Business Bankruptcy Prediction Based on Survival Analysis Approach



Similar documents
Simple Linear Regression

An Approach to Evaluating the Computer Network Security with Hesitant Fuzzy Information

Maintenance Scheduling of Distribution System with Optimal Economy and Reliability

Dynamic Two-phase Truncated Rayleigh Model for Release Date Prediction of Software

Average Price Ratios

Report 52 Fixed Maturity EUR Industrial Bond Funds

Application of Grey Relational Analysis in Computer Communication

Forecasting Trend and Stock Price with Adaptive Extended Kalman Filter Data Fusion

SHAPIRO-WILK TEST FOR NORMALITY WITH KNOWN MEAN

The simple linear Regression Model

CHAPTER 2. Time Value of Money 6-1

STATISTICAL PROPERTIES OF LEAST SQUARES ESTIMATORS. x, where. = y - ˆ " 1

Study on prediction of network security situation based on fuzzy neutral network

ANOVA Notes Page 1. Analysis of Variance for a One-Way Classification of Data

IDENTIFICATION OF THE DYNAMICS OF THE GOOGLE S RANKING ALGORITHM. A. Khaki Sedigh, Mehdi Roudaki

Applications of Support Vector Machine Based on Boolean Kernel to Spam Filtering

Chapter 3. AMORTIZATION OF LOAN. SINKING FUNDS R =

APPENDIX III THE ENVELOPE PROPERTY

1. The Time Value of Money

A New Bayesian Network Method for Computing Bottom Event's Structural Importance Degree using Jointree

An IG-RS-SVM classifier for analyzing reviews of E-commerce product

Optimizing Software Effort Estimation Models Using Firefly Algorithm

of the relationship between time and the value of money.

DECISION MAKING WITH THE OWA OPERATOR IN SPORT MANAGEMENT

Classic Problems at a Glance using the TVM Solver

Banking (Early Repayment of Housing Loans) Order,

Projection model for Computer Network Security Evaluation with interval-valued intuitionistic fuzzy information. Qingxiang Li

Optimal replacement and overhaul decisions with imperfect maintenance and warranty contracts

T = 1/freq, T = 2/freq, T = i/freq, T = n (number of cash flows = freq n) are :

Credibility Premium Calculation in Motor Third-Party Liability Insurance

The impact of service-oriented architecture on the scheduling algorithm in cloud computing

IP Network Topology Link Prediction Based on Improved Local Information Similarity Algorithm

The Gompertz-Makeham distribution. Fredrik Norström. Supervisor: Yuri Belyaev

AN ALGORITHM ABOUT PARTNER SELECTION PROBLEM ON CLOUD SERVICE PROVIDER BASED ON GENETIC

Beta. A Statistical Analysis of a Stock s Volatility. Courtney Wahlstrom. Iowa State University, Master of School Mathematics. Creative Component

10.5 Future Value and Present Value of a General Annuity Due

Performance Attribution. Methodology Overview

Regression Analysis. 1. Introduction

Statistical Intrusion Detector with Instance-Based Learning

ECONOMIC CHOICE OF OPTIMUM FEEDER CABLE CONSIDERING RISK ANALYSIS. University of Brasilia (UnB) and The Brazilian Regulatory Agency (ANEEL), Brazil

A particle Swarm Optimization-based Framework for Agile Software Effort Estimation

A particle swarm optimization to vehicle routing problem with fuzzy demands

6.7 Network analysis Introduction. References - Network analysis. Topological analysis

Settlement Prediction by Spatial-temporal Random Process

Report 05 Global Fixed Income

The Application of Intuitionistic Fuzzy Set TOPSIS Method in Employee Performance Appraisal

ADAPTATION OF SHAPIRO-WILK TEST TO THE CASE OF KNOWN MEAN

Statistical Pattern Recognition (CE-725) Department of Computer Engineering Sharif University of Technology

DYNAMIC FACTOR ANALYSIS OF FINANCIAL VIABILITY OF LATVIAN SERVICE SECTOR COMPANIES

Green Master based on MapReduce Cluster

A NON-PARAMETRIC COPULA ANALYSIS ON ESTIMATING RETURN DISTRIBUTION FOR PORTFOLIO MANAGEMENT: AN APPLICATION WITH THE US AND BRAZILIAN STOCK MARKETS 1

CIS603 - Artificial Intelligence. Logistic regression. (some material adopted from notes by M. Hauskrecht) CIS603 - AI. Supervised learning

Integrating Production Scheduling and Maintenance: Practical Implications

Reinsurance and the distribution of term insurance claims

Proceedings of the 2010 Winter Simulation Conference B. Johansson, S. Jain, J. Montoya-Torres, J. Hugan, and E. Yücesan, eds.

A Parallel Transmission Remote Backup System

How To Make A Supply Chain System Work

Software Reliability Index Reasonable Allocation Based on UML

CHAPTER 13. Simple Linear Regression LEARNING OBJECTIVES. USING Sunflowers Apparel

The paper presents Constant Rebalanced Portfolio first introduced by Thomas

Automated Event Registration System in Corporation

Common p-belief: The General Case

Optimal Packetization Interval for VoIP Applications Over IEEE Networks

Using Phase Swapping to Solve Load Phase Balancing by ADSCHNN in LV Distribution Network

Preprocess a planar map S. Given a query point p, report the face of S containing p. Goal: O(n)-size data structure that enables O(log n) query time.

Optimal multi-degree reduction of Bézier curves with constraints of endpoints continuity

Dynamic Provisioning Modeling for Virtualized Multi-tier Applications in Cloud Data Center

A DISTRIBUTED REPUTATION BROKER FRAMEWORK FOR WEB SERVICE APPLICATIONS

An Automated Selecting Subcontractor Model In E-Commerce by Pao-Hung Lin

The analysis of annuities relies on the formula for geometric sums: r k = rn+1 1 r 1. (2.1) k=0

Using Data Mining Techniques to Predict Product Quality from Physicochemical Data

Abraham Zaks. Technion I.I.T. Haifa ISRAEL. and. University of Haifa, Haifa ISRAEL. Abstract

A Study of Unrelated Parallel-Machine Scheduling with Deteriorating Maintenance Activities to Minimize the Total Completion Time

Developing tourism demand forecasting models using machine learning techniques with trend, seasonal, and cyclic components

The Digital Signature Scheme MQQ-SIG

Security Analysis of RAPP: An RFID Authentication Protocol based on Permutation

ROULETTE-TOURNAMENT SELECTION FOR SHRIMP DIET FORMULATION PROBLEM

Web Service Composition Optimization Based on Improved Artificial Bee Colony Algorithm

Curve Fitting and Solution of Equation

Optimization Model in Human Resource Management for Job Allocation in ICT Project

Mixed Distributions for Loss Severity Modelling with zeros in the Operational Risk losses

Transcription:

Busess Bakruptcy Predcto Based o Survval Aalyss Approach ABSTRACT Mg-Chag Lee Natoal Kaohsug Uversty of Appled Scece, Tawa Ths study sampled compaes lsted o Tawa Stock Exchage that examed facal dstress betwee 2003 ad 2009. It uses the survval aalyss to fd the ma dcators whch ca expla the busess bakruptcy Tawa. Ths paper uses the Cox Proportoal Hazard Model to assess the usefuless of tradtoal facal ratos ad market varables as predctors of the probablty of busess falure to a gve tme. Ths paper presets emprcal results of a study regardg 12 facal ratos as predctors of busess falure Tawa. It showed that t does ot eed may ratos to be able to atcpate potetal busess bakruptcy. The facal dstress probablty model s costructed usg Proftablty, Leverage, Effcecy ad Valuato rato varables. I the proposed steps of busess falure predcto model, t used detal SAS procedure. The study proves that the accuraces of classfcato of the mode overall accuracy of classfcato are 87.93%. KEYWORDS Busess Falure predcto; Survval Aalyss; Cox Proportoal Hazard model; Logstc model 1. INTRODUCTION Busess Falure Predcto (BFP) models are estmato of the bakruptcy probablty of a frm usg a set of covarates, such as facal ratos, Captal turover, Captal turover, etc [77]. I past decades, BFP has bee a topc research for busess ad corporate orgazatos. Ivestors or credtors, borrowg orgazatos ad govermets are creasg terest to predct of corporate bakruptcy [26]. BFP help to avod ledg to (or vestg ) busess lkely to fal, early detfcato of falg busess by regulatory bodes, ad more accurate scorg models for ratg ageces. Bakruptcy predcto models use statstcal aalyss ad data mg techque to ehace the decso support tool ad mprove decso makg [68]. Statstcal busess falure predcto models attempt to predct the busess falure or success. The Multple dscrmat aalyss (MDA) has bee the most popular approaches, but there eed a large umber of alteratve techques avalable ([18], [37], [42]). Such as the data mg techques clude decso tree, eural etworks (NNs), support vector mache (SVM), fuzzy system, rough set theory, geetc algorthm (GA) [68]. Varous researches have demostrated the artfcal tellgece (AI) techques such as artfcal eural etworks (ANNs) ca serve as a useful tool bakruptcy predcto [61]. Back propagato eural etwork (BPNN) was used bakruptcy predcto. Before that BPNN some of the techques followed such as k-earest eghbor ad the tree DOI:10.5121/jcst.2014.6207 103

algorthm (ID3) but offered better predctve compare tha compared models. Multvarate cumulatve sum (CUSUM) s a sequetal procedure to predct a busess tedecy towards falure. A survval aalyss (SA) techque s the term appled to a dyamc statstcal tool used to aalyss the tme tll a certa evet [18]. SA uses the Cox proportoal hazard model to aalyss survval probablty ad falure tmes; t s oe dyamc model approach [53]. SA techques have used to exame the drvers behd the survval of Iteret busess ([29], [30]). Dscrmat aalyss (DA) ad Logt aalyss (LA) were foud to be slghtly superor predctors to the Cox proportoal hazard model [27]. Nevertheless, Late ad Luoma [33] argued that the SA approach was more atural, flexble, ad approprate ad used more formato Busess Falure predcto. Keasey et al. [31] also recommeded that SA techques be used BFP. Yap et al. [69] use facal rato ad logstc regresso for evaluatg compay falure Malaysa. The models of eterprse credt rsk modes clude statstcal model, eural etwork, learg vector, soft-computg, ad hybrd models. Table 1 deoted as eterprse credt rsk model. Eterprse credt rsk evaluato models ths study are Neural etworks, Bayesa classfer, Dscrmat aalyss, Logstc regresso, K-earest eghbor, Decso tree, Case base reasog, Support vector mache, Software computg, Fuzzy rule-based system, Geerc algorthms, Grey relato, ad Hybrd models. Table 1: Eterprse credt rsk evaluato models Category Area Some Approach Statstcal model Neural Networks Learg vector Softcomputg Parametrc Statstcal Method No-Parametrc Statstcal Method Mache learg Mache learg Reducto attrbutes 1. Dscrmat aalyss 2. Lear mult dscrmat aalyss 3. Logstc regresso 4. Bayesa rsk Dscrmat aalyss 1. K- earest eghbor 2. Cluster aalyss 1. Multlayer percepto 2. Back propagato 3. Radal fucto eural etwork 3. Probablstc eural etwork 4. Self-orgazed competto Support Vector Mache 1. Rough sets of reducto kowledge 2. Grey relatoal of reducto kowledge Altma [3]; Ohlso [47]; Yap et al. [69]; Stefaescu et al. [56]; Tabachck ad Fdell [59]; Lag ad X [36] Ice, ad Akta [24]; Islam et al. [25]; Lau [34]; Su ad LI [58] Islam et al. [25; Che [13]; Lopez [39]; Mues et al. [44]; Sarkar ad Srram [50]; Stefaescu et al. [56]; Tam ad Kag [60] ; Che [12]; Odom ad Sharda [46] Zhou et al. [72]; Che et al. [11] ; Km ad Soh [32]; Sh et al. [52] Dmtras et al. [17]; Cheg et al. [14]; Ba ad Mazlack [6]; Hu [23]; Lu et al.[38] ; Tug et al. [62]; 104

Survval aalyss (SA) Hybrd models Tme to evet data aalyss Combato of two or more methods 3. Geetc algorthm of reducto kowledge 4. Fuzzy-Rough Sets Approach 1. Credt rsk modelg based o SA 2. Corporate credt rsk ad the macro ecoomy 1. Rough - K Nearest Neghbor 2. Rough Sets Neural Network 3. Fuzzy-Rough Sets - Nearest Neghbor 4. Fuzzy- Nearest Neghbor 5. Support Vector Mache wth Nearest Neghbor 6. GA-based eural etwork approach 7. At Coloy Algorthm based o quck-reduct algorthm We et al. [64]; Wog et al. [65]; Zhao [72]; Xhu ad Zhog [67] Stepaova, ad Thomas [57]; Atoaks ad Sfakaaks [4]; Cao et al. [8]; Soh et al. [55] Tug et al. [62]; Wag et al. [63]; We ad Zhag[64]; Wog et al. [65]; Xao et al. [66]; Chaduhur ad De [10]; Tam ad Kag [60]; Yu et al. [70]; Zhag et al. [71]; Zhou ad Ba [73]; Zhou et al. [75] The most useful beefts to SA are: (1) I the modelg process, SA s able to take tme-varyg varables to accout [22]. Ths s doe through proportoal hazard models [5]. (2) SA s ot restrcted by the assumpto that the dstrbutos of the varables the data eed to be ormal [54]. (3) SA oly produces postve predctos of tme [21]. The tme-varyg has the potetal to ot follow a ormal dstrbuto. It eeds to be postve predctos ad s flueced by tme-varyg varables. The major cotrbuto of SA methods s estmato procedures that cosder chages the value of covarates over tme [35]. Thus, SA approaches to BFP dfferet from the other approaches metoed above [18]. 3. MERHODOLOGY 3.1 Logt model I settg up the logstc regresso model, frst establsh the fudametal model for ay multple regresso aalyss. The outcome varable s assumed as a lear combato of a set of predctors. If outcome varable s Y, ad a set of predctor varables are X 1, X 2,..., X, the Logt model s: [1] Y 0 1X1 2 X 2... X 0 j X j (1) j1 105

Where 0 s the expected value of Y whe X s set 0. j s the regresso coeffcet for each correspodg predctor varable X j. s the error of the predcto. Defes (x) as the probablty that Y = 1. Smlarly, 1- (x) s the probablty that Y = 0. These probabltes are wrtte the followg form: ( x ) P ( Y 1 X, X 2,..., X 1 1 1 2 ) ( x) P( Y 0 X, X,..., X ) (2) Ths model for the atural logarthm of the ( x) 1 ( x) s: P( Y 1 X 1, X 2,..., X ) ( x) l l 0 j X 1 P( Y 1 X, X,..., X ) 1 ( x) 1 2 j1 j (3) Usg the verse of the Logt trasformato of (3), t obtas at the followg: P( Y 1 X, X 1 2,..., X ) e 1 e 0 j X j j1 0 j X j j1 1 e 1 ( 0 j X j ) j1 (4) Thus, (4) s a logstc regresso model, the codtoal mea s betwee 0 ad 1. Now, t wll ft the logstc regresso model to the data. Frstly, t must establsh a techque for estmatg the parameters. The maxmum lkelhood s the method of parameter estmato logstc regresso model. Ths method costructs the lkelhood fucto, whch expresses the probablty of the observed data as a fucto of the ukow parameters. Ths process wll have selected the estmators (4). For a set of observatos the data ( x, y ), the cotrbuto to the lkelhood fucto s ( x ), where y 1, ad 1 ( x ), where y 0. The followg equato results for the cotrbuto to the lkelhood fucto for the observato x, ) s x ) : ( y ( y ( ) ( ) 1 y x [1 ( ) x x ] (5) The observatos are assumed to be depedet of each other so t ca multply ther lkelhood cotrbutos to obta the complete lkelhood fucto l (B). The result s gve (6). l( B) ( ) (6) 1 x Where B s the collecto of parameters ( 0, 1,..., ) ad l(b) s the lkelhood fucto of B. 106

Maxmum lkelhood estmates (MLE s) ca be obtaed by calculatg the B whch maxmzes l (B). However, to smply the mathematcs, from the logarthm of (6) before fdg the value whch maxmzes the lkelhood fucto. As show (7). L(B) s deoted the log lkelhood expresso. L( B) l[ l( B)] ( y 1 l[ ( x )] (1 y ) l[1 ( x )]) (7) It employs the techques of calculus to determe the value of B based o maxmum of L (B). Ths s doe by dfferetatg (3) wth respect to 0, 1,..., ad settg the resultg dervatves equal to zero. These equatos are called lkelhood estmatos, ad there s +1 equato. They are of the followg form: 1 y ( ) 0, for the tercept 0, ad x,..., x 1. k 1 [ y ( x )] 0, for the predctor varables, The soluto ca be solvg by usg computer programs such as SAS or SPSS. It performs the logstc regresso aalyss of the data for ths study ad wll calculate the maxmum lkelhood estmates. 3.2 Cox s PH model [15] Accordg the assumpto about the relatoshp betwee the hazard (or survval) fucto ad the set (vector) of explaatory varables ( X ), there have varous models. Thus, the geeral T T regresso fucto ca be wrtte as h( t) g( t, X ), where X s the traspose of X. s the vector of explaatory varable coeffcets. I SA models, t s customary to estmate the hazard rate, ad the derve the survval rate are requred by usg regresso model. Two ma types of regresso models are SA model. These types are the proportoal hazards (PH) ad accelerated falure tme (AFT) models, both of whch have fully parametrc ad sem-parametrc verso. A parametrc regresso model based o the expoetal dstrbuto: logeh ( t) 1x1 2x2... x (8) Or equvaletly, h ( t) exp( 1x1 2x 2... k xk ) x e e e x 1 1 2 2... e k xk (9) Where dexes subjects; x 1, x2,... xk s the values of covarates for the th subject Ths model s parametrc because, oce the regresso parameters, 1, 2,..., k are specfed, the hazard fucto h (t) s fully characterzed by the model. The costat represets a kd of basele hazard, (9), sce log e h ( t), or equvaletly, h ( t) e whe all of the x s are 0. Other parametrc hazard regresso models are based o other dstrbutos (Gompertz ad k k 107

Webull dstrbuto) commoly used modelg survval data. The Cox model superseded full parametrc hazard regresso models, whch leaves the basele hazard fucto uspecfed: Or equvaletly, logeh ( t) ( t) 1x1 2x2... x (10) h t) h ( t)exp( x x... x ) (11) ( 0 1 1 2 2 k k To estmate the model parameters, the maxmum lkelhood estmates are derved by maxmzg a lkelhood fucto. Ths Cox model [17] s termed sem-parametrc because whle the basele hazard ca take by form, the covarates eter the model through the lear predctor... x (12) 1x1 2x2 I (12), there s o costat term (tercept) the lear predctor; the costat s absorbed the basele hazard. The Cox regresso model s also a proportoal hazard model. Cosder two observatos, ad ', that dffer ther x-values, wth respectve lear predctors... x ad 1x1 2x2 k k ' 1x ' 1 The hazard ratos for these two observatos are: k k k k (13) ( ) 0( ) t h t e ' e ' ( t) ' h0( t) e h h 2x '...k x 2 ' k (14) I (14), the rato s costat over tme. Therefore, the Cox model ca easly accommodate tmedepedet covarates. The Cox model accouts for survval tmes, thus, t uses more formato the the logstc model. The Cox PH model allows cesored observatos ad corporates survval tmes. A Cox PH model therefore uses more formato tha a logstc regresso model. 3.3 Goodess-of-ft test A set of covarates the Cox PH model ca be tme-depedet (or tme varyg) covarates. Used SAS (or SPSS) to perform the Cox PH model aalyss of the data for ths paper ad wll calculate the maxmum lkelhood estmates. Used lkelhood rato test to see the varables cluded the fal model are sgfcat explag some of varablty data. The Ch-Square statstc s the dfferece -2 Log Lkelhood (-2LL) betwee the fal model ad a reduced model. The ull hypothess s that all parameters ( 1, 2..., k ) of the effect are 0. Ths test s comparable to oval F test for regresso aalyss. The hypothess testg s as follows: H0 : 1 2... k 0; H 1 :~ H 0 Where s the parametrc estmato of explaatory varable The statstc quatty of the aforesad hypothess testg s -2 Log Lkelhood (= -2Log (L(0)- 108

L( )) whch observes 2 ( k ), where L(0) s the lkelhood fucto value uder the ull hypothess, whle L( ) s the lkelhood fucto value cosderato of the whole model. R 2 s a tutve measure of how well model predcts the values of the depedet varables [69]. R 2 the Cox regresso s a pseudo measure of assocato betwee the respose varable ad covarates. I geeral, hgher R 2 value meas the model s ft for aalyss of samplg. Sce lght that maxmum of 1 caot be obtaed usg Cox & Sell R 2 for measuremet; Nagelkerke [45] proposed a modfcato of Cox & Sell R 2. Cox & Sell R 2 : 2 L(0) R 1 [ ] 2 N cs 1 exp[( L(0) L( )) * 2 / N ] (15) L( ) Nagelkerke R 2 : R 2 2 / max 2 N Rcs Rcs (16) Where L (0) = the lkelhood fucto value cotag oly tercept; L ( ) fucto value cosderato of the whole model; N= sample sze, 4. COX MODEL PREDICTIVE ABILITY = the lkelhood 2 2 max R cs 1 [ L(0)] The most mportat characterstcs of a BFP model are ts producto of accuracy. Type Ⅰ error refer to the stuato whe actual falure compay s classfed as o falures compay, ad Type Ⅱ error refer to o falure compay s classfed as o falures compay. Type error s more mportat tha Type error. The objectves of predctve of accuracy should be to reduced Type error whle keep Type error. The reaso for ths s that Type Ⅱ error oly creates a lost opportuty cost from ot dealg wth a successful busess, for example, mssed potetal vestmet gas. I cotrast, due to volvemet wth a busess that wll fal, Type error results a realzed facal loss, for example, losg all moey vested a mpedg bakrupt busess [18]. The method used for calculatg the accuracy of classfyg dstressed compaes ad odstressed compaes s llustrated Table 2, whch C deotes the umber of TypeⅠ error, that s the umber of dstressed compaes the sample based o actual observato that were msclassfed as a o-dstressed compay. B deotes the umber of Type Ⅱ error that s the umber of o-dstressed compaes the sample based o actual observato that were msclassfed as a dstressed compay. A ad D represet respectvely the umber of odstressed ad dstressed accurately classfed by the models [34]. By determg accuracy of classfcato, we ca lear about whether the costructed model s the optmal predcto model. 109

Table 2: Robustess of model No-dstressed No-dstressed compay Dstressed compay Observed value No-dstressed compay A B E Dstressed compay C D F G Overall accuracy of classfcato Accuracy of Classfcato Note: 1. The accuracy of classfcato of o-dstressed compay s expressed by E A /( A B) 2. The accuracy of classfcato of dstressed compay s expressed by F D /( C D) 3. The oval accuracy of classfcato s G ( A D) /( A B C D) 5. EMPIRICAL RESEARCH I ths secto, the study frst performs descrptve statstc of the samplg ad Covarates, ad follows by the costructo of busess falure predcto model based o Cox model ad aalyss of emprcal results. I order to better aalyze the effect of the Cox model predcted, we radom select the stock market lsted compay's tradtoal maufacturg Tawa. I sub-secto, goodess-of-ft test s carred out ad robustess of the model s examed usg accuracy of classfcato. The proposed steps of busess falure predcto model are: Step 1: Defto of varables Step 2: Samplg ad data Step 3: Reduced the umber of facal rato Step 4: Goodess-of-ft test Step 5: Robustess of model predcto accuracy 5.1 Selecto of Varables The ma goal of ths research s to assess the emprcal classfcato ad predcto accuracy of the COX SA model whe appled to BFP. Karels ad Prakash [28] suggested a careful selecto of ratos to be used the developmet of bakruptcy predcto model. A set of covarates used ths study cludes a combato of facal ratos ad market varables [20]. I facal reportg aalyss, [19] suggest fve factors for evaluato eterprse facal falure. Facal ratos have bee wdely used explag the possblty of busess facal dstress ([3], [7], [9], [43], [47], [48], [49], ad [76]). Table 3 s The 12 ratos selected ths study. 110

Table 3: The 12 ratos selected ths study Table 3 shows the detals ad defto of covarates used ths study. 12 facal ratos are used ths study. The Proftablty ratos clude EBIT marg (EBT), Retur to equty (ROE), ad Retur o assets (ROA). Curret rato (CUR) ad Quck rato (QUK) wll be used ths study order to measure the lqudty of the frms. Two types of Leverage ratos are Debt rato (DET) ad Debt to equty rato (DER), two types of Effcecy ratos are Fxed asset turover (FAT) ad Captal turover (CAT) ad three types of Valuato ratos are Prce to sales rato (PSR), prce eargs rato (PER), ad prce to book value (PBV). 5.2 Data collecto ad Sample The sample ths research s radom selecto the stock market lsted compay's tradtoal maufacturg Tawa. I order to cosder the survval problem, the choce of lsted compaes lsted o the Tawa Stock Exchage usg aual data o facal ratos for the perod 2003-2009. I order to better aalyze the effect of the Cox model predcted, ths study estmated that from 2003 to 2009 sample was dvded to estmatg samples ad forecastg samples. Ths paper select sample lsted compaes from 2003 to 2006 for estmatg sample. There are 56 facally dstressed compay ad 154 actvty lsted compaes the aalyss. Ths paper select sample lsted compaes from 2007 to 2009 for forecastg sample. It radomly selected 46 facally dstressed compay ad 128 actvty lsted compaes the aalyss as forecastg samples. 5.3 Reduced the umber of facal rato There are two ways to reduce the large umber of facal rato (1) Pearso correlato (2) The model accepted has a good ft ad that the mult-learty level s acceptable. Accordg to Pearso correlato, the correlato betwee CUR ad QUK s 0.9876, whch are statstcally sgfcat wth p-value less tha 0.0001. Ths meas the postve relatoshp 111

betwee these par of varable. Based o the lkelhood rato resulted from Pearso correlato, the covarates QUK are selected to Cox proportoal hazards model. Table 4 s the Cox proportoal hazards model. Table 4: The Cox proportoal hazards model Varable D F Parameter Coeffcet Stadard error Chsquare pr> chsq Hazard Rato EBT 1 0.3272 0.0125 0.0785 0.6542 1.387 ROE 1-0.2631 0.1345 5.9682 0.0164** 0.769 ROA 1-0.5012 0.1877 6.5896 0.0068** 0.606 QUK 1-0.0165 0.0112 2.0125 0.0781 0.984 DET 1 0.2958 0.1432 5.2146 0.5606 1.344 DER 1-0.1285 0.1245 0.3352 0.0085** 0.879 FAT 1 0.1421 0.2198 1.5428 0.0109** 1.153 CAT 1 0.0026 0.0156 0.5976 0.1875 1.003 PSR 1 0.2415 0.0968 0.3524 0.2432 1.273 PER 1 0.1243 0.1265 0.1548 0.1861 1.132 PBV 1 0.2341 0.0065 0.2382 0.0235** 1.264 ** Sgfcat at 5 percet Usg Cox proportoal hazards model wth facal ratos, the proportoal hazards model are represeted Table 4. I SAS software, PROC RHREG s used to ft the Cox proporto hazards model ad to aalyze the effects of the facal o the survval of the compay. Table 4 s deoted as the coeffcet estmato, the Stadard error, Ch-square tests wth the relatve p-value for testg the ull hypothess that the coeffcet of each covarate s equal to zero. Hazard rato s obtaed by computg e, where s the coeffcet a proportoal hazard model. By cosderg the p-value, sx covarates are hghly sgfcat at 5 percet. These ratos are EBT, QUK, DET, CAT, PSR ad PER wth the coeffcet 0.3272, -0.0165, 0.2958, 0.0026, 0.2415 ad 0.1243 respectvely. Therefore, the early warg dcators are ROE, ROA, DER, FAT ad PSV. ROE ad ROA are egatve value dcatg that a crease ether covarate decreases the 0.2631 hazard of eterg to facally dstressed. Hazard rato of ROE s 0.769 ( e 0.769). It meas that a crease of oe ut ROE mples 0.769 decreases rsk facal dstress. For the sample ths study, proftablty (EBT), lqudty (QUK), leverage (DET), Effcecy (CAT) ad Valuato (PSR, PER) have ever foud statstcally sgfcat the model. The model s show as follow: Log h t) ROE( t) ROA( t) DER( t) FAT ( t) PSV ( ) 5.4 Goodess-of-ft test ( 1 2 3 4 5 t Oe measure of overall goodess-of-ft test s partal lkelhood-rato test. I SAS software, PROC RHREG s used to obta the lkelhood rato ch-square statstc from the model ft statstcs Table 5. I Table 5, the output produces cludes the value of -2log lkelhood for fttg, AIC (Akake Iformato Crtero) ad SBC (Schwartz Bayesa Crtero) for fttg a model wthout covarace ad fttg a model wth covarates. 112

Akake [2] troduced the cocept of the formato crtera as a tool optmal model selecto. AIC s a fucto of the umber of observatos, the sum of square errors (SSE), ad the umber of depedet varables k p 1 where k cludes the tercept, as show (17). SSE AIC l[ ] 2k (17) The frst term (17) s a measure of the model lack of the ft whle the secod ter (2k) s a pealty term for addtoal parameters model. Schwartz [51] derved from a Bayesa modfcato of the AIC crtero to develop a SBC model. SBC s a fucto of the umber of observatos, the SSE, ad the umber of depedet varables k p 1 where k cludes the tercept, as show (18). SSE SBC l[ ] k l (18) Table 5: Goodess-of-ft test Crtero -2 LOG L AIC SBC The PHREG Procedure Model Ft Statstcs Wthout Covarates 270.544 270.544 270.544 Wth Covarates 253.390 258.428 266.512 Testg Global Null Hypothess: Beta = 0 Test Ch-square DF Pr > ch-square Lkelhood Rato 17.154 4 <0.001 Score 19.218 4 <0.001 Wald 18.356 4 <0.001 The ch-square of lkelhood rato s 17.145 (270.544-253.390). Ths statstc s also show the Table 5 Testg Global Null Hypothess: Beta = 0. The lkelhood-rato test, Score test ad Wald test equals 17.154, 19.218, 18.356 respectvely, wth 4 degree of freedom. Thus, the ull hypothess s rejected (p<0.001). Aother measure of model performace may be some measure aalogous to R 2, as show the formula below. Keep the md that ths measure does ot expla the proporto of varablty of the respose varable by the explaatory varables as the lear regresso. 2 L(0) R 1 [ ] 2 N cs 1 exp[( L(0) L( )) * 2 / N ] = 0.406 L( ) R 2 2 / max 2 N Rcs Rcs = R cs 2 2 /( 1 [ L(0)] ) = 0.63 113

The valdato by Cox & Shell R 2 ad Nagelkerke R 2 shows that the explaatory varables of the predcto model process explaatory power for the cdece of facal dstress. After we have settled o assessg the adequacy of the model that seems a good-ft, we ca carry out statstcal ferece of a ftted model. The output below s produced by rug PROC PHREG wth 5 covarates, ROE, ROA, DER, FAT ad PBV. The RL (RISKLIMTS) opto the Model statemet provdes 95% cofdece terval for the hazard rato estmates. Table 6 s the PHREG procedure. Varable D F Parameter Coeffcet Table 6 The PHREG Procedure Aalyss of Maxmum Lkelhood Estmates Stadard error Chsquare pr> chsq Hazard Rato 95% Hazard Rato Cofdece lmts ROE 1-0.2584 0.1425 5.869 0.0158** 0.769 0.479 1.049. ROA 1-0.4952 0.1860 6.5765 0.0124** 0.606 0.324 0.888 DER 1-0.1308 0.1230 0.3254 0.0105** 0.879 0.456 1.302 FAT 1 0.1546 0.2065 1.6488 0.0209** 1.153 0.568 2.738 PBV 1 0.2438 0.0078 0.2412 0.0242** 1.264 0.642 2.886 ** Sgfcat at 5 percet Results of the aalyss dcate that fve covarates appear to add sgfcatly to the model. The p-value of the parameter estmates for the regresso coeffcets are hghly sgfcat for ROE, ROA, DER, FAT ad PBV. The coeffcet sgs of ROE, ROA DER covarates are egatve dcatg that a crease ether covarate decreases the hazard of eterg to facally dstressed. For example, Hazard rato of ROE covarate s 0.769 ( e 0.2584 0.769). It meas that a crease of oe ut ROE covarate mples 0.769 decreases rsk facal dstress. The terpretato of the estmated hazard rato of ROE s 0.769. It meas that a crease of oe ut the rato of Net come to Total equty wll shrk the hazard rate by 23.1% (1-0.769). The terpretato of the estmated hazard rato of (FAT) s that facal compaes ths study fal at about 1.153 tmes the rate of those o-facal sector. The 95% cofdece terval for hazard rato suggests a sector as low as 0.568 or as hgh as 2.738. The terval wdth equals 2.170 (2.738-0.568). Ths terval also cludes the pot estmate of 1.153 ad does ot cota the ull value of 1. 5.5 Accuraces of classfcato o the model The ft of PH model used ths study s valdated by comparg the predcted value of each sample wth the cutoff value [40]. If the predcted value s below ths cut value, the sample s classfes as a dstressed compay; otherwse, the compay s classfed as o-dstressed compay. Accordg to the suggesto of Mart [41], ths study uses emprcal cutoff value whch s the percetage of facal dstressed samples total sample at 0.264 (46/174). The accuraces of classfcato of the model are compled Table 7. 114

Table 7: The accuraces of classfcato of the mode No-dstressed No-dstressed compay Dstressed compay Accuracy of Classfcato Observed value No-dstressed compay 115 13 115/128 (89.84%) Dstressed compay 8 38 38/ 46 (82.60%) Overall accuracy of classfcato 153/174 (87.93%) Therefore, Type Ⅰ error s 13 /128 = 10.64% ad Type Ⅱ error s 8 /46 =17.50 % ad the overall accuracy of classfcato s 87.93% 6. CONCLUSION I ths paper, the lsted compaes o the Tawa Stock Exchage that expereced data betwee 2003 ad 2009 are employed as dstressed data set. I order to better aalyze the effect of the Cox model predcted, ths study estmated that from 2003 to 2009 sample was dvded to estmatg samples ad forecastg samples. Ths paper selected 56 dstressed compaes ad 154 o-dstressed compaes for estmatg samples data betwee 2003 ad 2006; 46 dstressed compaes ad 128 o-dstressed compaes for forecastg samples data betwee 2007 ad 2009. Form the proposed steps of busess falure predcto model; the facal dstress probablty model s costructed usg Proftablty, Leverage, Effcecy ad Valuato rato varables. Step 1 select the facal ratos for usg the developmet of bakruptcy predcto model. Step 2 cosder the survval problem, the choce of lsted compaes lsted o the Tawa Stock Exchage usg aual data o facal ratos for the perod 2003-2009. I Step 3, there are two ways to reduce the large umber of facal rato (1) Pearso correlato (2) The model accepted has a good ft ad that the mult-learty level s acceptable. I SAS software, PROC RHREG s used to ft the Cox proporto hazards model ad to aalyze the effects of the facal o the survval of the compay. Step 4, oe measure of overall goodess-of-ft test s partal lkelhoodrato test. I SAS software, PROC RHREG s used to obta the lkelhood rato ch-square statstc from the model ft statstcs. The valdato by Cox & Shell R 2 ad Nagelkerke R 2 shows that the explaatory varables of the predcto model process explaatory power for the cdece of facal dstress. I Step 5, cosder the robustess of model predcto accuracy, ths study the accuraces of classfcato of the mode overall accuracy of classfcato s 87.93%. ACKNOWLEDGEMENTS I would lke to thak the aoymous revewers for ther costructve commets o ths paper. 115

REFERENCES [1] Adaa, NHA., Halm, A., Ahmad, H., ad Roha, M. R. (2008), Predctg corporate falure of Malaysa s lsted compaes: Comparg multple dsrtrmat aalyss, logstc regresso ad the Hazard model. Iteratoal Research Joural of Face ad Ecoomc, Vol. 5, pp. 202-217. [2] Akake, H. (1973), Iformato theory ad a exteso of the maxmum prcple. Secod Iteratoal symposum o Iformato theory, pp. 267-281. [3] Altma, E. I., (1968), Facal Ratos, Dscrmat aalyss ad the predcto of corporate bakruptcy. Joural of Face, Vol. 23, No. 4, pp.589-609. [4] Atoaks, A. C. ad Sfakaaks, M. E.,(2009), Assessg aïve Bayesa as a method of screeg credt applcatos. Joural of appled Statstcs, Vol. 36, No. 5, pp. 537-545. [5] Barros, C. P., Butler, R. ad Correa, A. (2010), The legth of stay of golf toursm: A survval aalyss. Toursm Maagemet, Vol. 31, No.1, pp. 13-21. [6] Ba, H. ad Mazlack, L.,(2003), Fuzzy-rough earest-eghbor classfcato approach, 22d Iteratoal Coferece of the North Amerca, 24-26 July, 2003. [7] Bog, P., Ferr, G. ad Hahm, H.(2000), Corporate bakruptcy Korea: Oly the strog survval?. Facal Revew, Vol. 35, pp. 71-111. [8] Cao, R. Vlar, J. M. ad Deva, A., (2009), Modelg cosumer credt rsk va survval aalyss. Statstcs & Operatoal Research Trasactos, Vol. 33, No. 1, pp. 3-30. [9] Cataach, A. H. ad Perry, S. H. (2001), A evaluato of survval model s cotrbuto to thrft sttuto dstress predcto. Joural of Maageral Issue, Vol.13,No.4, pp. 401-417. [10] Chaduhur, A, ad De, K., (2011), Fuzzy support vector mache for bakruptcy predcto, Appled Soft Computg, Vol. 11, No. 2, pp.2472-2486. [11] Che W. Ma, C. ad Ma, L.,(2009), Mg the customer credt usg hybrd support vector mache techques. Expert Systems wth Applcato, Vol. 36, No. 4, pp. 7611-7616. [12] Che, D., (2009), Usg eural etworks ad data mg techques for facal dstress predcto model. Expert Systems wth Applcato, Vol. 36, No. 2, pp.4075-4086 [13] Che, M. Y., (2011), Predcto corporate facal dstress based o tegrato of decso tree classfcato ad logstc regresso. Expert Systems wth Applcato, Vol. 38, No. 9, pp.11261-11272. [14] Cheg, J. U., Yeh, C. H. ad Chu, Y. W.,(2007), Improvg busess falure predcto usg rough sets wth No-facal varables. LNCS, Vol. 4431/2007, 2007, pp. 614-621. [15] Cox, D. R. & D. Oakes, (1984), Aalyss of Survval Data. Lodo: Chapma ad Hall. [16] Cox, D. R. (1972), Regresso models ad lfe-tables, Joural of Royal Statstcal Socety B, Vol. 34, pp. 187-220. [17] Dmtras, A. L., Slowsk, R., Susmaga, R. ad Zopouds, C., (1999), Busess falure Predcto usg rough sets. Europea Joural of Operatoal Research, Vol. 114, pp. 263-280 [18] Gepp, A. ad Kumar, K. (2008), The role of survval aalyss facal dstress predcto. Iteratoal Research Joural of Face ad Ecoomc, Vol. 16, pp.13-34. [19] Gbso, C. H. (2006), Facal Reportg ad aalyss: usg facal assocatg formato. South-Wester College Pub, 10 edtos, Aprl, 27. [20] GIGS (2006): Global Idustry Classfcato Stadard (GICS@), Stadard & Poor s, August, 2006 [21] Gokoval, U., Bahar, Q. ad Kozak, M. (2007), Determats of legth of study: A practcal use of survval aalyss. Toursm Maagemet, Vol. 28, No.3, pp.736-746. [22] Golub, J. (2007), Survval aalyss ad the Europea Uo decso-makg. Europea Uo Poltcs, Vol. 8, No. 2, pp. 155-179. [23] Hu, X. T., L, T. Y. ad Ha, J., (2003), A ew rough sets model based o database systems, Lecture Notes Artfcal tellgece, Vol. 2639, pp. 114-121 [24] Ice, H. ad Akta, B., (2009), A comparso of data mg techques for credt scorg bakg: a maageral perspectve. Joural of Busess Ecoomcs ad Maagemet, Vol. 3, No. 2, pp. 233-240 [25] Islam, M. J., Wu, Q. M. J., Ahmad, M. ad Sd-Ahmed, M. A.,(2007), Ivestgatg the performace of Naïve- Bayes Classfer ad K-NN classfers. Iteratoal Coferece o Covergece 116

Iformato Techology, IEEE Computer Socety. 2007. [26] Javath, J., Suresh Joseph, K. ad Vashav, J. (2011), Bakruptcy predcto usg AVM ad Hybrd SVM survey. Iteratoal Joural of Computer Applcatos, Vol. 34, No.7, pp. 39-45. [27] Joes, S. ad Hesher, D. A. (2004), Predctg frm facal dstress: A mxed logt Model. Accoutg Revew, Vol.79, No. 4, pp. 1011-1038. [28] Karels,G. V. ad Prakash, A. J. (1987), Multvaate ormalty ad forecastg of corporate bakruptcy. Joural of Busess Face ad accout, Vol. 14, No. 4, pp. 573-592. [29] Kauffma, R. ad Wag, B. (2001), The success ad falure of dotcoms: A mult-method survval aalyss. I proceedg of the 6th INFORMS Coferece o Iformato System ad Techology (CIST), Mam, FL, USA [30] Kauffma, R. ad Wag, B. (2003), Durato the dgtal ecoomy: Emprcal bases for the survval of teret frms. I 36th Hawa Iteratoal Coferece o System Sceces (HICSS), Hawa. [31] Keasey, K., McGuess, P. ad Shot, H. (1990), Multlogt approach to predctg corporate falure: Further aalyss ad the ssue of sgal cosstecy. Omega, Vol. 18, No. 1, pp. 85-94. [32] Km, H. ad Soh, S. Y., (2010), Support vector maches for default predcto of SMEs based o techology credt. Europea Joural of Operatoal Research, Vol. 201, No. 3, pp. 838-846. [33] Late, E. K. ad Luoma, M. (1991), Survval aalyss as a tool for compay falure predcto, Omega, Vol. 19, No. 6, pp. 673-678. [34] Lau, S. T., Cheg, B. W., ad Hseh, C. H., (2009), Predcto model buldg wth clusterglauched classfcato ad support vector maches credt scorg. Expert Systems wth Applcatos, Vol. 36, No. 4, pp. 7526-7556. [35] LeClere, M. J. (2000), The occurrece ad tmg of evets: Survval aalyss appled to the study of facal dstress. Joural of Accoutg Lterature, Vol. 19, pp. 158-189. [36] Lag, Y. ad X, H., (2009), Applcato of Dscretzato the use of Logstc Facal Ratg. Iteratoal Coferece o Busess Itellgece ad Facal Egeerg, 24-26 July 2009, pp. 364-368 [37] L, L. ad Pesse, J. (2004), The detfcato of corporate dstress UK dustrals: A codtoal probablty aalyss approach. Appled Facal Ecoomcs, Vol. 14, pp. 73-82. [38] Lu, K., La, K. K., ad Guu, s. M., (2009), Dyamc credt scorg o cosumer behavor usg Fuzzy Markov model. Fourth Iteratoal Mult-Coferece o Computg the Global Iformato Techology. IEEE Computer Socety, 2009. [39] Lopez, R. F., (2007), Modelg of surers ratg determats. A applcato of mache learg techques ad statstcal models. Europea Joural of Operatoal Research, Vol. 183, No. 2. PP.1488-1512. [40] Lu, Y. C. ad Chag, S. L., (2009), Corporate goverace ad qualty of facal formato o predcto power of facal dstress of lsted compaes Tawa. Iteratoal Research Joural of Face ad ecoomcs, Vol. 32, pp. 114-138. [41] Mart, D. (1977), Early warg of bak falure: a Logt regresso approach. Joural of Bakg ad Face, Vol. 1, pp. 249-276. [42] Mcleay, S. ad Omar, A. (2000), The sestvty of predcto models to the o-ormalty of boud ad ubouded facal ratos, The Brtsh Accoutg Revew, Vol. 32, No. 2, pp. 213-230. [43] Mossma, C. E., Bell, G. G. Swartz, L. M. ad Turtle, H. (1998), A emprcal comparso of bakruptcy models. Facal Revew, Vol. 33, No. 2, pp. 35-53. [44] Mues, C. Baeses, B., Fles, C. M., ad Vathee, J., (2004), Decso dagrams mache learg: a emprcal study o real-lfe credt-rsk data. Expert Systems wth Applcatos, Vol. 27, No. 2, pp. 257-264 [45] Nagelkerke, N. J.D. (1992) axmum Lkelhood Estmato of Fuctoal Relatoshps, Pays-Bas. Lecture Notes Statstcs, Vol. 69, pp. 110 [46] Odom, M., Sharda, R., (1990), A eural etwork model for bakruptcy predcto. IEEE INNS Iteratoal Jot Coferece o Neural Networks, Vol. 12, 1990, pp.163-168. [47] Ohlso, J. A. (1980), Facal Ratos ad the Probablstc Predcto of Bakruptcy. Joural of Accoutg Research, (Sprg,1980), Vol. 14, No. 1, pp. 109-131. 117

[48] Rommer, A. D. (2005), A comparatve aalyss of the determats of facal dstress Frech, Itala ad Spash frms. Workg paper, Damarks Natoal Bak, Copehage. 18, May, 2005. [49] Routledge, J. ad Gadee, D. (2000), Facal dstress, reorgazato ad corporate performace. Accoutg ad Facal, Vol. 40, No. 3, pp. 233-259. [50] Sarkar, S., ad R. S. Srram., (2001), Bayesa Models for Early Warg of Bak Falures. Maagemet Scece, Vol. 47, No. 11, pp.1457 1475. [51] Schwartz, G. (1978), Estmatg the dmeso of a model. Aals of Statstcs, Vol. 6, pp. 461-464. [52] Sh, K. S., Lee, T. S. ad Km, H. J.,(2005), A applcato of support vector maches bakruptcy predcto model. Expert system Applcato, Vol. 28, No. 1, pp. 127-135 [53] Shumway, T. (2001), Forecastg bakruptcy more accurately: A smple hazard model. Joural of Busess, Vol. 74, pp. 101-124. [54] Sloot, T. ad Verschure, P. (1990), Decso-makg sped the Europea commuty. Joural of Commo Market Studes, Vol. 29, No.1, pp. 75-85. [55] Soh, S. Y.ad Shm, H. W.,( 2006), Reject ferece credt operatos based o survval aalyss. Expert Systems wth Applcatos, Vol. 31, No. 1, pp. 26-29. [56] Stefaescu, C. Tuaru, R., ad Tubull, S., (2009), The credt ratg process ad estmato of trasto probabltes: a Bayesa approach, Joural of Emprcal Face, Vol. 16, No. 2, pp. 216-234 [57] Stepaova, M. ad Thomas, L. C.,(2002), Survval aalyss methods for persoal loa data. Operatos Research, Vol. 50, No. 2, pp. 277-289. [58] Su, J. ad LI, H. L., (2009), Facal dstress predcto based o seral combato of multple classfcato. Expert Systems wth Applcato, Vol. 36, No. 4, pp.8659-8666 [59] Tabachck, B. G. ad Fdell, L. S., (2000), Usg Multvarate Statstcs. Ally ad Baco Press, UK. 2000 [60] Tam, K. Y. ad Kag, M., (1992), Maageral applcatos of eural etwork: the case of bak falure predctos. Maagemet Sceces, Vol. 38, pp.927-947 [61] Ta, C. N. W. ad Dhardjo, H. (2001), A study o usg artfcal eural etworks to develop a early warg predctor for credt uo facal dstress wth comparso to the probt model. Maageral Facal, Vol. 27, No. 4, pp. 56-77. [62] Tug, W. L., Quek, C., Cheg, P. ad EWS, G.,(2004), a ovel eural-fuzzy based early warg system for predctg bak falures. Neural Networks, Vol. 17, No.4, PP. 567-587. [63] Wag, Y., Wag, S. ad La, K. K., (2005), A ew fuzzy support vector Mache to evaluate credt rsk. IEEE Trasactos Fuzzy Systems, Vol. 13, No. 6, pp. 820-831. [64] We, L. L. ad Zhag, W. X., (2003), Probablstc rough characterzed by Fuzzy Sets. Lecture Notes Artfcal Itellgece, 2369, pp. 173-180 [65] Wog, S. K.M., Zalo, W. ad L, Y. R., (1986), Comparso of rough set ad statstcal methods ductve learg. Iteratoal Joural of Ma-Mache Studed, Vol. 24, pp. 53-72. [66] Xao, Z., Yag, X., Pag, Y. ad Dag, X.,(2011), The predcto for lsted compaes facal dstress by usg multple predcto etwork wth rough set ad Dempster -Shafer evdece theory. Kowledge Based Systems, Vol. 26, PP.196-206. [67] Xhu, C. ad Zhog, Q., (2009), Cosumer credt scorg based o mult-crtera fuzzy logc, Iteratoal Coferece o Busess Itellgece ad Facal Egeerg. IEEE Computer Socety. 2009. [68] Yag, Z., You, W. ad J, G. (2011), Usg partal least squares ad support vector maches for bakruptcy predcto, Expert Systems wth Applcatos. Vol.38, No.7, pp. 8336-8342. [69] Yap, BCF., Muaswamy, S., ad Mohamad, Z.,(2012), Evaluatg Compay Falure Malaysa Usg Facal Ratos ad Logstc Regresso. Asa Joural of Face & Accoutg, Vol. 4, No. 1, pp. 330-344. [70] Yu, L., Wag, S., We, F., La, K. K., ad He, S.,(2008), Desgg a hybrd tellget mg system for credt rsk evaluato. Joural of Systems Scece ad Complexty, Vol. 21, No. 4, pp. 527-539. [71] Zhag, D., Hf, M. Che, Q. ad Ye, W.,(2008), A hybrd credt scorg based o geetc 118

programmg ad support vector mache. 4th Iteratoal Coferece o Natural Computato, IEEE Computer Socety, 2008, pp. 8-12 [72] Zhao, H, (2007), A mult-objectve geetc programmg approach to developg Pareto optmal decso tree. Decso Support Systems, Vol. 43, No. 3, pp. 809-826. [73] Zhou, J. ad Ba, T., (2008), Credt rsk assessmet usg rough set theory ad GA-based SVM. The 3rd Iteratoal Coferece o Grd ad Pervasve Computg- Workshops, IEEE Computer Socety Kumg, 2008, pp. 320-325. [74] Zhou, L. ad La, K. K. ad Yu, L.,(2009), Credt scorg usg support vector mache wth drect search for parameters selecto. Soft Computg, Vol. 13, No. 2, pp. 149-155. [75] Zhou, L., La, K. K., ad Yu, L.,(2010), Least squares support vector maches esemble models for credt scorg. Expert Systems wth Applcatos, Vol. 37, No. 1, pp. 127-133. [76] Zulkarma, M. S., Mohamad, A. A. H., Auar, M. N. ad Zaal, A. M. (2001), Forecastg corporate falure Malaysa dustral sector frms. Asa Acdemy of Maagemet Joural, Vol. 6, No.1, pp. 15-30. [77] Buyamu, A. ad Bashru, S. (2014), Corporate falure predcto: A fresh techque for dealg effectvely wth ormalty based o quattatve ad qualtatve approach, Iteratoal Joural of Facal Ecoomc, Vol. 3, No. 1, pp. 1-12. Author Mg-Chag Lee s Assstat Professor at Natoal Kaohsug Uversty of Appled Sceces. Hs qualfcatos clude a Master degree appled Mathematcs from Natoal Tsg Hua Uversty ad a PhD degree Idustral Maagemet from Natoal Cheg Kug Uversty. Hs research terests clude kowledge maagemet, parallel computg, ad data aalyss. Hs publcatos clude artcles the joural of Computer & Mathematcs wth Applcatos, Iteratoal Joural of Operato Research, Computers & Egeerg, Amerca Joural of Appled Scece ad Computers, Idustral Egeerg, Iteratoal Joural ovato ad Learg, It. J. Servces ad Stadards, Lecture Notes computer Scece (LNCS), Iteratoal Joural of Computer Scece ad Network Securty, Joural of Covergece Iformato Techology ad Iteratoal Joural of Advacemets computg Techology. 119