release on the Bundesbank website.
In the euro-area ICASs are currently adopted by Banca d’Italia, Banco de Espana, Banco de Portugal, Banque de France,
Banka Slovenije, Bank of Greece, Deutsche Bundesbank, Oesterrichische Nationalbank. For a comprehensive overview of
central banks’ ICAS system in the euro area, see Auria et al. (2023).
Collateral valuation haircuts are determined according to the credit risk assessment of the debtor, the residual maturity
and the type of interest of the loan.
The PDs reported by banks in the context of the annual ECAF exercise are those used to calculate prudential capital
requirements under the European Banking Authority Common Reporting (CoRep) framework; as such, they comply both
with the Capital Requirements Regulation (CRR) and any Single Supervision Mechanism bank-specific requirements.
Additional credit claims (ACCs) are credit claims that do not fulfil all the eligibility criteria applicable under the general
collateral framework; every national central bank is free to set up a country-specific ACC framework. The possibility of
implementing ACC frameworks was introduced in December 2011, as part of the enhanced credit support measures to
support bank lending during the financial crisis. BoI has made full use of this possibility from the beginning, accepting as
collateral loans granted to debtors with lower creditworthiness and extending its ACC scheme several times. The ACC
scheme was introduced in 2012 in the BoI collateral framework. Within the BoI ACC framework individual claims may be
accepted if the one-year probability of default of the borrower is not higher than 1.5 per cent, while pools of credit claims
may be considered with a probability of default not higher than 10 per cent. In the context of the pandemic collateral easing
measures, from 25 May 2020 BoI has temporarily removed the 10 per cent PD threshold for pools of credit claims. Hence,
in our analysis, differently than Calza et al. (2021), we use the full range of ICAS’ PDs without censoring them.
The case of Italy is particularly fit to investigate the consistency across commercial and central
banks’ risk assessments systems, as a significant number of IRB systems are employed also for
monetary policy purposes and a large number of non-financial corporations are rated by the BoI
ICAS.10 Banks that choose to use their IRB as a primary source in the collateral framework must use
their IRB PDs for the mobilization of credit claims for all debtors rated by their IRB system.11 The PDs
are used both for the determination of the eligibility of the claim and for the calculation of the haircut.
The BoI ICAS performs an assessment of the borrower creditworthiness based on a statistical
model that employs a large set of variables, yielding ICAS ‘Statistical ratings’. For the largest
exposures the model assessment is complemented by BoI financial analysts’ evaluation, which yields
the ICAS ‘Full ratings’. ICAS Full ratings must be used for the sub-set of debtors whose claims can be
accepted also by the Eurosystem in the general collateral framework, while ICAS Statistical ratings
can only be used to assess debtors whose credit claims are accepted by BoI in the temporary collateral
framework.
IRBs and ICAS have different primary purposes. The primary purpose of IRBs is to compute
banks’ capital requirements to cover credit risk; they must be authorized by the banking supervisors
and abide to banking regulation.12 Besides, to be used for monetary policy purposes, IRBs must also
be specifically authorized by the Eurosystem. The primary purpose of ICAS is instead to assess the
credit quality of eligible loans to be used as collateral in monetary policy operations and thus to
determine the corresponding valuation haircuts. For the banks that do not manage an IRB, ICASs
constitute an important tool for expanding the sources of liquidity, allowing the use of bank loans as
collateral.
We contribute to the literature on credit risk assessment systems by bringing new evidence,
including for the post Covid-19 recovery. We study some features that to our knowledge have not been
explored yet, such as the use of credit rating systems within the ACC framework, which lends itself to
the analysis of loans of a lower credit quality, 13 and of ratings produced from NCB internal models
In 2023 eight out of ten Italian banks with IRB systems authorized for regulatory purposes had their IRBs approved also
for monetary policy purposes. The number of Italian banks with IRB systems authorized for ECAF purposes ranges from
seven in 2015 to eight in 2023, due to new authorizations and mergers. They were nine in 2020.
Whenever a debtor is not covered by the IRB, the secondary source, namely the ICAS in our analysis, is used.
The IRB approach to assign risk weights to exposures was introduced in 2004 by the Second Accord of the Basel
Committee on Banking Supervision (Basel II) as an alternative to the standardized approach (SA). In Europe the use of the
IRB approach for regulatory purposes was allowed since June 2006 by the Capital Requirements Directive; its adoption by
banks started to spread from 2008 onwards. The use of IRB models is conditional on supervisory authorities’ validation. In
the euro area validation of the models used to be granted by national supervisors until the end of 2014, when the Single
Supervisory Mechanism (SSM) established that IRB models of ‘significant’ banks must be validated by the European
Central Bank.
See note 9 for further details.
(ICAS Statistical) compared with those given by IRB quantitative models (IRB Statistical), 14 which
significantly increase the sample size and provide more robust results.15 Overall, the evidence and the
analysis of the drivers of the small differences between the credit risk assessment provided by IRBs
and ICAS lead us to conclude that the evidence of an economically significant strategic management
of the IRB credit systems found by Behn et al. (2022) and Calza et al. (2021) is not supported by the
data for Italy.
The remainder of this paper is organized as follows. Section 2 discusses the literature. Section
3 describes the data and stylized facts about the two credit assessment systems. Section 4 presents the
results of the performance analysis of the rating systems. Section 5 presents the results of the empirical
analysis. Section 6 discusses the robustness analysis. Section 7 concludes.
2. Literature review
With regard to the analysis of the framework of monetary operations, Calza et al. (2021) find
that, based on data between 2014 and 2018 from Austria, Belgium, France, Germany, Italy, Portugal
and Slovenia, for the set of debtors whose loans are eligible as collateral, banks’ ratings are on average
more conservative than NCB ratings, while the opposite occurs for the subset of borrowers whose loans
are pledged with the Eurosystem. Banks’ ratings become less conservative as loan size and the level of
central bank liquidity utilization increase.16 These results rely mainly on observations from France,
Italy and Austria; in the authors’ view, they support the hypothesis that banks strategically manage
their IRBs to maximize access to central bank funding. As such, IRB leniency by banks might generate
an unwarranted risk exposure for the central bank.
From the regulatory perspective, a set of studies investigate with pre-Basel II data the
consistency of banks’ IRBs to ascertain the role played by banks’ policies within the debate on the
merits of model-based capital regulation, with mixed results. Lenders may disagree on borrowers’
riskiness reflecting different views on credit quality (Carey, 2002) and also because of different lending
We group the IRB models which the banks use for ECAF purposes into Statistical (IRB Statistical) and Full (IRB Full)
ones. In the first group PDs are calculated with statistical or econometric models, while in the second PDs are calculated
complementing the models with the banks’ financial analysts’ assessment (potentially leading to so called override). We
compare ICAS Full PDs with IRB Full PDs and ICAS Statistical PDs with IRB Statistical PDs.
In 2023 ICAS Full ratings were in excess of 3,000 and ICAS Statistical ratings of 330,000.
In contrast, the degree of capitalization of banks does not seem to have explanatory power for the difference between the
ratings of the two systems for bank loans used as collateral.
or risk management styles (Jacobson et al., 2006). Estimates of PDs can diverge significantly, but not
systematically across banks.17
A second set of studies investigates the effect of the introduction of the model-based approach
with post-Basel II data, focusing on the relation between the variability of PD estimates across IRBs
and banks’ characteristics. These works find that in Germany from 2008 to 2012 weakly capitalized
banks may report lower PDs (Berg and Koziol, 2017) and some banks may systematically under-report
risk to lower their capital requirements (Behn et al., 2022). However, there is also evidence that
validated IRB models are accurate and robust, and that the introduction of the IRB approach promotes
the adoption of stronger risk management practices among Western European banks from 2008 to 2015
(Cucinelli et al., 2018).18
3. The Italian banks’ IRBs and the Bank of Italy ICAS features
General description – IRBs and ICASs are both designed to evaluate the probability of default
associated with credit exposures.19 However, they differ in their objectives, regulatory context,
methodologies. In terms of objectives and regulatory context, IRBs’ valuations are mainly used for the
calculation of capital requirements and as important inputs in approving and pricing loans, and in
managing the portfolio of a wide range of credit exposure which does not entail only the corporate
ones; they are developed in line with the Basel framework. ICASs are instead specifically aimed at
evaluating the creditworthiness of non-financial corporations whose credit claims are used as collateral
in monetary policy operations; they are developed by central banks in line with the Eurosystem
requirements.
As concerns the methodologies, both systems use statistical techniques and historical data to
estimate the likelihood that a borrower will default within a specific time horizon, i.e. one year;
however, the exact statistical methodology, the set of information taken into consideration and the
length of the time series used in the models may differ among the systems.20 Another important element
A feature of the Basel rules is that banks can have differences in opinions and approaches to managing and measuring
credit risk, that imply different risk parameters. Persistent differences in loss-given-default (LGD) may be attributed to
banks’ policies (Firestone and Rezende, 2016).
From Austria, Belgium, Denmark, Finland, France, Germany, Great Britain, Ireland, Italy, Netherlands, Norway,
Portugal, Spain, and Sweden.
IRB models are also used for the estimation of other parameters apart from the PD, i.e. the Loss Given Default (LGD)
and the Exposure at Default (EAD). For a detailed overview of the regulatory requirements that IRB models have to comply
with see EBA (2017) and ECB (2024).
These elements may of course also differ (for the same class of debtors) from one IRB bank to another. As concerns
different ICASs, while all the ICASs developed within the Eurosystem are similar in their general characteristics and
comply with Eurosystem requirements, some differences can be observed, either in the sources of information or in the
methodology adopted in order to calculate the ratings.
which may differ among IRBs and ICASs is the rating “philosophy”. In the context of rating systems,
two approaches can be adopted, one that includes cyclical effects and one that does not. The two
approaches generate different rating types, commonly known as point-in-time (PIT) and through-thecycle (TTC).21 While often models cannot be classified as purely PIT or TTC, but are rather a hybrid
combination of the two approaches, it can be said that the rating philosophy of BoI ICAS is point-intime, while it is expected to be more through-the-cycle oriented for IRBs.22
The BI-ICAS rating process is based on a two-stage procedure, which combines a statistical
module assessment with a judgmental model. 23 Stage 1 (Statistical Module) consists of a system of
logit models that determines a one-year default probability (“ICAS Statistical PDs”). 24 Stage 2 (Expert
Assessment) involves the financial analysts’ assessment through the use of a wider range of
information sources.25 The PDs output of this process are the “ICAS Full PDs”.
Data – We use yearly data between 2015 and 2023 on ratings collected from ten Italian IRB
systems and the BoI ICAS during the annual Eurosystem performance monitoring exercises and on
credit claims used as collateral in credit operations with BoI. 26 In order to ensure consistency, accuracy,
and comparability of the credit assessment systems used for monetary policy purposes, the Eurosystem
established a framework to monitor their performance.27 The managers of each NCB’s credit
assessment system are required to send to the ECB data on all firms whose bank loans are assessed to
be eligible as collateral in the Eurosystem monetary policy operations. The data for ICASs and IRBs
are reported to the respective NCB and include PDs assigned to debtors over a one-year horizon.28
Accepted credit assessment sources can use their own individual rating scales and grades. The
Eurosystem maps these different grades into a harmonized rating scale to make the credit ratings
comparable across systems and sources. The scale has a number of Credit Quality Steps (CQSs) linked
PIT ratings aim at evaluating the current situation of an entity by taking into account both cyclical and permanent effects.
In contrast, TTC ratings focus mainly on the permanent component of default risk and are essentially independent from
cyclical changes in the entity’s creditworthiness.
See Paragraphs 105-106, Credit Risk Chapter of the ECB guide to internal models. Common statistical methods used by
IRBs include traditional credit scoring models and machine learning algorithms. In particular, IRB models often use
traditional credit scoring techniques which include discriminant analysis models and regression models (linear, logit,
probit).
See Giovannelli et al. (2020) for further details.
The system exploits two sets of variables: indicators derived from the National Credit Register (NCR) and indicators
based on financial statements data. Model parameters are estimated using the observed defaults derived from NCR as
dependent variable. Rating levels are associated to estimated PDs according to a scale mapping
The analysis can either confirm the rating derived from Stage 1 or modify it by notching the master score up or down.
For the last performance exercise, conducted in 2022, the new ICAS statistical model (Narizzano et al., 2024) was not
yet available.
The basic principles of the framework are included in Article 126 and Annex IX of the Guideline (EU) 2015/510 of the
ECB of 19 December 2014 on the implementation of the Eurosystem’s monetary policy framework (ECB/2014/60) (recast).
See Appendix 1 for details on the variables employed.
to maximum PDs over a one-year horizon (Table 1). The CQSs are relevant for the eligibility of assets
and to determine the valuation haircuts of the collateral.
The Eurosystem considers a PD over a one-year horizon of up to 0.1 per cent as equivalent to
a credit assessment of CQS 1 and 2 in the Eurosystem harmonized rating scale, while a PD from 0.1
up to 0.4 per cent is equivalent to CQS 3. All assets accepted by the Eurosystem as collateral must
meet the minimum requirement of a credit assessment of CQS 3. Probabilities of default from 0.4 up
to 1 and from 1.0 to 1.5 per cent correspond to CQS 4 and 5, respectively, and are relevant for the
acceptance by NCBs (among which BoI) of individual loans as collateral within temporary collateral
frameworks. Similarly, NCBs may accept as collateral pools of ACCs up to CQS 8 (as in the BoI
temporary collateral framework).
Data for ICAS and IRB systems transmitted for the annual Eurosystem monitoring exercise
comprise PDs at the beginning and at the end of the year.29 We focus on ratings assigned to nonfinancial corporates that are jointly rated by BoI ICAS and by at least one IRB system. We first consider
the observations for all eligible loans from borrowers jointly rated by the BoI ICAS and by at least one
IRB system; we then construct a narrower data set focusing only on those loans actually mobilized as
collateral.30 Table 2 shows the number of firms rated by at least one IRB, by BoI ICAS and jointly by
both systems, with the detail of the number of firms whose credit claims are used as collateral by Italian
banks.
Stylized facts – We compare the assessments of IRBs and BoI ICAS collected annually for the
Eurosystem performance monitoring exercise.31 IRB Full PDs have been slightly less conservative than
ICAS Full ones for eligible borrowers since 2017 (Fig. 1A): the difference has been mildly larger for
debtors with loans pledged as collateral. IRB Statistical PDs, instead, since 2016 have been overall
more conservative than ICAS Statistical PDs (Fig. 1B), to a smaller extent for debtors with loans
pledged as collateral. The differences between IRB PDs and BoI ICAS PDs are statistically significant,
according to the Student’s t-test and the Wilcoxon signed-rank test, in nearly all the years and samples.
In our analysis we use beginning-of-year PDs. Calza et al. (2021) use end-of-year PDs, in order to benefit from greater
rating variability, because beginning-of-year data are bounded by the ex-ante eligibility threshold (while end-of-year PDs
may also refer to debtors that have lost eligibility throughout the year). However, in jurisdictions (like Italy) in which pools
of credit claims are accepted, the credit assessment sources are requested to report debtors belonging to the full rating
spectrum also at the beginning of the year; therefore, we are able to use beginning-of-year PDs without incurring in the
problem of losing rating variability.
All the Italian banks with IRBs authorized for monetary policy purposes have chosen to use the PDs calculated by their
IRB systems when pledging their loans to the ECAF.
In our analysis firms rated by more than one IRB are considered as different debtors, i.e. if a firm is rated by two different
IRBs and by the BoI ICAS it is considered twice in the sample of debtors rated in common.
The circumstance that IRB Full PDs are only slightly less conservative than ICAS Full ones for
both eligible borrowers and those with loans mobilized as collateral, indicates the absence of a relevant
underestimation of credit risk by the banks in the use of IRB to maximize the pledgeability of their
loans. Moreover, the differences between the ratings of the IRBs and BoI ICAS, which are more
pronounced for the debtors with mobilized loans, may reflect the selection made by banks when
assembling the collateral portfolio, contrary to the negative effect associated to the hypothesis of a
strategic management of collateral. Banks may have in the first place an incentive to pledge loans with
higher credit quality according to their genuine assessment to secure lower haircuts and more stable
central bank liquidity. 32
To appreciate how a sub-sample with a higher share of debtors with lower PDs may affect the
difference between IRB and ICAS PDs, we focus on the three sub-samples of the general collateral
framework, of the individual claims and of the pools of credit claims in the temporary collateral
framework of Additional Credit Claims (ACC).33
Within the general collateral framework sub-sample of pledged debtors (Fig. 2) the differences
between IRBs and ICAS Full PDs are negative, as in the full sample of pledged debtors; this is in line
with the finding of Calza et al. (2021). However, when considering individual ACC claims, IRB ratings
are on average less conservative than ICAS ratings only in half of the years analyzed and IRB ratings
are on average more conservative than ICAS ones for ACC pools.34 The comparison between IRBs
and ICAS Statistical ratings provides a similar result, with larger differences for the ACC pools (Fig.
A collateral valuation haircut is the deduction of a certain percentage from the valuation of an asset for the purpose of
calculating the amount of liquidity that can be backed by the asset in case of counterparty default. The calibration of haircuts
aims to ensure the equivalence of risk across different types of collateral assets. The loss in value of collateral that the
Eurosystem expects to incur – with low probability – in an adverse scenario should be the same for the different assets and
asset types. An adverse, but still reasonable scenario, is defined by the Eurosystem as the average loss occurring within the
worst percentile of the loss distribution. For calibration purposes, an adverse scenario is set to correspond to the average
loss in the worst one per cent of the cases, i.e. to expected shortfall at a 99 per cent confidence level (ES[99%]). An
important risk component for the haircut calibration is represented by default risk, which is embedded in rating assessments,
even though it is not the only one. Market and liquidity risk are also other modelled sources of risk. For more detail on
haircut calibration see ECB (2015) and, with specific reference to marketable assets, Adler et al. (2023).
We proxy the three frameworks (ECAF, ACC single claims and pool of ACCs) splitting the sample into three subsamples
based on IRB PD ranges; this is due to the fact that debtors can overlap between the three frameworks, especially for ACC
pools, where there is not a lower PD boundary and PDs can also assume values below 1.5 per cent. Therefore, in this
paragraph when we refer to ECAF, ACC single claims and pool of ACCs we intend, respectively, credit claims with PD
values: i) up to 0.4 per cent; ii) from 0.4 to 1.5 per cent; iii) greater than 1.5 per cent.
In the context of the collateral easing measures introduced in 2020 to cope with the effects of the Covid-19 pandemic,
the 10 per cent upper limit has been removed in the Italian ACC framework. This means that banks can pledge loans in the
ACC pools with debtors’ PDs greater than 10 per cent, provided that such loans are performing. This measure is supposed
to be in place until March 2024, when the ECB should, to this moment, phase-out all the pandemic collateral easing
measures, following a comprehensive review of the ACC frameworks.
2). The fact that IRB PDs are less or more conservative than ICAS PDs is mainly an effect of the
considered sub-sample’s selection criteria, which is based on the variable of analysis (i.e. the PD).35
The selection induced by the criteria of the collateral frameworks arguably leads to observed
differences between the ratings of the two systems that are more pronounced in the sub-sample of
pledged debtors than in the sample of eligible debtors (Fig. 3).36 Indeed we find that the share of IRB
banks’ loans in CQSs 1, 2 and 3 in the sample of borrowers assessed with full ratings is 48 per cent for
pledged debtors and 36 for eligible debtors (Fig. 3, left graph). As for the sample of borrowers assessed
with statistical rating, the share of banks’ loans in CQSs 1, 2 and 3 is 29 per cent for pledged debtors
and 22 per cent for eligible debtors (Fig. 3, right graph). These figures suggest to control for IRB credit
quality rating in the empirical analysis since the credit quality distribution is more skewed towards
higher CQSs for pledged firms than for eligible firms.
4. The performance of the credit assessment systems
AUROC curve – One of the most widely used metric to test the performance of the credit
assessment system is the Area Under the Receiver Operating Characteristic (AUROC) curve.37 The
ROC curve shows the trade-off between the true positive rate (TPR) and the false positive rate (FPR)
across different decision thresholds, where:
??? =
???? ????????? (??)
???? ????????? (??) + ????? ????????? (??)
??? =
????? ????????? (??)
????? ????????? (??) + ???? ????????? (??)
The calculation of the area under the ROC curve is a popular way of testing the discriminatory
power of the rating systems, which measures the degree with which the system is capable of assigning
every entity to the correct class (in our case, the system performs a binary classification between
The PDs of borrowers whose loans are pledged by IRB banks have to be lower than 0.4 per cent, while ICAS PDs are not
upper-bounded. As for the ACC individual claims sample and the ACC pool sample, instead, a censoring of lower IRB
PDs occurs, since IRB PDs cannot be lower than 0.4 and 1.5 per cent, respectively, while ICAS PDs are not lower-bounded,
contributing to the circumstance that IRBs are more conservative than ICAS in this sample.
Eligible debtors are those whose loans are eligible as collateral in monetary policy operations; pledged debtors are those
whose loans are pledged as collateral in such operations.
See Engelmann (2006) and Tasche (2006) for further details.
defaulted and non-defaulted debtors).38 The area under the curve ranges from 1, corresponding to
perfect discrimination, to 0.5, corresponding to a model with no discrimination ability (naïve model).
Figure 4 shows the ROC curves relative to the Italian IRB systems accepted for monetary policy
purposes (considered as a whole) and to BoI Full ICAS.
The AUROC is quite high for both the IRBs and the ICAS Full, with IRBs showing a steeper
ROC curve (0.87 and 0.83 in Fig. 4A and 4B, respectively). The good results in terms of AUROC are
confirmed also considering smaller samples of data for individual banks, years, and for debtors with
loans pledged in monetary policy operations. Similar evidence can be detected also regarding the
AUROC of IRBs and ICAS Statistical (0.87 for both systems in Fig. 4C and 5D). This is another
element which confirms that the ability of the two credit assessment sources to properly evaluate the
credit risk of the debtors they rate is satisfactory and that it does not raise concerns with respect to
underestimation of credit risk.
Back-testing – In the previous sections we have verified that ICAS and IRB ratings are
dissimilar to a certain extent, especially for pledged loans and in the more recent period. However, it
is rather normal that ratings issued by two different credit assessment sources are not identical
(Firestone and Rezende, 2016), as this may be due to different information sets, diverse statistical
models, divergences in the expert assessment of the analysts. The true PD of a debtor is an unobservable
variable; therefore, it is not possible to establish ex-ante whether a credit assessment system is correct
in assigning a certain rating to a debtor, but it can be verified ex-post. From the perspective of a central
bank, which receives bank loans as collateral in monetary policy operations, the satisfactory
performance of the credit assessment systems evaluating these loans is of the utmost importance.
We performed a back-testing exercise with the purpose to confirm that the two credit
assessment systems under investigation do not underestimate the defaults in the sample of the debtors
with bank loans pledged as collateral in monetary policy operations, as for the AUROC, we focus on
the full rating sample in line with the goal of the paper. 39 The methodology we use follows a ‘trafficlight approach’ (TLA) with a green zone, a yellow (‘monitoring’) zone and a red (‘trigger’) zone. These
zones indicate different levels of significance of the deviations of the number of defaulted entities and
A system with a very high discriminatory power has a ROC curve which goes closer to the top left hand corner of the
plot, whereas a model performing poorly (naïve model) has a ROC curve close to a 45 degree line.
The results of the tests are not reported due to confidentiality reasons.
of the corresponding default rates from the PD thresholds of the respective Eurosystem CQSs as
defined in Table 1.40
We first calculated the back-testing on eligible debtors, as if we had a whole static pool which
includes all the static pools of each IRB system. We then repeated the test only on the sub-sample of
debtors in common with ICAS with loans pledged in monetary policy operations. The results indicate
that, apart from some performance issues between 2015 and 2017, due to the unsatisfactory
performance of some IRBs, in the last five years the IRB systems in aggregate showed a satisfactory
performance both for the whole perimeter of eligible debtors and for the sample of debtors with pledged
loans, which is the most relevant from a central bank risk management perspective.
We repeated the performance monitoring exercise for ICAS ratings on the whole sample of
eligible debtors and then we focused on the sample of debtors in common with IRB banks with loans
pledged as collateral; we detected some performance issues in the first years under analysis, the startup phase of the ICAS system, which were solved from 2018 onwards.
In addition, we conducted the performance monitoring exercise also for IRB and ICAS
Statistical ratings. The results for IRBs indicate that, apart from some performance issues in 2015, IRB
Statistical systems show a good predictive capacity both on the universe of eligible debtors and on the
sample of pledged debtors in common with ICAS Statistical. The performance is satisfactory also for
ICAS Statistical ratings for almost the whole period; the model had some minor performance issues
only in 2022, 41 then overcome with a new statistical model, developed and implemented also to address
these issues, as explained in Narizzano et al. (2024). In the development of the new statistical model
features of traditional models (i.e. the logit regression) and machine learning components for some
variables are combined, in the attempt to solve the trade-off between predictability and transparency.
In conclusion, despite some issues emerging mainly in the first years of the analysis, in the
more recent period both IRBs and ICAS systems show a good performance, even when tested on bank
loans pledged as collateral.
The trigger zone indicates that the deviation is considered very significant, i.e. the probability that the credit assessment
system is mis-calibrated is very high. The monitoring zone, in turn, indicates a degree of deviation that is significant, i.e.
the probability that the credit assessment system is mis-calibrated is high, but not large enough for classifying this situation
as a breach of good performance by the system. The green, yellow and red zones are determined on the basis of a statistical
binomial test: for further insights see Coppens et al. (2007).
The performance issue in 2022 was related to the micro firms class, due to the fact that a number of such firms disclosed
a simplified version of the financial statements, without the breakdown required for a complete model estimation.
5. Empirical analysis
Empirical approach – In this section we investigate the factors driving the differences between
the ratings provided by banks’ IRB systems and those provided by BoI ICASs in the collateral
framework. First, we estimate the following equation in a panel setting:
??
log (???????? ) = ?? + ?? + ?? ? ?? +? ? ?????????? + ????
??????
??
where i indicates the individual firm, j the bank and t the year examined, log (???????? ) is the log ratio
??????
between PDIRB and PDICAS, Dpledgeijt is a dummy that is active when the loan is pledged in either of
the general collateral, individual or pool temporary collateral frameworks. We include year (Dt), firm
and bank fixed effects, to control for business cycle, structural changes, businesses’ and banks’
unobservable characteristics and we use the log ratio between the PDs of the IRBs and that of BoI
ICAS to address the non-linearity in the absolute differences between PDs.
As in Calza et al. (2021), we assume as a benchmark for the risk assessment of the borrower
the PD assigned to the debtor by the NCB, since it has no incentive to manage strategically its rating
models parameters, while commercial banks may have an incentive to optimize them to save on capital
requirements (this would be the primary goal of a hypothetical strategic behaviour in the management
of the internal rating systems), or to save on the collateral posted. According to this hypothesis, banks
could underestimate the credit risk of the debtors whose loans are pledged to the NCB to save on the
haircuts imposed on the value of the collateral in the monetary policy operation to maximize the
liquidity obtained. The effect of this behaviour would materialize in a lower credit risk assessed by
IRBs, with respect to ICAS, for the borrowers whose loans are pledged, while it would not alter the
creditworthiness assigned by banks to the other borrowers, relative to that calculated by ICAS.
In our analysis, thanks to the richness of the data set received from the banks and used to
monitor the compliance with the collateral frameworks, we can also differentiate the IRB models of
banks among Statistical ones, that do not entail a credit analysts’ judgment before issuance of the rating,
and Full ones where analysts’ judgment is layered on the outcome of the statistical model. We also
have the benefit of being able to match those data with the respective ratings produced by the BoI ICAS
Statistical and Full (which include BoI analysts’ judgment) models. This wealth of data allows us to
differentiate the models among those theoretically amenable to some strategic management of the risk
assessment via the intervention of the analysts’ judgment, the Full ones, and those who are not, the
Statistical ones. IRB models are validated and supervised by the Single Supervision Mechanism (SSM)
both for supervisory purposes and for the use in the Eurosystem monetary policy purposes and their
parameters are rigorously calibrated excluding the possibility that they systematically underestimate
the credit risk.
Our analysis thus differs from that of Calza et al. (2021), who only consider ICAS Full models
in comparison with undistinguished Full and Statistical IRBs and interpret as strategic underestimation
of the credit risk any observation of a debtor to which an IRB assigns a PD lower than the one assigned
by the ICAS. 42
We also argue that to ascertain a hypothetical strategic behavior of banks on the basis of the
comparison between IRB and ICAS PDs for the firms whose loans are pledged (pledged firms) and for
those whose loans are not (eligible firms), the effects of the criteria for acceptance of loans in the
collateral frameworks on the credit quality composition of the sample of pledged loans have to be duly
considered. Indeed, when we compare IRB and ICAS assessments of eligible debtors the observed
borrowers are not selected and represent the universe of the debtors assessed.43 Differently, when we
consider the assessments of pledged debtors the observed sample is filtered by the selection made by
the banks and, presumably, the pledged debtors are those with the best IRB rating available responding
to the incentive to secure the lowest haircuts for the loans.44
Hence, in order to correctly measure the effect of the pledging of a loan on the difference
between IRB PDs and ICAS PDs (which proxies the degree of relative conservativeness in credit risk
assessment of IRB banks with respect to the NCB) the credit quality of the pledged loan has to be
controlled for. In equation (2) we therefore control also for the credit quality of the loans, where k
indicates one of the eight CQSs in which loans are classified according to the ECAF rating scale (Table
??
log (???????? ) = ?? + ?? + ?? +?? ? ?? +? ? ?????????? + ????
??????
Before the introduction of the pandemic easing measures, BoI ICAS was the only ICAS in the euro area to use both a
Full and a statistical model (the latter only in the ACC framework). Therefore, in Calza et al. (2021) only ICAS Full models
are considered. Since 2020, Statistical ICAS have been also used in the ACC framework by Oesterreichische Nationalbank,
Banco de España, Banco de Portugal and Banca Slovenije.
We considered all the debtors jointly assessed by at least one of the IRB systems of the banks and the ICAS system.
The effects of the selection can be seen in Fig. 3, which shows that the CQS distribution is more skewed to the left in
the pledged sample with respect to the distribution for all eligible borrowers.
We then aim to investigate what drives the differences among the IRBs and ICAS ratings for
pledged debtors. In particular, we analyze the characteristics of the debtors rated by IRBs and ICAS
and of the banks which rate them estimating equation (3) for debtors with loans pledged as collateral.
log (
????????
????????
) = ?? + ?? + ?? + ?? + ?? . ?? +?1 ? ??????2 + ?2 ? ?????? + ?3 ? ??????? +?4 ? ???1??? + ?5 ? ????? + ????
where REV is the log of the borrower’s net revenues, that proxies the size of the borrower,45 and DEB
denotes the share of the loans received by the company from bank j, to proxy the relationship lending
intensity. CVAH is the collateral value after haircuts of the loans of the company i pledged by the bank
j in monetary policy operations. CET1R is the bank Common Equity Tier 1 Ratio of bank j, and LTC
is the loan-to-collateral ratio for bank j, i.e. the ratio of the central bank liquidity received in the
monetary policy operation and the collateral posted by the bank.46 We also include time, bank and firm
fixed effects.
Results – The results for equation 1 indicate that IRB Full PDs are lower than ICAS Full PDs
for the eligible debtors, as the coefficients of the time dummies are negative (Table 3a),47 also
controlling for firm fixed effects, indicating that IRB models are less conservative than ICAS in credit
risk assessment.48 The (negative) difference between IRB and ICAS PDs increases when a loan is
pledged, also controlling for firm fixed effects. Calza et al. (2021) find a different result for the eligible
debtors, since IRB Full PDs are higher than ICAS Full ones, while their results for pledged debtors are
in line with our evidence, i.e. a negative difference between IRBs and ICAS Full PDs; they argue that
this large decrease (comparing results on eligible debtors and pledged debtors) in the difference
between IRB and ICAS PDs indicates a strategic underestimation of the credit risk by commercial
banks to increase the pledgeability of loans and banks’ access to central bank liquidity.
The results are different for the sample of Statistical ratings (Table 3b). The positive coefficients
of the year dummies, with the exception of 2015, indicate that for eligible debtors IRBs are more
The variable is used with a two-year lag with reference to the PDs to conform to the data available when ratings are
assigned.
The rationale behind the introduction of this covariate is to test for the hypothesis that banks with a higher level of central
bank liquidity utilization are more keen to strategically underestimate credit risk when pledging credit claims.
The results are discussed using the same metric of Calza et al. (2021).
The only exception is the 2016 coefficient for eligible loans.
conservative than ICAS, and, similarly to what observed for Full ratings, the (positive) difference
decreases when considering pledged debtors.
The coefficients estimated for the dummy pledged indicate to what extent the circumstance that
a debtor is pledged affects the difference between IRB PDs and ICAS PDs. The estimated effect for
Full ratings is -0.40 and, controlling for bank fixed effects, it is halved, to -0.23. The inclusion of credit
quality steps in equation 2 yields a significant reduction by a factor of 10, to -0.027 (Table 4) for Full
ratings, and from -0.18 to -0.08 for Statistical ratings, in both cases with an improvement in the
explanatory power of the model.
This evidence further indicates that the hypothesis of Calza et al. (2021) of a sizeable strategic
management behaviour of banks is not supported by the data as the coefficient of the dummy for
pledged debtors significantly decreases controlling for bank fixed effects and CQS by a factor of almost
20. Indeed, the small change (i.e. -0.027) in the gap between IRB and ICAS PD associated to the
pledging of a loan contrasts with the sensibly larger variation in Calza et al. (2021), who find that IRB
PDs double the ICAS PDs for eligible debtors,49 and amounts to two-thirds of ICAS PDs for pledged
debtors.
The analysis of the drivers of the difference between the assessments of IRBs and ICAS can be
beneficial to the calibration of the credit risk models and it can also shed light on how the features of
the rating systems interact with banks’ and firms’ characteristics that influence the transmission of the
monetary policy.
We find that rating disagreements become less likely the larger the borrower (Tables 5a, 5b),
as the coefficient of REV is positive, reducing the negative gap between the IRB and the ICAS PDs:
this result is in line with the fact that public information is more easily available for larger firms,
inducing a likely convergence of the ratings from different CASs, as found by Carey (2002). The results
indicate also that the intensity of the lending relationship, proxied by the variable DEB, affects
differently the two kinds of assessment to a very limited extent and only in the Statistical ratings
sample, possibly as the use of loans as collateral is not perceived relevant enough to activate banks’
behavior aimed at protecting their relationships by increasing the pledgeability of the loans via a more
favorable credit assessment.
We also find that more capitalized banks tend to assign more conservative ratings to their
debtors, as the Common Equity Tier 1 coefficient is positive in the Full ratings sample, where
assessments are constituted of the banks’ analysts’ judgment overlaid on the statistical models’
With positive average log-differences ranging from 0.8 to 0.95.
outcome, while it is mostly not significant in the statistical ratings sample, that relies only on models’
ratings. This finding suggests that, correspondingly, banks with less capital and more limited balance
sheet capacity are incentivized to be less conservative in credit risk assessment to strengthen primarily
their capital base, and, secondarily, their access to central bank funding, in line with Calza et al. (2021)
and the literature on regulatory capital and credit risk (Berg and Koziol, 2017, Behn et al., 2022).
On the contrary, an increase in the amount of the pledged loan, proxied by the variable CVAH,
increases the negative difference between the ratings of the IRBs and ICAS in the Full ratings sample,
in line with the results found by Calza et al. (2021).50 This result could suggest a strategic behaviour
of the banks in managing the collateral.
The effect of the loan-to-collateral ratio in the monetary policy operations in the Full rating
sample is positive, when also bank fixed effects are introduced, indicating that higher NCB liquidity
utilization by banks makes the banks produce more conservative ratings.51 This evidence contrasts with
the negative effect of high liquidity utilization of banks on banks’ conservativeness in the assessment
of credit risk found by Calza et al. (2021), who infer that banks interested to save on their capacity to
access central bank’s liquidity underestimate the credit risk of the collateral.
We then introduce among the regressors the classes of credit quality steps to control also for
the effect of the criteria for acceptance of loans in the collateral framework. Such approach represents
an innovation with respect to previous works, making our results not directly comparable with them.
In the Full rating sample, the effect of net revenues remains positive and that of the intensity of the
banking relation is not significant (Table 6a). The effect of bank capital remains positive, confirming
that more capitalized banks tend to assign more conservative ratings to their debtors. The effects of the
collateral value after haircuts and of the LTC ratio, instead, turn out to be not significant, when all
unobservable effects are accounted for, confirming the absence of a strategic behaviour of banks.52 The
results for the Statistical rating sample (Table 6b) confirm those commented above.
The residual differences between IRB and ICAS PDs not explained by the abovementioned
factors could be driven also by diversity across rating models along the following dimensions: i) the
In the Statistical rating sample, the effect, statistically weaker, is of a reduction of the positive difference between IRB
and ICAS ratings; we do not view this as a sign of strategic management of collateral either since banks’ analysts cannot
override the ratings produced by the models.
The effect of higher liquidity utilization in the Statistical rating sample, statistically weaker, is of a reduction of the
positive difference between IRB and ICAS ratings; we do not view this as a sign of strategic management of collateral
either since banks’ analysts cannot modify the ratings produced by the models. We also note that higher values of loan-tocollateral ratio do not necessarily indicate more liquidity-constrained banks, instead they can be linked to the liquidity
management practices or to the specific business models of the banks under analysis.
Finally, we explore whether the disagreements could be lower if a debtor is rated by many banks (i.e. by three or more
banks). The results suggest that this variable is not significant and therefore it seems not to have a role in explaining the
differences among IRB and ICAS ratings.
rating quantification approach, i.e. the assignment of PD values to different rating classes; ii) the set of
information and statistical methodology used for the rating assignment; iii) the length of the time-series
used for PD calibration; iv) the rating “philosophy”, which is more through-the-cycle oriented for IRBs
and point-in-time for ICAS, as explained in Section 2.
6. Robustness analysis
In order to allay the concern on the dependence of the comparison between eligible and pledged
debtors on the IRB PD distribution in the two samples we estimate equation (2) separately, for both
Full and Statistical ratings, on the seven sub-samples of eligible debtors grouped by credit quality step
class according to their IRB PD, which allows also for controlling more accurately for the effects of
the different distribution of loans across lower and higher credit quality classes. The results confirm
those found controlling directly for credit quality steps in equation (2) (Tables 7, 8).
As a further robustness check, we run equations (1), (2) and (3) using the absolute differences
between IRB PDs and ICAS PDs as the dependent variable. The results (tables 9a, 9b, 10, 11a, 11b)
broadly confirm the main evidence found when running the same equations with log PD ratios as the
dependent variable.
Finally, we test the differences between the IRBs and ICAS in terms of ranking of borrowers’
creditworthiness using proximity indicators (Appendix 2). Since the relative ordering of the debtors’
assessments is similar and the statistics indicate an agreement of the two kinds of systems, the evidence
supports the assessment that the differences between the IRB and ICAS systems are quite low.
7. Conclusions
This work aims at understanding whether there are systematic differences in the ratings issued
by Italian IRB systems accepted for monetary policy purposes and by the internal rating system (ICAS)
developed by BoI. The data indicate that small differences exist and that IRBs tend to be less
conservative than ICAS when considering the Full models, i.e. those that include banks’ financial
analysts’ assessment, while the opposite occurs when considering the purely statistical models. The
difference between IRB and NCB PDs mildly increases for firms with loans pledged as collateral.
We bring new evidence on credit risk assessment systems, including for the post Covid-19
recovery. We claim to contribute to the literature with the study of features of the internal rating models
not yet explored to our knowledge, such as the pandemic related extension of the Eurosystem collateral
framework, which allows to analyze loans of lower credit quality, and ratings produced by the NCB
internal models and by IRB purely quantitative models, extending the analysis to a vast number of
small non-financial corporates that account for a large part of the Italian economy.
We observe that IRB and BoI systems show an overall satisfactory predictive capacity and a
good discriminatory power.
We find that, for borrowers whose corporate loans are eligible as collateral, IRB Full PDs are
slightly less conservative than those assigned by ICAS, especially in the more recent years. We also
find that the difference between IRB and ICAS Full PDs increases very modestly when restricting the
sample to the debtors with loans mobilized as collateral and fully accounting for borrowers’
creditworthiness. This evidence suggests that, when taking into account banks’ common liquidity
management practices in the context of the collateral framework, a strategic management of the rating
of the collateral, if any, is negligible and of low economic significance. Furthermore, ratings produced
by quantitative models without the judgment of analysts, IRBs Statistical PDs, are often more
conservative than ICAS Statistical PDs, but to a lesser extent when considering the borrowers whose
loans have been pledged.
We also characterize the factors driving the differences between the ratings issued by IRBs and
by BoI ICAS. We find that the difference between PDs issued by IRBs and by ICAS decreases with
the size of the borrower. We view this finding as reflecting the consideration that for larger companies
and for those accessing market financing the availability of public information is larger, contributing
to align the assessment of rating systems. In addition, we find that the difference between PDs is not
affected by the size of the loan pledged and by the levels of central bank liquidity utilization, allowing
us to conclude that IRB banks do not underestimate the PDs assigned to borrowers to mobilize
strategically the largest amount of collateral. Residual differences between IRB and ICAS PDs may be
ascribed also to methodological diversity across rating systems.
We also evaluate the ratings produced by the banks and NCB systems as similar according to
proximity measures.
Overall, the evidence discussed above and the analysis of the drivers of the small differences
between the credit risk assessment provided by IRBs and ICAS leads us to conclude that the hypothesis
of an economically significant strategic management of the IRB credit systems is not supported by the
data for Italy and that, while IRBs in some instances appear less conservative than ICAS, both IRBs
and ICAS evaluate credit risk adequately, in this way contributing to minimize the risks borne by the
Eurosystem and Bank of Italy within the monetary policy collateral frameworks.
References
M. Adler, G. Camba-Méndez, T. Džaja, A. Manzanares., M. Metra and G. Vocalelli (2023), “The
valuation haircuts applied to eligible marketable assets for ECB credit operations”, ECB Occasional
Paper No. 312.
P. Antilici, G. Gariano, F. Monterisi, A. Picone, and L. Russo (2023), “The Eurosystem Collateral
Framework and the Measures Introduced in Response to the Pandemic Emergency”, in Financial Risk
Management and Climate Change Risk: the Experience in a Central Bank, edited by A. Scalia, 55-72.
Springer.
L. Auria, M. Bingmer, C.M. Caicedo Graciano, C. Charavel, S. Gavilà, A. Iannamorelli, A. Levy, A.
Maldonado, F. Resch, A.M. Rossi, S. Suaer (2021), “Overview of central banks’ in-house credit
assessment systems in the euro area”, ECB Occasional Paper No. 284.
M. Behn, R. Haselmann, and V. Vig (2022), “The Limits of Model-Based Regulation”, Journal of
Finance, Vol. LXXVII, No.3, 1635-1684.
T. Berg and P. Koziol (2017), “An Analysis of the Consistency of Banks’ Internal Ratings”, Journal of
Banking and Finance, 78: 27-41.
A. Calza, J. Hey, A. Parrini and S. Sauer (2021), “Corporate Loans, Bank’s Internal Risk Estimates
and Central Bank Collateral: Evidence from the Euro Area”, ECB Working Paper Series No. 2579.
M. Carey (2002), “Some Evidence on the Consistency of Banks’ Internal Credit Ratings”, in Credit
Ratings: Methodologies, Rationale and Default Risk, Michael K. Ong (Ed.), 449?467, London: Risk
Books.
G. Cascarino, R. Gallo, F. Palazzo and E. Sette (2022), “Public guarantees and credit additionality
during the Covid-19 pandemic”, Working Papers No. 1369, Bank of Italy.
F. Coppens, F. González and G. Winkler (2007), “The Performance of Credit Rating Systems in the
assessment of Collateral Used in Eurosystem Monetary Policy Operations”, ECB Occasional Paper
No. 65.
D. Cucinelli, M. L., Di Battista, M. Marchesi and L. Nieri (2018), “Credit Risk in European Banks: the
Bright Side of the Internal Ratings Based Approach”, Journal of Banking and Finance, Vol. 93, issue
C, 213-229.
EBA (2017), Guidelines on PD estimation, LGD estimation and the treatment of defaulted exposures,
November.
ECB (2015), “The financial risk management of the Eurosystem’s monetary policy operations”, July.
ECB (2024), Guide to Internal Models, February.
B. Engelmann (2006) “Measures of a Rating’s Discriminative Power — Applications and Limitations”,
The Basel II Risk Parameters, 263-287.
S. Firestone and M. Rezende (2016), “Are Banks’ Internal Risk Parameters Consistent? Evidence from
Syndicated Loans”, Journal of Financial Services Research, 50:211-242.
F. Giovannelli, A. Iannamorelli, A. Levy, M. Orlandi (2020), “The in-house credit assessment system
of Banca d’Italia”, Occasional Paper Series, No 586, Bank of Italy, November.
K. Hornik, R. Jankowitsch, M. Lingo, S. Pichler and G. Winkler (2006), “Validation of Credit Rating
Systems Using Multi-Rater Information. Available at SSRN: https://ssrn.com/abstract=871213 or
http://dx.doi.org/10.2139/ssrn.871213
T. Jacobson, J. Lindé and K. Roszbach (2006), “Internal Ratings Systems, Implied Credit Risk and the
Consistency of Banks’ Risk Classification Policies,” Journal of Banking and Finance, Vol. 30(7), 1899?
1926.
S. Narizzano, M. Orlandi and A. Scalia (2024), “The statistical credit assessment model for nonfinancial firms of the Bank of Italy”, Markets, Infrastructures, Payment Systems, No 53, Bank of Italy,
October.
Tables and Figures
Table 1
Mapping between Eurosystem Credit Quality Steps (CQSs) and probabilities of default (PDs) over
one-year horizon
?0.1%
?0.4%
?1%
?1.5%
?3%
?5%
ECAF; individual
Employed for
loans and pools of
individual loans
and pools of ACCs
pools of ACCs
Table 2
Distribution of firms rated by IRBs and BoI ICAS
(2015-2023)
Total
No. of
No. of firms
firms
rated by at rated by
least an IRB BoI ICAS
776 480
738 471
725 815
722 134
792 090
803 290
832 210
721 533
695 785
1 192
2 057
2 357
2 374
2 444
2 850
2 991
2 925
3 036
22 226
No. of firms rated by BoI ICAS Full and by at
No. of firms rated by BoI ICAS Stat and by at
No. of
least 1 IRB
least 1 IRB
With mobilised credit
firms
claims
With mobilised credit claims
rated by
Total
Total
BoI ICAS
assessed
assessed
Total
Total
single
single
ECAF claims
pools
claims
pools
256 452
4 616
278 813
8 649
3 988
2 464
2 197
275 096
7 935
1 658
7 837
4 563
10 086
290 300
9 386
2 628
1 306
4 382
25 307
284 526
10 463
3 376
1 253
1 626
2 792
29 183
285 349
12 103
4 247
1 611
2 032
3 400
37 799
268 943
14 239
4 571
1 714
2 234
3 055
34 949
273 100
14 634
5 506
1 655
3 134
4 474
48 361
309 889
12 849
4 400
1 328
2 490
3 411
58 850
334 370
12 568
3 832
1 080
2 321
9 547
3 087
49 842
98 793
30 886
11 173
4 607
Note: firms rated by more than one IRB are considered as different debtors, i.e. if a firm is rated by two different IRBs and
by the BoI ICASs it is considered twice in the sample of debtors rated in common (this explains why the figures in the
fourth column exceed those in the second column).
Table 3a
Estimation of equation (1) – Full ratings
Dependent Var:
Year:
??
??
-0.021
(0.018)
-0.055 ***
(0.015)
-0.271 ***
(0.014)
-0.149 ***
(0.012)
-0.181 ***
(0.011)
-0.218 ***
(0.010)
-0.391 ***
(0.010)
-0.341 ***
(0.010)
-0.275 ***
(0.010)
??
??
0.007 ***
(0.021)
0.045 ***
(0.018)
-0.178 ***
(0.017)
-0.107 ***
(0.016)
-0.180 ***
(0.015)
-0.253 ***
(0.013)
-0.404 ***
(0.013)
-0.396 ***
(0.013)
-0.322 ***
(0.010)
dummy_pledged
??
??
0.068 *
(0.021)
-0.124 ***
(0.021)
-0.032 ***
(0.022)
-0.095 ***
(0.021)
-0.179 ***
(0.021)
-0.307 ***
(0.021)
-0.307 ***
(0.021)
-0.244 ***
(0.021)
-0.396 ***
(0.008)
??
??
0.044 *
(0.019)
-0.163 ***
(0.019)
-0.083 ***
(0.020)
-0.115 ***
(0.019)
-0.195 ***
(0.019)
-0.333 ***
(0.019)
-0.333 ***
(0.020)
-0.282 ***
(0.020)
-0.231 ***
(0.007)
Fixed – Effects —————–Firm FE
Bank FE
————-Yes
————-Yes
————-Yes
___________
Observations
90,015
R-sqr
R-sqr_adj
*** p