SVM-based Abnormal Account Monitoring Model of Bank

Abstract

The account transaction flow of bank loan customers reflects the capital transaction of customers and represents their management level. Normal account transaction flow shows regularity to some extent, while abnormal account transaction flow manifests itself differently in transaction amount and transaction frequency et al. When compared with those of normal account transaction flow, abnormal account usually shows abnormal change of fund flow caused by money laundering, hot money, private lending or bad management. In this study, based on identification approach between normal account and abnormal account, the characteristics of different industry transaction flow are drew out, and then the classification for normal and abnormal accounts of different industry is well established by the Support Vector Machine model. Thus, the loan monitoring and risk management can be conducted more appropriately and efficiently. Introduction The capital flow of bank account is needed for normal demand of the business activities. However, many abnormal or illegal capital flows may be mingled, which may bring challenges to perfecting finical supervision system and improving supervision method. Until to now, some studies have been conducted on monitoring finical capital flow: Liu et al.[1] distinguished account transaction types through scan statistic method. This method had an effect on distinguishing abnormal transaction. The accuracy of the model was strongly relative with the parameter, and the model parameter setting is strongly subjective, Yang[2] studied the credit card fraud detection model based on Radial-Basis Function Support Vector Machine (RBF-SVM). This model has some flaw that the performance of the model is unstable, Through the RBF (Radial-Basis Function) neutral network technique, Lv et al.[3] proposed a RBF neutral network model for anti-money laundering based on APC-III clustering algorithm and RLS (Recursive Least Square) algorithm. The sensitivity of the accuracy of the model to the parameter variation was strong. For commercial banks, Wang[4] suggested to use the theory of decision tree to identify suspicious customers. Results showed that this method is effective for generating anti-money laundering rules from customer information, while the setting of rules is relatively subjective. Zhu[5] proposed the EMD method for the decision of the abnormal capital flow based on historical information in the same group. The accuracy and confidence level for the model was not calculated. Taking Bank of the Malay Archipelago as an example, Aspalella[6] studied the influence of setting reporting system for suspicious transactions on bank risk monitoring. Pramod et al.[7] structured the regulation frame of anti-money laundering for the bank, and they considered that the information system based on the process control played a critical role in finding fraudulent transactions. Current techniques effectively used in the monitoring of abnormal capital are the Support Vector Machine (SVM) model and neutral network technique et al[8]. Some drawbacks of the neutral International Conference on Economic Management and Trade Cooperation (EMTC 2014) © 2014. The authors Published by Atlantis Press 274 network algorithm are discovered: (1) the model assumption is too strict, for example, uniform distribution of the trained samples is usually required, (2) dimension disaster: overabundant characteristic numbers may cause the algorithm poorly efficient or inefficient, (3) the accuracy of the results: the results from the neutral network algorithm are highly dependent on the prior knowledge and experience of the user. The advantages of using SVM model in the monitoring of bank account flow are as follows: (1) SVM model can solve multidimensional sample issues through linear method. This feature is fit for the multidimensional characteristics of bank account. (2) The assumed conditions of SVM model is not as strict as that of the neutral network technique: the sequence of samples is not required by SVM model. (3) Compared with the neutral network technique, fewer parameters are needed for SVM model. Users can obtain satisfactory classification results from raw data provided that they have some prior knowledge and experience. SVM Model of linearly separable sample set Assume 0 0 ( ) T g x x b w   , thus 0 0 ( ) 0 T p p g x x b w    , 0 0 0 0 0 ( ) || || T T T p g x x b x x r w w w w      . (1) where ( ) g x represents the distance of sample x to the separable hyperplane. The decision function