Computational Intelligence-Based Anomaly Detection For Financial Data

There is no doubt that finance is core of the modern economy in the 21st century. And the healthy development of the financial industry is required by the sustained economic growth and the steady social progression. It's one of the main objectives for governments and investment organizations at all levels to explore disciplines hidden in the constantly changing financial market so as to manage the financial market effectively and improve the financial investment efficiency. However, the financial anomalies emerging in recent years, such as insurance and credit fraud, money laundering, accounting information forging, etc. (negative anomalies), seriously damage the healthy development of the financial industry and interests of the general people. Therefore, it has great practical significance to enhance financial supervision and timely identify and prevent the occurrence of such anomalies. In addition, it also has important investment significance for financial investment institutions or individual investors to study the other kind of anomalies in the financial industry-the turning point that the price trending is reversed in the price sequence of financial products (positive anomalies).Basis on the studying of the related literatures, this paper explores new ideas and applies the theory of computational intelligence to two kinds of financial anomaly detection, so as to provide decision-making references for the relevant departments or individuals. And the credit data set from UCI is used for the research on the negative anomalies in the financial field, while the stock data set is used for the research on the negative anomalies.The existing studies on the two datasets are as follows:(1) After studying the related literatures, we find that the current literatures mainly use the credit data as an experimental data set to verify the classification algorithm without considering the characteristics of the data itself, so it is difficult to provide references for the relevant departments.(2) The existing empirical research literatures on stock anomalies mainly propose a new anomaly detection algorithm, and then share the data as an experimental data set to validate the algorithm; or just study the stock market characteristics before and after anomalies, not quantify the characteristics of this discovery to predict anomaly yet; or the defined anomaly in the literature has no great reference value for the real decision-making.Aiming at the shortages above, this paper does the following research:(1) For the credit data set, firstly the rough set theory is applied for the feature selection to get a feature subset after reduction; and then the Bayesian analysis is applied to find which features leading to the credit anomalies occurring; finally the Naive Bayes classifier is used for classifying research and the algorithm is improved to increase the correct recognition rate of the abnormal credit.(2) Anomaly detection research on the banking share data from the A shares of the Shanghai stock. First, the relevant data is collected and collated; then based on the related theory of the stock technical analysis and through analyzing the development trend of the individual share price, a new kind of outlier:buying or selling points-points shortly before a substantial price rise or fall define, and these outliers are found out in the data set according to the definition; finally, in according with the research needs, the data is divided into two groups based on the average daily turnover, and the F-score based SVM classification model is trained on each group. Experimental results show that the SVM classifier can achieve good results, and the result of the second group which is more active in transaction is better than the first group:the classification accuracy rate can reach about 80%, the correct recognition rate of the plus outlier (buying point) is 79%, and the correct recognition rate of the negative outlier (selling point) is 75.56%.