skip navigation

Sunday, November 23, 2014

[ – ] Text Size [ + ]  |  Print Page

SRC Insights: Third Quarter 2009

Retail Credit Risk Modeling: A Help or Hindrance to Effective Bank Management?

"...financial innovation is good for the economy, but, as demonstrated in the current crisis, the benefits of innovation are usually understood well before the risks come to light." 1

This quote addresses the complex financial products that played a central role in today's financial crisis, which resulted, in part, from advances in financial engineering. It also clearly echoes the challenge that senior bank managers face in leveraging the information that comes out of analytical models used in banking—models are supposed to illuminate risks before the benefits of key bank practices can be realized. If credit risk models, in particular, are to be useful in practice, the inherent drawbacks from using these models must also be well understood. As the latest credit crisis and recession continues to unfold, the importance of credit risk modeling in financial markets has become increasing clear; it is no longer a "backwater" topic among banking professionals.

In today's global recession, a few banking institutions are experiencing defaults on some assets that are unprecedented and could be categorized as "tail events." Industry practice has led to the adoption of a solvency standard used by rating agencies for historically observed default rates for AA-rated companies. To have the risk profile of an AA-rated bank, a bank must hold an amount of capital that is sufficient to weather all but the three worst of 10,000 possible loss scenarios for a one-year period and remain solvent. Or equivalently, the bank must remain solvent in 9997 out of 10,000 possible loss scenarios. In other words, this equates to a solvency standard of 99.97 percent.

It is understood that most banking models could not predict credit losses at or beyond the 99.97 percent threshold or tail of the loss distribution with such benign historical data. Most banks' loan portfolios, however, are not experiencing losses in the tails during this recession, and what remains important is that modeling credit losses is the starting point for a conversation that will consider all significant factors that affect the collectability of a portfolio at any given point in time.

This article will provide insight into the various types of modeling techniques commonly used in risk management practices, with a special focus on retail credit, and will assess the state of those models within the context of the current recession.

Why Do Banks Use Credit Risk Models?

Banking has long evolved from a relationship-centered business model to more of a market-based aggregator of assets that can distribute credit risks and returns to a global investor base. The key to this evolution has been the advancement in risk measurement tools that are more accurate and better validated. As the business of banking has become more complex, so have both the environment in which banks operate and the level of rigor embedded in credit risk models.

One of the most critical risk modeling functions for banks is estimating credit losses that serve as inputs to the allowance for loan and lease losses (ALLL). The ALLL covers estimated credit losses on individually impaired loans and loans evaluated at a segment level with similar risk characteristics, and it reflects adjustments for relevant qualitative and environmental factors (i.e., economic). In other words, a loan loss model for the ALLL must be conditional on the state of the economy, as it is used to determine estimated losses as of the evaluation date. Credit loss models for the ALLL are not usually statistic-based models like those used for making retail credit decisions (e.g., scoring models). Credit scoring models are generally built as tools to rank order the performance characteristics of the population, rather than to accurately forecast the incidence or the dollar amount of loss. Credit scoring has transformed the retail business by contributing to the dramatic loan growth through automated decision mechanisms. Scoring models are also important in evaluating credit acquisition and account management strategies once an account is booked.

Another critical risk modeling function for banks is capital estimation. Lenders price for their expected losses by incorporating a credit risk component into their pricing models, along with important components, such as yield, cost of funds, fixed costs, etc. In any given year, however, the credit environment may be such that actual losses exceed expectations. These unexpected losses generally arise as a result of changes in economic conditions or policy (e.g., a change in bankruptcy laws). To calculate how much capital is needed to cover unexpected losses, it is helpful to estimate what losses would be in several possible states of the economy. The various loss outcomes in these different states provide a loss distribution that associates various loss levels with probabilities that each loss level will occur. Economic capital is calculated as the difference between the expected loss and a much higher loss amount, i.e., at the solvency threshold noted above, that has only a .03 percent chance of occurring.

Although these modeling activities provide critical results, bank managers must still conduct more forward-looking analyses to better understand loss estimates, revenue, and reserve needs under specific and more adverse macroeconomic conditions. This type of analysis is called stress testing, and what bank managers have learned in the current crisis and recession is that stress test results can help avert potential financial distress if testing is done in enough time to implement risk mitigation strategies (such as raising more capital).

Common Credit Risk Modeling Frameworks

Unlike in wholesale credit modeling, retail loan portfolios are made up of individual small loans, and limited resources are devoted to analyzing the idiosyncratic risk of an individual borrower. To fully utilize economies of scale associated with risk assessment, statistical tools (credit scoring), and account management, retail loans are generally grouped into segments that have homogenous risk characteristics. Every institution will have a slightly different view of its risk segmentation, but credit risk models are commonly applied at the segment level if the data permit, while modeling estimated losses for the ALLL and unexpected losses for allocated economic capital is commonly performed at the portfolio level.2 The following points briefly describe fundamental modeling frameworks for retail portfolios.

Scorecard Models

  • Scorecard model development is primarily used for rank-ordering purposes. Scorecards can include prediction of delinquency, default, bankruptcy, attrition, profitability, and account acquisition, as the data reflect portfolio risk characteristics.
  • Scorecard development requires statistical techniques that include logistic/probit regression, decision tree methods, neural networks, and linear regression.
  • Macroeconomic information is rarely considered in scorecard modeling, but with some adjustments, scorecards could be augmented with economic variables to address causal relationships.

Roll Rate/Markov Chain Models

  • Roll rate models measure the percentage of accounts or dollars that "roll" from one stage of delinquency to the next until the accounts meet contractual default criteria.
  • Individual accounts are not tracked in the model. The stages of delinquency reflect a pool of accounts at the segment or portfolio level.
  • Markov chain models are similar to roll rate models in that they track the transition of a pool of accounts into other stages of delinquency; however, these models can account for all types of transitions. For example, Markov models will not only reflect the average probability that a delinquent account will become further delinquent, but also the probability that a delinquent account will become current in the next period. This allows for bank managers to account for different assumptions around collection trends and attrition.
  • Like scorecard models, Markov and roll rate models are based on portfolio risk characteristics and ignore economic factors. With considerable augmentation of the reference data, roll rate and Markov models could be adjusted to ensure that loss estimates are conditional on different economic conditions.

Vintage Models

  • Vintage models normally segment the portfolio by either year-on-book (YOB) or month-on-book (MOB) that an account is booked on a bank's balance sheet. Once the vintage criterion is determined, the loss performance is tracked over time.
  • Vintage models can be further segmented to reflect more granular levels of risk, such as delinquent/nondelinquent and bankrupt/nonbankrupt populations.
  • Annual loss rates by vintage usually provide fewer data points, so nonparametric smoothing methods (such as weighted averages) are useful for estimation purposes.
  • Assumptions regarding account management strategies and economic conditions can be incorporated into the smoothing algorithms.

Credit Risk Model Performance in the Financial Crisis

Credit risk models were severely hampered by the speed at which financial, economic, and borrower behaviors were changing over the course of the crisis and into the current recession. Retail credit conditions worsened rapidly in 2007, as credit performance trends in credit card, prime mortgages, and home equity lines of credit (HELOCs) became more adverse. At that time, many market participants believed that the observed market turbulence would stay contained within the subprime and near-prime mortgage business. By 2008, a series of escalating events triggered by the failures and near failures of some of the world's largest financial institutions severely eroded confidence in the U.S. financial system, shut down capital markets, and ultimately affected the real economy. The official announcement of the onset of a recession and the freezing of credit markets set the stage for an unprecedented policy response by the U.S. government and the Federal Reserve. As a result, it is difficult to assess credit risk models under these stress conditions.

It is important to note that much of the risk that was mounting in the mortgage market was known by financial institutions, as it was clearly outlined in regulatory guidance and accessible in mainstream publications and research reports. One could argue that if the emerging risks were known, model frameworks and assumptions could have been changed to reflect the heightened risks. Federal regulators issued interagency guidance on subprime lending in March 1999, while expanded supervisory guidance was issued in January 2001. Under this guidance, the regulatory agencies asserted their belief that responsible subprime lending can expand credit access for consumers and offer attractive returns, provided that institutions recognize and manage the unique risks associated with this activity.

The Federal Reserve Bank of Atlanta published a financial update for the third quarter of 2005, noting the inherent risk in a growing subprime mortgage market to holders of securities backed by subprime mortgages. Robert A. Eisenbeis, director of Research at the Atlanta Fed at that time, warned that observers, regulators, and markets did not yet fully understand the risks (of subprime lending and securities backed by subprime mortgages) because the phenomenon was relatively new. The article also noted that at the time, at least 60 percent of the rates on subprime mortgages would reset, beginning in 2006 and continuing through 2014. These early warnings signs went largely unnoticed, as we now fast-forward nearly four years and find ourselves in one of the most severe recessions since the Great Depression.

The chart below best captures how a large, sophisticated banking institution with significant modeling data history and deep risk management expertise demonstrated an inability to accurately estimate near-term losses on its credit card portfolios. In particular, as the unemployment rate began its sharp ascent, the bank formulated loss rate projections that were less adverse due, in part, to assumptions around the correlation between card losses and unemployment. As higher actual unemployment trends became realized, the bank subsequently needed to increase loss projections to reflect deteriorating economic conditions.3 As a result of the bank's inability to accurately capture losses, inadequate reserving has led to greater pressure on earnings and may have exacerbated the downward pressure on equity prices in the midst of the financial crisis.

Net Charge-off Forecast - Credit Card
VIEW LARGER IMAGE

Conclusion

After nearly 20 months into the current recession, the importance of credit risk models to help inform risk management decisions is absolutely clear. In particular, banking firms need to better incorporate economic variables into credit risk models in order to provide bank management greater insight on the direction and accuracy of loss estimates. It's not clear, however, that, had many of the credit risk models discussed above been conditional on the state of economy, it would have had an important impact on estimating the severity of losses experienced by institutions in this recession and helped to mute the severity of the crisis.

It is clear that there is room for enhancement of credit risk models at even the most sophisticated and largest institutions. Enhancement might come in the form of more rigorous statistical models, but not at the cost of predictive accuracy. Benchmarking existing modeling frameworks with alternative or new models would significantly strengthen banks' modeling efforts. Additionally, more work on stress testing and model validation should become standard practice for banks.

Bank examination staff should continue to strive to enhance stress testing and model validation, using actual bank data, in order to support more robust supervisory discussions that might help inform a bank's capital decisions in the future.

  • 1   Collins, Michael E., "Restoring Confidence in the Banking System," SRC Insights, Second Quarter 2009, Vol. 13, Issue 4
  • 2   For large complex banking institutions that are mandatory Basel II institutions, the advanced approach requires minimum required regulatory capital to be estimated at the segment level. The "use test" also suggests that the advanced approach should mimic the bank's standard risk management practices; therefore, risk segmentation should be a fundamental component of risk management.
  • 3   This might also suggest that bank models may have been enhanced with better economic forecasts of key macroeconomic variables.

The views expressed in this article are those of the author and are not necessarily those of this Reserve Bank or the Federal Reserve System.