B1728
Title: Specification and estimation of mixtures with dynamic weights
Authors: Marco Bee - University of Trento (Italy) [presenting]
Abstract: Mixture distributions with weights depending on the magnitude of the observations are flexible models for non-negative, skewed and heavy-tailed data. However, estimation is not trivial, mainly because the density contains an intractable normalizing constant, and the number of parameters is large. So far, in all versions of this model studied in the literature, the functional form of the mixing weight is the Cauchy cumulative distribution function. The statistical properties of dynamic mixtures are analyzed based on different specifications of the weight function, exploring the trade-off between the larger flexibility granted by distributions with more parameters and the more efficient estimation that characterizes less flexible models with fewer parameters. Three estimation methods, namely maximum likelihood, approximate maximum likelihood and noisy cross-entropy, will be employed. The comparison will be based on both classical measures of statistical performance, such as root-mean-squared error and information criteria, and on considerations of computational burden.
