8/2/2019 Loan Losses Provision
1/36
Preliminary Draft. For Comments Only
BANK LOAN-LOSS PROVISIONING,
Methodology and Application
J. Dermine and C. Neto de Carvalho*
April 9 2004
* INSEAD (Fontainebleau) and Universidade Catolica Portuguesa (Lisbon),
respectively. The authors are grateful to Professor Santos Silva for his insights inthe econometrics of fractional responses, to M. Suominen for comments on a first
draft, to J. Cropper for editorial assistance, and to Banco Comercial Portugus(BCP) for access to internal credit data.
8/2/2019 Loan Losses Provision
2/36
1
BANK LOAN-LOSS PROVISIONING,
Methodology and Application
Abstract
A fair level of provisions on bad and doubtful loans is an essential input in mark-to-
market accounting, and in the calculation of bank capital and solvency. Fewacademic studies have analyzed the adequacy of loan-loss provisioning because, dueto the private nature of bank loan transactions, few micro data are readily available.Access to data on recovery over time on bad and doubtful bank loans allowsdeveloping and applying two methodologies to calculate a fair level of loan-loss
provisions. Empirical estimates are then compared to a regulatory provisioningschedule imposed by a central bank.
8/2/2019 Loan Losses Provision
3/36
2
Introduction
Fair provisioning on bad and doubtful loans is of great importance for bankregulators. If there has recently been intense discussion on the merits of Basel 2, the
revised capital accord that would much better capture the actual risks taken by banks
(Basel Committee, 2003), it is quite evident that this accord will not have much
relevance if the measurement of bank capital is not satisfactory. A key input in the
measurement of bank capital is the amount of loan-loss provisions1 on bad and
doubtful loans. Well-known cases of significant under-provisioning include the
French Credit Lyonnais in 1993, Thailand in 1997, Japan in late 1990s (Genay,
1998), and, more recently, China.
As bank loans are, by their economic nature, private, there is not much market-based
information to estimate their current value, so that loan-loss provisions must be
estimated. In several countries, central banks have provided a provisioning schedule
for non-performing loans that takes into account the number of days since the default
date and the quality of collateral support. A segment of the academic literature has
studied whether or not loan-loss provisioning is used to manage or smooth reported
earnings and capital, its impact on investors returns, and its contagious effects on
other banks. To the best of the authors knowledge, not much has been written on
normative methodologies that could be used to estimate a fair level of loan-loss
provisioning at a given point in time, and its dynamic evolution over time. The
purpose of this paper is to provide a methodology to calculate dynamic loan-loss
provisions. Having access to a unique set of micro data on losses on loans to small
and medium size firms over the period 1995-2000, the methodology can be illustrated
with real bank data. In terms of contribution to the academic literature, this is the first
paper to apply mortality analysis and multivariate statistics to calculate dynamic loan-
loss provisioning, and to compare a micro databased empirical estimate of
provisioning schedule to a regulatory schedule enforced by a central bank.
1 Note that European terminology is being used. Loan-loss provisions represent the expected losses
on a portfolio of impaired loans. They constitute a contra-account, which is deducted from the value ofgross loans in a balance sheet. Provisions are referred to as loan-loss allowance or loan-lossreserves in US bank accounts.
8/2/2019 Loan Losses Provision
4/36
3
The paper is structured as follows. The literature on bank loan-loss provisioning is
reviewed in Section 1 of the paper. In Section 2, the database on individual loan-
losses is presented. The mortality-based approach to analyzing fair provisioning on
bad and doubtful loans is discussed in Section 3. Empirical univariate estimates of
provisioning over time are presented in Section 4, and these estimates are compared to
those imposed by a central bank. Finally, a multivariate statistics approach to loan-
loss provisioning is developed in Section 5. Section 6 concludes the paper.
Section 1. Review of the Literature
The accounting and finance literatures have analyzed three main provisioning issues
related to private information held by banks on loans, one of the fundamental
characteristic which explains the economic role of banks (Diamond, 1984): the extent
of earnings and capital smoothing, the impact of reported provisions on a banks stock
returns, and the systemic impact on the banking industry of disclosure on loan
provisions by one bank.2
A series of papers have analyzed the extent of earnings and capital smoothing through
a pro-cyclical loan-loss provisioning, with high provisions in good times and lower
provisions in bad times. For instance, Laeven and Majnoni (2003) and Hasan and
Wall (2004) report empirical evidence throughout the world consistent with the
earnings smoothing hypothesis.3 It is argued that the Basel Accord cannot achieve a
level playing field if it does not address adequately the loan-loss provisioning
practices around the world. Consistent with the above analysis, Gunther and Moore
(2003) find that supervisory examinations in the US precipitated a significant numbers
of adverse revisions of financial statements during their sample period. Similar
studies, documenting managerial discretion in loan-loss provisioning, include Wahlen
(1994), Wetmore and Brick (1994), Beattie et al. (1995), Soares de Pinho (1996), Kim
and Kross (1998), and Ahmed, Takeda and Thomas (1999). It must be noted that none
of the above studies propose a methodology to calculate fair provisions on bad and
doubtful loans, but rather report evidence that the actual level of aggregate
2 See the survey on loan-loss accounting by Wall and Koch (2000).3 A theoretical argument for earnings smoothing can be found in Degeorge et al. (1999).
8/2/2019 Loan Losses Provision
5/36
4
provisioning is correlated with accounting income data.4
A second stream of the literature has analyzed the impact of reported provisions on
bank stock returns. Whalen (1994) finds that unexpected provisions have a positive
impact on stock returns. This, at first counter-intuitive, result is interpreted as a signal
that future earnings will be good, allowing the build-up of provisions by bank
managers. Musumeci and Sinkey (1990) report similar results in the context of
provisioning and the international debt crisis.
A third stream of the loan-loss reserve literature has analyzed the signalling impact of
loan-loss reserves announcements on the stock returns of the announcing bank and
that of other institutions (for instance, Grammatikos and Saunders, 1990, and Docking
et al., 1997). Again, this stream of research is related to aggregate loan-loss reserves
reported in financial statements. The purpose of the current study is quite different
from that of the above literature. It proposes a micro-based methodology to calculate
fair provisions on bad and doubtful loans.
Two approaches have been suggested to measure credit risk and provisions on a loan
portfolio (Crouhy et al., 2000): a mark-to-market mode and a default mode. In a
mark-to-market mode, the value of loans (both performing and non-performing) is
estimated, either with actual market prices (available for large loans in the USA, Allen
et al., 2003), or with an estimate of their fair value, which will integrate the
probability of default over time, of losses-given-default, and a risk-adjusted discount
rate. In the default-mode approach, provisions are estimated only for non-performing
loans. With the exception of cases with readily available market prices, one notices
that, in both the mark-to-market mode and the default-mode, one must estimate the
expected losses-given-default. This is the objective of the paper. It is an important
issue as, according to the above literature, bank managers use discretion in loan-loss
provisioning. Bergeret al. (1989 and 1991), in their critical review of market value
accounting, suggest that value adjustment for credit risk should be based on a loan-
loss reserve on nonperforming loans. According to regulators (Basel Committee,
1998, or Wall and Koch, 2000), these reserves should be prudent and conservative,
falling within an acceptable range of expected losses. The paper aims to provide a
methodology to calculate over time loan-loss provisions on bad and doubtful loans.
4 Related papers on bank earnings management include Barth et al. (1990) who document income
smoothing through the realization of capital gains and losses, Scholes et al. (1990) who analyze therealization of capital gains and losses in the context of tax planning, and Galai et al. (2003) whoanalyze, in a theoretical model, the timing of realized capital gains on real assets (such as buildings).
8/2/2019 Loan Losses Provision
6/36
5
Section 2. Bank Loan-losses, Database and Measurement Issues
The largest private bank in Portugal, Banco Comercial Portugus, provided the
database used in this study. It consists of financial credit given to 4,000 small and
medium size companies5 over the period June 1995 to December 2000. All these
companies have a turnover of more than 2.5 million. Table 1 (Panels A and B)
provides information on the number of defaults per year, and on the amount of debt
outstanding at the time of default. One observes that the series of 371 default cases is
distributed evenly over the six-year period, and that the distribution of debt
outstanding at the time of default is highly skewed towards the low end. Half of the
debt is less than 50,000.6 The available information in the database includes the
industry classification, the interest rate charged on the loan, the history of the loan
after a default has been identified, the type of collateral or guarantees, and the internal
rating attributed by the bank.7
In Table 1 (panel C), the various forms of guarantees, or collateral, are reported along
the classification used by the Bank for International Settlements (Basel Committee,
2003). These include:
Personal guarantee
Real estate collateral
Physical collateral (inventories)
Financial collateral (bank deposits, bonds or shares)
In 35.6% of the cases, there is no guarantee or collateral. Personal guarantees, which
are used in 58.2% of the cases, refers to written promises made by the guarantor
(often the owner or the firms director) that allows the bank to collect the debt against
5 The data gathered for this study did not include any reference to the identity of the clients or any otherinformation which, according to Portuguese banking law, cannot be disclosed. The data set concernsSME loans, mostly from the region of Lisbon.6 This sample includes loans of much smaller size than those used in the reported US bank studies. Theface value of loans was higher than US$ 100 million in Altman and Suggitt (2000), and the averagecommercial and industrial loan was US$ 6,3 million in Asarnow and Edwards (1995).7
The database used in this study did not exist in the appropriate format. The bank provided help toidentify all these variables, and to recover the data located in various databases. The history of eachloan, after a default had occurred, has been carefully analyzed.
8/2/2019 Loan Losses Provision
7/36
6
the personal assets pledged by the guarantor. Collateral is used in 15.1 % of the
cases.8
Any empirical study of credit risk and loan provisioning raises two measurement
issues. Which criterion should be used to define the time of the default event? Which
method should be used to measure the recovery rate on a defaulted transaction?
The criterion used for the classification of a loan in the default category is critical
for a study on provisions, as a different classification would lead to different results.
Three default definitions are used in the literature:
i) A loan is classified as doubtful as soon as full payment appears to be
questionable on the basis of the available information.9
ii) A loan is classified as in distress as soon as a payment (interest and/or
principal) has been missed.
iii) A loan is classified as default when a formal restructuring process or
bankruptcy procedure is started.
In this study, because of data availability, we adopt the second definition, that is, a
loan is classified as in default as soon as a payment is missed.10 For information, the
reporting to the Central Bank of Portugal takes place after thirty days, if the loan
remains unpaid or unrestructured.
The second methodological issue relates to the measurement of recovery on defaulted
loans, as provisions will attempt to measure the amount that will not be recovered.
There are two methodologies:
8 Jimenez and Saurina (2002) also observe, in the case of Spain, that a very large proportion of bankloans are not collaterized.9 Hurt and Felsovalyi (1998).10 For the sake of comparison, the definition of default adopted by the Basel Committee is as follows(Basel Committee, 2001, p 52): A default is considered to have occurred with regard to a a particularobligor when one or more of the following events has taken place:a) It is determined that the obligor is unlikely to pay its debt obligations (principal, interest, or fees) infull;
b) A credit loss event associated with any obligation of the obligor, such as a charge-off, specificprovision, or distressed restructuring involving the forgiveness or postponement of principal, interest or
fees;c) The obligor is past due more than 90 days on any credit obligation; ord) The obligor has filed for bankruptcy or similar protection from creditors.
8/2/2019 Loan Losses Provision
8/36
7
i) The price of the loan at the default date, defined most frequently as the trading price
one month after the default. This approach has been used in studies on recoveries on
corporate bond defaults.
ii) The discounted value of future cash flows recovered over time after the default
date.
As no market price data are readily available for defaulted bank loans in Portugal, the
second methodology -the present value of actual recovered cash flows- is the only
feasible alternative. In any case, this approach is preferable as it allows the study of
the actual recovery of cash flows over time, whereas the price of a loan after the
default date is only representative of the value of expected cash flows. Moreover, the
actual recovery approach avoids eventual problems of price biases created by a
lack of liquidity on the defaulted loan market. This approach was adopted, in studies
on recovery on bad and doubtful loans, by Asarnow and Edwards (1995), Carty and
Lieberman (1996), and Hurt and Felsovalyi (1998).
In order to measure the cash flows recovered after a default event, we tracked, each
month, the post-default credit balances. Capital recovery is a reduction in the total
balance. The total cash flow recovered is this capital recovery plus the interest on the
outstanding balance.
To estimate the present value of the cash flows recovered, one must choose a discount
rate. We follow Asarnow and Edwards (1995), Carey (1998), Hurt and Felsovalyi
(1998), and the Basel Committee (1999) who suggest using a contractual pre-default
interest rate on the loan to calculate the present value. This approach has the
advantage that, if a loan is fully repaid, the present value of actual cash flows
recovered will be equal to the outstanding balance at the default date. While the above
authors did not have access to the interest rate charged on individual loans and had to
rely on an approximation of credit-risk adjusted yield curve, data on interest rates
charged on the loans are available in this study.
If, at first, the tracking of cash flows after a default event would appear a relatively
simple (but time consuming) exercise, special cases did require some judgmental
adjustment. Two such cases are discussed hereafter.
8/2/2019 Loan Losses Provision
9/36
8
i) First, there are fifty-two cases of multiple defaults. This refers to situations
in which a company enters a default category, returns to performing
status at a later date after paying the loan fully, and then falls back into the
default category. A first option could have been to consider each default as
a separate event, but this might have biased the econometric analysis with
several defaults linked to the same borrower. In the econometric results
that will be reported, we have kept the first default cases.
ii) Second, in the case of formal loan restructuring, with the loan returning to
a performing status, we did not consider these cases as a 100%
recovery. Indeed, in many cases, these loans subsequently fell back into
the default category. The cash flows received on restructured loans were
carefully identified.11
Two approaches will be used to analyze loan-loss provisioning, a univariate mortality-
based approach, and a multivariate statistical analysis of the determinants of recovery.
Section 3. Loan-Loss Provisioning, a Mortality-based Approach
To calculate provisions, we shall, in a first step, analyze recovery on bad and doubtful
loans. Having access to the history of cash flows on these loans after default, we can
study the time distribution of recovery. With reference to studies by Altman (1989)
and Altman and Suggit (2000), we apply the mortality approach. It must be noted that
the mortality approach was applied to measure the percentage of bonds or loans that
defaulted n years after origination. The application of mortality to loan recovery rates
and provisions is, to the best of our knowledge, novel. It examines the percentage of a
bad and doubtful loan, which is recovered t months after the default date. This
methodology is appropriate because the population sample is changing over time. For
some default loans, those of June 1995, we have a long recovery history (66 months),
while for the recent 2000 loans in default, we have an incomplete history of
11 Note that this approach is more conservative than the one adopted by Hurt and Felsovalyi (1998)who considered a 100% recovery when the debt was reclassified to a performing status.
8/2/2019 Loan Losses Provision
10/36
9
recovery. The actuarial-based mortality approach, based on the Kaplan-Meier
estimator (Greene, 1993), adjusts for changes over time in the size of the original
sample.12
To define the concepts used to measure loan recovery rate and provisions, it is, for
expository reasons, useful to refer to a simple example. Consider a loan of 100 that
enters the default category in December 2000. We track the subsequent payments on
this loan, assuming, for expository convenience, that all payments take place at the
end of the year, with a final payment in December 2003. The interest rate is 10%.
Dec. 2000 Dec. 2001 Dec. 2002 Dec. 2003
Loan outstanding 100 110 66 44(before cash payment)
Cash payment 0 50 26 14
Loan balance 100 60 40 30(after cash payment)
Let us define theMarginal Recovery Rate at December 2001, MRR1, as the proportion
of the outstanding loan in December 2001 that is being paid, one period (in the
example, one year) after default:
MRR1 = Cash flow paid 1 / Loan balance 1
= 50/110 = (50 / 1.10) / 100 = 5/11
The marginal recovery rate can also be interpreted as the percentage repayment on the
loan outstanding, in present value terms.
Let us define thePercentage Unpaid Loan Balance after payment in December 2001,
PULB1, as the proportion of the December 2001 loan balance that remains to be paid
one period after default :
PULB1 = 1 MRR1 = 1 5/11 = 6 /11
12
The different approaches (cohort analysis used by Moodys, static pool employed by Standard andPoor, and mortality rate) to measure credit risk are discussed and contrasted in Caouette, Altman, and
Narayan (1998).
8/2/2019 Loan Losses Provision
11/36
10
Similarly, one can define theMarginal Recovery Rate at December 2002 =
MRR2 = Cash flow paid 2 / Loan 2
= 26/66
ThePercentage Unpaid Loan Balance after payment in December 2002, two periods
after default is equal to :
PULB 2 = 1 MRR2 = 1 26/66 = 40/66
And the Cumulative Recovery Rate in December 2002 for a loan defaulting in
December 2000, CRR0,2 , is defined as:
CRR0,2 = 1 (PULB1 x PULB2)
= 1 (6/11 x 40/66) = 1 240/726 = (1 40/121) = 81/121
= (81/1.12) /100 = 66.9%
Similarly, one can calculate aMarginal Recovery Rate, MRR3, aPercentage Unpaid
Loan Balance, PULB3, and a Cumulative Recovery Rate, CRR0,3, at December 2003
=
MRR3 = Cash flow paid 3 / Loan 3
= 14/44
PULB3 = 1 MRR3 = 1 14/44 = 30/44
CRR0,3 = (1 (PULB1 x PULB2 x PULB3)
= 1 (6/11 x 40/66 x 30/44) = 1 7,200 / 31,944 = 1 30/133.1
= 103.1/133.1 = (103.1/1.13) / 100 = 77.5%
Similar to Carey (1998), Asarnow and Edwards (1995), Carty and Lieberman (1996),
Hurt and Felsovalyi (1998), and La Porta et al. (2003), the Cumulative Recovery Rate
at time T on a loan balance outstanding at time 0, CRR0,T , represents the proportion
of the initial defaulted loan that has been repaid (in present value terms), T periods
after default. Note that these authors report only the total cumulative recovery rate
over a long (unidentified) period, whereas we, adopting the Altman mortality-based
8/2/2019 Loan Losses Provision
12/36
11
approach, report the extent of cumulative recovery over time. This information will
allow calculating dynamic provisions on the remaining loan balance, several months
after the default date.
The Loan-Loss Provision, LLP0, on a loan balance outstanding at the default date,
December 2000, is equal to:
LLP0 = 1 - CRR0,3 = 1 - 0.775 = 22.5%
Moving forward, one can define a Cumulative Recovery Rate,CRR1,3 , on a loan
balance outstanding at December 2001, and a dynamicLoan-Loss Provision, LLP1 ,on
a loan balance outstanding at December 2001. These are defined as :
CRR1,3 = (1 PULB2x PULB3) = 1 - 40/66 x 30/44 = 58.7%
LLP1 = 1 CRR1,3 = 41.3%
The Cumulative Recovery Rate, CRR2,3, and the dynamic Loan-Loss Provision,
LLP2, on a loan balance outstanding at December 2002 are defined as :
CRR2,3 = (1 PULB3) = 1 - 30/44 = 31.8%
LLP2 = 1 CRR2,3 = 68.2%
To generalize for an individual loan i in default, we define four concepts, t denoting
the number of periods after the initial default date 0:
MRRi,t = Marginal Recovery Rate in period t
= Cash flowi paid at the end of period t / Loani outstanding t
PULBi,t = Percentage Unpaid Loan Balance at the end of period t = 1 MLRRi,t
The Cumulative Recovery Rate evaluated from the default date 0 until infinity,
CRRi,0, , and the Loan Loss Provision, LLPi,0, are equal to:
8/2/2019 Loan Losses Provision
13/36
12
CRRi,0, = Cumulative Recovery Rate periods after the default = 1 ,1
i t
t
PULB
=
LLPi,0 = Loan-loss provisions = 1 CRRi, 0,
Note that for the sake of presentation, the loan-loss provision was calculated on a loan
balance outstanding at the default date, 0. In a more general dynamic provisions
setting, the provision can be calculated on a loan balance outstanding at any date n
after the default date. For instance, one can compute the cumulative recovery on loan
balances outstanding 4 or 13 months after the default date.
Having computed the cumulative recovery rate on individual loans, one can compute
an unweighted average cumulative recovery rate for the sample of loans. This is the
approach adopted in the losses-given-default literature. Alternatively, one can
compute a sample weighted average recovery rate that will take into account the size
of each loan. This is defined as follows:
SMRRt = Sample (weighted) Marginal Recovery Rate at time t
= ,1 1
m m
i t i t
i i
Cash flow received Loan outstanding = =
, , where i stands for each
of the m loan balances outstanding in the sample, t periods after default.
SPULBt = Sample (weighted) Percentage Unpaid Loan Balance at period t
= 1 SMLRRt
SCRR = Sample (weighted) Cumulative Recovery Rate, an infinite number of
periods after the default
=1
1 tt
SPULB
=
A comparison of the Sample (weighted) Cumulative Recovery Rate with the average
of recovery rates on individual loans will be indicative of a size effect.
8/2/2019 Loan Losses Provision
14/36
13
Section Four. Loan-Loss Provisioning, Empirical Evidence with aMortality-based Approach
The application of the mortality-based approach to calculate dynamic provisions is
followed by a comparison with the mandatory dynamic provisioning imposed by the
Central Bank of Portugal.
Dynamic Loan-Loss Provisioning, Empirical EvidenceThe sample marginal and cumulative recovery rates for the sample T periods after the
default date, SMRRT and SCRRT, respectively, are reproduced in Figures One and
Two. One observes, in Figure One, that most of the marginal recovery rates in excess
of 5 % occur in the first five months after the default. In Figure Two, one observes
that the cumulative recovery rate is almost completed after 48 months, and that the
weighted cumulative recovery rates are lower than the unweighted cumulative
recovery rates. For instance, the unweighted cumulative recovery rate after 48 months
is 70%, while the weighted recovery rate is 56.3 %. This is indicative of a size effect,
further analyzed and found statistically significant in Neto de Carvalho and Dermine
(2003). Cumulative recovery on large loans appears to be significantly lower.
It is also of interest to analyze the distribution of cumulative recovery rates across the
sample of loans. The distribution of cumulative recovery rates after 48 months is
reproduced in Figure Three. This figure shows a bi-modal distribution with many
observations with low recovery, and many with complete recovery. These results are
quite similar to those reported by Asarnow and Edwards (1995) for the US, and Hurt
and Felsovalyi (1998) for Latin America. Loan portfolio models which incorporate a
probability distribution for recovery rates should take into account this bi-modal
distribution.13
13Current commercial credit portfolio models do not incorporate the bi-modaldistribution. For instance, CreditRisk+, developed by Crdit Suisse FinancialProducts, assumes a fixed expected recovery rate within each band, whileCreditMetrics uses a beta distribution (Crouhy et al., 2000).
8/2/2019 Loan Losses Provision
15/36
14
Cumulative recovery on outstanding balances on bad and doubtful loans are reported
in Table 2. These dynamic cumulative recoveries are calculated on loan balances
outstanding n months after the default date, with n varying from 0, the time of
financial distress, to 37 months after the default date. Estimates of dynamic
recoveries will allow the calculation of dynamic provisioning on expected loan-losses.
One first observes, for the total sample, that the reported difference between
unweighted and weighted recoveries is significant at time zero, the time of default,
(72.6% vs. 63.8%, respectively), but less significant at later dates (for instance, 33%
vs 35.8%, 19 months after the default date). This is expected as, in Figure Two, the
difference between the weighted and the unweighted cumulative recovery rates does
not increase through time. Over time, the weighted (and unweighted) cumulative
recovery rates on the unpaid loan balance are decreasing monotonously over time,
55.7% (57.8%) in period 4, to 23.3% (17.1%) in period 37. More complete
information is obtained, in Table 2, by dividing the sample into three groups,
according to the absence of personal guarantee/collateral, the existence of personal
guarantee, or the existence of collateral support. As expected, the cumulative
recovery rates (unweighted, so as not to be affected by the size effect) is the highest
for the case of loans with collateral for every period, for instance 85.3% at time 0, the
default date. A counterintuitive and significant observation is that the (unweighted)
cumulative recovery rates on loans without any type of real or personal guarantee
dominates, in every period, the cumulative recovery rate on loans with personal
guarantees. For instance, at time 0, (unweighted) recoveries on loans without any type
of guarantee/collateral is 80.6%, vs.63.9% for recoveries on loans with personal
guarantee (or 47.0% vs. 20.8%, 19 months after the default date). A director of the
bank explained that this result is likely to be due to two factors. First, guarantee or
collateral support is not usually requested from reliable clients, so that the existence
of a guarantee is an indicator of greater risk. Second, some borrowers are able to
shift ownership of personal assets to other persons, so that, when the bank tries to
execute the debt, there is not much left. This empirical result, if confirmed in other
studies, implies that regulatory provisions should not penalize loans without personal
guarantee, as the absence of a guarantee might be justified by higher expected
recoveries.
8/2/2019 Loan Losses Provision
16/36
15
In Figures 4a and b, we show the dynamic cumulative recoveries on unpaid balance at
time 4 and 25. As expected, the weighed and unweighted curves are very similar,
confirming that the size effect is mostly at work in the early recoveries. The frequency
of cumulative recoveries on unpaid balance at times 4 and 25 are reported in
Figures 5a and b. These show again a bipolar distribution of cumulative recoveries on
outstanding balance at time 4, but that there are few cases of full recovery on
balances outstanding at time 25.
Loan-Loss Provisioning: Central Bank Mandatory Provisions vs. Mortality-based
Provisions
In Table 3, we report the provisions rules of the Bank of Portugal for bad and doubtful
loans for applications other than consumer credit or residence credit. The 1%
provisions rule for time 0 to 3 months after default reflects the central banks
policy of imposing a conservative general provisions of 1% on all outstanding bank
loans in Portugal. A first observation is that specific provisioning on bad and doubtful
loans starts 3 months after the default date, that is, two months after the loan is
reported to the Bank of Portugal. A second observation is that provisioning is
increasing through time with 100% provision required at month 13 for loans with no
personal guarantee/collateral, at month 19 for loans with personal guarantee, and at
month 31 for loans with collateral. A third observation is an implicit assumption,
according to which recoveries on loans without guarantee/collateral are the lowest at
all points in time. It is quite interesting to compare the regulatory provisioning
schedule of a central bank, from the one that could be inferred from the empirical
estimates of cumulative recoveries.
In Figures 6a,b,c, we report the provisioning schedule of the central bank several
months after the default date and the provisioning schedule implied by the estimates
reported in Table 2, for loans with no guarantee/collateral, loans with personal
guarantee, and loans with collateral support. A first observation is that the regulatory
practice of enforcing specific provisions three months after the default event does not
seem justified by the data. Provisioning should start at the time of the default event.
A second observation is that the provisioning on loans without guarantee/collateral is
excessively conservative. It calls for 100% provisioning 13 months after the default
date, when our data shows an expected (weighted) recovery of 66.7%. A third
8/2/2019 Loan Losses Provision
17/36
16
observation is that the provisioning on loans with personal guarantee is too
optimistic. It calls for provisioning of 50%, 13 months after the default date, when
the data shows an expected recovery of only 20.5%. It is only in the case of loans
with collateral support that the divergence between the estimated and the regulatory
schedule is minimal.
Section Five. Loan-Loss Provisioning, a Multivariate Analysis
In the previous section, univariate mortality-based estimates of cumulative recoveries
and provisions were provided. In this section, we attempt to estimate empirically the
determinants of the cumulative recovery rates and dynamic provisions. A discussion
of the choice of explanatory variables and the econometric specification is followed
by the empirical results.
Explanatory variables and econometric specification
Explanatory variables include the size of the loan, the type of guarantee/collateral
support, past cumulative collection, and the default year.14 The size of the loan is
included because some empirical studies and the sample univariate weighted and
unweighted average cumulative recovery data have pointed out the effect of the loan
size. Past cumulative collection is included on the assumption that a good level of
collection could indicate a genuine effort by the borrower to repay the loan fully. A
year dummy is included so as to have a better understanding of the volatility of the
recovery rate over time. In the case of Banco Comercial Portugus, it was indicated
by the bank that a reorganization of the workout unit could have an impact on
recoveries. Finally, it is of interest to know the impact of guarantee/collateral, as, if
statistically significant, this variable could be taken into account in calculating loan-
loss provisions on bad and doubtful loans.
14 As we had access to information on the banks internal rating on a subset of loans,we also attempted to test the impact of ratings with this subset of data. As the rating
explanatory variable was found not to be statistically significant, the results are notreported. Similarly, the fifteen industrial sectors variables were not significant.
8/2/2019 Loan Losses Provision
18/36
17
The size of loan excepted, the explanatory variables will be represented by dummy
variables. The dependent variable, the cumulative loan recovery rate, is a continuous
variable over the interval [0-1]. Due to the boundaries of the dependent variable, one
cannot use the ordinary least square (OLS) regression,
1 2 2( | ) . ... .k kE y x x x x = + + + = (1)
as it cannot guarantee that the predicted values from the model will lie in the bounded
interval (Greene, 1993). A common econometric technique is to use a transformation
G(y) that maps the [0-1] interval onto the whole real line [- , + ] (McCullagh and
Nelder, 1989). For instance, one can model the log-odds ratio transformation as a
linear function of explanatory variables. In other words, instead of having y as the
dependent variable, the model uses the following transformation of y:
log1
yy
y
and the regression becomes:
log1
yE x
yx
=
where x is a k*1 vector of explanatory variables, and is a k*1 vector of slope
coefficients. The drawback of the log-odds ratio transformation is that it is not defined
for the dependent variable, y, taking values of 0 or 1, so that an approximation must
be used. Since our data set includes a sizeable series of cases with extreme loan
recoveries of 0% or 100%, the log-odds transformation is not an attractive solution.
Instead, we follow Papke and Wooldridge (1996) who studied econometric methods
for fractional response variables in the context of employee participation rates in 401
(k) pension plans. A model is specifically developed to consider the fact that the
dependent variable, y, is bounded between 0 and 1, with positive probabilities at the
extremes.
Assuming a sequence of observations (xi, yi ; i=1,2,N), where 0y1 andNis the
sample size:
8/2/2019 Loan Losses Provision
19/36
18
1 2 2( | ) ( . ... . ) ( )i i k k iE y x G x x G x = + + + = (2)
where G(.) is a known function satisfying 0
8/2/2019 Loan Losses Provision
20/36
19
of provisioning. Second, the coefficient of collateral is positive at all time intervals,
although statistically significant at two horizons only, 4 and 7 months. Finally, and
most interestingly, is the sign of the past recovery variable. It is positive and
statistically significant at all horizons indicating that past recovery is a good indicator
of future recovery.
To assess further the magnitude of the impact of past recovery on fair loan-loss
provision, we use the estimated regression to simulate the level of fair provisions for
different levels of past recovery. These are reported in Table 5. If, for instance, one
considers the balance outstanding seven months after the default date on a loan with
no guarantee/collateral, one observes that the level of provisions should be 53.2%
when past recovery has been zero, but that this provisions falls to 16.5% in the case of
a past recovery of 50%. Although one would need additional empirical research to
confirm these results, it appears that, for this sample of loans, past recovery has a
significant impact on loan-loss recoveries and on the level of fair loan-loss provisions.
Section Six. Conclusions
A fair level of provisions on bad and doubtful debt is an essential part of capital
regulation and bank solvency. Micro data on recovery overtime on bad and doubtful
loans allows us to provide two methodologies to compute dynamic provisions. First a
univariate mortality-based approach allows us to compute provisions for three classes
of loans: no guarantee/collateral, personal guarantee only, and collateral with or
without guarantee. A multivariate approach facilitates analyzing more precisely the
determinants of loan recoveries and provisions over time. Three significant results are
as follows. First, bad and doubtful loans with no guarantee/collateral exhibit better
recoveries than loans with personal guarantee. This could be due to the fact that the
decision to lend without guarantee took into account the higher expected recovery
rates. Second, the past recovery history has a highly significant positive impact on
future recovery. Third, a comparison with the Bank of Portugal mandatory
provisioning rules indicates some conservatism in calling for 100% provision 31
months after the default date, when, in fact, significant amounts are still recovered
after that date. But, more stringent provisions could be enforced in the shorter run,
8/2/2019 Loan Losses Provision
21/36
20
before the 90-day- provisioning trigger date. Additional empirical studies are
needed to validate the empirical findings of the paper, but the two methodologies
presented in this paper provides a basis to compute fair loan-loss provisions on bad
and doubtful bank loans.
8/2/2019 Loan Losses Provision
22/36
21
Table One: Descriptive Statistics for the Sample of Doubtful Loans16
Panel A: Number of Defaults per yearNumber of cases
1995 651996 891997 591998 57
1999 472000 54
Total 371
Panel B: Debt Outstanding at the Time of Default (Euros)
Number ofobservations
Percentage
0
8/2/2019 Loan Losses Provision
23/36
22
Table 2: Mortality-based estimate of cumulative recoveries on loan balances outstanding n months af
Total sample No Collateral /
No Guarantee
Personal Guarantee only
Number of
months after
default (n)
Unweighted Weighted Unweighted Weighted Unweighted Weighted
0 72.6% 63.8 % 80.6% 80.0% 63.9% 45.9%
4 57.8% 55.7% 69.6% 75.3% 46.8% 33.4%
7 50.2% 47.7% 63.3% 70.8% 38.5% 27.0%
13 41.03% 42.1% 56.6% 66.7% 27.6% 20.5%
19 33.0% 35.8% 47.0% 59.4% 20.8% 13.6%
25 27.1% 32.2% 41.3% 56.7% 16.3% 11.6%
37 17.1% 23.3% 28.3% 43.2% 8.6% 6.9%
Note: These are cumulative recovery rates up to month 68, calculated on the loan balance outstanding n mont
8/2/2019 Loan Losses Provision
24/36
23
Table 3: Bank of Portugal Provisions Rules for Applications other than
Consumer Credit or Residence Credit
Risk Class Time period If the credit has no
guarantee/collateral
If the credit has
a personal
guarantee
If the credit
has a real
guarantee
Class I From 0 to 3 months 1% 1% 1%
Class II From more than 3
months to 6 months
25% 10% 10%
Class III From more than 6
months to 9 months
50% 25% 25%
Class IV From more than 9
months to 12 months
75% 25% 25%
Class V From more than 12
months to 15 months
100% 50% 50%
Class VI From more than 15
months to 18 months
100% 75% 50%
Class VII From more than 18
months to 24 months
100% 100% 75%
Class VIII From more than 24
months to 30 months
100% 100% 75%
Class IX From more than 30
months to 36 months
100% 100% 100%
Class X From more than 36
months to 48 months
100% 100% 100%
Class XI From more than 48
months to 60 months
100% 100% 100%
Class XII More than 60 months 100% 100% 100%
Source: Aviso de Banco de Portugal n 8/2003, Dirio da Repblica- I Srie-B, N 33,
8 February 2003.
8/2/2019 Loan Losses Provision
25/36
24
Table 4: Statistical estimate of cumulative recoveries on loan balances outstanding n months after defa
(The estimate of the parameter is followed by the p-value)
Number of Months after Default (n)
4 7 13 19
Constant 0.39
(0.115)
0.55
(0.049*)
0.28
(0.302)
0.04
(0.858)
Past Recovery 5.23
(0.000*)
2.88
(0.001*)
2.34
(0.00*)
1.88
(0.000*)
Loan Size 0.02
(0.911)
0.02
(0.962)
- 0.285
(0.604)
- 0.58
(0.176)
Year Dummy 0.10
(0.666)
-0.28
(0.226)
-0.22
(0.324)
-0.19
(0.416)
Personal Guarantee -0.47
(0.048*)
-0.67
(0.005*)
-0.81
(0.001*)
-0.73
(0.002*)
Collateral 1.19
(0.002*)
0.92
(0.038*)
0.805
(0.078)
0.61
(0.186)
R2 0.286 0.26 0.339 0.359
CHSQ 35.04
(0.000
*
)
28.74
(0.000
*
)
38.0
(0.000
*
)
32.9
(0.000
*
)Reset - 1.88
(0.05*)
-1.30
(0.191*)
-0.83
(0.404)
0.14
(0.883*)
Number of 125 112 96 85
8/2/2019 Loan Losses Provision
26/36
25
observations
*: Significant at the 5% level.
Note: The table presents the result of the log-log regression. The dependent variable is the cumulative recoveoutstanding at 4, 7, 13, 19 , 25, and 37 months after the default date. The explanatory variables include a coninitial loan already recovered, the loan size, a dummy for personal guarantee, and collateral. Because of datarecovery is calculated up to 48 months after the default date. This is unlikely to create a bias as Figure Two inachieved 48 months after the default date.
8/2/2019 Loan Losses Provision
27/36
26
Table 5: Multivariate-estimate of provisions vs. Central bank (CB) mandatory provisioning (in percent
Period No Guarantee Personal Guarantee
Level of past recovery Level of past recovery
CB 0% 20% 50% CB 0% 20% 50% CB
4 25 44.8 18.8 4.2 10 61.6 28.5 6.7 10
7 50 53.2 34.8 16.5 25 77.5 56.8 29.8 25
13 100 62.4 45.7 26.1 50 89.0 74.8 49.5 50
19 100 71.1 57.3 38.3 100 92.5 83.0 63.5 100
25 100 81.2 71.2 55.0 100 96.1 91.0 78.7 100
37 100 95.2 89.8 77.4 100 99.9 99.4 96.5 100
Note: The level of provisioning (provision = 1 cumulative future recovery) is estimated with the regr
8/2/2019 Loan Losses Provision
28/36
27
Figure One: Sample Unweighted Marginal Recovery Rate at time t+n
(SMRRt+n)
0%
5%
10%
15%
20%
25%
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65
Time (number of months after default)
Note: This figure presents the marginal recovery n- months after default. The
mortality-based approach is used to calculate the marginal recoveries.
Figure Two: Sample Unweighted and Weighted Cumulative Recovery Rate
at time t+n (SCRRt+n) on Balances outstanding at time 0
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65
Time (number of months after default)
Weighted Unweighted
Notes: The figure presents the cumulative weighted and unweighted recovery
rates n- months after default. They have been calculated with the mortality-based
approach.
8/2/2019 Loan Losses Provision
29/36
28
Figure Three: Distribution of individual loan cumulative recovery rates on
balances outstanding at time 0.
0%
10%
20%
30%
40%
50%
60%
0%-5%
5%-10%
10%-15%
15%-20%
20%-25%
25%-30%
30%-35%
35%-40%
40%-45%
45%-50%
50%-55%
55%-60%
60%-65%
65%-70%
70%-75%
75%-80%
80%-85%
85%-90%
90%-95%
95%-100%
Recovery rates
Loans
frequency
0%
10%
20%
30%
40%
50%
60%
0%-5%
5%-10%
10%-15%
15%-20%
20%-25%
25%-30%
30%-35%
35%-40%
40%-45%
45%-50%
50%-55%
55%-60%
60%-65%
65%-70%
70%-75%
75%-80%
80%-85%
85%-90%
90%-95%
95%-100%
Recovery rates
Loans
frequency
Note : The figure presents the frequency of cumulative recovery on individual loans,
48 months after default.17
17
Due to data limitation (five years), the cumulative recovery is calculated up to 48 months afterdefault. This does not seem too restrictive as Figure Two indicates that most of the recovery is achieved48 months after default.
8/2/2019 Loan Losses Provision
30/36
29
Figure Foura Cumulative recoveries as a % of loan outstanding in month 4
Time 4
0%
10%
20%
30%
40%
50%
60%
70%
4 10 16 22 28 34 40 46 52 58 64
Number of months after default
Total Recovery
Weighted
Total Recovery
Unweighted
Figure Fourb Cumulative recoveries as a % of loans outstanding in month 25
Time 25
0%
10%
20%
30%
40%
50%
60%
70%
25 28 31 34 37 40 43 46 49 52 55 58 61 64 67
Number of months after default
Total Recovery
Weighted
Total Recovery
Unweighted
8/2/2019 Loan Losses Provision
31/36
30
Figure Fivea Frequency of cumulative recoveries on loans outstanding 4 months
after the default date.
0%
5%
10%
15%
20%
25%
30%
35%
40%
0
%-5%
5%
-10%
10%
-15%
15%
-20%
20%
-25%
25%
-30%
30%
-35%
35%
-40%
40%
-45%
45%
-50%
50%
-55%
55%
-60%
60%
-65%
65%
-70%
70%
-75%
75%
-80%
80%
-85%
85%
-90%
90%
-95%
95%-100%
Recovery rates
Loans
frequency
Figure Fiveb Frequency of cumulative recoveries on loans outstanding 25
months after the default date.
0%
5%
10%
15%
20%
25%
30%
35%
40%
45%
50%
0%-5%
5%-1
0%
10%-1
5%
15%-2
0%
20%-2
5%
25%-3
0%
30%-3
5%
35%-4
0%
40%-4
5%
45%-5
0%
50%-5
5%
55%-6
0%
60%-6
5%
65%-7
0%
70%-7
5%
75%-8
0%
80%-8
5%
85%-9
0%
90%-9
5%
95%-10
0%
Recovery rates
Loans
frequency
Note : The figure presents the frequency of cumulative recovery on individual loans,
48 months after default.
8/2/2019 Loan Losses Provision
32/36
31
Figure Sixa: Comparison of empirical (weighted) estimates of loan-loss
provisions to those imposed by the central bank, for the loans without collateral
or personal guarantee.
No collateral/ no guarantees
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
0 3 6 9 12 15 18 21 24 27 30 33 36 39 42 45 48
Bank of Portugal Provisioning rules Estimated Provisioning
Figure Sixb : Comparison of empirical estimates to those imposed by the central
bank for the loans with personal guarantee
Personal guarantee only
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
0 3 6 9 12 15 18 21 24 27 30 33 36 39 42 45 48
Bank of Portugal Provisioning rules Estimated Provisioning
8/2/2019 Loan Losses Provision
33/36
32
Figure Sixc: Comparison of empirical estimates to those imposed by the central
bank for the loans with collateral (with or without personal guarantee)
Collateral with or without guarantee
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
0 3 6 9 12 15 18 21 24 27 30 33 36 39 42 45 48
Bank of Portugal Provisioning rules Estimated Provisioning
8/2/2019 Loan Losses Provision
34/36
33
References:
A.S. Ahmed, C. Takeda, and S. Thomas (1999): Bank Loan-Loss Provisions: aReexamination of Capital Management, Earnings Management, and SignalingEffects, Journal of Accounting and Economics, 28, 1-25.
Allen L., G. DeLong, and A. Saunders (2003): Issues in Credit Risk Modeling ofRetail Markets, mimeo, 1-46.
Altman E. (1989): Measuring Corporate Bond Mortality and Performance, TheJournal of Finance, 44, 909-922.
Altman E. and H. Suggitt (2000): Default Rates in the Syndicated Bank LoanMarket: A Mortality Analysis, Journal of Banking and Finance, 24.
Asarnow E. and D. Edwards (1995): Measuring Loss on Defaulted Bank Loans: a
24-Year Study, The Journal of Commercial Lending, 11-23.
Basel Committee on Banking Supervision (1999): Sound Practices for LoanAccounting and Disclosure, 1-44.
Basel Committee on Banking Supervision (2003): The New Basel Capital Accord.
Barth M., W. Beaver, and M. Wolfson (1990): Components of Earnings and theStructure of Bank Share Prices, Financial Analysts Journal , May-June, 53-60.
Beattie V.A., P.D. Casson, R.S. Dale, G.W. McKenzie, and M.J. Turner (1995):Banks and Bad Debts, Accounting for Loan-Losses in International Banking,Chichester: John Wiley & Sons.
Berger A., K. Kuester, and J. OBrien (1989): Some Red Flags Concerning MarketValue Accounting, Finance & Economics Discussion series, 85, Federal ReserveBoard, 1-31.
Berger A., K. Kuester King , and J. OBrien (1991): The Limitations of market valueAccounting and a More Realistic Alternative, Journal of Banking and Finance,15,753-783.
Caouette J., Altman E., J. and P. Narayan (1998): Managing Credit Risk: The Next
Great Financial Challenge, New York: John Wiley a Sons.
Carey M. (1998): Credit Risk in Private Debt Portfolios, The Journal of Finance,
LIII, August, 1363-1387.
Carty, L. and D. Lieberman (1996): Defaulted bank Loan Recoveries, Moodys
Investors Service, November.
8/2/2019 Loan Losses Provision
35/36
34
Crouhy M., D. Galai and R. Mark (2000): A Comparative Analysis of Current CreditRisk Models, Journal of Banking and Finance, 24, 59-117.
Degeorge F., J. Patel, and R. Zechauser (1999): Earnings Management to ExceedThresholds,Journal of Business, 72, 1-34.
Diamond D.W. (1984): Financial Intermediation and Delegated Monitoring,Review of Economic Studies, 393-414.
Docking D.S., M. Hirschley, and E. Jones (1997): Information and Contagion Effectsof Bank Loan-Loss Reserve Announcements, Journal of Financial Economics, 43,219-239.
Frye J. (2000a): Collateral Damage, Risk, April, 91-94.
Frye J. (2000b): Depressing Recoveries, Risk, November, 108-111.
Frye J. (2003): A False sense of Security, Risk, August, 63-67.
Galai D., E. Sulganik, and Z. Wiener (2003): Accounting Values versus MarketValues and Earnings Management in Banks, mimeo, Hebrew university ofJerusalem, 1-26.
Genay H. (1998): Assessing the Condition of Japanese Banks: How Informative areAccounting Earnings ?, Economic Perspectives, Federal Reserve Bank of Chicago,12-34.
Gourieroux C., A. Monfort and A. Trognon (1984): Pseudo-Maximum LikelihoodMethods : Theory, Econometrica, 52, 681-700.
Grammatikos T. and Sauders A. (1990): Additions to Bank Loan-Loss Reserves:Good News or Bad News, Journal of Monetary Economics, Vol. 25, 289-304.
Greene W.H. (1993): Econometric Analysis, Second edition, MacMillan.
Gunther J.W. and R.R. Moore (2003): Loss Underreporting and the Auditing Role ofBank Exams, Journal of Financial Intermediation, 12, 153-177.
Hasan I. And L.D. Wall (2004): Determinants of the Loan Loss Allowance: SomeCross-country Comparisons, Financial review, 39, 2004.
Hurt L. and Felsovalyi A. (1998): Measuring Loss on Latin American DefaultedBank Loans, a 27-Year Study of 27 Countries, The Journal of Lending and Credit
Risk Management, October.
Jackson P. and D. Lodge (2000): Fair Value Accounting, Capital standards, ExpectedLoss provisioning, and Financial Stability, Financial Stability Review, June 105-125.
8/2/2019 Loan Losses Provision
36/36
35
Jimenez G. and J. Saurina (2002): Loan Characteristics and Credit Risk, Bank of
Spain, 1-34.
Kim M.S. and W. Kross (1998): The Impact of the 1989 Change in Capital Standards
on Loan-Loss Provisions and Loan Write-offs, Journal of Accounting andEconomics, 25, 69-99.
Laeven L. and G. Majnoni (2003): Loan-Loss Provisioning and EconomicSlowdowns: too Much, too Late ?, Journal of Financial Intermediation, 12, 178-197.
La Porta R., Lopez-deSilanes F., and G. Zamarripa (2003): Related Lending, TheQuarterly Journal of Economics, February, 231-268.
Musumeci J. and J.F. Sinkey (1990): The International Debt Crisis and Bank Loan-Loss Reserve Decisions: The Signalling Content of Partially Anticipated Events,
Journal of Money, Credit and Banking, 22, 370-387.
Neto de Carvalho C. and J. Dermine (2003): Loan-losses-Given-Default, EmpiricalEvidence, mimeo, 1-30.
Papke L.E. and J. M. Wooldrigdge (1996): Econometric Methods for FractionalResponse Variables with an Application to 401 (K) Plan Participation Rates, Journalof Applied Econometrics, 11, 619-632.
Ramsey J.B. (1969): Tests for Specification Errors in Classical Linear Least SquaresRegression Analysis, Journal of the Royal Statistical Society, Series B 31, 350-371.
Scholes M.S., G.P. Wilson, and M.A. Wolfson (1990): Tax Planning, RegulatoryCapital Planning, and Financial Reporting Strategy for Commercial Banks, Reviewof Financial Studies, 3, 625-650.
Soares de Pinho, P. (1996): "An Empirical Analysis of the Determinants of theProvisions for Loan Losses in Portuguese Banks", Economia Vol. XX, Janeiro -Maio - Outubro , Universidade Catlica Portuguesa.
Wahlen J.M. (1994): The Nature of Information in Commercial Bank Loan-loss
Disclosures, The Accounting Review, 69 (3), 455-478.
Wall L. and T. W. Koch (2000): Bank Loan-Loss Accounting: A Review of theTheoretical and Empirical Evidence, Economic Review, Federal Reserve Bank ofAtlanta, 1-19.
Wetmore J. L. and J.R. Brick (1994): Loan-loss Provisions of Commercial Banks
and Adequate Disclosure, a Note, Journal of Economics and Business, 46, 299-305.