At the heart of credit risk measurement and management is a fear: what has
gone shall not return. The process of risk management, therefore, becomes a series
of mechanisms to pre-empt such an occurrence. The process of mitigation, in
whatever form expressed, starts as early as the disbursement stage and continues
till the money comes back to the lender—along with the interest.
The business of lending for single, large ticket transactions (usually referred to as
‘wholesale lending’) is structurally different from large volume homogeneous
lending (usually referred to as ‘retail lending’). It is important to understand the
distinction since the modelling structures applied for risk measurement and
monitoring of exposures take different forms to accommodate the differences.
The following peculiarities differentiate the management of retail credit risk visà-vis wholesale credit risk.
 The vehicles for retail lending are products that are homogeneous in nature
while wholesale lending tends to have products tailor-made for custom
purposes. The homogeneity of retail business products makes it imperative
that demographic segmentation is accurate to achieve construction of
portfolios which, in turn, are homogeneous.
The customer takes a front seat in the understanding of retail credit risk,
while product characteristics drive the understanding of the embedded credit
risk in wholesale lending. Thus, capturing customer information (and possibly
superimposing it with localized economic indicators) at every step of a
relationship and integrating such information become the keys to effective
credit risk management in the retail lending business.
The “explanatory variable set” – and its myriad combinations – which
explains risk at different phases of the credit lifecycle is much more for retail
credit than it is for wholesale. This is because the demographics associated
with a consumer (or consumer group) are many. This has created enormous
scope for quantitative modellers and, as a result, the application of quantitative techniques in the measurement of retail credit risk has increased
manifold. Unlike wholesale credit risk, where the concentration of the
monitoring activity lies primarily in the account maintenance function, retail
credit risk management and monitoring cuts across the entire credit lifecycle.
Table 1 represents the retail credit lifecycle and the different activities that are
part of each phase.
| CREDIT SCORING |
CREDIT QUALITY MONITOR |
CREDIT COLLECTIONS |
| LOSS FORECASTING |
| External Internal |
FORTFOLIO ANALYSIS |
CREDIT RECOVERIES |
Origination Acct Maintenance Collection and Recovery
The predominance of credit bureau information on retail credit scores has
converted credit risk measurement at origination from a specialized function to a
‘commoditized’ task. The emergence of specialized organisations using tested odels
to generate scores like FICO have found wide acceptance. However, prudent
lenders tend to blend in-house rating scores with the external ratings to arrive at
scores that are finally used in the credit approval process. Such techniques become
important especially under the following situations.
The lender, from its modelling of change of state or otherwise, has reasons to
believe that the current demographic characteristics of the customer has a
predictive risk profile (that is, a forecast of how such demographic accounts
have behaved post credit disbursement) that merits adjustments to the current
score.
‘Look-back’ analyses demonstrate that the scores given by external agencies
were not representative of current status. Lenders regularly apply such
techniques to gauge the efficacy of scoring models.
While the benefits of blending internal with external scores cannot be understated,
the challenge lies in the ability to provide accurate and consistent demographic
information to the scoring models. This has to be systematic in order to understand
the changing nature of the demographic segmentation input and interpret the
credit scores accordingly.
Monitoring the health of an account is the main task of the Account Maintenance
function at the lender institution. There is no one technique of measuring ongoing
quality of credit but prudent lenders choose from multiple representation of the
information to estimate the health of portfolios. This task is more onerous than it
seems, largely because of the volume of the retail lending business. Lenders tend
to create and monitor the health of homogenous portfolios rather than that of
individual accounts. This poses a challenge in terms of drilling through from
aggregated data (portfolios) to granular data (accounts) seamlessly.
The following are popular methods of monitoring credit quality:
Delinquency-based analysis (both coincidental and lagged delinquencies)
State Transition matrices (both volume and accounts)
Credit Score Transition matrices
Loan-to-value analysis
The emerging solution has to provide a choice of models, possessing a proven
power for explaining the dependant state.
Risk assessment techniques and methodology, as applied to retail portfolios, are
quite different from those applied to corporate portfolios mainly because of the
non-efficacy of a case-by-case judgmental effort and the unavailability of a rich
default history to base statistical models of potential loss.
Banks, especially those with advanced systems and techniques, tend to adopt
most loss concepts (Probability of Default, Expected Loss and Delinquency), and
a computation of Expected Loss (EL) based on Probability of Default (PD), Loss
Given Default (LGD) and Exposure at Default (EAD) is clearly the norm amongst
sophisticated banks. Table 2, excerpted from a study by ISDA, shows the relative
popularity of different loss forecasting methods employed by a sample of
internationally diversified banks with significant retail exposures.
 |
Note: Bank names have been masked to retain anonymity
From the table, it is evident that Historical Loss analysis still remains the most
popular method of loss forecasting. This poses a challenge to data management.
Data management implies ensuring appropriate information at demanded
granularity over user-chosen time periods, pre-modelled for timely and accurate
analysis.
In addition to the techniques mentioned, Delinquency Flow models and Segmented
Vintage analysis are now commonly used to identify portfolio dynamics and
behaviour patterns. A large measure of the credit for ushering in this sophistication must lie with credit card companies with their massive segmentation profiles
and advanced analytical models.
The last logical step in loss forecasting, and certainly not least in terms of
importance, is the validation of the loss forecasting model. In the case of retail
credit risk, this process is relatively simple as the loss behaviour is modelled with
internal historical information, making ongoing validation and calibration of
models that much easier. Several statistical techniques are popular as validation/
calibration models. For example, scorecard performance is monitored against
predicted performance to enable review of the choice of parameters by means of
well-known statistical techniques such as Discriminant Analysis.
Until recently, lenders and regulators have tended to push for loss modelling of
commercial and industrial portfolios, with relatively little focus on retail credit
risk modelling. The over-emphasis on modelling idiosyncratic risks of a large
commercial portfolio possibly made more economic sense. However, over the last
decade, a push from regulators and the dominance of the retail business, as a
consequence of lower industrial activity, has seen hectic activity in the retail
portfoliomodelling scene. Based on their levels of sophistication, lenders follow
either of the two methods for credit portfolio modelling.
Being a simple model, this is popular as well. In this method, portfolio loss
standard deviation is computed either based on historical distribution or based
on standalone computation of PD and LGD (most lenders do not complicate
computations by introducing correlations between the two). Subsequently, a
multiplicative factor is chosen to represent a confidence interval and the resultant
is recognized as the economic capital.
Some lenders do adopt techniques that use fixed form distributions for default
events (e.g.,beta or gamma distributions) and deriving loss distributions
considering LGD as a separate variable. From this distribution, the loss
corresponding to a given percentile is extracted.
Though not a necessity for lenders adopting this approach, calibrating models
for greater accuracy demands that initiatives are undertaken to develop
independent causal models for default and LGD rates.
Lenders at the highest grade of sophistication employ causal modelling that
invariably tends to take the form of factor models. These models allow lenders to
anticipate changes in default rates and their distributions due to macro-economic
changes, and model default correlation across portfolio segments.
Similar causal relationships can be developed for LGD. For example, a codependency
between LGD and default probabilities may be introduced by
modelling LGD using risk drivers identical to those (or at least some of those)
retained for default rates. This is a particularly desirable feature for those classes
of assets where a general dowturn in the economy affects both default and loss
rates (for instance, via loan to value ratios).
Validating deployed models against the actual (back-testing) and subjecting the
model assumptions to extreme situations (stress-testing) go a long way in
understanding out-of-tail situations and model efficacy. In fact, allowing
subsequent state-representation analysis by stressing the factors could be a big
incentive for lenders to develop factor models. Validating retail portfolios as
compared to corporate lending is simpler because default events are more frequent
in retail portfolios, portfolios are more homogenous and stable through time, and
longer loss time series are available as data is predominantly internal.
The Federal Reserve Board of Philadelphia, in its recent initiative to foster greater
sophistication in retail credit risk monitoring observes: ‘…the future of consumer
credit risk management lies in organizing portfolio performances and account
level details into databases and then applying refined analytical models to discern
pattern or trends.’
Herding is seen as a common phenomenon amongst retail lenders—that is, the
propensity to follow the herd in pricing of loans. However, information computed in some of the credit lifecycles described earlier, especially in the area of loss
forecasting and credit portfolio modelling, has the potential to provide strategic
inputs to pricing. Economic capital understanding for portfolios assumes
significant importance in order to be able to deliver superior risk adjusted pricing
to different segments within portfolios or sub-portfolios. Such demographic
segmentation can only be achieved if the information repository is deep and rich,
and able to support multiple segmentation models. For an information ecosystem,
the challenge lies in the possibility of analytical applications ‘closing the loop’
with the origination information systems so that credit cycle end-state economic
insights are pumped back towards origination (Figure 1). The success of such a
‘virtuous cycle’ would, in-turn, result in superior risk-segmented pricing.
 |
Profit motives and regulatory directives (notably Basel II, which provides incentives
by way of lower capital set-asides for adopting superior risk management
techniques) have and shall continue to drive best practices in credit risk
management for retail lending. Advances in technology provide the necessary
impetus for these models to bridge the gap between academic articulation and
actual deployment in production environments.
was formerly the Head - Functional Solutions Expert
Group, Reveleus, i-flex Group of Companies, India
|
|
