“…For example, Haas et al [47], Alexander and Lazar [48], Haas et al [49], Haas and Paolella [50] and Haas et al [51] use discrete normal mixture (MixN) GARCH models, with the MixN being a short-tailed distribution; Broda and Paolella [52] use a normal inverse Gaussian (NIG) structure, this having so-called semi-heavy tails (and existence of a moment generating function); several authors, including Mittnik and Paolella [53], Kuester et al [54], and Krause and Paolella [55] use an asymmetric Student's t distribution (this having heavy tails but such that the tail index lies in p0, 8q); Mittnik and Paolella [56] and Broda et al [57] use the asymmetric stable and mixtures of symmetric stable, respectively, these having heavy tails such that the tail index lies in p0, 2q. The fact that all of these methods can deliver highly accurate VaR forecasts confirms that the choice of asymptotic tail behavior and the (non-)existence of the second moment are not highly relevant for risk forecasting, but rather the use of a leptokurtic, asymmetric, bell-shaped distribution, in conjunction with a GARCH-type process.…”