Preparing for Basel II modeling requirements; Part 1: model development

RMA Journal, The, May, 2003 by Jeffrey S. Morrison

Gain the benefit of a regional bank's experience. In this first of four articles, Jeff Morrison discusses SunTrust Banks' approach to statistical modeling. Part II details steps taken to validate the model. Part III pulls it all together within a GUI software interface. Then Part IV moves into the realm of stress testing.

The Basel II Capital Accord, currently planned for implementation in 2007, sets out detailed analytic requirements for risk assessment that will be based on data collected by banks throughout the life cycle of the loan. The purpose of Basel II is to introduce a more risk-sensitive capital framework with incentives for good risk management practices. Many banks are examining or implementing models now to help enhance their risk management efforts. And it can get pretty confusing.

Models

Remember that old statistics book in college and what you said about it? "I'll never use that stuff in the real world!" Well, never say "never." That old book and this article can serve as a refresher.

Let's start by defining the word model. Webster's more statistical definition of the word is "... a system of postulates, data, and inferences presented as a mathematical description of an entity or state of affairs." Basically, think of a model as a mathematical representation of reality. It's not going to be perfect and will definitely be oversimplified, but the aim of such a representation is to gain insight into behavior so predictions can be made that are both reasonably accurate and directionally correct.

Quantitative models in consumer credit have been used for many years. Models developed from the application data on new accounts are called front-end or application models. These models do not use the prospective lender's payment history information for a potential new borrower because that information is simply not available. Once these accounts begin to become seasoned, different models can be developed to yield behavioral scores, that is, algorithms designed to include payment history as well as other factors associated with loan origination, geography, and the demographics of the borrower. In contrast, scores developed from pools of data typically obtained from credit bureaus are called generic models. These models reflect credit behavior across a variety of financial institutions and capitalize on the assumption that a consumer will exhibit behavior around some average risk level. Customized scores developed with the payment history of a single institution can often outperform generic models because they are tailored to the specific credit issuer.

Models for Easel. Similar models may be developed for Base!. The models used in SunTrust's Risk Rating System have been built specifically for Basel II on a two-dimensional structure. The first dimension reflects the probability of default (PD) for the obligor. The second reflects the loss given default (LGD) associated with a particular loan or facility. Therefore, for each loan, the expected dollar loss is simply the product of the dollar Exposure at Default X PD X LGD.

Let's begin by looking at developing a PD model for the obligor and then move toward developing a facility-based model for LGD. We can construct these types of models for the commercial side of the business, but to make it simpler, think in terms of retail portfolios, such as residential mortgage, as you read further.

Typically, bank models for Easel requirements come in two flavors--vendor and custom. In the commercial world, models may have to come from vendors because only they have invested the resources to collect data robust enough for modeling. This is because the number of commercial defaults for any single bank in a given year is so small. Based on the sheer size of the loan volume, the retail side is just much riper for custom modeling, where a bank can use its own data and not rely on costly vendors. Even if a bank does not yet have enough historical data to develop a statistical model, it can begin with one derived from judgment and consensus until the more sophisticated models are available.

Judgmental models are simply a set of rules that quantify assumptions about the portfolio's risk level without the use of statistical approaches. Examples might include a mapping of risk grades according to loan-to-value or debt-to-income ratios. Others might provide a rough mapping of FICO score bands to PD. Although judgmental models definitely have their place, the remainder of this article will focus on the development of statistical models that are reflected in both custom and vendor efforts. And because Easel requires all loans to be rated with these models for a certain minimum amount of time before the advanced approach may be used, integrating vendor and custom solutions into the process should begin as soon as possible.

The current school of thought surrounding the models mentioned in Easel is that banks should have separate models for the obligor and the facility. The obligor model should predict PD--usually defined as 90-plus days delinquent, or in foreclosure, bankruptcy, charge-off, repossession, or restructuring. Models on the facility side should predict LGD, or 1 minus the recovery rate. The recovery rate is simply the amount of dollars recovered divided by the dollars owed at the time of default.