Semiparametric Structure Guided by Prior Knowledge with Applications in Economics

Abstract

Among many developments in statistical modelling in recent years, non- and semiparametric methods have proved to be a particularly powerful data-analytic tool. Nevertheless, there still exist justified doubts regarding there forecasting performance, for example in the context of financial time series. The aim of this thesis is to demonstrate that, by suitable modification, these techniques can perform well in different economic fields, like empirical demand analysis or prediction of stock returns, if they are adapted to the specific application under investigation.This thesis makes use of the semiparametric nature of economic problems to reduce dimensionality, and is based on the structure that is inherent in the economic process that generates the data. A key feature is to show how prior knowledge can guide the modelling process. This is done either by directly applying economic theory or by examining simple parametric models to identify the coarse features of the relationships. The use of prior knowledge not only improves the plausibility of the model but also the interpretability of the results. Furthermore, it can be used to address some well-known problems associated with fully nonparametric approaches. For example, the curse of dimensionality can be circumvented, the estimation accuracy on boundaries can be improved, or the bias can be reduced by applying a semiparametric approach.The first part of this thesis is thereby a contribution to the analysis of consumer expenditure and price micro-data, while the second part addresses the prediction of excess stock returns. The use of nonparametrically generated bond yields is proposed and prior information about the shape of the unknown conditional mean function is used in the estimation process.