طبقه بندی چند مرحله ای با استفاده از رگرسیون لجستیک و شبکه های عصبی برای ارزیابی وضعیت مالی شرکت

The paper presents the new approach to the automatic assessment of the financial condition of the company. We develop the computerized classification system applying WOE representation of data, logistic regression and Support Vector Machine (SVM) used as the final classifier. The applied method is a combination of a classical binary scoring approach and Support Vector Machine classification. The application of this method to the assessment of the financial condition of companies, classified into five classes, has shown its superiority with respect to classical approaches.

The problem of assessment of the financial condition of company is very crucial to avoid customers, who may cause problems with financial liquidity or are the potential bankrupts [1], [2] and [19]. To decrease the risk of transaction the assessment of financial condition of the company is necessary. Company credit ratings are very costly to obtain, since they require to invest large amount of time and human resources to perform analysis of the company risk status. Such report is based on various aspects, ranging strategic competitiveness to the operational level details [2] and [10]. Although rating agencies emphasize the importance of analysts' assessment in determining credit ratings, there is a parallel way of developing the automatic methods relying on the artificial intelligence approach. Nowadays the most often used methods apply the artificial neural networks, being the universal tools able to do both classification and prediction tasks [6], [11], [13] and [17]. The important point in such application is the availability of large amount of historical data of the company, regarding the financial reports and embedded valuable expertise of the agencies in evaluating companies' credit risk levels. The objective of credit rating prediction is to build the mathematical models that can extract knowledge of credit risk evaluation from past observations and to apply it to evaluate credit risk of companies with much broader scope. However besides the prediction the modeling of the process can deliver another valuable information to the user. These studies can help the user to capture the fundamental characteristics of the most important dependencies between different financial ratings and the company risk status. Such analysis can also simplify the process of the assessment of the risk status of the company, by eliminating some parameters that are loosely associated with the level of the company risk. The most important point in the credit report is the assessment of the financial condition of company by using many factors (not only financial). Important are also such information as changes of board of directors, status of the company, location and other registry information. Very often this information can be available online in Internet. The other source of fresh information about company is the interview via phone with subject or other companies with which our subject cooperates e.g. suppliers, sister company, parent company, affiliates etc. On the basis of this we can obtain additional information about e.g. payment history. Other sources of data are the agents distributed all over the world. This additional information can be added to the set of attributes, enriching in this way the input information taken into account at taking decision. On the basis of all gathered information we can create the diagnostic features describing the financial state of the company, and then associate them with one of few classes, representing the level of insolvency risk. The insurance companies apply different number of classes. In this paper we assume 5 classes of insolvency risk [10]: •excellent (without any risk) •good •satisfactory •passable • poor. The problem that arises is to provide the unified way of representing the information as an input to the computer system performing the role of automatic extraction of knowledge and undertaking the final decision of assessment of insolvency risk of the company. In most application the numerical data is either represented directly in numerical form or converted to some classes, while the other (non-numerical) data is somehow coded in a binary way. In this paper we apply the unified way of representing data by using the weight of evidence (WOE) concept [17] associated with each feature. The data represented by WOE will be classified by us in two step procedure. In the first step we apply the binary logistic regression associating the data with seven models of 2-class classification. The results in the form of probability of membership to these 7 models are applied as the input attributes for the second stage of classification recognizing the final class (one of 5 already defined). As the classifier we apply the support vector machine (SVM), generally regarded as the most efficient classification tool [6] and [15]. The novelty of the paper may be characterized in the following points. • Development of continuous representation of financial data, very efficient in practical application and leading to the improvement of the quality of the classification system. • Proposing the 2-step classification of financial data by applying the binary classification systems in both stages. We will show that such solution leads to the significant improvement of the accuracy of classification.

In this study we have proposed the novel 2-stage approach to the problem of credit rating prediction of the companies. The most important point of this approach is the combination of the logistic regression method with WOE representation of data, forming the first stage of binary classification, and application of the second stage of final classification, in which the results of the first stage form the input information for second stage classifier. Application of WOE, reflecting the information about the proportion of the classes and their association with the input attributes, provides the natural way of implementing the human knowledge of the process into the classification methodology. The presented approach split large classification problem into few smaller models, each recognizing two classes only. Such method of solution is much easier to control. This way of data processing enables to apply at each level the dedicated feature selection methods such as stepwise or forward and backward regression, leading to the simplification of the classification decision and at the same time to better understanding the basic relationships between the attributes and the classes under recognition. The proposed approach automates the procedure of the assessment of the financial condition of the company applying for the loan, enabling to simplify the whole process and reduce the cost of its preparation. All steps of data processing are done automatically without intervention of the human operator. The numerical experiments have been performed using the data set of few thousand financial entries corresponding to many international companies. The results of these experiments have shown that our system provides acceptable results and is superior over the classical, one stage classification systems.