Outliers:The Dangers of Not Being One of the Pack

Abstract

Outliers are ubiquitous in modern datasets. Due to noise, uncertainty, and adversarial behavior, such observations are inherent to many real-world problems. In financial applications, outliers are very common not only in time series, but also in multivariate data used for the construction of fundamental factor models of stock returns. In many cases, failing to identify outliers may lead to wrong results and poor performance of the various models. Discriminating outliers from correct data have been studied extensively, in both statistics and machine learning. However, most financial practitioners rely on standard approaches like trimming and winsorization, which are trivial to implement and can easily be applied to large datasets. The aim of this article is to draw attention to the drawbacks of the standard methods and to present alternative ways by which outliers can be detected effectively in both the univariate and multivariate case.