The Subscriber's Credit Evaluation Of The Telecommunication Business Enterprise

Credit consumption is characteristic of the delay pay service provided by telecommunication business enterprises for subscribers, the subscriber's arrearage occurs continually. Informing the subscribers owing fee, stopping service to them and restricting unreliable person to become a subscriber are mainly adopted as control means, which caused subscriber's disaffection and revenue reduction. Different arrearage control measure to every subscriber is taken to reduce enterprise's arrearage as a whole, to make most subscribers satisfied, and to increase effective revenue, is the exclusive means.Actualizing credit management of subscribers is effective method to solve above problem. As the credit criterion, an index attached to every subscriber can be used directly or in directly as bounds of his or her overdraft consumption. The study object of this paper is to found a mathematical model to calculate credit index of every subscriber base on the subscriber information, service property, and consumption quantum and dealing records. The model can be applied to manage subscriber's arrearage in bill system.The subscriber's credit gradation is a kind of technique, about which some ecumenical rules do exist. But there isn't any universal model, which can be used directly as a software system. Befitting model ought to be opened out according to own data information, ready-made production from other country or organization can't be copied directly. The actual modus operandi is that the characteristics of violators and credible subscribers are found out based on some historical samples of two categories, about those who pay off on schedule and those who fell back, so the classified rules are summed up and the model is built to measure breach risk or probability of subscribers. Establishing credit gradation using data mining technique commonly contains ten steps, they are confirming business target, identifying data sources, collecting data, choosing data, auditing data quality, transforming data, mining data, explaining result, applying advice and result application. In this paper, binary logistic regression and classification tree are used with spss statistics software package to establish and evaluate models, and other software is used to process and transform data in seedtime. If penal sum start to be calculated, bad subscriber is defined, else good subscriber is defined. An exclusive credit evaluation database connected with bill system and business management system through ORACLE database links is established in a PC server with ORACLE installed and spare resources. The basic business data is copied to PC server with same or tidy storage format and correlation to be analyzed offline, and to be converted into storage format that can be analyzed with SPSS directly.The data about independent variables was copied on Dec.30,2005, and the data about dependent variable was copied on Mar.20,2006, the interval is eighty days. To make modeling variables is to create one and only sample for every subscriber, including a series of independent variables and dependent variable. Since modeling variables are from interior data source of the enterprise, the method to deal with missing values is all and singular to look on the missing value as another new category.There is maybe zero or many deal records for a subscriber, they must be gathered into a derivative variable at subscriber level, the breach degree is described with the derivative variable, which is called owing variable in this paper. It can be defined the sum of owing duration multiplied by money divided by duration to now, and other similar arithmetic should be examined. The key is the correlation between the derivative variable and the dependent variable, the arithmetic by which breach frequency increases most monotonously along with the derivative variable should be selected. When the owing duration and duration to now is converted by one of ten monotonic increasing elementary functions, 121 independent variables are constructed. Seeing about Kendall's Tau-b rank correlation coefficient between every owing variable and dependent variable, the best result is the arithmetic with most rank correlation coefficient. There are 14 polytomous variables, 36 dichotomous variables and 8 continuous variables in the samples. Since the relation between continuous variable and breach frequency is usually nonlinear, the continuous variables need to be dispersed into snippets. The breach frequency between the contiguous two snippets should be different significantly, there are definite amount of samples in each snippet. An original partition is confirmed first, by which the breach frequency of all groups keeps variety trend in the mass. In order to reduce redundancy levels of polytomous variables and to improve efficiency of model disposal, the amount of categories should be decreased furthest. This work can go along through combining categories with small difference of breach frequency between them. Whether the contiguous two groups can be combined lies on whether their effect of breach is distinct. Seeing about the breach contingency table of the contiguous two groups,χ2 value is calculated according to the definition of Pearson chi-square, using special PL/SQL program,χ2 value decides whether the contiguous two groups can be combined. Theχ2 critical value is enlarged and significant level is diminished to confirm that the amount of the levels of every variable brought into models isn't more than ten.The samples are subdivided to two parts through half sampling, one part for modeling, another for confirmation. The data stored in ORACLE is transformed into .sav file by SPSS Database Wizard. Firstly, independent variables are screened with forward stepwise selection, and logistic model is built. Secondly, classification tree model is built using CHAID arithmetic with three and four layer depth. Finally, in stocks and branches of classification tree model interaction effect is detected to add derivative variables, a logistic model including interaction terms is built using forward stepwise method. After model estimation, suitability must be considered, for example, goodness of fit, predictive accuracy and model chi-square statistic. Before model application, different models should be compared to find the best one. Comparing with logistic model, indexes classification tree model can provide are fewer. Using Hosmer-Lemeshow test to evaluate goodness of fit of the logistic models, the statistic is statistic significant, i.e. the model doesn't fit the sample data well. This conclusion can also be drawn from measure of analogous R2 of 0.432. ROC curves made from predicted probabilities, the area below every curve is similar to each other and different from 0.5 significantly.In this paper, model chi-square statistic and fit statistic of logistic model are all significant, the incompatible phenomena shows that on the one hand, the model doesn't fit the sample data very well, on the other hand, independent variables can explains the dependent variable very well. The credit model in this paper belongs to interior model established from interior business data mastered by enterprise.Among the independent variables brought into the model, owing variable is the strongest forecast independent variable, i.e. payment practice is the primary factor deciding breach. Service state, bill type and business area are also import factors deciding breach. This paper only provides some operable means and steps foe credit evaluation, before it can be applied formally, new data should be used, new variables must be constructed and the whole data mining process should be operated over again.