Big data has been one of the business buzz-words of the last couple of years. However, techniques for analyzing these data sets still lag behind this development, since sophisticated statistical models often require enormous computer resources to be applied to data sets of this size. In his dissertation Tim Salimans develops new computational methods that make it easier to apply sophisticated statistical models to the data sets of today. Tim also applies these new techniques to various statistical problems, including forecasting newspaper sales, locating dark matters in space, predicting the outcomes of chess matches, and recommending Xbox games.
Tim has shown the practical value of his research. He has won several statistical prediction competitions held by the San Francisco based competitive predictive modelling platform Kaggle.com This platform ranks him as one of the best data scientists in the world. As a result Tim has gained media coverage in several countries. He advises various Dutch companies through his predictive analytics consulting firm Algoritmica. Tim has published in several journals, including the Journal of Econometrics and Bayesian Analysis.
The dissertation will be defended on May 23 at 13.30 hours at Erasmus University Rotterdam. Salimans’ supervisor is Professor Richard Paap and co-supervisor is Professor Dennis Fok. Salimans conducted his doctoral research in economics at the Tinbergen Institute.
About Tim Salimans
Tim Salimans (1985) obtained a BSc in Science from University College Utrecht in 2007 and received an MPhil from Tinbergen Institute in 2009. He continued his PhD research at TI / ESE in Rotterdam. During this period, he spent three months at Microsoft Research Cambridge as part of his dissertation research. Currently, Tim works at the consulting form Algoritmica, which he co-founded in 2012. Tim is also a regular competitor in statistical prediction competitions, held by Kaggle.com. Organizations supply these competitors with a data set and a modelling problem, while holding back part of the data. The challenge is to build the best statistical model to predict the values of the data that is held back. Tim is the winner of several of these competitions and received cash prizes, including:
- The Deloitte/FIDE Chess rating challenge, predicting the outcomes of chess matches based on past results.
- Observing Dark Worlds competition, organized by the University of Edinburgh and sponsored by Winton Capital: discovering the location of dark matter halos in space, based on the distortions they cause in astronomical observations.
- Currently in second place, out of 1.500+ teams, for the Heritage Health Prize: a contest to predict hospitalizations with over 3 million dollar in total prize money.
Abstract
Econometrics relies on probabilistic models to describe and analyze how observed data relates to economic hypotheses and to provide a rigorous framework for reasoning under uncertainty. Statistical analysis of such models can be performed using likelihood-based inference methods such as Bayesian analysis and the method of maximum likelihood. This dissertation deals with the computational challenges associated with these methods. It aims to solve these computational problems using Monte Carlo methods as well as deterministic approximations of the (marginal) likelihood. By developing efficient approximate inference algorithms, this work addresses various problems in economics for which likelihood-based econometric inference was previously difficult of infeasible.