The most important component of the FDIC’s integrated model will be its model of projected losses from bank failures. DIR is currently beginning development of a credit risk model for this purpose. This will be a sophisticated and flexible econometric model that will generate a loss distribution – the likelihood of experiencing any given insurance loss over a particular time horizon – based on a simulation engine and a set of assumptions about failure probabilities, loss rates, and correlation structures. The credit risk model will also provide the ability to run a broad range of scenario and ‘what if’ analyses quickly and easily without extensive human intervention, capabilities that can facilitate resource planning and allocation across the organization. DIR has already committed some initial resources to developing its credit risk model. As detailed below, these efforts should be accelerated and expanded over the next 18 months.

In order to complete its integrated model of financial risk management, the FDIC will need to complement the credit risk model with additional models of investment income, deposit growth, and premium income. These three auxiliary models should be developed and refined in parallel. Once all four pieces are in place, the integrated model will give a comprehensive picture of the FDIC’s overall risks, allowing it to manage and mitigate those risks in the future.

The credit risk model concept

A credit risk model is a method of producing a statistically valid distribution of credit losses over various time horizons. In FDIC’s case, such a model should incorporate four essential elements:

• A statistical model of the probability of failure for individual institutions. The FDIC should develop a state-of-the-art econometric model that estimates bank failure probabilities on a per bank basis using CAMELS ratings, pertinent financial metrics drawn from Call Report data (e.g., capital ratios), and relevant market data (e.g., bond yield spreads and equity volatility).



• A statistical model of loss given failure for individual institution. The model should incorporate institution size, asset composition (e.g., cash, securities, consumer loans, commercial loans), liability structure (e.g., quantity of insured deposits), and speed of deterioration of the institution.



• A correlation structure reflecting the relationships between bank failures, as such events are not truly independent. A deterioration of consumer credit, for example, is likely to cause subprime institutions as a group to face an increased chance of failure. An understanding of how and to what extent the fortunes of particular groupings of banks are correlated is crucial for accurately estimating the likelihood of the most severe systemic outcomes (i.e., the tail of the distribution), itself a critical component of scenario analysis.



• A rigorous Monte Carlo-based simulation engine to integrate these three components into a distribution of credit losses over any given time period. The simulation engine must be capable of handling correlations of arbitrary complexity, in a reasonable amount of time, with a workable user interface, and thorough integration into FDIC IT infrastructure.

Specifics of Recommendation 2.1 (Accelerate development of the integrated model)

DIR has in place plans to develop a basic credit risk model to replace the CLR and 2-year failed-asset models. These plans represent a significant first step toward Horizon 2, and this effort should be accelerated and expanded into the full-blown credit risk model, such as that described above. The FDIC should develop the credit risk model in three phases over the next 18 months: A working prototype by the end of 2003, an “intermediate” model by June 2004, and an “advanced” model by December 2004. The FDIC should simultaneously develop the three auxiliary models of investment income, deposit growth, and premium income.

2.1.a. DIR should build a working prototype capable of generating a basic loss distribution by December 2003. Although the tails of the resulting distribution will contain limited information due to the need to employ a basic correlation structure at this stage, the prototype will demonstrate the feasibility of the model and serve as a platform for future improvements. Failure probabilities should be based on an implementation of the work now being performed through collaboration between the FDIC and Robert Jarrow, an economics professor at Cornell University, but extended to include CAMELS ratings as a variable. (The ongoing work is aimed at developing estimates of failure probabilities using market and Call Report data.) Loss rates should be derived not from the current methodology but from the Research Model, as revised and recalibrated in Horizon 1 and extended to cover not only institutions on the problem list (i.e., those rated CAMELS 4 or 5, and failing institutions) but all banks. A simple correlation structure that includes a limited degree of bucketing (e.g., up to 25 categories) should be employed. Similarly, the simulation engine should be basic – either an inexpensive off-the-shelf solution or an in-house approach (e.g., built with SAS or MATLAB). The ongoing work with Professor Jarrow is relevant to all of these components, and should enable a working prototype to be created in 90 days with only a modest shift of existing DIR staff and within the existing DIR budget.

2.1.b. DIR should develop an intermediate credit risk model to replace the current CLR methodology by June 2004. The prototype will need to be backtested, and any obvious problems resolved to develop an intermediate model suitable for use in operations. Integrated and flexible, the intermediate model will combine updated approaches to failure probabilities, loss rates, and the other components into a single architecture with an enhanced user interface and convenient report-generation capability. The integrated model will provide a greatly enhanced ability to conduct “what if” analyses through its use of moresophisticated correlation structures, while the associated loss distributions – the tails in particular – will provide additional insight into risks to the insurance funds. The model should be used for several months to shadow the current CLR methodology before being adopted for that use by June 2004.

Within the model itself, failure probabilities should represent an extension of those used in the prototype, with additional testing and calibration to more systematically incorporate market-based offerings such as KMV for public institutions and RiskCalc for private institutions. Loss rates should be refined to incorporate institution size as a variable.

The limited correlation structure used by the prototype should be expanded to include multiple scenarios and refined to incorporate additional data. A more sophisticated correlation structure could, for example, shed light on how proposed consumer-bankruptcy legislation would affect the likelihood of subprime lenders failing as a group. Another possible correlation structure might reflect the possible progression of a liquidity crisis through the banking sector and how that might impact failure probabilities. These efforts to improve the model’s correlation structures will require both statistical analysis (e.g., examining historical data for potential relationships between depository institutions) and qualitative, forwardlooking analysis of potential future correlations.

Even with these improvements, enhanced credit loss distributions will not be assured without a more robust simulation engine. The increasing need to run simulations frequently and the “curse of dimensionality” (i.e., the size of the problem increases exponentially with the number of correlations) will require a solution more sophisticated than the ad hoc simulation engine crafted out of SAS, MATLAB, or similar software for the prototype. The FDIC should purchase an advanced commercial Monte Carlo simulation package or build a similarly capable simulation program in-house.

2.1.c. By December 2004, DIR should develop and adopt an advanced credit risk model capable of reliably estimating loss distributions over multiple years. Once firmly established as the basis for estimating necessary one-year reserves (the CLR), the intermediate model should be updated with new and ongoing FDIC research into model inputs, such as the impact of region and size on correlations among bank failure probabilities. Loss rates similarly should be refined to add new variables, such as the rate at which an institution deteriorates and its asset-toliability ratio. Also, the model should be extended to project losses over multiple years.

Pursuing this coordinated three-phased approach to the development of a credit risk model will require a considerable degree of project management and planning that may strain DIR’s existing organizational structure. DIR should create a risk management modeling group, with detailed work schedules and regular progress review meetings, to ensure the delivery and continuous improvement of the various models used in Horizon 2. Much as Exhibit 2-2 presents a high-level summary of the credit risk model schedule, detailed supporting schedules should be created for each of the model components. In addition, the risk management modeling group should be charged with conducting further research into available outside solutions – not as replacements for internal products, but as additional benchmarks for the FDIC’s own work on an ongoing basis in the future.

Key characteristic Capability to systematically generate loss distributions, but supporting models unintegrated and relatively inflexible Models partially integrated, improved user interface allow staff with different skills to produce reports in less time Models are increasingly integrated/interactive and run in closer to real time  Default  rates Implementation of the Jarrow work including CAMELS ratings Continued extensions, testing and calibration against RiskCalc, KMV and others Continuous improvement of models, with particular emphasis on speed, flexibility  Loss  rates Implementation of the Research Model Extension of loss rates model to incorporate new variables, testing of model against AVR database Continued enhancement of loss rate estimates  Correlation  structure Assumed that all banks are independent or limited bucketing (e.g.,as many as 25 buckets) Increasing use of correlation structures, 2-3 standard variations developed by statistical analysis Procedures and computer programs to facilitate creating correlation structures to express views of DSC and others  Simulation  engine MATLAB, SAS, or similar 'Cheap and cheerful' off-the-shelf solution Limitations of 'cheap and cheerful' solution increasingly become binding constraint Increasing need for fast commercial Monte Carlo package Output/ capability Basic loss distribution, tails will contain limited information due to simple correlation structure Improved CLR and multi-year projections Tails of loss distribution provide real insights for FDIC decision making Detailed analysis at least quarterly Detailed analysis now routinely available, perhaps at daily RAC meeting and on intranet Some risk reports can be created at intranet