Strengthening Financial Risk Management at the FDIC
Improving Financial Reporting – Horizon 1
RECOMMENDATION 1.2: REPLACE THE 2-YEAR PROJECTION
WITH A CONFIDENCE INTERVAL AROUND THE CLR AND A
2-YEAR LOSS ESTIMATE
The 2-year Projection of failed bank assets is the FDIC’s estimate of the sum of the
assets of all FDIC-insured institutions whose failure is reasonably possible in the
coming 24 months. Compared with the CLR, the 2-year Projection provides a
slightly longer-term perspective and a somewhat weaker threshold for the likelihood
of failure. Until recently, its use was twofold: as a key component of the semiannual
rate case, and as a basis for a footnote to the Annual Report describing
“reasonably probable” losses to the deposit insurance funds. Today, it is used only
for the latter purpose.
The methodology of the 2-year Projection could be improved, but it is
fundamentally not the best estimate for FDIC’s needs, and it should be replaced.
The first step in this reasoning is to explain how the 2-year Projection is calculated.
How the 2-year Projection is calculated
The 2-year Projection is based on the DSC’s 8-Quarter List and three DIR models,
with six different scenarios between them. Each of these seven inputs (DSC’s list
and six model scenarios) yields a list of institutions expected to fail in the coming
2 years, which in turn yields seven estimates of the sum of 2-year failed-bank
The FRC creates a “model range” as the band between the largest and smallest of
these seven estimates. The FRC participants then use their judgment to determine
whether that range should be altered, for example by raising the upper end of the
range. It is this modified range, the “reported range,” that the FRC disseminates to
the rest of the FDIC and that is used in the FDIC’s annual report. As depicted in
Exhibit 1-6, the current model range for BIF-insured institutions is $0 to $22
billion, and the FRC currently has raised and widened its reported range, to
$2 billion to $35 billion.
The seven input estimates of failed-bank assets are derived from the DSC 8-Quarter
List and three DIR models. Those models are the Pro Forma Model, the Stress
Analysis Model (SAM), and the Proportional Hazards and Logistic Model (treated
here as a single model, the “PH Model”).
DSC’s 8-Quarter List contains the FDIC’s best supervisory assessment of which
institutions are likely to fail in the coming 24 months, along with the quarter of
expected failure. An institution is put on the list whenever supervisors judge its
likelihood of failure to be more than 50 percent.
The Pro Forma Model was discussed above. This model provides fairly
straightforward forecasts of the balance sheets of depository institutions. It uses
accounting ratios and rules to estimate an institution’s future financial condition
based on its current financial condition. For the 2-year Projection, Pro Forma is run
in two different scenarios: “Optimistic” and “Pessimistic.” The primary difference
between the two scenarios is in their assumptions about the share of an institution’s
assets that are non-performing. In either scenario, an institution is projected to fail
if the Pro Forma forecast of the institution’s capital falls below two percent of its
SAM, like Pro Forma, is a balance-sheet simulation model, but SAM incorporates
additional financial data and has more-sophisticated rules about how balance sheets
evolve. Furthermore, SAM has 14 parameters (e.g., loan charge-off rates) that are
calibrated to fit historical data. Like Pro Forma, SAM projects that an institution
will fail whenever its capital falls below 2 percent of assets.
The PH Model is based on a set of statistical regressions about the causes and
timing of failures of depository institutions. Specifically, it uses a logistic
regression to estimate the probability that an institution will fail, based on nine
measures of its current condition. These variables range from CAMELS ratings to
the change in its capital over time. For the 2-year Projection, the PH Model is run
in three different scenarios: “Optimistic,” “Baseline,” and “Pessimistic,” where the
optimistic and pessimistic scenarios are 90 percent confidence bounds around the
baseline scenario. In each case, the PH Model reports an institution as expected to
fail whenever the failure probability from the regressions is above 50 percent.
Performance of the 2-year Projection
The performance of the 2-year Projection should be measured based on how the
estimate is actually used. The 2-year Projection is referred to in the Rate Case and
is an input to a DOF calculation that is reported in Note 6 of the FDIC’s Annual
Report. In crafting Note 6, the Division of Finance uses the upper end of the
reported range ($35 billion for the BIF example), multiplies by an assumed 20
percent loss rate on assets ($35 billion x 0.2 = $7 billion) and subtracts the current
CLR ($7 billion – $1 billion = $6 billion). The result is explained as follows:
“Due to the uncertainty surrounding future economic and market conditions,
there are other banks for which the risk of failure is less certain, but still
considered reasonably possible. Should these banks fail, the BIF could incur
additional estimated losses up to $6.0 billion.”14
The 2-year Projection and the calculations described above are not an ideal way to
arrive at “possible” losses over and above the CLR, for three reasons. First, the
seven input estimates are not equally valid measures of possible losses, and using
the largest of them takes no account of the meaningful information contained in the
others. As depicted in Exhibit 1-7, all the input estimates except SAM generally
underestimate actual losses, so they would be poor estimates of the FDIC’s
downside risk. Second, the 1-year loss estimate in Note 6 is derived from a longer,
2-year estimate. Third, the current method for arriving at Note 6 assumes a flat 20
percent loss rate on failed assets, while average loss rates are typically lower and
depend on an institution’s size, asset composition, and liability structure.
PREDICTIVE POWER OF INPUT MODELS TO 2-YEAR PROJECTION Assets in failing depository institutions in subsequent 2 years, 1997-2001, $B
Note: Scale for SAM . Mild Stress differs from other three scatter plots
Source: FDIC; McKinsey analysis
Specifics of Recommendation 1.2 (Replacing the 2-year
Since the 2-year Projection is not well suited to the FDIC’s needs, the FRC should
not expend its limited resources improving the estimate. Instead, the 2-year
Projection should be abandoned, and two better-suited estimates calculated in its
1.2.a. The FDIC should no longer calculate the 2-year Projection. The 2-year
Projection should be replaced with two other estimates, detailed below, that will
better serve FDIC’s needs.
1.2.b. DIR should decide whether to keep, revise, or eliminate the three models
supporting the 2-year Projection:
DIR should keep the Pro Forma model as it is today. That model is
inexpensive to maintain and is used appropriately in the risk
groupings for the CLR calculations.
SAM would be better suited to offsite monitoring and should be
migrated to that use. SAM is a useful research model and shows
promise as a tool to screen for at-risk institutions that deserve further
attention, which is precisely the objective of offsite models.
The PH Model should be abandoned. The statistical methods that it
uses have some merit, but those methods can and should be
implemented afresh with the new credit risk model that DIR is
developing as its next-generation risk management model (described
in the second chapter). While the methods embodied should be
retained in this manner, the model itself should be abandoned as a
vehicle for supporting risk reporting requirements.15
1.2.c. DIR should calculate a confidence interval around the CLR, and DOF
should use the upper end of this confidence interval in Note 6 to the FDIC’s
annual report. There are two stated objectives in Note 6 of the FDIC’s annual
report. The first is to report losses that are “probable and reasonably estimable.”
This is met by the CLR itself. The second objective is to report additional losses
that are “possible.” A confidence interval about the CLR would be the natural basis
for such possible losses. For example, suppose the CLR was $1 billion and the
upper 90 percent confidence bound for the CLR was $7 billion. Then 90 percent of
the time the true 1-year loss would be less than $7 billion, so DOF would report
$6 billion ($7 billion less the $1 billion CLR) as possible additional losses.
There are a variety of ways to calculate a confidence interval around the CLR. The
simplest and most robust method would be to track the differences between the
CLR and the FDIC’s actual losses over time. Sometimes the CLR will be very
close to the actual loss. Other times, the CLR and actual losses will diverge. The
distribution of these errors gives a sense for how likely it is for the actual loss to be
far from the reported CLR. Exhibit 1-8 depicts this methodology in more detail.
The first step in creating such a confidence interval will be for DIR to compute pro
forma CLRs for as far back as possible, e.g., to 1990, using the revised
methodology described in the previous section. DIR should then tabulate the
distribution of errors between this CLR and the FDIC’s actual losses over that time
period.16 This computation would then form the basis for a robust confidence
interval around the CLR going forward.
1.2.d. DIR should create a 2-year loss estimate using methodology similar to that
of the CLR. The calculations will be identical to those of the CLR, but using 2-year
rather than 1-year failure rates. Such an estimate can be used to inform the internal
budgetary planning process, the semi-annual rate assessment, and/or other
organizational processes that may benefit from an extended loss outlook.
1.2.e. DIR should investigate the sensitivity of such a 2-year loss estimate to
changes in the reserve list. Since the CLR methodology estimates losses only for
institutions with current CAMELS ratings of 4 and 5, the proposed 2-year loss
estimate would not account for failures that were unanticipated a full two years in
advance. As such, it may be necessary to adjust the proposed 2-year loss estimate
upward to account for institutions that may have their CAMELS ratings
downgraded from 1-to-3 to 4-to-5 after the first year.
14FDIC 2002 Annual Report, at 51.
15 This model and others may well have value in other contexts, but a decision by DIR to retain such models should
hinge on an explicit identification of what those contexts may be and a clear description of what corresponding value
the models will create.
16 The methodology in Exhibit 1-9 assumes that the distribution of errors in percentage terms does not depend on the
level of the CLR, e.g., that the FRC is just as likely to be 10 percent off when the CLR is $1 billion as when it is
$1 million. In statistical terms, Exhibit 1-9 assumes that the errors are homoskedastic, when in fact they may be
heteroskedastic. After calculating a larger set of historical errors (e.g., by back-testing the CLR to 1990), DIR should
explore whether there is a systematic relationship between the level of the CLR and the size of the error. If such
differences exist, DIR should control for them, either by a) creating error distributions for CLRs in different size
bands or b) using statistics to fit a functional relationship between the variance of the error and the level of the CLR
(e.g., the variance of the errors might be a linear function of the CLR).