Date Added: Mar 2010
Generalized AutoRegressive Conditional Heteroskedasticity (GARCH) processes have become very popular as models for financial return data because they are able to capture volatility clustering as well as leptokurtic unconditional distributions which result from the assumption of conditionally normal error distributions. In contrast, Bollerslev (1987) and several follow-ups provided evidence that starting with leptokurtic and possibly skewed (conditional) error distributions will achieve better results. Parallel to these flexible but to some extend arbitrary chosen parametric distributions, recent years saw a rise in suggestions for maximum entropy distributions (e.g. Rockinger and Jondeau, 2002, Park and Bera, 2009 or Fischer and Herrmann, 2010).