A Bayesian network based early warning model that estimates the probability of non-performing corporate credits

Early warning systems are systems that generate signals to predict the probability of the occurrence of a problem and the size of the potential impact of the problem whether it occurs. From the oldest part of the history, human are tend to survive and protect their current savings as a part of human being. This instinct is also mentioned in Maslow's Hierarchy of Needs pyramid as the secondary need of human being that must be fulfilled after vital activities like food and beverage. For that reason, human are tend to know the possible negative situations that might be face off and take precautions about those problems. When viewed from this aspect, it is not hard to understand why it is need to develop early warning systems about negative situations. Early warning systems are defined as the capacity of providing a meaningful information about a problem. These information is about the probability of the occurrence, size, impact and occurrence time of the problem in order to avoid or reduce the effect of the problem on organizations, society, people who might be affected. For that reason, early warning systems have a wide range area of usage such as health industry, disaster recovery and emergency management, predictions of environmental factors and finance industry. Early warning systems in finance industry are grouped in two main area; prediction of financial crisis and prediction of business failure. When it is considered the size of the impact and the importance of the problem, prediction of the financial performance of the companies is not only important for the company itself but also crucial for the employees, managers, shareholders, customers, contractors, creditors of the company and for the other companies in the sector and finally it has a critical importance for the economy of the country. For that reason, in this thesis study, it will be tried to model and predict the probability of business failure of companies. When considering the previous studies about business failure, it can be seen that the first studies are started with single variable models and continued with such different and various methods used in that area such as linear models, probability based models, artificial intelligence based models. The increasing uncertainity level of the environment, the increasing number of competitors in activity area of a company and the hard environmental conditions of a company make the estimation of the probability of business failure of a company a critical problem. For that reason, as a success criterion, it is needed to be that the estimation method of this problem must be useful about modeling the uncertainity of the environment and must be adaptive on changing conditions. Comparing with other methods that are used in the estimation of business failure probability, Bayesian Networks is preferred and used in thesis study due to its advantages of usage that are its usefullness, easiness, apparentness, not needing large amount of data sets at model generating phase, ability of definition and observation of all variables in model, ability to model the uncertainity, ability of generating signals with missing values of variables. One of the most important characteristic of this thesis study that differenciates the study from other studies made in this area is not using only variables about company itself, but also taking into account variables about company's shareholers and environment. These new variables taken into account due to the assumption that it is not possible to consider a company seperately itself and neither without its shareholders nor its environmental conditions. Secondly, more than one source of information about variables are used in the thesis model, so the risk of mistaken results due to the usage of wrong or untrusting data sets is decreased. Third, the variables that are acquired from interviews with speacialists on this subject, beside the literature review are used in the model. So, in that way, it has been possible to combine different points of view together and to use new variables which are so important and not used in the problem of estimation of the probability of business failure before. The data set used in the model is one of the financial companies' corporate customers' data, which 55 of them signed as "non performing", totally 150 companies' financial data. The data set is composed of companies' balance sheets, income tables and the risk level on other financial companies and the age of the company, the financial background of the shareholders' and the experience level of the shareholders' and some other criteria that can be distincively shows the uncertainity of the environment. Due to the variables are continious in the model such as financial ratios, one of the main questions in the thesis study was which method would be used to discritize the variables and which reference value would be used to compare the level of these variables. To solve this problem, sectoral balance sheets and income tables which show the annual sectoral averages that are published by Central Bank are used for compariastions. The deviation of financial ratios from sectoral average values are calculated and the discritization process is done with these values. Variables are converted to discrete forms into three main groups "Low", "Middle" and "High" due to their deviation values. This process made the model independent from sectoral limitations such that the companies' variables are compared with their own sectoral averages. Another issue that is taken into account in thesis study is the determination of cut-off points of the target variable that model generates. Determination of the point that is from which value of the target variable, it is defined to be "performing" or "non-performing" is made by solving the minimization problem of the equation of the cost of type I and type II misclassificaiton error. After all, importance level of the variables on the estimation of the target variable is determined, so it would be possible to develop the estimation power and simplication of the model. At last part of the thesis study, the model was established with logistic regression method with same variables and same data set to validate that the Bayesian networks was the correct and strong method for modeling the probability of a company's failure. The results acquired from two methods are compared. The comparison results show that the percentage of correct estimations of two method were pretty similar to each other. But, Bayesian Networks had notably successful performance on estimating the "non-performing" companies from logistic regression. For that reason, when comparing the cost of the mistaken clasifications of the two model, it was said that Beyesian Networks had better performance on estimating the probability of business failure.