Corporate Ratings and a Model Proposition for the Manufacturing Industry at Borsa Istanbul

In this study it is aimed to develop a model using logistic regression analysis for the forecasting the rating grade of a manufacturing firm that form the basis to expert evaluation. Under the scope of this study 35 financial ratios are used as the independent variables, which are calculated on the grounds of annual financial statements and their notes during the period of 2007-2013 which are disclosed by the 206 listed manufacturing firms on Borsa Istanbul, and the status of the firms being “good” or “bad” based on financial capability is used as the dependent variable. Percentage of correct classification of developed model is at acceptable levels. By using the developed model, probability of a firm being "good" or "bad" can be estimated and using the proposed scale rating grade can be appointed of the firm that rating wanted to be performed.


Introduction
Rating agencies have been criticized severely due to recent bankruptcies and financial crisis. The critics have been focusing on the inadequate regulations at the national and international level and rating agencies have been criticized on losing impartiality, giving better or worse grades than what is deserved. However, taking increment in number of issuers, securitizations and complexity of financial instruments and level of globalization into consideration, it is clear that the importance of rating activities have gradually increased from the point of market and market participants (Öcal, 1997). In this manner, it is thought that the rating agencies will continue to exist and operate in the market.
On the other hand, there is no single procedure regarding rating methodology. Each rating agency has its own unique methodology and the methodology can be differentiated in itself by sectors, type of firm and type of financial instrument.
Using statistical methods solely while carrying out rating activities have been criticized. On the other hand, as rating based on solely experts' judgments, can be a waste of time and resource and as they are open to human error and have potential deviation from objectivity, the benefits of using statistical methods come into prominence. However it shall not be assumed that using only statistical methods are adequate. Accordingly, the critique of statistical methods shall be considered as well. Therefore statistical techniques and experts' judgments shall be used jointly in order to remove the limitation of a human, selecting data, processing and concluding with a result from a wide range of data and inadequacy of non-numeric data processing of statistical models shall be eliminated.
In this study it is aimed to develop a model using logistic regression analysis for the forecasting the rating grade of the manufacturing firms that will form basis to Expert evaluation. It should be kept in mind that the rating grade that will be determined by the developed model, is not final.
Furthermore, the rating requires usage of non-public data and/or information besides public data and/or information. In this study, only public data and/or information of the firms are used.
Examining especially some foreign studies, it is understood that rating grades that are given by independent rating agencies are used as dependent variable and therefore dependent variable is considered depending on number of rating grades. In Turkey, since the rating is not compulsory, and optional/voluntary, rating results are not disclosed to public unless they are adequate, usage of rating grades that are determined by rating agencies as dependent variable are not sufficient to run the analysis. Therefore our dependent variable that is used in this study constitutes from two groups (the status of the firms being "good" or "bad" due to financial capability).
Overviewing studies in this area it is seen that discriminant analysis highly used at early times. Afterwards, probit and logit analysis have been started to be used and it is understood that analysis based on "artificial intelligent" have started to be used in some of the latest studies.
In our study, instead of running highly used discriminant analysis which break the assumption of normal distribution, continuity and equality of the deviation matrix, logistic regression analysis is used which gives better results than discriminant analysis and gives opportunity to classify results by producing different probability values for each observation.
Reviewing the rating studies that are conducted in Turkey, no prior study that used same methodology, data and period is found.
Under the scope of this study 35 financial ratios are used as the independent variables, which are calculated on the grounds of annual financial statements and their notes during the period of 2007-2013 which are disclosed by the 206 listed manufacturing firms on Borsa Istanbul equity market.
In this study, the status of the 206 firms being "good" or "bad" based on financial capability is collected by searching news and disclosures at Borsa Istanbul web site for the period of 2007-2009 and Public Disclosure Platform (PDP) web site for the period of 2010-2013. To identify "good" or "bad", all news and disclosures of sample firms is evaluated by taking into consideration the information about market changes and delisting made by Borsa Istanbul on PDP. Delisted firms or firms those trading activities are held due to financial distress, firms that changed trading market (in negative manner), firms obliged monthly declaration due to financial distress, firms that are warned to take measures by CMB or Borsa Istanbul due to losing capital, or the firms that applied to court by itself due to financial distress are counted as "bad" and crated unique data. This manually created data makes our study original and contributing.
Abstract of the studies regarding the rating activities and methods that are used is stated under the second section of our study, where the data and the method that is used in this study is stated under third and findings that are received and interpretations of them are stated under the fourth section of our study. Finally the conclusion part is stated under the last section.

Literature Review
Some of the main studies and methodologies on corporate rating are summarized as follows. The studies which are conducted by using ordinary statistical methods are summarized in Table 1 and the studies which are conducted by using artificial intelligent based analysis are summarized in Table 2 (Hajek, 2010).  Horrigan (1966) Linear Regression 9 200 6 56.0 West (1970) Linear Regression 9 -4 62.0 Mingo (1973 and1975) Mult The variables and their definitions that used on related previous studies are given in Table 3 (Hajek, 2010). As the number of variables increase in analysis, the interpretation of model becomes more difficult and applicability of model decreases. Therefore the elimination of insignificant or non-explanatory variables in models should be done. In case of high correlation between variables, it is possible to decrease number of variables by using variable selection methods. On the other hand, this approach may cause excluding important or significant variables from the model due to election criteria (Özdinç, 1999).
Some studies on rating have also conducted in Turkey last decades. The main difference between this domestic and international studies arises from dependent variable of models. Except for rating on banks, as the firm does not have rating or disclosed rating grades in Turkey the rating grades of firms cannot be used as a dependent variable. Some of the related studies on rating in Turkey can be summarized as in Table 4. In rating activities both quantitative data (financial ratios) and qualitative information shall be taken into consideration. Qualitative informations may enter to the model as a dummy variable or as an expert's judgment. It is argued that using dummy variable for qualitative data is a more appropriate way and it is seen as increasing the success of the model. For example, delay in disclosure of financial reports, independent audit opinion, age of firm, number of employee at managing level, the duration of work of the managers in the firm, the mortgage on the firms' assets may be included to model as the dummy variables. Expert's judgment may include fairness and correctness financial reports and data of the firm, since financial reports and data of the firm had been manipulated (Kadıoğlu, 2014). Keasey and Watson (1997) argues that including qualitative information to model as the dummy variables will increase the success of forecasting financial failure in small and middle sized firms.

Data and Methodology
In this study, annual financial statements and their notes of firms that conduct their activities in manufacturing sector and which are prepared according to International Financial Reporting Standards ( Due to data prepared, 35 financial ratios in 5 groups were calculated that may have an effect on solvency of the firms. While determining the financial ratios that serve basis to independent variable of this study, the ratios that are used prior studies were also taken into consideration. Additionally, as IFRS was in force during the period of examination, it became possible for us to use the information that obtained from cash flow from operating and investing activities and foreign exchange position which weren't taken into consideration by prior studies conducted in Turkey.
The financial ratios that are prepared to be used within the scope of the study and their definitions are as follows. Since there is no rating obligation for Turkish firms except for banks, it is not possible use rating notes as the dependent variable that are given by independent rating agencies. Therefore in this study, the status of the companies being "good" or "bad" based on financial capability is used as the dependent variable. In other words, if the firm is financially in a bad situation or in case of failure then dependent variable takes the value of "0" and otherwise it takes the value of "1".
According to Özdemir (2011), quantitative and qualitative indicators can be used in the determination of financial failure and quantitative indicators can be classified as the book value based indicators and market value based indicators.
In the case of using qualitative indicators, determining the class of the firm is easier and market value based indicators give more fair and accurate results when the market is efficient (Özdemir, 2011). Taking consider Özdemirs' idea into account, we used qualitative indicators to classify bad or good firms 2 .
For this manner, the status of the 206 companies being "good" or "bad" based on financial capability is collected by searching news and disclosures at Borsa Istanbul web site for the period of 2007-2009 and KAP web site for the period of 2010-2013. To identify "good" or "bad", we searched all news and disclosures of sample firms by taking into consideration following criteria and the firm, matched following criteria, is classified as "bad" firm.
i) Firms that are delisted by Borsa Istanbul due to financial distress, ii) Firms those trading activities are held by Borsa Istanbul due to financial distress, iii) Firms those trading market are lowered by Borsa Istanbul, iv) Firms that are obliged monthly declaration due to financial distress, v) Firms that are warned by CMB or Borsa Istanbul to take measures to recover the capital, vi) Firms that are applied to court by the firm itself due to financial distress.
The firm, being "bad", is checked yearly bases and whenever the information stated above is disseminated we accepted that year as the starting year for "bad" for the firm. If the firm counted as "bad" and if there is new reversal information in following years then we changed the firm as "good".
Depending on financial distress, being "good" or "bad" is constitute our dependent variable and it takes value of "1" for "good" and "0" for "bad".
In our data outliers and having much missing observations have been eliminated by basic sorting and filtering applications. Additionally, since the base of the number is different, all independent variables normalized by subtracting mean and dividing to standard deviation. As result, 88.5 % (1149 observations) of total sample (1298 observations) is classified as "good" and 11.5 % (149 observations) of total sample is classified as "bad".
To avoid weakness and critics on discriminant analysis and least square regression (not fulfilling normal distribution assumptions), we chose logistic regression to run for our model. The studies also show that logistic regression gives better results in case of dependent variable is discrete. Furthermore in our case, logistic regression gives different probability values for each observation depending on the variables in the model and this enable us to determine rating notes depending on tranches of probability.
In our logistic regression analysis, SPSS 18 Portable and SPSS Clementine 11 software packages have been used.
The classification studies on unbalanced data such as an unequal number of "bad" and "good" observation has the disadvantages. Because, it is argued that correct classification success for proportionally high number of observations (in our case being "good") is higher than correct classification success for proportionally lower number of observations (in our case being "bad"). It is also the case for our sample. In order to overcome this biasness, we run the analysis on balanced sample. In our study, to create balanced sample we took all "bad" observations and randomly selected 15% of "good" observations by using SPSS Clementine 11 software. At the end, our subsample consist of 49% of "bad" observations and 51% of "good" observations and total subsample size became 306 observations. In order to use as much as observations, we designed to our sample consisting 49% "bad" and 51% "good" observations. The descriptive statistics of variables are given Table 6.

Empirical Results
As it mentioned before, in order to balance our sample we restructured sample by taking all "bad" observations and randomly selected 15% of "good" observations of 1298 observations. In our subsample there are 149 "bad" observations and 157 "good" observations and the subsample is 24% of total sample. 266 observations of subsample are used for estimating the model and 40 observations of subsample are used for testing the model. In logistic regression model that run by exclusion of variables BK7, L7, L8, L11 and BK4, and that run with 5 statistically and academically/scientifically significant variables (KSE5, BK5, SY2, L4, L10) it has found out that coefficient of L10 (Assets Turnover Rate (Net Sales / Total Assets)) is insignificant. Accordingly, a model consisting of variables KSE5, BK5, SY2 and L4 and excluded variable L10, has formed and the results are stated under Table 8 (See also Appendix 1). All variables are significant which are used in model. As seen from   Accordingly, it is seen that, our model classifies "bad" firms 77.85% correctly and classifies "good" firms 79.64% correctly due to application of the model to whole 1298 samples and overall correct classification success is found as 79.42%.
As mentioned before, logistic regression gives different probability values for each observation. This characteristic of the method enables us to make ad libitum classification of firms regarding their success probability.
Within the scope of this study firms are classified as 5 groups based on the probability values for each observation. The group size, rating grades and the probability tranches are given under Table 11. As it is seen from the table, according to probability tranches of being good, firms were rated as follows where A represents the best grade and where E represents the worst. Firms success probability which are -greater than 0.80 is "A", -greater than 0.60 and less than 0.80 is "B", -greater than 0.40 and less than 0.60 is "C" -greater than 0.20 and less than 0.40 is "D" -less than 0.20 is "E"

Conclusion
There is no single definition and methodology for corporate rating. Each regulative authority or the rating agency has its' own unique definition and own unique methodology where these are differentiated due to sectors, and regarding the rating of firms or financial instrument. Furthermore, it won't be possible to use same methodology forever that once formed and the rating methodologies have to be reviewed and improved in time.
In this study a model is developed using logistic regression analysis for the forecasting the rating grades of the manufacturing firms that form the basis to expert evaluation. It should be kept in mind that the appointed rating grades developed by models, are not conclusive and they need to be evaluated by the experts in view of the subjective facts.
The status of the firms being "successful/good" or "unsuccessful/bad" based on financial capability is tried to be determined within the study by using more than one defining variables. In other words, our dependent variable consists of two groups. Accordingly, in our study, logistic regression method is used which is one of the most appropriate model that enables a dependent variable to take two values. sample. Accordingly, publicly disclosed balance sheets, income statements and cash flow statements and their notes of 206 listed firms are collected and 35 financial ratio in 5 groups have calculated. Mentioned ratios form independent values of the study and consist of 88.5 % (1149 observations) that are classified as "good" and 11.5 % (149 observations) that are classified as "bad".
While forming the model instead of using whole observations (149 "bad", 1149 "good" of total 1298 observations), 306 ( 157 "good", 149 "bad") observations which forms circa 24% of the total data were used. In order to balance our sample remaining observations that belong to good firms were excluded.  (Y)). By accepting calculated probabilities that are greater than 0.5 as "good" (1) and otherwise as "bad" (0) Y values are predicted. Y values that are predicted and observed were compared and it has found that the model classifies "bad" firms 77.85% correctly and classifies "good" firms 79.64% correctly and the classifies overall 79.42% correctly.
Within the logistic regression model different probability values are calculated for each observation. This feature of the model enabled us to distribute firms depending on tranches of probability. In this study, firms are classified into 5 group based on the probability values for each observation. If the probability of being good is greater than 0.80 then firm's grade is classified as "A", if it is greater than 0.60 and less than 0.80 then firm's grade is classified as "B", if it is greater than 0.40 and less than 0.60 then firm's grade is classified as "C", if it is greater than 0.20 and less than 0.40 then firm's grade is classified as "D" and if it is less than 0.20 then firm's grade is classified as "E", where A group represents the best grade firms and E group represents the worst grade firms.
As a result, by running the four variable logistic regression model to the datas of a selected firm, Y value of the chosen firm can be calculated with a success ratio of 78%-80%, by using the formula Exp(Y)/(1-Exp(Y)), the probability of being "good"-"bad" of the chosen firm can be determined and rating grade can be given by using a scale that is recommended by us or a scale that will be developped by the users.
It should be kept in mind that the appointed rating grades developed by models, are not final and they need to be evaluated by the expert in view of the facts that are not present in models. Additionally, running model directly to the firm's data may not always result accurately. Users must evaluate the financial reports of the firm on the grounds of fairness and correctness and examine them whether they represent a permanent time-period or not and then give a final rating grade. Step 1

Appendices
Step

Model Summary
Step -2 Log likelihood Cox and Snell R Square Nagelkerke R Square 1 266.819(a) .402 .536 a Estimation terminated at iteration number 7 because parameter estimates changed by less than .001.