_II_JS
Journal of Philippine Development Number34, VolumeXIX, No. 1, FirstSemester1992
P_
IDENTIFICATION AND CLASSIFICATION OF PROBLEM BANKS Gilbetto.M. Llantoand Aniceto C. Orbeta,Jr.'
INTRODUCTION The 1980s proveda difficulttime for local banks, principally the small unit banks, when several of them shut down. To this day, a number are still in various stages of receivership and liquidation. Monitoring and evaluation of bank performance are principal concerns of bank regulators because of the financial system's critical role in development. Careful monitoring of the financial condition of banks allows regulators to take precautionary action againstthe possibility of financial instability due to bank failures. For this reason, a system -- commonly called "early warning system ''2(EWS) -- has gained popularity and recognition among bank regulators as an aid in bank surveillance (Bovenzi 1983, Korobow and Stuhr 1983, West 1985). Initial attempts at constructing an early warning system consisted of establishing critical financial ratios which picture the state of health of a particular bank, and making corresponding judgments based on relative values of such ratios..More recent failure prediction models have taken advantage of developments in estimation techniques and have used more sophisticated statistical methods to warn regulators of an impending bank failure. The EWS, currently used by the Supervision and Examination Sector Department III of the Central Bank of the Philippines, consists of: (a) a set of financial ratios with defined acceptability levels and (b) a description of compliance to certain Central Bank (CB) rules, regulations and directives as well as banking laws, A bank is categorized as "strong", "average" or "weak" based on actual values of its financial ratios and compliance with banking laws, rules and regulations. Table 1provides these ratios and their critical or acceptable values.
'Executive Directorand Deputy ExecutiveDirector, AgriculturalCredit Policy Council, respectively. =Thistermshouldnotbe confusedwiththe 'leading indicators system'inthe senseof over time analysesof variablessincethese modelsprimarilyuse cross-section data.
124
JOURNAL OF PHILIPP;NE DEVELOPMENT
One difficulty, however, with this EWS is thatthe procedure assigns no weights to financial ratios and other criteria to show their relative importance as early warning measures. The current EWS assumes each of the criteria is as important as the other. This assumption implies the probability of not satisfying one criterion is correlated with the likelihood of not satisfying the other criteria, and vice-versa (Lamberte 1987). The weakness of this approach lies in the danger of ignoring the performance of a particular bank with respect to other criteria so long as performance according to one criterion has been ascertained. There is, therefore, no systematic way of determining which criterion or set of criteria (financial or otherwise) is critical or important in assessing bank performance. This approach will naturally lead to classification problems. One solution to this problem and the risk of misclassifying a particular bank is to employ formal statistical techniques. Such techniques identify the variable or variables critical for evaluating bank performance. They are also able to classifybanks according to performance categories which, as presumed in the CB's EWS, will predict bank performance as well.3 These statistical techniques have become popular in the developed countries. They are used not only as an EWS for banks but also as a method for predicting debt default, performance of manufacturing firms and consumer-borrower behavior? This paper reports the empirical results of an attempt to develop a statistical classification system for rural banks, using common financial ratios as predictor variables. Specifically, it tests several statistical techniques for classifying problem and non-problem banks and discusses applicable early warning techniques for rural banks. The statistical techniques used are the following: (a) multiple discriminant analysis and (b) multinomial Iogit analysis. Discriminant analysis searches for a mathematical rule or "discriminant function" which classifies an observation into known classes. The rule is based on a combination of quantitative variables that best reveals the difference among the stated categories. Multinomial Iogit analysis, on the other hand, fits a linear model with the transformation of the classification variable as the dependent variable
_The use of current financial ratios in developing classification models may fail to consider the role of previous financial position of banks in determining their current status. In addition, the economic environment may have a role in determining the current financial state of banks. These are valid points all models using cross-section data face since they depend only on the properties of the distributionof variables at a point in time rather over time for predictive perfomance. These factors shall be considered as weaknesses of the methodology being presented. However, unless a full model of bank behavior over time is developed, one could not readily introduce these concerns in bank classification models. 4A review of the literature, which includesthe different statistical techniques and their relative merits, is made in Llanto and Suleik (1988). Because of space limitations, the review is omitted in this paper.
LLANTOAND ORBETA:PROBLEMBANKS
RATING
125
Table 1 PROFILE FOR BANKS
CRITERIA
RATING Strong
Average
Weak
risk asset ratio (b) past due ratio
over 10%
over 10%
below 10%
25% and below
over 25% to 50%
over 50%
(c)
complied with
deficient in minimum required capital
deficient in minimum required capital
deficient but not chronic with arrears
deficient and chronic with arrears
alternate profit and loss but
alternate profit and loss but
average results in profit
average results in loss
none
none
serious
none none
not serious not serious
serious serious
1. Solvency (a)
2.
minimum paid in capital
Liquidity (a)
legal reserve deficiency
(b) arrearages to CB
none
3. Profitability
4. Management
none
continuous profit
- violations
(a) laws (b) rules and regulations (c) CB directives
of:
Source: Supervision and Examination Sector III, CentralBankof the Philippines.
126
JOURNALOF PHIUPPINE DEVELOPMENT
given a vector of predictor variables, The fittedmodelserves as the basis for classifyingobservationsintodifferentcategories. This paper hasfour sections:Followingthis introductorysectionis a discussionof the statisticalmodels, methodologyand data used in this paper.The empiricalresultsare presented in the thirdsection. Conclusionsabouttheusefulnessof EWS for banksbasedonthisexperimentare containedin sectionfour. STATISTICAL MODELS, METHODOLOGY AND DATA Methodology Classifyingbanks intospecificcategories requiresresolutionofthree issues.The firstisnatureof classification.As notedin Table 1,the Department of Supervision and Examination of the CB uses a three-bank category method, namely: (a) strong, (b) average, and (c) weak. To classify a bank into one of these three categories, the method uses a combinationof financial ratios and judgmentalfactors based on personal examination of bank performance. We follow this three-category classification of banks,s A variant of this, a two-category classification system, was also studied. The second issue concerns the characteristics of banks relevant in assessing their viability. The CB uses both financial ratios and the evaluation of examiners from a direct personal examination of a bank. An impersonal method based solely on financial ratios is presented in this paper, This approach is functional and simple. The third issue pertainsto the selection ofthe method of classification. We usetwo statistical methods, namely: (a) multiple discriminant analysis (MDA), and (b) multinomial Iogit analysis (MLA). The MDA is noted to perform better than the MLA where the independent variables are multivariate normal, The MLA, on the other hand, is known to be more robust to non-normally distributed predictor variables (Maddala 1983). For the MDA, if the within-classification covariance matrix of the predictor 5The paperdoesnotassumethe correctnessof CentralBank'sjudgmentandclassification of the banks.The CB'sclassificationis basedboth on a personaljudgmentmade by the bank examinersand a considerationof relevantvariablesin the balancesheet and profitand loss statements.Thepaperattemptsto showstatisticaltechniquescanmakean impersonalclassificationbysimplyusingfinancialdatareportedbybankstothe CB. Thepaper,however,takesthe classificationmade by CB as a standard for evaluatingthe classificationperformanceof the statisticalmodels.Sincethereisnoknownevaluationofthe performanceofCB'sEWSinpredicting actualbank pedormance,no absoluteevaluationcan also be conductedfor the modelsbeing presented. An attemptwasalsomade bytheauthorstoclassifyorclusterbanksintodistinctcategories withoutreferencetotheofficialCBclassification.Using*,heprincipalcomponentsofvariables,we failedto producegoodclustersbothinthe two-or three-bankcategories.Almostall bankshave beenclassifiedintoa singlecluster,We usedthe rawvaluesof the financialratiosandobtained similarresults.
LLANTO AND ORBETA: PROBLEM BANKS
127
variables is not equal across categories, the quadratic discriminant function rather than the linear discriminant function is appropriate. The paper initially follows the three-category classification of the CB. Results show that relative to CB's EWS classification output the MDA performs better in classifying strong banks while the MLA performs well in classifying the weak banks. Given these initial results, we attempted atwocategory classification: problem and non-problem banks. Banks were classified as non-problem banks if the CB examination classified them as strong;, otherwise they were considered as problem banks, Results are discussed below. Description of the Data We used balance sheet and profit and loss data of 867 rural banks as ofthe end of 1986. The banks that yielded invalid ratios-- mainly because of zero values for the denominator in the financial ratios -- were dropped. The final data set came only from 816 rural banks classified by the CB as weak (495), average (245), and strong (76), In the two-category classification, 76 were considered non-problem and 740 problem banks. The Predictor Variables Sixteen financial ratios were initially considered as candidate predictor variables --five capital adequacy/solvency ratios, six liquidity ratios and five profitability ratios, Table 2 lists these financial ratios and their definitions. A multivariate ANOVA test using the 16 candidate predictor variables was performed to determine if the bank categories really come from different populations, i.e. strong, average and weak or problem and nonproblem. This provided a basis for developing and testing statistical models for bank classification using the predictor variables. The Multiple Discriminant
Analysis
If the data are from a random sample of a population, knowledge of the probability that a bank belongs to a certain category can improve the performance of the classification model (Maddala 1983). This study virtually utilizes the whole population of rural banks. We used two methods to select predictor variables. The first method is the ANOVA test for the difference of means. Variables that exhibit significantly different means are deemed useful as predictor variables. The other method is stepwise discriminant analysis. Variables were selected according to a significance level of an F-test from an analysis of co-variance.
128
JOURNALOFPHlUPPINEDEVELOPMENT
Table 2 FINANCIAL RATIOS AND THEIR DEFINITIONS
Ratio
Description
Formula
A. Capital Adequacy/Solvency Total Capital (accounts) 1,
NWDEP
Net Worth/Deposits Total Deposit Liabilities Total Capital (accounts)
2.
NWBOR
Net Worth/Borrowings Total Borrowings Total Capital (accounts)
3,
NWDAB
Net Worth/Deposits & Borrowing
Total Deposits+Total Borrowings Total Capital (accounts)
4,
NWASS
Net Worth/Assets Total Assets Total Liabilities
5.
DEQTY
Debt/Equity ,
Capital Accounts
B. Liquid#j{ Cash and Due from Banks 6.
LAR
Liquid Assets Ratio Total Deposit Liabilities
7.
CAL
Current Assets/ Current Liabilities
Cash and Due from Banks + Loans & Discounts + Investment Total Deposit Liabilities + Bills Payable Cash on hand and due from banks + Investment in bonds & other debt instruments
8.
TRTA
Total Reserves/ Total Assets
Total Assets
LLANTOAND ORBE-TA: PROBLEMBANKS
129
Table 2 (Continued) Ratio
Description
Formula
Total Loans and Discounts (net) 9.
LD
Loans/Deposits Total Deposit Liabilities Total Loans and Discounts
10. LNW
Loans/Net Worth Total Capital Accounts Total Loans and Discounts
11. LB
Loans/Borrowings Total Borrowing
C. Profitability Net Inco me 12. NYGY
Net Income/ Gross income
Total Income Net Income
13, NYNW
Net Income/Net
Worth Total Capital Accounts Net Income
14, NYA
Net Inco me/Assets Total Assets Net Income
15. NYE
Net Income/Expenses Total Expenses
16. TCRA
Total Capital/Risk Assets
Net Worth Total Assets-Non-R_sk Assets where: Non-risk assets = cash on hand from CB + Investment in securities + Bank premises, furniture, fixtures and equipment
The univariate ANOVA test of all the 16 variables yielded nine with significantly different means among the bank categories. These are NWDEP, NWBOR and NWDAB (capitaladequacy variables); CAL, TRTA, LD (liquidity variables) and NYGY, NYA and NYE (profitability ratios). The
130
JOURNAL OF PHILIPPINE DEVELOPMENT
stepwisemethodselected9 predictorvariables.Three are capitaladequacy/ solvency variables (NWDEP,NWDAB, NWASS),three are liquidityvariables (LAR, TRATA, LD), and three are profitability variables (NYA, NYE, TCRA). Table 3 provides details of the test resultsfor both three- and twocategory models. The actual proportion of rural banks classified by the CB into the different categories may be used as a prior probability estimate that a bank belongs to a specific category. This knowledge may improve the classification performance of the MDA. To ascertain whether this knowledge improves the classification performance, models with and without this information were estimated. The model not using this information performed better. There is no prior information to justify the assumption of homogeneity of within-classification covariances of predictor variables among bank categories. It was therefore necessary to test for homogeneity of withinclassification covariances. If covariances are not homogeneous,then a quadratic multiple discriminant function is used, otherwise, a linear discriminant function is appropriate. As an example, for the three-category model using ANOVA variables and equal probabilities, the test yields a chi-square value of 3342.82 with 90 degrees of freedom. This is significant at a 99 percent level of confidence, indicating a non-homogeneous within-classification covariance. Thus, a quadratic discriminant function is more appropriate. We therefore used the following quadratic discriminant models for the classification of a rural bank given itsfinancial ratios; where xj isthe vector of financial ratios,_ isthe vector of sample means for each category j, and the covariance matrix of the financial ratios for each category j is COVj. Generalized Squared Distance Function D2i(x)= (x- x-)l'COVl" (x-_j)+ InlCOVil
(1)
Posterior Probability of Membership in Each Category (k) Pr (jlx)= exp
(-.5D_i
(X))/ sum,exp (-.5 D_,(x))
(2)
The classification rule is to consider a rural bank a member of category ik) where the computed generalized distance function (DZ_is smallest (minimum), or alternatively, where the probability of being in category j, Pr (jlx) is largest (maximum).
LLANTOAND.ORBETA:PROBLEMBANKS
PREDICTOR
1.31
Table 3 VARIABLES
SELECTED
• 3 CATEGORIES Variable
Description
2 CATEGORIES
ANOVA STEPWlSE* ANOVA STEPWlSE*
A. CAPITAL ADEQUACY/SOLVENCY 1, NWDEP 2.
NWBOR
3, NWDAB 4. NWASS 5. DEQTY
Net Worth/ Deposits Net Worth/ Borrowings Net Worth/ Deposits/Borrowings Net Worth/Assets Debt/Equity
*
5
** * 6
2
* 3
B. LIQUIDITY 6.
LAR
7, CAL 8. TRTA •9. LD 10, LNW 11. LB
Liquidity Assets Ratio Current Assets/ Current Liabilities Total Reserves/ Total Assets Loans/Deposits Loans/Net Worth Loans/BorrowingS
9
4
* *
8
*
**
3
**
2
C. PROFITABILITY 12. NYGY 13. NYNW 14. NYA 15. NYE 16. TCRA
Net Income/ Gross Income Net Income/ Networth Net Incomes/ Assets Net Income/ Expenses Total Capital/ Risk Assets
*
*
*
4
*
*
1
*
* Significantat the 1 percentlevel. ** Significantat the 10 percentlevel. Sequencenumber of variables selected,
7
1
5
132
JOURNALOF PHILIPPINEDEVELOPMENT
LOGIT STATISTICS
Variable
Coefficient
Table 4 FOR THREE-CATEGORY
MODEL
Std, Error
T-ratio
(Slg. Lvl)
Mean of X
Std. Dev. of X
.3574 ,1758E-02 .1503E-03 ,1720E-02 .2862E-02 .9773E-02 .4316EL03 .1553E-01 .9370E*01 .1053E-01
-.778 3.654 -.472 1,104 -.667 ,061 -3,922 4.564 .288 -2,952
(.43657) 01.000 000.000 (.00026) 43,629 302.470 (.63724) 105.820 618,030 (.26948) 40.366 84.647 (,50483) 121.660 52.765 (.95096) 14.626 12.369 (.00009) 419.450 632.250 (.00000) -7.4351 61.474 (.77339) 0.99510E-01 1.9468 (.00316) 4.3644 31.435
.4828 .2715E-02 ,5405E-03 .2811E-02 .2891E.02 .1403E-01 .9187E-03 .4776E-01 ,1687 .3205E-01
-4.270 1.754 -,962 .814 -.114 1.600 -2.544 3.636 1.595 -3.457
(.00002) 01.000 000.000 (.07779) 43,629 302.470 (.33584) 105.820 618.030 (.41567) 40.366 64,647 (.90886) 121.660 52.765 (.10965) 14.628 12,369 (.01096) 419.450 632.250 (.00026) -7,4351 61.474 (,11080) .99510E-01 4.9466 (.00055) 4.3644 31.435
Average Banks ONE NWDEP NWBOR NWDAB CAL TRTA LD NYGY NYA NYE
-.278058 .642306E-02 -.708982E-04 .189900E-02 -.192178E-02 ,601064E-03 -,169313E-02 .711721E-01 ,269797E-01 -,310700E-01
Strong Banks ONE -2.06161 NWDEP .478750E-02 NWBOR -.520220E-03 NWDAB .226803E-02 CAL -,330930E-03 TRTA .224476E-01 LD -.233727E-02 NYGY .173635 NYA .268969 NYE -.110786
Log-Likelihood.........................-584.66 Restricted(Slopes=0Log-L) ...-722.60 Chi-Squared(18) .....................275.89 SignificanceLevel ...................0,32173E-13
The Multinomial Loglt Model An alternative method of classifyingrural banks, given a set of predictorvariables, is the multinomialIogitmodel. Our multinomialIogit modelis givenas: Y=j = 1, wherej = 2 ifthe bankisstrongand j = 3 if the bank is average = 0 otherwise _lj = Pr (Yll = 1)
LLANTOAND ORBETA:PROBLEMBANKS
133
Table 5 CLASSIFICATION SUMMARY USING THE QUADRATIC DISCRIMINANT FUNCTION FOR THREE BANK CATEGORIES
Predicted Numbers and Percent Classified Under Bank Categories
Actual CB Classification
Average
Strong
Weak
Average
56 (22.86)
169 (68,98)
20 (8.16)
245 (100, 00)
Strong
8 (10,53)
67 (88.16)
1 (1.32)
76 (100.00)
Weak
122 (24.65)
187 (37.78)
186 (37.58)
495 (100.00)
Total
186
423
207
816
(22.79)
(51.84)
(25,37)
(100.00)
The multinomial
Iogit model is then simply _i2 log
_il
_ _T=Xl
t[i3
log
_il
- 13Ty_
where X= is the vector of predictor variables coefficients to be estimated.
and the i_T` are vectors of the
Table 4 shows the estimation results of a three-category multinomial Iogit model, using ANOVA selected predictor variables. Four variables (NWDEP, LD, NYGY, NYE) are significant at 99 percent level in predicting average banks. Three variables (LD, NYGY, NYE) are significant at 95 percent level in predicting strong banks. EMPIRICAL Three
RESULTS
Bank Categories
Table 5 shows the classification mated three-bank category quadratic
summary obtained, using the estidiscriminant function and the actual
134
JOURNALOF PHILIPPINE DEVELOPMENT
Table 6 CLASSIFICATIONSUMMARYUSINGTHE MULTINOMIAL LOGITMODEL FORTHREE BANK CATEGORIES Predicted Numbers and Per Cent Classified Under Bank Categories Weak
Actual CB Classification
Average
Strong
Average
1O0 (40,82)
3 (1.22)
142 (57.96)
245 (100.00)
Strong
37 (48.68)
2 (2.63)
37 (48,68)
76 (100.00)
Weak
54 (10.91)
2 (0.40)
439 (88.69)
495 (100.00)
Total
191 (23,41)
7 (0.86)
618 (75.73)
816 (100.00)
CB bank classification. The estimated discriminant function seems to classifythe strongbanks quite well. It is ableto properlyclassify88 percentof the strongbanks. On the other hand, there is only a very modest degree of success in correctly classifyingthe weak and average •banks.The model identifiesonly 37 percentand 23 percentof the weak and average banks, respectively. The classificationresultsfor the multinomialIogitmodelare shownin Table 6. Th$ model performswell in classifyingweak banks but is only •moderatelysuccessfulin classifyingaverage banks.Itfailsto identify the strong,banks. Eighty-eight percent of weak banks and 41 percent of average banks are successfullyclassifiedwhile only 3 percentof strong banks are successfullyidentifiedby our Iogitmodel. Table7 presentserrorcountestimatesrelativetotheCB's EWS output for the three-category models__-_under different selection methods for predictor variables and useof WJot_-probability estimates. The error counts are defined as one minus the proportion of banks correctly classified by the model. The proportions under the total column are weighted averages of the error counts for each category with the assumed prior probabilities as weights. They can be used then as scalar measures of the classification ability of the models relative to the standard -- the CB EW$.
LLANTOAND ORBETA:PROBLEMBANKS
135
Table 7 ERROR COUNT ESTIMATES (THREE BANK CATEGORIES) BANK CATEGORIES Weak
Average
Strong
Total
STEPWISE VARIABLES* Priors Equal Priors Prop ANOVA VARIABLES** Discriminant:
0.7495 0.7232
0.7918 0.7878
0.1053 0.1316
0.5489 0.6875
PriorsEqual PriorsProp. Logit
0.6404 0.6242 0.1131
0.7959 0.7714 0.5918
0.0921 0.1184 0.9737
0.5095 0.6213 0.5596
* NWDE P,NWASS,NWDAB,TCRA, TRTA,LD,LAR,NYA, NYE ** NWDEP,NWBOR,NWDAB,CAL,TRTA,LD,NYGY, NYA,NYE Two-Bank Categories Next we groupedthe ruralbanks intotwocategories:(a) problemand (b) non-problembanks,Bankswere considerednon-problem banks if the actual CB examinationclassifiedthem as "strong"banks, otherwisethey are classifiedprob/em banks. Averagebankswere lumpedwithweak banks.Bothwere considered as problem banks. This is because the average banks are borderline cases. They can have any of the followingproblems:required minimum capitaldeficiency,legalreservedeficiency,arrearages,andsomeviolation of CB rules, directivesor regulations. Again, Table 1 providesthe basis for this classification. The quadraticdiscriminant functionperformedwellfor thenon-problem (strong)bankswhilegivinga moderatelysuccessfulclassification of problem banks,Itwas abletosuccessfullyclassify95 percentof non-problembanks and only 40 percentof problembanks (Table 8). This performancewas consistentwiththe earlierresultsusingthree-bankcategories. The multinomialIogitmodel failed to classifyeven one of the nonproblembankssinceall76 non-problembanksby CB'sclassificationwere erroneouslyclassifiedas problem banks. The model was designed to identify non-problembanks and to classifyany bank it cannot identifyas suchtobe problembanks,Itwas,therefore,erroneoustoconcludethatthe multinomial Iogit model has an excellentjprediction performance for problembanks as Table 9 seemsto imply. Table 10showstheerrorcountestimatesforthetwo-categorymodels,
138
JOURNALOF PHILIPPINEDEVELOPMENT Table 8 CLASSIFICATION SUMMARY USING THE QUADRATIC DISCRIMINANT FUNCTION FOR TWO BANK CATEGORIES
Predicted Numbers and Percent Classified Under Bank Categories Non-Problem
Non-Problem
Problem Total
Problem
Our Classification Based on CB Data
72
4
76
(94.74) 440 (5g. 46) 512 (62.75)
(5,26) 300 (40.54) 304 (37.25)
(100.00) 740 (100.00) 816 (100.00)
Table 9 CLASSIFICATION SUMMARY LOGIT MODEL FOR TWO-BANK
USING THE CATEGORIES
Predicted Numbers and Per Cent Classified Under Bank Categories Non-Problem
Non-Problem Problem Total
Problem
(0100) (100.00) (0.00) (100.00) (0, 00)
76 (100.00) 740 (100.00) 816
(loo.oo)
(loo.oo)
Our Classification Based on CB Data
76 740 81 S
CONCLUSION Our results indicate the development of statistical models for classification and identification of problem banks holds promise. Investment of time and research effort will produce improved statistical models useful for banks and bank regulators alike. For local rural banks, the multiple discriminant model seems to be able to give a decision rule for classifying banks. It is also clear that for simplicity and economy we can use the two-
LLANTOANDORBETA:PROBLEMBANKS
137
Table 10 ERRORCOUNTESTIMATES (TWO-BANKCATEGORIES) BANK CATEGORIES
STEPWISEVARIABLES* PriorsEqual PriorsProp. ANOVA VARIABLES** Discriminant PriorsEqual PriorsProp. Logit
Problem
Non-Problem
Total
0.7797 0.7149
0.0526 0.0658
0.4162 0.6544
0.5946 0.5230 0 0000
0.0526 0.0789 1.0000
0.3236 0.4816 0 5000
* NYA,TRTA,NWASS,LAR,TCRA ** NWDAB,TRTA,LD,NYGY, NYA,NYE
bank category model.Some amountof parsimonyis achievedwith use of onlynineoutof16predictorvariablesinestimatingthediscriminantfunction. There is a need, however,to work out a theoretical model of bank behaviorwhichwillmotivatethe specificationof a model of classification andtheidentificationtechniquestobe used.Thiswillimprovetheexpedient tackadoptedhere of takingthe CB's classificationof banksas a given and proceeding to estimate the classificationfunctions, Whilethe statisticaltechniquesusedinthisstudyprovidea formaland fast way to classifybanks, they should not be treated as an absolute substituteto direc_nd on-sitebankexamination.However,theyserveas a useful aid in analyzing banks. In. a rapidly changing economic and financial environment,these statisticaltechniques help bank examiners make intellige_ntbankanalysesand more efficientlyallocatescarce bank examinationresources. As a final-note,itshouldbe emphasizedtheevaluationofthe proposed classification models was done with outcome of the CB's EWS as standard.Tothe authors'bestknowledge,nostudyonthe performanceof the CB's EWS in predictingactual bankfailureshas been documenti_d.A comparativeevaluation of the performanceof acombined personalized examinationandfinancialratiosevaluationsystemsuchasthe CB'sEWS, on the one hand, and an impersonalstatisticalclassificationsystembased solelyon financialratios,on the other, willbe an interestingsubjectof a futurestudy.Thisstudy,however,requiresauthoritytouseCB'sexamination reportsas well as knowledgeof actual bank failures. :
138
JOURNALOF PHILIPPINE DEVELOPMENT REFERENCES
Altman, Edward I. "Financial Ratios, Discriminant Analysis and Prediction of Corporate Bankruptcy." The Jouma/of Finance (September 1968): 589739. Booker, Irene_ "Tracing Banks from Afar: A Risk Monitoring System." Economic Review (November 1983): 36-39. Bovenzi, John; Marino, James; Failure Prediction Models."
and McFadden, Frank. "Commercial Bank Economic Review (November 1983): 14-26.
Graham, David and Hamprey, David• "Bank Examination Data as Predictors of Bank Net Loan Losses." Journal of Money, Credit and Banking (November 1978)_ 491-504. Joy, O. Maurice and Tollefson, John. "On the Financial Applications of Discrimi =' nant Analysis." Journal of Financial and Quantitative Analysis (December 1975): 723=739. Korobow, Leon and Stuhz, David. "Performance Measurement of Early Warning Models." Journal of Banking and Finance 9 (1985): 267-273.
Economic Lamberte,
• "The Relevance of Peer Groups in Early Warning Review (November 1983): 27-34.
Mario. "on the SES-III Warning
System_" (Unpublished,
Analysis."
1986).
Llanto, Gilberto and Suleik, Mercedes." "Early Warning of Bank Failure: The Role of Predictive Models (Unpublished)• November 1988• Maddala, G. Limited-Dependent and Quafitative Variables in Econometrics• York: Cambridge University Press, 1983.
New
Martin, Daniel• "Early Warning of Bank Failure - A Logit Regression Journal of Banking and Finance (1977):
Approach."
Pindyck, R, and D. Rubinfeld, Economic Models and Economic Second Edition. New York: McGraw-Hill, 1981.
Forecasts,
Putnam, Barron. "Concepts of Financial Monitoring•" Economic Review(Novembe 1983): 6-13. Sinkey, Joseph F. "Identifying Problems Banking (May 1978): 184-193.
r
Banks." Journal of Money, Credit and
Sinkey, Joseph F. "A Multivariate Statistical Analysis of the Characteristics Problem Banks." Journal of Finance (March 1975): 21-36• West, Robert Craig. "A Factor-Analysis Approach to Bank Condition•" Banking and Finance 9 (1985): 253-265.
of
Journal of