CCAR and stress-testing segmentation insights: Account-level risk segmentation and correlations to macroeconomic factors November 16, 2015 Share Subscribe RFP Regulatory scrutiny has significantly increased following the financial crisis. The previous lack of attention to low-probability, high-impact events that could strain a firm’s capital adequacy and spread to other financial institutions prompted banks to develop complex models at the lowest level of granularity to capture the impact of economic cycles. Dodd-Frank Act Stress Testing (DFAST) requires bank holding companies (BHCs) with assets over $10 billion to run supervisory stress-testing scenarios to provide a forward-looking assessment of the capital adequacy under different economic cycles. Comprehensive Capital Analysis and Review (CCAR) is a broader supervisory program that includes supervisory stress testing but also assesses a BHC’s unique stress-testing scenarios. The purpose of CCAR is to assess whether the largest BHC’s operating in the United States have sufficient capital to continue operations during times of economic and financial stress and to ensure that they have robust, forward-looking capital-planning processes that can account for their unique risks. CCAR regulatory guidance requires financial institutions to establish a quantitative basis for enterprise-wide scenario analysis. However, the guidance does not prescribe the use of particular quantitative approaches to estimate losses, revenues and expenses across the industry, leaving BHCs to select the approach that best fits their type of portfolio and level of complexity. On the other hand, CCAR guidance provides clear expectations and a common ground for how losses, revenues and expenses are calculated. The guidance states: “While the Federal Reserve does not require BHCs to use a specific estimation method, there is the expectation that each BHC should estimate its losses, revenues and expenses at sufficient level of granularity so that it can identify common, key risk drivers and capture the effect of changing conditions and environments.1” Segmentation plays a critical role in determining the right level of granularity for stress modeling. Practically speaking, the objective of segmentation is to define a set of sub-populations that, when modeled individually and then combined, rank risk more effectively than a single model tested on the overall population. Stress-testing modelers understand that finding the right level of granularity to segment the bank’s portfolio represents one of the first steps in establishing a quantitative basis for the enterprisewide scenario analysis of stress testing. What is segmentation? Segmentation was originally a marketing strategy used to aggregate prospective buyers into groups — or segments — that have or are perceived to have common needs, interests, priorities, or buying behaviors, and designing and implementing strategies to target them. In other words, they are homogeneous. Market segmentation enables companies to target categories of consumers who perceive the value of certain products and services differently from one another. This allows organizations to design customized marketing campaigns and customer treatments. Segmentation approaches have also been successfully applied in many different fields including economics, business, finance and risk management. Granular segmentation and its application to stress tests Granular segmentation is a key requirement for meeting the expectation that BHCs are able to identify risk and forecast change under different economic scenarios. To accomplish this, modelers will need to identify relevant homogeneous segments or sub-portfolios that are sensitive to similar risk drivers and are consistently correlated to comparable macroeconomic factors over time. When effectively selected, these smaller segments, or cohorts, should be sensitive to material changes in risk exposure or portfolio characteristics over the planning horizon. Good segmentation yields greater modeling granularity and captures risk drivers with more precision. Segmentation provides the foundation for banks to develop models that meet regulatory expectations to estimate revenues and expenses at a sufficient level of granularity so they can identify common risk drivers and capture the effects of dynamic economic conditions and environments. Models used for stress-testing purposes should account for portfolio changes under different credit cycles by considering correlations with macroeconomic variables. An example of using the right level of aggregation to determine correlation is the application of loss-estimation methodologies to retail and commercial portfolios. Typically, bank’s retail portfolio is composed of credit cards, personal loans, home equity lines of credit and consumer mortgages. When different sub-portfolios are aggregated together, the assumption is that they will create a homogeneous segment, while in reality each sub-portfolio exhibits a different probability of default (PD). As a result, the combined retail portfolio will show correlations to macroeconomic variables that are less intuitive than when each sub-portfolio is modeled separately. Granularity is also a requirement for a bank’s commercial portfolio. For example, commercial real estate — one of the largest material components of a commercial portfolio — should use granular segmentation schemes that capture specific risk drivers, such as geography and property type, and risk measures, such as loan-to-value and coverage ratios, debt service and net operating income. The correlations with macroeconomic variables are stronger at the granular level, where the sub-portfolios exhibit similar loss rates under the same credit cycle. Banks often develop their models at a more aggregated level and thus get less reliable results. These models typically represent the first generation of stress-testing models, with weaker correlations to macroeconomic variables. Multifactor approach to granular segmentation An effective approach to identifying homogeneous sub-segments with a higher level of predictability is to use multifactor segmentation. This method differentiates risk better than the traditional segmentation scheme using a single risk characteristic. Multifactor segmentation approaches are based on a combination of quantitative and qualitative characteristics. For example, once a risk score is assigned to an account (e.g., PD), a multifactor segmentation methodology will first group accounts based on similar risk scoring and then apply a second level of segmentation using qualitative characteristics such as geography, industry, products, past due date, etc. Granular segmentation challenges BHCs adopt multifactor segmentation approaches to forecast losses, revenues and expenses that correspond with the size and diversity of their portfolios. In doing this, institutions of different sizes confront different technical challenges. The development of a dynamic multifactor segmentation approach involves data-intensive modeling work that starts with scoring the lowest level of granularity (e.g., account level). However, for most institutions, developing account-level models is a difficult undertaking for three reasons: (1) a lack of access to granular account-level data, (2) the need to use sophisticated modeling software and data processing technology, and (3) the need for more advanced modeling techniques. Typically, CCAR BHCs — with heterogeneous customer bases and product portfolios — have many potential segment combinations. This makes it very difficult to identify the right number of segments to account for the heterogeneity of their portfolio. While most of the large CCAR banks have achieved a high level of sophistication in their modeling approach, some have developed complex quantitative models using a large number of predictive parameters without sufficient data to support the level of granularity of the modeling framework. On the other hand, the vast majority of DFAST banks tend to identify a limited number of segments. Many of these institutions use top-down methodologies to identify the correlation with macroeconomic factors at the portfolio level. Even those DFAST institutions that use a segment-level approach tend to use segments that are not granular enough to identify portfolio changes. These organizations find themselves managing at the portfolio level as they are unable to identify the relevant cohorts that drive the behavior of the entire portfolio. Thus, many CCAR and DFAST banks struggle with determining the right level of segmentation to achieve sufficient granularity in their models. To determine the right level of segmentation, banks should consider: The ideal level of segmentation is achieved when the strength of the correlation of a sub-segment with the macroeconomic variables is strong enough to drive results for the entire portfolio. The level of granularity of the segmentation is determined by the granularity of the data. Multifactor segmentation for a credit card portfolio Many CCAR institutions that have adopted advanced account-level modeling face different segmentation challenges because they need to aggregate accounts into relevant, homogenous segments. Multiple statistical techniques are available for segmenting a population. Traditional techniques rely on using a single characteristic to define a segment. However, the use of multiple characteristics yields more predictive models. For example, credit card companies segment their portfolios based on vintages, paying behaviors (i.e., transactors vs. revolvers) and type of credit card product. Leading credit card companies have developed account-level models at the lowest level of granularity by applying the following segmentation design as the foundation of their modeling process. Model PD at the account level using competing hazard survival models.2 Examples of hazards include contractual at 180 days, bankruptcy, deceased and attrition. Aggregate accounts into relevant segments based on common PD scores or a similar type of risk score. Apply a second segmentation logic approach using qualitative factors such as ZIP code, age and education. Apply other aggregation criteria, such as buying behaviors or vintages. Finding the right number of relevant segments is an art and requires trial and error combined with sophisticated modeling and data mining techniques. Each segment’s level of granularity will depend on the business objective. This hybrid approach can be illustrated using the example of a hypothetical credit card portfolio. Conclusion — takeaways Granular segmentation is a key element for meeting the CCAR/DFAST regulations that require banks to identify relevant segments or sub-portfolios that are sensitive to similar risk drivers and are consistently correlated to similar macroeconomic factors over time. Many CCAR and DFAST banks struggle with how to determine the right level of segmentation to enable sufficient granularity for their models. For most institutions, developing account-level models is a difficult undertaking that is negatively affected by three main factors: – The lack of access to granular account level data – The need to use sophisticated modeling software and data processing technology – The need for more advanced modeling techniques Multifactor segmentation approaches based on a combination of both qualitative and quantitative risk characteristics produce better segmentation and a higher correlation to risk drivers. To determine the right level of segmentation, banks should consider: – The right level of segmentation is achieved when the correlation between a sub-segment and the macroeconomic variables is strong enough to drive the results for the entire portfolio. – The level of granularity of the segmentation is determined by the data’s granularity. The adoption of more granular segmentation carries a set of implications to multiple areas including: – Data management – Reporting – Model validation – Audit traceability – Scenario building Segmentation and the Pareto rule The Pareto rule, also known as the 80/20 rule, can be applied in risk management for portfolio management, where 20% of the cohorts usually generate the highest concentration of risk exposure under stress. The identification of relevant sub-populations that are sensitive to macroeconomic factors — and subsequently have a higher propensity to generate exposure — becomes one of the most critical tasks of the risk identification process. Many questions need to be answered: What sub-segments/sub-portfolios are driving the largest levels of exposure in the portfolio? What sub-segments are more sensitive to macroeconomic factors under different economic cycles? Banks should take a closer look at their segmentation approaches. The Pareto rule applies to most portfolios where the correlation between certain sub-segments and macroeconomic variables is strong enough to drive the correlations for the entire portfolio. Identifying these smaller sub-populations can have a major impact, as management actions can become more effective by targeting the sub-segment with appropriate pricing actions.How Grant Thornton can help Grant Thornton’s Model Risk Management Center of Excellence has extensive industry experience in model development, implementation, validation and use. We have professionals with business and risk/regulatory acumen and advanced quantitative modeling knowledge to address the risks that arise from model use. We can also provide deep insights into analytical solutions pertaining to risk modeling (credit, market, operational, AML, etc) and CCAR, DFAST, Basel I, II, and III, Solvency II. Download the PDF. Contacts Ilieva Ageenko Managing Director Financial Services Advisory Leader, Model Risk Management T +1 704 632 6820 Paul Makowski Managing Director Financial Services Advisory T (mobile) +1 240 463 0550 Nigel Smith National Leader Financial Services Advisory T +1 212 542 9920 1 Board of Governors of the Federal Reserve System. Capital Planning at Large Bank Holding Companies: Supervisory Expectations and Range of Current Practice, p. 19, August 2013. 2 Hosmer, David W.; Lemeshow, Stanley; and May, Susanne. Applied Survival Analysis: Regression Modeling of Time to Event Data, 2nd edition, John Wiley & Sons Inc., March 2008.