Nonparametric density estimation: A comparative study

Motivated by finance applications, the objective of this paper is to assess the performance of several important methods for univariate density estimation focusing on the robustness of the methods to heavy tailed target densities. We consider four approaches: a fixed bandwidth kernel estimator, an adaptive bandwidth kernel estimator, the Hermite series (SNP) estimator of Gallant and Nychka, and the logspline estimator of Kooperberg and Stone. We conclude that the logspline and adaptive kernel methods are superior for fitting heavy tailed densities. Evaluation of the convergence rates of the SNP estimator for the family of Student-t densities reveals poor performance, measured by Hellinger error. In contrast, the logspline estimator exhibits good convergence independent of the tail behavior of the target density. These findings are confirmed in a small Monte-Carlo experiment.