Multivariate cumulants in outlier detection for financial data analysis

TL;DR

This paper introduces a novel method using 4th order multivariate cumulants to detect outliers and predict financial crises by analyzing tail dependencies in asset returns modeled with t-Student copulas.

Contribution

The paper presents a new outlier detection algorithm based on multivariate cumulants, specifically designed for identifying tail dependencies and potential crises in financial data.

Findings

Successfully detected financial crises using artificial data

Demonstrated the effectiveness of 4th order cumulants in tail dependency analysis

Potential application for early crisis prediction

Abstract

There are many research papers yielding the financial data models, where returns are tied either to the fundamental analysis or to the individual, often irrational, behaviour of investors. In the second case the bubble followed by the crisis is possible on the market. Such bubble or crisis is reflected by the cross-correlated extreme positive or negative returns of many assets. Such returns are modelled by the copula with the meaningful tail dependencies. The typical model of such cross-correlation provides the t-Student copula. The author demonstrates that the mutual information tied to this copula can be measured by the 4th order multivariate cumulants. Tested on the artificial data, the 4th order multivariate cumulant approach was used successfully for the financial crisis detection. For this end the author introduces the outliers detection algorithm. In addition this algorithm displays the potential application for the crisis prediction, where the cross-correlated extreme events may appear before the crisis in the analogy to the auto-correlated ones measured by the Hurst Exponent.