Statistical estimation of composite risk functionals and risk optimization problems (Q2409393)

The authors analyze the asymptotic properties of empirical estimators for optimized risk functionals. Let \(X\) be an \(m\)-dimensional random vector, \(U\) a compact set in a finite dimensional space, and let \(f_j:U\times\mathbb{R}^{m_j} \times \mathbb{R}^m \to \mathbb{R}^{m_{j-1}}\) \((j=1,\dots,k)\) with \(m_0=1\) and \(f_{k+1}:U\times\mathbb{R}^m\to\mathbb{R}^{m_k}\). Set \[ \varrho=\min_{u\in U}\mathbb{E}\Big[f_1\Big(u,\mathbb{E}\big[f_2\big(u,\mathbb{E}[\;\cdots f_k(u,\mathbb{E}[f_{k+1}(u,X)],X)]\;\cdots,X\big)\big],X\Big)\Big] \] with empirical estimator \[ \varrho_{n}=\min_{u\in U}\sum_{i=1}^n \frac{1}{n}\Big[f_1\Big(u,\sum_{i=1}^n \frac{1}{n}\Big[f_2\Big(u,\sum_{i=1}^n \frac{1}{n}[\cdots~~~~~~~~~~~~~~~ \] \[ ~~~~~~~~~~~~~~~f_k(u,\sum_{i=1}^n \frac{1}{n}[f_{k+1}(u,X)],X)] \cdots,X\Big)\Big],X\Big)\Big]. \] Using Donsker's theorem, the authors obtain that if the \(f_{j}\)s satisfy some smoothness conditions, \(f_{j}(u,\eta,X)\) has finite second moment for all \(j,u,\eta\), and the optimization problem in the definition of \(\rho\) has a unique solution then \(\sqrt{n}(\varrho_{n}-\varrho)\) converges in distribution to a normal distribution. The special cases of this result where either \(U\) is a single point or \(k=1\) are discussed separately. The authors note that their results are applicable for the average value-at-risk \[ \rho_{\alpha}(X) = \min_{z\in\mathbb{R}} \bigg\{z + \frac{1}{\alpha} \mathbb{E}[(X- z)_+]\bigg\}, \] the mean-semideviation risk measure \[ \rho_{\kappa,p}(X) = \mathbb{E}[X] + \kappa \Big[\mathbb{E}\big[\big(\max\{0,X- \mathbb{E}[X]\}\big)^p\big]\Big]^{\frac{1}{p}}, \] and the higher order inverse risk measure \[ \rho_{\alpha,p}(X) = \min_{z\in \mathbb{R}} \Big\{ z + \frac{1}{\alpha}\big\|\max(0, X- z)\big\|_p \Big\}, \] and spell out the limit distribution of \(\sqrt{n}(\varrho_{n}-\varrho)\) in these special cases. Simulation results are also presented to illustrate how for the higher order inverse risk measure with \(p=2\) and \(\alpha = 0.05\) the quality of the normal approximation depends on \(n\) and the tail behavior of \(X\).