Professor John F. Geweke, Harlan E. McGregor Chair in Economic Theory at the Department of Economics and Statistics at University of Iowa, will visit UNSW's School of Economics and give us a lecture on April 13th, titled "Bayesian Modeling of Conditional Distributions." Professor Geweke is one of the most prominent scholars in theoretical as well as applied econometrics, with specific expertise in Bayesian econometrics, time series analysis, and financial economics. He holds a PhD from University of Minnesota (1975) and held positions in University of Wisconsin, University of Minnesota, and Duke University, to name a few. He has been the fellow of the Econometric Society since 1982 and currently a Co-Editor at the Journal of Econometrics. He has published numerous papers in the top five journals in Economics and a number of other papers in top field journals.
There will be a lunch and dinner hosted by the School of Economics with limited number of availability. If you are interested in attending the lecture and the lunch/dinner, please contact Denzil Fiebig.
Date and time: 13 April (Friday), 3:30-5:00 pm
Room: John Goodsell Building, Room 119 (tentative)
In econometric modeling the problem of inferring the distribution of a variable of interest conditional on a vector of covariates arises repeatedly. Examples include distributions of future asset returns conditional on available information and the assessment of changing inequality in earnings conditional on characteristics like age and education. Typically the conditional distribution cannot reasonably be confined to a known parametric family of models. Since the entire conditional distribution, and not just the conditional mean, is important the voluminous parametric and nonparametric regression literature does not address this problem. This paper introduces a new variant of mixture of experts models developed in the neural computation literature over the past decade, using probit gating functions. As the number of mixture components increases the Kullback-Leibler distance to any conditional distribution in the exponential family becomes arbitrarily small. The paper shows that these models are practical for data sets commonly used in applied econometrics, and they outperform widely used alternatives.