Main Question: which is the best model that is based on the classic statistical analysis to predict the distressed companies in Chinese listed manufacturers?　To elaborate my research, this paper tries to collect the data from public annual report of each company and use the statistical analysis software-SPSS to select the best financial variables and the most optimal indication for the prediction model.
Sub-problems:
A. Why this article chooses the Chinese listed manufacture companies as research object?
The first reason is that majority of the listed companies publish complete financial information because governments take strict safety measure to protect benefits of investors. To be more specific, China Securities Regulatory Commission (CSRC) asks listed companies to provide detailed and accurate information about the IPO and daily operation phase, so the financial data from annual report is convenient to collect and can be trusted. Secondly, Platt's study (1990) shows that "industry effects on corporate failure were significant". In other words, the statistic model that only focuses on one industry possesses better predictive ability than models which reflect the whole market in general. Thirdly, manufacturing industry is also considered to be the most crucial pillar which braces the China's economy rising. In other words, the brilliant economic development in china is because of the transferring of global manufacturing. Specifically, the open policy and relatively lower labor and environment cost attract global investment, especially in manufacturing industry. Thus, the importance of manufacturing industry can not be neglected and the research is meaningful.
The relationship between bank lending decision and distressed companies
Firstly, this question relates to banking credit risk management, as we know, credit risk for banking system normally concerns loan risk and security risk. Considering the particular situation in China, loans are still the main assets of commercial banks and the major sources of banks' profits. Meanwhile, China has not developed the securitization market well (and thanks for that, China survived in the period of sub-prime crisis). To sum up, management deficiency and lack of innovation of banks might lead to failure due to the uncertain competition environment. Furthermore, it is essential for banks in China to build a warning system that foreign financial institutions have already established.
Considering lending decision includes two participate parties: lender-banks and borrower-enterprises, there are also studies discussing both behaviors and trying to reduce risk exposure from different angle. In this paper, the risk mitigation question is translated to financial failure prediction of enterprises using statistical assessment. Because the current economic circumstances Chinese banks are facing which introduced in previous section, the priority task in banks' risk management is to establish a continuous analysis of the financial situation of loan companies. In other words, financial distress prediction can be an efficient measurement of credit to detect possible obligation default at early stage. Furthermore, many scholars do researches follow this thread and this will be further expanded in the literature review section.
C. Why the statistic method has been chosen
Over the last 40 years, business failure prediction has been studied by scholars. Generally, there are two method categories used: statistical methods and non-statistical methods (Balcaen & Ooghe, 2006). Balcaen and Ooghe also give a sketch of statistical methods which contain the univariate models (Beaver, 1967 et al), risk index models (Tamari, 1966 et al), multiple discriminant analysis (MDA) models (Altman, 1968 et al), and conditional probability models (Zavgren, 1983 et al), and details have been given in literature review part. But here I want elaborate why I use statistical method to build the prediction model.
First of all, statistic methods have been used for a long time and the modules are more developed than non-statistical methods which have just been studied in the past 10 years, so the maturity of non-statistical methods have been questioned. Meanwhile, those non-statistical methods are mostly complicated and difficult for practical utilization. Secondly, statistical method is considered as objective measurement criteria used to ensure that business financial failures are satisfactorily predicted because the quantitative data collection method. As we know, financial information from annual report can provide insight into a company's risk position, furthermore, the indicators based on annual report tend to be reviewed on periodic basis (quarterly or annually) to alert banks to change their lending decision. Moreover, Chinese scholars who studied business failure are more focused on the whole security market rather than one sector, and what is more, the previous researches about Chinese manufacturing industry (Chen et al, 2000; Wu et al, 2001; Li, 2001; et al) have some sampling issues, for instance, the samples are not sufficiency enough. These drawbacks of early researches would be avoided in my research. In other words, this paper will try to perfect the prediction model in the near future.
2.2 The limitations
The limitation of sampling and conventional statistic method will be further discussed in the literature review. Although the LA statistic method used in this report is the most prevailing method currently, this research not cross-compared with the latest neural networks method (Odom and Sharda, 1990) or others new development hybrid research methods because the those method is too complicated for regular utilization and they are still theoretical modes.
Furthermore, although the samples used in this research are all listed manufacturing companies, it is still difficult to avoid the issue of sample selecting bias which mentioned in the literature review. Additionally, the samples are all chosen from listed companies, this brings limited prediction scope to the model. Specifically, the private section has not been calculated.
2.3 The definitions of terms
Because the studying object is ST listed companies, the research method uses conventional statistical model. Only the definition of the ST Company will be given here, the definition of the research method will be further given in literature review.
This research uses the Chinese criterion, special treatment (ST) to define the financial distress. According to the definition applied by China Securities Regulatory Commission (CSRC), the ST will be only added as a title of listed company under such circumstances as followed:
"(1) the company has been in the red in the most recent two consecutive years (based on the audited net profits disclosed in the latest two annual reports);
(2) After correcting the serious errors or falsehoods in its financial report either on its own initiative or by the order of the CSRC, the company adjusts its previous financial reports retroactively and as a result, the company goes into red for the most recent two consecutive years;
(3) The company has been ordered by the CSRC to correct the serious errors or falsehoods in its financial report but fails to mend its way within the specified time limit, and the company's stocks have been suspended from trading for two months;
(4) The company fails to disclose its annual report or interim report with the statutory period and the company's stocks have been suspended from trading for two months;
(5) The company is likely to be dissolved;
(6) The court has accepted the company's application for reorganization, settlement or bankruptcy liquidation;
(7) As equity distribution as prescribed hereof renders the company unsuitable for listing, the company submits to the exchange a plan for addressing the equity structure problem and obtains approval for exchange; or
(8) Other circumstances as recognized by the Exchange" (Shanghai Stock Exchange, 2008).
Majority of overseas studies hold a different opinion about the definition of the financial distress. Normally, they regard broken or nearly broken companies as companies in financial distress. However, considering the specific condition in china, using ST listed companies as observed objects is applicable to business distress studying. Firstly, major commercial banks and the Ministry of Finance of China tend to support firms in severe financial condition (especially state owned listed companies) decided by their traditional role (the last stander for state owned property) or political considering (The presentation of economic stability is one of standards for assessing Chinese officials, and theoretically, share market stability also attract investments), and this helps keep certain companies alive (not bankruptcy) in the Chinese market; secondly, it is extremely rare to see the handling of broken companies, as approved IPO activities are strictly under the control of Chinese government, therefore, even if a company is in financial distress, other capitals will be invested due to the limited possible number of listed companies. Thus, it is inappropriate to measure financial distress of Chinese companies using brokerage, instead, a better standard with more Chinese characteristic, ST listed companies, should be utilized to measure the status of financial distress of Chinese companies.
Additionally, based on the definition of ST companies given by Shanghai Stock Exchange, ST listed company is also similar to some western scholars' definitions of financial distress, for instance, it also uses interest cover ratio (Asquith, Gertner and Scharfstein, 1994); cash flow ratios (Whitaker, 1999) and the change of equity price (Netter, 1992) as measurement criterion.
2.4 The importance of the study
Relationship between establishing loans warning system of banks and forecasting the financial dilemma of enterprises and clear reasons why listed manufacture companies have been chosen as subjects all show the practical significance of research. Generally, the idea is that a company does not fall into financial distress immediately and it is a gradual process which is appropriate for developing an effective method to predict the financial crisis before it occurs, and it is especially meaningful to banks' lending decision.
The meaning may reveal itself more clearly if I demonstrate the drawbacks of Chinese banking system, especially in making a lending decision. Firstly, although china has made a huge progress in the marketization, the government, or precisely the communist party, controls economic functions, especially financial department-banks. In such circumstances, administrations may have an impact on banks lending decisions. For example, during the 1980s, China continually kept tremendous GDP growth by increasing infrastructure projects and export surplus. In this period, bankers are reluctant to subordinate themselves to unify government control. It means that, banks may lend loans to the state-owned enterprises (some of them are listed companies) even they are heavily in debt and hardly to pay off the loans. Since Chinese financial institutions are increasingly in line with international standards, it is essential for banks to establish an objective and quantitative verification system to manage loans. Secondly, when China joins WTO, it also means open the financial market to foreign companies. As competition become fierce, Chinese banks also need to implement a quantificational inspection to the loan enterprise for eliminating risk exposures. Generally, because the change of circumstance those Chinese bankers are gradually free from the administrative guidance and change their strategies which are obeying market principle. Building a quantitative model to loans management seems urgent for banks in China.
3. The review of the related literature
The literature review can be divided into two parts. The first section will review the development of using statistical model to predict companies' financial failure. In this part, the history of model developing will be given and there is also critical thinking evaluation of various models. And then the logit analysis (LA) model (Ohlson, 1980) will be analyzed clearly, the reason why using LA model for this research will also be given. The second part will focus on contributions of Chinese scholars. With compared analysis, the drawbacks of those previous researches will be concerned and improvements might be given in methodology section.
A.1 Brief introduction to the development of statistic modules
In the recent 40 years, business distress prediction model has become a major issue in corporate finance studies. Many scholars are dedicated to build an ideal model to predict business financial failure. The brief history of traditional statistical methods development has been shown in this section and critical analysis also be conducted to different parts.
A.1.1 Univariate model and risk index model
Beaver (1966) pioneers a business distress model with financial ratios. He selects samples by dichotomous classification test-it means the financial data of 79 failure cases and 79 corresponding successful enterprises have been collected. From these two samples sets, He uses mean comparisons and broken down likelihood analysis methods to develop a univariate model-several financial ratios have been used as discrimination standards to predict business failure separately. In the situation of excluding industry factor, Beaver (1966) finds there are three valid financial ratios to forecast financial failure, among these three options, the indicator of cash flow/total debt represent the best. Using this index to predict the business failure in the following period (t+1), the accuracy rate is as high as 87%; it is a significant contribution at that time. However, the defects of this method are also obvious, (1). The accuracy rate is not high enough for current utilization. (2). It is hard to show the reason why it causes enterprise into financial dilemma. In other words, this index only works like a trigger but do not analyze the situation in detail. (3). this model is based on the whole stock market and neglects the industry effect. It is easy to understand that manufactory is so different from other businesses (from the production life cycle to cash flow ratio). Thus, the statistic model which only focuses on one industry might provide better predictive capability than the model that only reflects the whole market. (4).The model is also based on the stringent assumption of a linear relationship between all measures and the failure status. This feature is hard for practical utilization. (5). Finally, each univariate model has fatal defect, so that any single index is unable to reflect the business characteristic comprehensively. However the advantage of using univariate model is also clear, that the technique is extremely simple and the application does not require any statistical knowledge.
Therefore, Tamari (1966); Moses and Liao (1987) develop a risk index model to eliminate the negative impacts of using univariate model, they try to use easy handled point system based on several financial sets to predict precisely. However, the index system is just simply sum up the unvariate ratio that means they are naive to simplify the relationship between each individual ratio or they simplify the correlation coefficient in the formulation. Secondly, the weight of each variable is allocated by subjective assumption that causes the model unreliable. Finally, they do not consider the multicollinearity between selected ratios. That may causes the some factors are exaggerated for double counting the related financial ratios.
A.1.2 Multiple discriminant analysis (MDA)
A.1.2 Multiple discriminant analysis (MDA)
Considering the inefficiency of previous model building, Altman (1968) first introduces a statistic multivariate analysis (MDA) to the problem of business failure prediction. MDA is a "statistical technique used to classify an observation into one of several a priori groups dependent upon the observation's individual characteristics…it attempts to derive a linear combination of these characteristics which 'best' discriminates between the groups" (Altman, 1968). Altman (1968) observes financial ratios of 33 bankrupt enterprises and same amounts of successful businesses from 1946-1965, and he selects five ratios from twenty two observed indicators which might be filtered out to provide more information of the financial health and established a distinction function. When independent variables have been decided, Altman (1968) puts them into the discriminant function for linear MDA model which is as follows (Lachenbruch, 1975):
where is the discriminant score for the firm i; (with j=1,…..,n) is the variables' value for the firm i; is the intercept; and is the linear discriminant coefficient for attribute j. Therefore, the formulation generates several firm financial characteristics into one single discriminant score- and can clear indicate business' financial health. Back to research of Altman (1968), he uses Z-score model to demonstrate the financial situation of observed company which as
Z=0.012+0.014+0.033+0.006+0.999 (Altman, 1968)
Where Z is overall index for the enterprise;
=net working capital / total assets which measures liquidity in the relation of company size.
= retained earnings/ total assets which reflect profitability in period basis.
=earnings before interest and taxes (EBIT)/total assets which indicates business' operating efficiency apart from tax and leverage factors. It also demonstrates long term performance of enterprises.
=market value of common and preferred stock/book value of debt which considers security price changing as a market dimension.
=sales/total assets which shows sales turnover that might varies greatly in different industry.
Furthermore, Altman calculates the 50-50 point is 2.675 which means that the company might has the same probability as to survive or to bankrupt; and his zone of ignorance is from Z=1.81 to Z=2.99, that means it is uncertain in this scope about how the observed companies should be classified; meanwhile, he points out that the larger Z score can indicate more financial health. He collects the financial ratios from 1970 to 1973 and the accuracy rate of this model is up to 82% in the following period (t+1). To pursue a better prediction ability, Altman et al. (1977) adjust Z-score model to better performing Zeta analysis model, which concludes more financial variables; based on the longer period data collection and more enterprises are included in the sample set. Thus the accuracy rate of prediction provided from Zeta model is up to 91% in the following period (t+1). Because the high reliable of MDA technique, it dominated the literature of business distress prediction until the 1980s.
However, the biggest limitation of MDA technique is that it us based on several strict assumptions (Edmister, 1972; Eisenbeis, 1977; Zavgren, 1983; Karels and Prakash, 1987). Firstly, the samples are dichotomous. It means that both bankruptcy and successful companies are equal; furthermore, Balcaen and Ooghe (2006) argues that it makes "groups are discrete, non-overlapping and identifiable." But it also leads to a more restrictive mathematics hypothesis (Balcaen & Ooghe, 2006): (1), "multivariate normally distributed independent variables"; (2). "equal variance matrices across the failing and non-failing group"; (3). "specified prior probability of failure and misclassification costs". If the data collection cannot satisfy these assumptions which may cause MDA method to be used in an inappropriate way and the results from MDA analysis are not suitable for further prediction (Joy and Tollefson, 1975; Eisenbeis, 1977; Richardson & Dividson, 1984; Zavgren, 1985).
However, these assumptions are too strict, which lead to difficult sampling application in the reality. First of all, Balcaen and Ooghe argues that multivariate normality is often violated (Deakin, 1976; Taffler and Tisshaw, 1977; Barnes, 1987), this conclusion is based on significance tests with the influences of error rates (Eisenbeis,1977; Richardson and Davidson, 1984; Mcley and Omar, 2000). Some researchers want to test univariate normality instead of multivariate normality, these researchers tend to believe that when a variable can satisfy the requirement of univariate normality, it will also satisfy the requirement of multivariate normality when used in equation. These researchers might simplify the issue too much, they tend to ignore the following problem: firstly, univariate normality is not a necessary condition of multivariate normality, or should we say that the logical relationship between these two are uncertain; secondly, in an equation, all the variables and their relationships might be changed under different conditions, thus the MDA model might be distorted (Eisenbeis,1977; Ezzambel and Mar-Molinero,1990). There are other researchers who try to estimate multivariate normality by trimming the outliners, but it will also cause a significant information loss (Ezzambel and Mar-Molinero,1990).
Secondly, scholars find that the data is difficult to satisfy the assumption of equal dispersion matrices through significance testing. Under such circumstance, a quadratic MDA model needs to be used (Joy and Tollefson,1975; Eisenbeis,1977; Zavgren,1983). But due to the complicity of quadratic MDA model, satisfying conclusion can only be drawn when sampling size is significant enough. In reality, scholars tend to transform data into a format which is suitable for equal dispersion matrices and then apply linear MDA (Taffler, 1982).
Thirdly, a lot of the scholars do not consider the issue of prior probability of failure and misclassification costs, and this might lead to false setting of the optimal cut-off score of MDA model, thus misleading the estimation of model accuracy. (Edmister, 1972; Eisenbeis, 1977; Deakin,1977; Zavgren,1983; Hsieh,1993;Steele, 1995).
There are also some limits of statistic method that Altman has not concerned: firstly, he tries to use market data collecting and analyzing (for instance, he adds fluctuating of stock value in formulation) to predict market event (i.e. business failure), but that market value might be influenced by whole share market performance which is hard to predict. Secondly, the model is relying on unfeigned data from annually or quarterly report and that data can be fiddled by managers for concealing true financial position. Generally, the opacity of balance sheet can cause the inefficiency of model. Thirdly, he dose not consider there are increasing using of off-balance sheet items for business. That omission of monitoring can mislead regulators and bankers and suddenly bankrupt of Baring bank provides footnote for this problem. Furthermore, there might be severe multicollinearity among independent variables but actually this assumption is still being questioned. (Doumpos and Zopoudinis, 1999).
A.1.3 Conditional probability models (especially Logit analysis)
Because the limitations of MDA methods, the utilization of it has benn decreasing since 1980s (Dimitras et al. 1996), but it remains the main method for comparative researches and it is still generally accepted as the standard for distress predicting. MDA was gradually replaced by conditional probability models afterward, such as, logit analysis (LA), probit analysis (PA) (Zavgren, 1983; Zavgren, 1985; Doumpos & Zopoudinis,1999). Among all these methods, Ohlson (1980) pioneers using LA model in business distress prediction. Even till now, due to the sampling requirement of LA model is relatively lower than other models, it is still the prevailing model in the academic research.
Generally, conditional probability model permits the use of the non-linear maximum likelihood assumption to predict business distress. Each model is based on their probability distributions respectively: logit model (LA) bases on logistic distribution assumption (Maddala, 1977; Hosmer & Lemeshow, 1989); probit analysis (PA) is built on a cumulative normal distribution assumption (Theil, 1971). Non-linear probability models are very different from linear probability models which assume that the relationship between variables and distress probability is linear (Altman et al. 1981), and non-linear probability models provide more accuracy of prediction and they only requires loose numeral statistics. Furthermore, as the most popular conditional probability methods, LA is widely used for business failure prediction, therefore, LA method will be used in this paper. In the following part, the comparative advantage of LA model will be introduced compared with PA.
As stated above, LA model obtains the parameter estimates based on logistic distribution assumption through non-linear maximum likelihood method. (Hosmer & Lemeshow,1989; Gujarati, 2003). The formulation is as followed:
Whereas "" is the probability of failure given the vector of attributes; is the value of attribute j (with j=1,…, n) for firm i; is the intercept and is the coefficient for attribute j and it also shows the importance of each independent variables separately. Thus LA model combines several financial characteristics into a general score which predict the probability of business distress. In practical use, higher logit score indicates a higher probability of enterprise failure and the LA model sets up a cut-off point that assigns firms to the failing or non-failing group.
LA model is considered as a less demanding in prior assumption than MDA and LA model becomes popular in current literature due to the special characteristics of it, that LA model does not require "multivariate normal distributed variables or equal dispersion matrices" (Ohlson, 1980).
However, it should be emphasized that LA models are still extremely sensitive to multicallinearity (Ooghe et al. 1993; Doumpos and Zopoudinis,1999). The cause of multicallinearity problem is because that LA models are based on financial ratios collected from annual report. These ratios have very tight relationships, serious multicallinearit disturbance will occur if these highly related ratios are used in models without further handling and this will lead to inappropriate cut-off point.
Additionally, probit analysis model (PA) is considered as an advanced model. It is presented as followed:
However, PA model is the same as MDA model; they both require normal distributed variables. This does not simplify the modeling process, instead, it. Thus, this method is less employed than LM (Dimitras et al., 1996).
A.2 critically analyze the classic statistical methods
Although the advantages and disadvantages of all models are discusses beforehand, the general problems related to the classic statistical methods will be critically analyzed too in this paper.
Balcaen and Ooghe (2006) points out that the false prediction of classic statistical models are due to various reasons: first of all, the use of classical paradigm leads to several problems because the classical paradigm is unable to consider certain important business failure prediction; second of all, there are certain problems that are related to the neglect of time dimension; third of all, some problems is due to the wrong modeling focus; last but not least, problems might also be caused by the "use of linear classification rule, the use of annual account information, and the neglect of the multidimensional nature of failure."
A.2.1 problems related to the classical paradigm
Classical statistical methods, such as MDA models and conditional probability models categorized listed companies into failing and un-failing group which means they all starts from classical paradigm and Hand (2004) defines classical paradigm as 'given a set of firms with known descriptor variables and known outcome class membership, a rule is constructed which allows other companies to be assigned to an outcome class on the basis of their descriptor variables.' However, classical paradigm had internal defect itself (Hand, 2004):
First of all, it assumes that financial distress of companies can all be presented from model, it might lead to neglect of the real individual cause of corporate failure;
Second of all, Due to the interpretation of financial distress by different scholars, the categorization of samples might be rather subjective. For instance, in most of the researches, financial distress is defined as bankruptcy (Dirickx and Van Landeghem, 1994; Ward and Foster,1997; Van Caillie,1999; Daubie and Meskens,2002; Charious et al. 2004). On the other hand, some scholars categorize failure-related events into financial distress, for example, cash insolvency (Laitinen, 1994), Loan default (Ward and Foster, 1997), capital reconstructions and loan covenant renegotiations with bankers (Taffler and Agarwal, 2003). additionally, Balcaen and Ooghe (2006) argues that there are several specific problems related to the use of bankruptcy to define business failure. 1，bankruptcy might be one of the strategic decisions of firms, for example, companies might choose to broke in order to avoid part of the debt so they could restart their activities with a clean sheet. 2，some bankruptcy might be caused by unexpected events, e.g. natural disaster or war. Therefore, Hill et al.(1996) and Huang (2004) apply a concept of 'accident bankruptcy'. 3，some scholars neglect the fact that bankruptcy is just part of a large range of possible endings of distress procedure. For example, company financial distress might display itself in the form of merger or absorption. 4，some scholars believe that rescue attempted by managerial level might lead to the actual timing of bankruptcy is much later than the real moment of financial failure, this will also lead to the inaccuracy of modeling (Theodossiou, 1993; Ooghe et al., 1995; Pompe and Bilderbeek, 2000). additionally, based on the understanding to the definition of ST company, it can be found that the essence of ST company might lead to more possibilities to the cause of financial distress (compared with only use bankruptcy to measure financial distress), ST companies are only active under the confirmation of state authority - Shanghai Stock Exchange, therefore, to use ST listed companies as observed objects is adequate and it simplifies the procedure of screening of data.
Thirdly, another major defect of classical paradigm is due to the actual non-stationarity and data instability which lead to the lose of efficiency of models。 Classical statistical methods require the relationships between variables are stable and remain the same in future sample, which means stability. (Edister,1972; Zavgren,1983; Mensah, 1984; Jones, 1987). However, there are numerous of actual data prove that non-stationarity and data instability (Barens,1982; Richardson and Davidson, 1984; Zmijewski,1984). And the data instability might be caused by inflation or the phases of business cycle (Mensah, 1984), or due to the change of competitive environment the company is in and the change of marketing strategy, or the utilization of new technology (Wood and Piesse, 1987). In summary，it can be seen that classical statistical methods generally suffer from stationarity problems (Moyer,1977; Mensah, 1984; Chariton et al. 2004).
Non-stationarity might cause the following severe consequences. 1. Poor predictive capacity in future-dated samples might be caused, even if a model has good previous classification outcomes (Mensah, 1984). 2. Data instability causes the models to be fundamentally unstable over time, thus classic statistical failure prediction models may need undating from time to time (Joy and Tollefson, 1975; Taffler, 1982; Mensah, 1984; Keasey, and Watson, 1991; Dirickx and Van Landeghem, 1994; Ooghe et al. 1994). 3. Due to non-stationarity and data instability, classic failure prediction models based on pooled estimation samples may be based on temporarily distorted data and this may result in inconsistent coefficient estimates and low accuracy levels (Platt et al. 1994; Back et al. 1997). 4. Classical paradigm has effect on sampling selectivity too because classical paradigm is established based on random sampling selecting. However, majority of the classic failure prediction modeling is established based on non-random samples (for instance: Altman, 1968; Deakin, 1972; Blum, 1974; Taggler and Tisshaw, 1977; Van Frederikslust, 1978; Ohlson, 1980 et al. just names a few), therefore, samplings chosen under these conditions might be subjective.
Forth, Non-random samples selected method also should be concerned; it might be emerged due to the following reasons. For example: 1, the probability of company bankruptcy actually covers a little part of the entire economy entity, therefore state-based sample is adopted by many researchers. It might lead to over-sampling of the failure firms and it may cause a choice-based sample bias. (Zmijewski, 1984; Platt and Platt, 2002). 2. the fact that companies that are bankrupted has the issue of incomplete data compared with companies that are not bankrupted, therefore several bankrupted companies cannot be included in the sampling and this might also lead to a sample selection bias (Taffler, 1982; Zmijewski, 1984;Ooghe and Verbaere, 1985; Declerc et al. 1991). 3. Matched pairs sampling method is used in previous researches in different industries and companies, artificially chosen sampling might lead to selection bias.
Non-random samples might also cause severe consequences，in anther words, it is impossible to build a model with strong prediction capacity if all the sampling and data are inaccurate. If the estimation samples are not chosen randomly, the result will be biased parameters which will eventually lead to difference between the overall ex-post classification accuracy and the ex-ante predictive performance. That is to say, the stated accuracy might not be generalized and this will mislead further researches.(Zmijewski, 1984; Piesse and Wood, 1992).
A.2.2 problems related to the neglect of the time dimension of failure
A 2.3 problems related to variables and modeling selection
A.2.4 other problems (2000 words still in editing)
B.1 The contributions of Chinese scholars
Recently, financial distress prediction is becoming increasingly popular among Chinese scholars. Chen (1999) first applies MDA model to distress prediction, she chose 27 ST companies as samples to study companies in financial distress and another 27 non-ST companies are also chosen as pair sample, it is to satisfy the requirement of normal distributed variables. Research result suggests that the accuracy of using models to predict financial distress of companies in the next 1 to 3 years is 92.6%, 85.2% and 79.2%
Zhang (2000) further advanced this study by choosing 120 ST companies as sample and a compare study was conducted in order to choose a better reacted parameter to the company financial status. Her research result demonstrates that MAD model has a 4 year prediction capacity, which means companies can be aware of the situation 4 years before the financial distress. However, due to the limitations of MDA model mentioned before, especially that the sampling requires normally distribute and equal variance matrices across the failing and non-failing group which makes sampling rather difficult, and this raises questions to the accuracy and adequacy of models.
Chen et al (2000) applies LA model to predict the business distress of listed companies. Their contribution is to screen the variable combination with obvious prediction function from the following six combinations, profitability ratios, cash flow ratios, asset utilization ratios, liquidity ratios, leverage ratio and annual growth ratios. And the variable sampling of this paper is based on research result of Chen et al. (2000).
Wu & Lu (2001) and Li (2001) both compare LA model with other analysis method and draw similar conclusion, and they furthers proves that the false-positive rate of LA model is the lowest, which is to say that LA model has a very strong prediction capability and it can be used in practical operation. At the same time, Jiang and Sun (2001) also uses LA model to conduct their research and they discuss the issue how to obtain the most optimal cut-off point, and this will also be concerned in this paper.
In general, however, the inadequate sample data and insufficient focus on a certain industry are the main problems of LA model research in China, and this is what this paper will be correcting. Furthermore, LA model is highly time-bounded, that is to that the newer the data is, the higher the prediction capacity the model will be, therefore, the research in this paper has high operation significance.
Total words: 5630 + 2000 ( in editing)
