Normex, a new method for evaluating the distribution of aggregated heavy tailed risks (Q482083)

The paper develops a theoretical mixed method (called \textit{Normex}) to approximate the cumulative distribution function of the sum \(S_n\) of \(n\) independent and identically distributed random variables \(X_1,\dots,X_n\), having the associated order statistics \(X_{(1)}\leq \dots\leq X_{(n)}\). The idea is to write \(S_n\) as the sum of two terms, \(T_k=\sum_{j=1}^{n-k}X_{(j)}\) and \(U_{n-k}=\sum_{j=n-k+1}^{n}X_{(j)}\), where the threshold \(k\) is defined as the smallest positive integer \(j\leq n-1\) such that \(\mathbb{E}[X_{(n-j)}^4]<\infty\). The choice of the threshold \(k\) is dictated by the convenient use (on the basis of the Berry-Esseen inequality) of the central limit theorem for a normal evaluation of the distribution of \(T_k\). The study focuses on the case of \(\alpha\)-Pareto distributed random variables \(X_i\), when the precise determination of \(k\) is immediate. Thus, by using a refined technique based on the conditional decomposition, the author obtains a reasonable approximation of the cumulative distribution function of \(S_n\). Further, the paper provides a bound for the Normex approximation error, an estimate of risk measures and a numerical study involving the comparison of Normex with the classical asymptotic approximation methods. Some results of the extreme value theory are evoked during this study. Finally, it is mentioned that the new proposed method seems to find important financial/actuarial applications.