Multiple Source Based Post Loan Risk Indicator Learning And Prediction

Credit business is main source of bank profit. Post loan risk management is mainly borne by the provincial and municipal branches. The main risk is detected based on the head office issued risk, credit the customer’s financial information, customer transaction information, without considering customers loan in a number of financial institutions and lack of analysis of industry conditions and other relevant external data. These essential factors on the user’s ability to repay funds. On the other hand, post loan risk assessment of bank related to time. Selection of time segments is varied year or quarter. But the emergence of abnormal indicators often difficult to restore the loss. Also ignores the life cycle of the credit risk of the post loan. Different life cycle stages may face different risks. Thus, post loan risk indicator selection, learning with external data and post loan management process optimization is an important issue of concern banks.For post loan risk indicators do not reflect the current industry risk and the overall analysis of credit customers. In this paper, the analysis of post loan indicator combine with existing indicators data of bank, the original data, and external sectors and regions data. Multi source data learn risk indicators, through information gain a wraparound feature selection method based on the probability of multi-source data analysis features. By calculating the similarity regional data, divide industries and regions. Experimental complete the extraction of features on two data sets. Verify forecast accuracy after the increase data, and based on the similarity of the external sector and region data divide the potential associated industries and regions.Considering Post loan risk prediction time correlation, a dynamic model based on time window is proposed in this paper. Through modeling loan life cycle and collecting credit data of different phases according to dynamic divided time window, data based on time window is used in the training of classification learner. The importance of data on different phase in loan life cycle is allocated dynamic according to the performance of time based classifier. And then the risk prediction is made. Experiment verify the time window and the relevant parameters, and compare the two data sets using different methods of risk prediction accuracy.For unbalanced of positive instance, the number of instance that is far less than the number of non-performing loans of instance makes regular loans covered and positive instance misclassified. We take two aspects to improve:Firstly, to increase the proportion of bad instance by determining the positive instance, combined with external data analysis filter related data collection. Secondly, optimize the risk assessment process for positive instance points for iterative analysis, building dynamic feedback mechanism. We use Particle swarm optimization for characterized learning. For two data sets, experiment verify the effect of feedback mechanisms, by optimizing the learning algorithm parameters, analyzing risk prediction accuracy of the post loan.