Probability of Default: A Modern Calibration Approach: Stefano Bonini and Giuliana Caivano

Probability of Default: A Modern Calibration
Approach
Stefano Bonini and Giuliana Caivano
Abstract An extensive academic and practitioner’s literature exists on rating mod-

els development with well-structured statistical methods, however these models do
not estimate PDs aligned with the economic scenario, then it is necessary a calibra-
tion. During the last years the effect of not well calibrated models has been observed
on the credit market: actually they show a high level of procyclicality that let them
loss market credibility and banking usability. The aim of this paper is to show a
modern structured calibration approach, based on Bayesian techniques, taking into
consideration specific economic factors. The calibration approach has been applied
on real data of a Corporate portfolio of a top tier European Bank and a new calibra-
tion test, adjusted by the economic cycle, has been performed.
Keywords Rating models · Credit risk modeling · Bayesian econometric

methods · Economic cycle
1 Introduction
Today, even because of the financial crises, banks need more and more reliable and
usable risk management tools. Moreover within Basel2 and Basel 3 Accord the es-
timation of Probability of Default (PD) plays a key role for an efficient allocation of
capital, pricing, client sanctioning, credit monitoring, and finally regulatory compli-
ance.
A typical feature of PD models across countries is that they are often based on
individual characteristics of clients or they use some information related to the type
of specific products, but no information are commonly used for taking into account
macroeconomic variables.
Typically the only way to align rating models with the economic scenario (as in
[1] and [3]) is to apply a sort of “addendum” to the model itself: that’s what is com-
S. Bonini (B) · G. Caivano

University of Rome “Tor Vergata”, via Columbia 2, 00133 Rome, Italy
e-mail: [email protected]
G. Caivano
e-mail: [email protected]
C. Perna, M. Sibillo (eds.), Mathematical and Statistical Methods for Actuarial Sciences 41
and Finance, DOI 10.1007/978-3-319-05014-0_9,
© Springer International Publishing Switzerland 2014
42 S. Bonini and G. Caivano
monly known as calibration. When speaking about calibration, it is a best practice

among banks to refer to Through-the-Cycle (TTC) PDs as forecasts aligned with
the average long run historical default rate typically used for capital requirement
calculation. In this context the use of models that reflect the average long run his-
torical default rate is required in order to stabilize capital ratios and optimize capital
management policies: in positive economic conditions a run up of capital should be
kept for negative economic phases. Given the relevance of final PDs in reflecting the
economic conditions this paper proposes a modern structured calibration approach
using Bayesian techniques for the estimation of average long run historical default
rate (so called Central Tendency, CT). The added value of this work is related to
the relevance itself of the topic in the last year: given the current context of crisis,
it is very important for banks to forecast PDs that are able to ensure stable capital
ratios. The methodology here proposed is easy to understand and also applicable to
each kind of portfolio, as demonstrated by the application on real data of a Corpo-
rate portfolio of a top tier European Bank. In addition, the paper proposes a new
binomial calibration test using PDs adjusted by the economic cycle.
2 Calibration Framework: Model Development and Validation

Tests
As above mentioned, the goal of calibration is to assign implicit PD’s to each grade
or score defined by the rating model: in this way it is possible to define a consistent
long-term PD based on the underlying scores defined by the statistical model.
In this context the Central Tendency has a key role, that is—among researchers
and practitioners—quite often defined as the average of historical defaults. Here
a new methodology is presented, in which the impact of a cyclical downturn in the
future is embedded. For this scope, it is quite common [4] the use of macroeconomic
variables such as GDP and their correlation with corporate defaults of each country
in order to define a cut-off between positive and negative cycles [2]. Here we then
define the correlation between quarterly Italian GDP, Italian corporate default rates,
European corporate default rate and bank corporate portfolio default rate in order to
forecast the future default and define the true picture of the portfolio in the current
as well as in the future scenarios, even under stress hypotheses.
We here define the calibration function by solving an optimization problem with
one objective function subject to the constraint that Central Tendency has to be equal
to the implicit probability of default derived from calibration, as in (1):

#rating
PDEstimated
t × %Popi = CT (1)
i=1
where %Popi represents the percentage of observations in the bucket i after propor-
tioning.
Probability of Default: A Modern Calibration Approach 43
The objective (target) calibration function has been defined according to (2):
n
1 − PDEstimated 1 − DRi 2
CF = ln i
− ln (2)
i=1
PDEstimated
i
DRi
where:
• i = score bucket i;
• n = number of buckets;
• RDF i = observed default rate post re proportioning in the bucket i;
• PDestimated
i is the PD calculated as logistic transformation of scores of statistical
model.
According to the Bayesian approach [5] the default rates of each bucket have
been proportioned in order to align them with the Central Tendency as shown in (3).
DRbucket DRCT
sample
Adj.DRbucket = (3)
DRbucket DRCT
sample
+ (1 − DRbucket ) 1−DR
1−CT
sample
A model calibrated according to cyclical effect generate PDs that could be

slightly different from the observed default rates during the most recent period
(e.g. the last year of observation) in this case the standard calibration binomial tests
(based on the null hypothesis H0 that the PD of each rating grade is correctly esti-
mated)1 could “wrongly” fail [6]. The authors propose then a new calibration test
called “cycle adjusted” in which PDs are “adjusted for the economic cycle trend, as
defined in (4):
−1 √
adj N (PD1 ) − AC × N −1 (Y t )
PD1,t = N √ (4)
1 − AC
where:
• PDi = Average PD of each rating class;
• AC = Asset correlation of defaults;
• Y = percentile of empirical GDP variation at time t.
The calibration function has been estimated on 7 years of historical data (2004–
2011) of a Corporate portfolio of a top tier European Bank: the gradient (β) of curve
has been considered as a constant, while the intercept (α) has been changed in order
to ensure that the average PD of the portfolio will be equal to the CT. In particular,
taking into account the function constraint (average PD equal to CT), the final values
of α (−50.02) and β (−115.97) have been found as the values that could minimize
the difference between estimated PD and proportioned default rate, as in (3).
1 k∗
√
= φ −1 (q) nPD(1 − PD) + nPD
where:
q = is the confidence level of test;
PD =: is the theoretical PD of each rating grade;
n = number of observations.
44 S. Bonini and G. Caivano
This function has been applied on a portfolio of Corporate loans existing in 2012,
thus only the value of intercept has been changed in order to align the average PDs
to CT, obtaining a new value of α equal to −32.20.
Finally, an adjusted binomial test has been performed, in which the adjusted PDs
have been used (as defined in (4)) for assessing the goodness of the overall model
after the calibration process.
3 Conclusion
In this paper a new approach of calibration has been proposed for aligning “tra-
ditional” rating models to the economic cycle by avoiding, at the same time, the
pro-cyclical effect of using long run estimates. In particular, a new approach for
the definition of Central Tendency has been proposed by forecasting the historical
default rates on a “long run” period (10 years) to be used for fitting the calibration
curve of the application portfolio represented by Corporate exposures of a top tier
European bank between 2004–2012. The authors have also performed a binomial
test “adjusted” for the state of the economy, in order to avoid the underestimation of
PDs during the last recent years: the results of the test show that the estimated PDs
are conservative.
References
1. Engelmann, B., Porath, D.: Do not forget the economy when estimating default probabilities.
Willmott Mag. (2012)
2. Iqbal, N., Ali, A.: Estimation of Probability of Defaults (PD) for low default portfolios: an
actuarial approach. In: ERM Symposium (2012)
3. Kiff, J., Kisser, M., Schumacher, L.: Rating through-the-cycle: what does the concept imply for
rating stability and accuracy? IMF Working Paper (2013)
4. Konrad, M.P.: The Calibration of Rating Models. Tectum, Marburg (2012)
5. Tasche, D.: The art of PD curve calibration. J. Risk Manag. (2013)
6. van der Burgt, M.: Calibrating low-default portfolios, using the cumulative accuracy profile. J.
Risk Model Valid. (2008)

Probability of Default: A Modern Calibration Approach: Stefano Bonini and Giuliana Caivano

Uploaded by

Copyright:

Available Formats

Probability of Default: A Modern Calibration Approach: Stefano Bonini and Giuliana Caivano

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Probability of Default: A Modern Calibration Approach: Stefano Bonini and Giuliana Caivano

Uploaded by

Copyright:

Available Formats

Probability of Default: A Modern Calibration

Stefano Bonini and Giuliana Caivano

Abstract An extensive academic and practitioner’s literature exists on rating mod-

Keywords Rating models · Credit risk modeling · Bayesian econometric

S. Bonini (B) · G. Caivano

monly known as calibration. When speaking about calibration, it is a best practice

2 Calibration Framework: Model Development and Validation

A model calibrated according to cyclical effect generate PDs that could be

You might also like