The search for outliers : a data mining approach

Abstract

Millions of companies’ shares are traded on a daily basis at financial markets as traders’ attempt to generate returns. As a result, vast amounts of data is generated and stored. Among this data, one may encounter an outlier: an observation within a dataset which deviates considerably from the remaining data. While an outlier may be caused by incorrect data, the search for these anomalous cases may indicate the presence of illegal activities in the stock market, such as insider trading and market manipulation. Therefore, one needs to analyse the valuable information that may be contained and associated in an outlier prior to potential removal. In this project, the Voting-Based Outlier Mining for Multiple time series (V-BOMM) algorithm is applied to 36 datasets of financial time series data from different companies and indices to elect those dates where potential outliers, in each respective dataset, is earmarked. The various steps involved, from loading data to voting, are outlined in this document. The outcome of this project is to further reinforce the belief that data mining outlier detection techniques should be used in financial markets. In this project, outliers are sought after for their potential in identification of illegal activity in through the use of attributes found in datasets. The successful development and implementation of this project identified five (5) candidate outliers across five (5) companies, each found in different sectors. These cases were calculated from a total of thirty-six (36) over the span of two hundred and fifty-one (251) trading days. These candidate outliers should therefore be further investigated for the possible evidence of illicit activities.