Many Types of Models
Are Used in Bank
The models discussed in this article
have compliance risk implications,
and their governance falls under the
auspices of risk management and/
or the compliance function. For example, the compliance officer may
be the assigned owner of a bank’s fair
lending model used to identify exposures due to credit underwriting
and pricing decisions, a CRA data
assessment system used to evaluate
lending patterns, or models used to
identify BSA/AML suspicious activity. Credit scoring or complaint
analysis systems are other models
that may be owned by other functional areas but warrant compliance
risk management oversight. And, an
example of models helping compliance officers deal with the growing
regulatory burden is the coming
analysis of the expanded HMDA
data. These models may be developed internally or purchased from
an external vendor depending on the
preference of the institution.
Trends in Model Usage
In the past 25 years, we have witnessed an analytical explosion and
a concomitant growth in modeling.
In fair lending, the initial modeling
effort involved differential treatment with respect to mortgage applicant race. Then, modeling expanded to include other prohibited
basis groups such as gender, age, and marital status. Additional
expansion included other credit products such as consumer secured
and unsecured loans, HELOCs, direct and indirect autos, small
business, and credit cards. Furthermore, the analytics expanded
from the credit and pricing decisions into other types of risk
such as redlining. We now witness expansion beyond banks to
mortgage companies and the like. Moreover, the recent amendments to HMDA that expanded the scope of data fields suggest a
heightened analytical capacity for models using such data.
Outside of the fair lending explosion, BSA/AML monitoring has
evolved from a largely manual process into a significant modeling
effort that is well beyond point-and-click processes. The urgency
of detecting terrorism financing post-9/11 certainly provided an
impetus for the increased use of automated monitoring processes.
Analysis now may involve machine learning or similar approaches,
especially among service providers.
As a direct result of the development of the Consumer Financial
Protection Bureau’s (Bureau’s) database, a new area of modeling
is “complaints”. At the start, modeling was applied on an ad hoc
basis as a way to discover causes for observed results. More re-
cently, the manual reviews of the past have been supplemented
with basic analytics employing big data using machine learning
or similar techniques. Big data can be used to supplement tradi-
tional statistical analysis to detect patterns not otherwise seen in
the data. A particular area of interest is unstructured data (which
many analysts have estimated comprises 80% of all data), such
as raw text and voice recordings. Big data can be implemented
using machine learning techniques which are available from sev-
eral external vendors using your institution’s data and perhaps
external data. You may also have an enterprising individual in
your institution that has the ability to create models internally.
If so, the experienced compliance officer knows to connect as
soon as possible with that individual to help shape the approach,
understand the risks, and avoid the pitfalls. Let’s consider a few.
Practical Considerations in Model Usage
While there are many benefits to using models, there are also
some drawbacks or risks. The benefits are:
■ ■ ■ Ability to review all applications/transactions. In fair
lending analysis prior to using models, the applications to be
reviewed were typically chosen by sampling. With regression
modeling, all applications are reviewed as part of the analysis.
■ ■ ■ Quick identification of outliers. As a result of the all-inclusive modeling effort, outlier applicants are quickly identified
as part of the analysis output.
■■ ■ Easy identification of matching applications/
transaction. Not only are outliers identified, but given a set
of matching criteria, matching applications for those outliers can
be identified, when they exist. Thus, the output of matched pairs
in the past are now matches to all similarly situated applicants,
if they exist.
■ ■ ■ Increased precision in matching. The operating protocol
for models minimizes the judgment associated with some of
the old matched pair analyses and permits matching across
■ ■ ■ Efficiency and cost control. Compliance officers (not to
mention management) like this process since it is efficient (i.e.
cost effective) compared to other, more manual analyses and
it establishes a reproducible process that is not dependent on
the availability of a staff person running the process.
■■ ■ Identification and explanation of factors highly
correlated with the target of the analysis. It can
identify what key factors highly correlate with credit and pricing
decisions, SAR decisions, money laundering and complaint
■ ■ ■ Early warning control. Complaint analysis can give the
institution an early warning on potentially unfair, deceptive,
or abusive acts and practices.