The principles under the analyses are to suggest a new attack by a two-stage intercrossed theoretical account of logistic regression-ANN, to research if the proposed theoretical account outperformed the traditional unreal nervous web ( ANN ) , and to build a recognition hazard warning system for banking industry in emerging market during 1998-2006. The proposed two-stage intercrossed theoretical account integrates the models of logistic arrested development and ANN. The important difference from traditional theoretical accounts is that this survey adopts the “ optimum cutoff point ” attack proposed by Hosmer and Lemeshow ( 2000 ) to find the cutoff point for recognition hazard. Additionally, cross-validation ( Stone, 1974 ; Efron and Tibshirani, 1993 ) is used to measure the anticipation power of the proposed theoretical account. The consequences find the factors of liquidness, capital, and plus quality are important factors related to the recognition hazard of Bankss in emerging market. In the anticipation of financially hard-pressed Bankss, the two-stage intercrossed theoretical account gives better public presentation and demonstrates stronger anticipation power than conventional ANN attacks. The two-stage intercrossed theoretical account is proved to be a utile and promising tool to supply a recognition hazard theoretical account which has appraisal deductions for analysts, practicians, and regulators.

## Introduction

The appraisal of the recognition hazard warning system of Bankss is an of import issue for oversing governments and investors. Credit hazard[ 1 ]induces fiscal hurt on Bankss, and its appraisal requires advanced patterning techniques that will associate it to the beginnings of uncertainness generated. Consequently, recognition hazard remains one of the major menaces that fiscal establishments face, and it is indispensable to pattern the recognition hazard of fiscal establishments. Therefore, there is necessity for systems to measure recognition hazard for Bankss.

Previous researches on recognition hazard patterning refers to traditional multivariate statistical and econometric techniques, for case inactive univariate analysis ( Beaver, 1966 ) , multivariate discriminant analysis ( Altman, 1968 ) , logit theoretical account ( Gulsun and Umit, 2010 ) , probit theoretical account, nervous web, dynamic Merton theoretical account, multivariate logit attack ( Demirguc-Kunt and Detragiache, 1998 ) and CUSUM, which are among the most widespread for recognition hazard mold ( Altman, 1981 ; Altman and Saunders, 1998 ; Gordy, 2000 ; Lin and Chang, 2006 ; Lin, 2009 ) , every bit good methodological analysiss from the field of unreal intelligence, operational research ( Dimitras et al. 1998 ; Zopounidis, 2002 ; Pastor, 2002 ) , multidimensional grading ( MDS, Mar-Molinero and Carlos Serrano-Cinca, 2001 ) and conditional chance analysis ( CPA ) theoretical accounts ( Lin and Piesse, 2004 ) . Over the past three decennaries, more sophisticated methodological analysiss have been developed to back up corporate fiscal hurt appraisal which included gambler ruin attack and option pricing theoretical account. The theoretical foundation about how a steadfast gets bankrupted when the settlement value of its assets falls below its debt duties. Such theoretical account was developed and modified by Wilcox ( 1972 ) , Merton ( 1974 ) and Scott ( 1981 ) . Recent surveies have considered new non-parametric methods such as mathematical scheduling, categorization trees, nervous webs ( Yu, et al. , 2008 ; Yeh and Lien, 2009 ; Abdou, et al. , 2008 ; Angelini, et al. , 2008 ) and support vector machines. As Baesens et Al. ( 2003 ) concluded that recognition hazard informations are decrepit non-linear, therefore proposing that extremely complex or non-linear theoretical accounts are non expected to hold a considerable predicting ability compared to simpler theoretical accounts.

On the quantitative attack of fiscal hurt warning system, Lee, et Al. ( 2002 ) and Lee and Chen ( 2005 ) explore the public presentation of recognition hiting utilizing a two-stage intercrossed mold process with unreal nervous webs and multivariate adaptative arrested development splines ( MARS ) , which outperforms the consequences utilizing discriminant analysis, logistic arrested development, unreal nervous webs and MARS. Furthermore, Tam and Kiang ( 1992 ) find that ANN performs better than logistic arrested development in the anticipation of fiscal hurt is consistent with that of old plants.

Most of the bing surveies on the development of recognition hazard theoretical accounts involve non-financial houses ( Chen, 2009[ 2 ]; Rauterkus, 2009[ 3 ]) . However, fiscal establishments, such as Bankss, investing houses, security agents, etc. , are besides disposed to default and major involvement due to their important function in the economic and concern activity in fiscal markets. Therefore, developing recognition hazard theoretical accounts for fiscal establishments, particularly for Bankss, is of major involvement to analysts, practicians, and oversing governments. Sahajwala and Van den Bergh ( 2000 ) reexamine the current patterns on the development and usage of recognition evaluation theoretical accounts for Bankss.

The part of this survey is to suggest a new attack by a two-stage intercrossed theoretical account of logistic regression-ANN, to research if the proposed theoretical account outperformed the traditional unreal nervous web ( ANN ) , and to build a recognition hazard warning system for banking industry in emerging market during 1998-2006. In separating financially sound and financially hard-pressed Bankss, this survey referred Lin ( 2009 ) and used the standard of Bankss ‘ Entire Capital Ratio is less than 8 % , Tier I Capital Ratio is less than 4 % , or Non-Performing Loan Ratio is greater than 5 % during the 1998-2006 period. The proposed two-stage intercrossed theoretical account integrates the models of logistic arrested development and ANN. First, act uponing variables are selected utilizing logistic arrested development, and so the influencing variables are taken as the input variables of BPN. The important difference from traditional theoretical accounts is that this survey adopts the “ optimum cutoff point ” attack proposed by Hosmer and Lemeshow ( 2000 ) to find the cutoff point for recognition hazard. Additionally, cross-validation ( Stone, 1974 ; Efron and Tibshirani, 1993 ) is used to measure the anticipation power of the proposed theoretical account.

The balance of this paper is structured as follows. Section II discusses the theoretical account and research design issues. Section III nowadayss and analyzes the consequences while the concluding subdivision concludes the survey.

## Methodology

## Two-Stage Hybrid Model

In phase I, act uponing variables are selected utilizing logistic arrested development. In phase II, the influencing variables are taken as the input variables of BPN. It is expected that by supplying the ANN with a good starting point, a more precise theoretical account can be developed on the strength of its learning ability. The consequences of the proposed two-stage intercrossed theoretical account are so compared to those of ANN. The process is as follows: First, utility dependant variable Y and independent variables, ,aˆ¦ , into logistic arrested development. Use logistic arrested development with Wald-forward method to place independent variables with important influence on hurt chance, ,aˆ¦ , and a significance theoretical account.

( 1 )

Furthermore, obtain the prognostic value for distress chance ( ) of fiscal hurt chance for each dataset utilizing the aforementioned significance theoretical account. Find the fiscal hurt cutoff point for prognostic values for hurt chance ( ) based on the point of intersection of sensitiveness and specificity harmonizing to Hosmer and Lemeshow ( 2000 ) and compare the consequences with existent values ( Y ) . Convert the hurt chance into unity or hurt to bring forth a new dependant variable and a new ANN theoretical account. Finally, replacement important variables, ,aˆ¦ , and the new dependant variableobtained in phase I into the ANN theoretical account as the independent variables and dependent variable of the input bed to bring forth prognostic values for fiscal hurt ( ) . Find the fiscal hurt cutoff point for prognostic values for hurt chance ( ) based on the point of intersection of sensitiveness and specificity harmonizing to Hosmer and Lemeshow ( 2000 ) and compare the consequences with existent values ( Y ) .

## Artificial Neural Network ( ANN )

Artificial nervous web ( ANN ) is a alone statistical technique that has monolithic calculating power, powerful memory, larning ability and mistake tolerance ability. That is, ANN is a computer science system[ 4 ]that emulates the interconnectedness of nerve cells in beings for complex information processing. It is an adaptative system with the ability to larn. Through different algorithms, ANN can be trained to supply coveted end product.

The basic back-propagation web ( BPN )[ 5 ]algorithm uses the gradient steepest descent method to minimise the mistake map between existent end product and mark end product of web. BPN trains the web through uninterrupted accommodation of weights ( ) in the steepest descent way and repeats the provender forward and back extension stairss.

= – ( 2 )

Tocopherol = ( 3 )

Where is the larning rate which controls the magnitude of each weight accommodation, E is error map, Tj is existent value, and Yj is end product value. Learning rate is a really of import parametric quantity in the preparation procedure of ANN. It affects the convergence velocity of ANN. Higher larning rate agencies faster convergence and smaller acquisition rate slows down the ANN convergence.

## Variables Definition

Dependent Variable: The dependant variable depicts whether a bank is in fiscal hurt ; 1 means a financially hard-pressed bank and 0 agencies a financially sound bank. This survey referred Lin ( 2009 )[ 6 ]and used the standard of Bankss ‘ Entire Capital Ratio is less than 8 % , Tier I Capital Ratio is less than 4 % , or Non-Performing Loan Ratio is greater than 5 % during the 1998-2006 period[ 7 ]. This survey adopts more rigorous criterions, that is, if a bank meets the three in one, so determined for the hard-pressed bank. These three variables can be more specific description of a bank ‘s capital construction and loaning quality.[ 8 ]

Independent Variables: Through overview of anterior literatures, this survey derives a broader coverage of important variables affects fiscal hurt of Bankss, including fiscal ratios which deem to be important forecasters of fiscal hurt for Bankss that is summarized in Table 1. Furthermore, the figure of Bankss that is classified as financially distress and financially soundness in 11 emerging states during 1998-2006 is shown in Table 2[ 9 ].

Insert Table 1 & A ; 2 Here.

## Period, Sample and Data Source

This survey targets publically listed Bankss for the period of 1998-2006 in emerging market, including China, Taiwan, Indonesia, Hong Kong, Thailand, Turkey, Pakistan, India, Russia, South Africa and Mexico. The Bankss are classified into financially distressed group and financially sound group based on the nine-year norm of sample informations and the definitions for fiscal hurt. The fiscal information of the sample Bankss over the period of 1998 to 2006 are obtained from BankScope database. Table 3 shows the statistics summary for the Bankss in emerging market during 1998-2006.

Insert Table 3 Here.

## Consequences

## Collinearity Test

First, this survey uses point-biserial correlativity coefficient to acquire rid of the extremely relevant variables. The correlativity coefficient of each variable is compared to the following 1. If the absolute value of the correlativity coefficient is comparatively little, the matching variable is so deleted. Furthermore, variables after testing by point-biserial correlativity coefficient have been tailored to cipher the tolerance, VIF and status index. The VIF of those screened variables in Table 4 are less than 10 and tolerance greater than 0.1, which indicates that they have no serious collinearity jobs. The logistic arrested development theoretical account takes merely those screened variables.

Insert Table 4 Here.

## Consequences of ANN[ 10 ]

As shown in Table 5, the ANN survey is summed up as follows: For the Bankss of China and India, two old ages prior to the default ; the ANN theoretical account achieves a prognostic truth of 65.1 % and 72.2 % in financially hard-pressed group, and 51.4 % and 72.7 % in financially sound group, individually.

For the Bankss of Taiwan, Indonesia, Hong Kong, Thailand, Turkey, Pakistan, Russian, South Africa, and Mexico, three old ages prior to the default ; the ANN theoretical account achieves a prognostic truth of 60.3 % , 51.9 % , 63.4 % , 70.5 % , 61.7 % , 64.4 % , 53.8 % , 47.6 % , and 65.4 % in financially hard-pressed group, and 49.3 % , 39.1 % , 56.4 % , 57.0 % , 65.8 % , 69.2 % , 44.8 % , 46.7 % , and 46.6 % in financially sound group, individually.

Obviously, ANN theoretical account performs significantly better in the anticipation of fiscal hurt than the anticipation of fiscal soundness for the Bankss in emerging market.

## Consequences of Two-Stage Hybrid Model

First of all, logistic arrested development is used to test important independent variables in emerging market. In add-on, Wald frontward method is employed to choose variables for the building of recognition hazard warning theoretical accounts. The logistic arrested development equation derived for the Bankss in 11 emerging states are as follows:

For the Bankss of China:

For the Bankss of Taiwan:

For the Bankss of Dutch east indies:

For the Bankss of Hong Kong:

For the Bankss of Siam:

For the Bankss of Turkey:

For the Bankss of Pakistan:

For the Bankss of India:

For the Bankss of Soviet union:

For the Bankss of South Africa:

For the Bankss of Mexico:

As shown above, Current Ratio ( X7 ) , Working Capital Ratio ( X9 ) , Current Asset Ratio ( X10 ) , Tier-I Capital Ratio ( X12 ) , Impaired Loans Net / Equity ( X13 ) , and Non-Performing Loan Ratio ( X15 ) are important at the degree of 1 % related to the recognition hazard of Bankss in 11 emerging states.

Furthermore, the predicted chances are calculated utilizing the above significance theoretical account in emerging market. Then, predicted chances are converted to a dichotomous variable ( ) through fiscal hurt cutoff. The important independent variables and the born-again dependant variable are substituted into the ANN theoretical account to build a warning theoretical account of recognition hazard. As shown in of Table 5, the consequences of two-stage intercrossed theoretical account are summed up as follows:

For the Bankss of Taiwan, Indonesia, Hong Kong, Thailand, Turkey, Pakistan, Russian, South Africa, and Mexico, two old ages prior to the default ; the logistic-ANN intercrossed theoretical account achieves a prognostic truth of 66.8 % , 61.1 % , 64.5 % , 73.0 % , 72.4 % , 80.8 % , 67.7 % , 93.8 % , and 73.8 % in financially hard-pressed group, and 56.4 % , 61.0 % , 79.7 % , 65.6 % , 73.8 % , 81.7 % , 70.0 % , 100 % , and 75.0 % in financially sound group, individually.

For the Bankss of China and India, three old ages prior to the default ; the logistic-ANN intercrossed theoretical account achieves a prognostic truth of 67.0 % and 73.9 % in financially hard-pressed group, and 63.2 % and 77.1 % in financially sound group, individually.

It is clear that the current two-stage loanblend theoretical account has better truth in foretelling financially hard-pressed Bankss, financially sound Bankss, and overall samples in 11 emerging states by utilizing informations of two old ages prior to the default. Furthermore, the two-stage intercrossed theoretical account performs significantly better in the anticipation of fiscal hurt than the anticipation of fiscal soundness for the Bankss in Thailand, Indonesia, Taiwan, Japan and China.

Insert Table 5 Here

## Reasoning Remarks

This survey proposes a two-stage intercrossed theoretical account of logistic regression-ANN. The important difference from traditional theoretical accounts is that this survey adopts the “ optimum cutoff point ” attack proposed by Hosmer and Lemeshow ( 2000 ) to find the cutoff point for fiscal hurt. Additionally, cross-validation ( Stone, 1974 ; Efron and Tibshirani, 1993 ) is used to measure the anticipation power of the proposed theoretical account. The consequences of logistic arrested development showed that factors of Current Ratio, Working Capital Ratio, Current Asset Ratio, Tier-I Capital Ratio, Impaired Loans Net / Equity, and Non-Performing Loan Ratio are significantly related to the fiscal hurt position of Bankss in emerging market. The important variables identified above warrant the devouring attending of banking regulator and may function as early warning signals of fiscal hurt for Bankss.

The ANN theoretical account achieves a anticipation truth of 51.9 % ~72.2 % in financially hard-pressed group. On the other manus, the proposed two-stage intercrossed theoretical account achieves a anticipation truth of 61.1 % ~93.8 % in the financially hard-pressed group in emerging market. In the anticipation of financially hard-pressed Bankss, this gives the best public presentation of 93.8 % for Bankss of South Africa. In the anticipation of full samples, two-stage intercrossed achieves the highest rate of 94.7 % for Bankss of South Africa. Comparing to ANN theoretical account, the current two-stage loanblend theoretical account has better truth in foretelling financially hard-pressed Bankss, financially sound Bankss, and full samples.

As the consequences reveal, the proposed two-stage intercrossed theoretical account is good suited for set uping the warning system of recognition hazard of Bankss in emerging market. The groundss of prognostic truth showed that the proposed two-stage intercrossed theoretical account outperforms the conventional ANN theoretical account. Therefore, the two-stage intercrossed theoretical account is a utile and promising tool to supply a recognition hazard theoretical account which has appraisal deductions for analysts, practicians, and regulators.

## Recognition

The writer acknowledge the fiscal support from National Science Council ( NSC96-2416-H-030-011 )