تشخیص کلاهبرداری سهام با استفاده از تجزیه و تحلیل گروه همتا

This study proposes a method to detect suspicious patterns of stock price manipulation using an unsupervised data mining technique: peer group analysis. This technique detects abnormal behavior of a target by comparing it with its peer group and measuring the deviation of its behavior from that of its peers. Moreover, this study proposes a method to improve the general peer group analysis by incorporating the weight of peer group members into summarizing their behavior, along with the consideration of parameter updates over time. Using real time series data of Korean stock market, this study shows the advantage of the proposed peer group analysis in detecting abnormal stock price change. In addition, we perform sensitivity analysis to examine the effect of the parameters used in the proposed method.

Stock price manipulation has been traditionally classified into three different categories (Allen & Gale, 1992). First, stock prices can be manipulated by actions that change the actual or perceived value of the assets (action-based manipulation). Second, manipulation can occur when false information or false rumors are released (information-based manipulation). Finally, traders attempt to manipulate stock prices simply by buying and then selling, without making any publicly observable actions to alter the value of the firm or releasing false information to change the price (trade-based manipulation). While the eradication of action and information based manipulation has been fairly successful, trade-based manipulation is still difficult to detect. This has been exacerbated since the online trading system was adopted. Moreover, it has been reported that trade-based manipulation has shown more diverse patterns and thereby the rules and patterns derived from historical data regarding previous attempts at manipulation may quickly become outdated. This study proposes a method to detect a suspicious symptom of stock price manipulation only using the change of the stock market data itself. For this, we apply an unsupervised learning technique, peer group analysis, which detects individual objects (target) that begin to behave significantly different from the other objects to which they had previously been similar (peer group). That is, we detect the abnormal behavior of a target by comparing it with its peer group members and measuring the deviation of its behavior from the peer group. The advantage of this approach is that we can find local outliers which cannot otherwise be detected when compared to the whole population. Based on the general peer group analysis, this study moreover proposes a method to improve the general technique by using weighted means as a summarizing statistic of peer group as well as updating the weights over time. Using time series data of stock prices of companies listed in the Korean stock market, we examine the application of the peer group analysis to the detection of stock price manipulation. Finally, we conduct a sensitivity analysis to examine the effect of the parameters used in the proposed method. The remainder of this paper is structured as follows. Section 2 reviews related literature and Section 3 describes peer group analysis. Section 4 applies peer group analysis to the real stock price data and performs sensitivity analysis for the parameters used in the peer group analysis. Finally, Section 5 concludes the paper.

In this study, we showed how peer group analysis can be used to detect stock price manipulation. To detect manipulation, the peer group analysis monitors stocks and compares them with other stocks that exhibit a similar pattern of price change. We moreover proposed a method to improve the general peer group analysis by incorporating the weight of peer group members into summarizing their behavior and also updating the weight based on newly-observed data. From the empirical analysis, we found that the proposed method showed better performance than the general peer group analysis and the approach that monitors single time series only. The method proposed in this study can be used as a technique for effective implementation of peer group analysis. Because some peer group members show systematic price changes over time, peer groups should be reconstructed to track their respective target more closely. However, this process requires numerous calculations and a considerable amount of time to reorganize peer groups for each target. Thus, without reconstructing peer groups, our approach can continuously adjust the structure of peer groups so that they exhibit similar behavior with their target. This study examined suspicious cases of stock manipulation utilizing public data, stock price, which is open to public in real time. Although it will be more accurate to detect manipulation cases using the detailed trading information which can be accessed only by the financial authorities, we believe that examining manipulation cases only with public data can be helpful to general investors. Likewise, trading volume is another important public data of time series, but it was not included in this study. It is expected that peer group analysis utilizing both stock price and trading volume shows better performance. Besides, this study did not examine when the flag should be raised for suspicious manipulation cases. It is also important to detect and raise an alarm for suspicious cases of stock manipulation at an appropriate time. We expect that peer group analysis can be also advantageous in deciding an accurate time to raise an alarm since it is effective to find local outliers. Thus, determining appropriate time points to flag manipulations should be of the goal of future research.