Sample selection bias and the evidence from a financial market

The aim of the diploma thesis is to introduce the methodology and several approaches previously used to account for the sample selection bias. The focus of the thesis is primarily on the consumer lending where the selection bias is unavoidable. Financial data about consumers are provided by a Czech commercial bank. The data are split into rejected and accepted accounts which is a source of the selection bias. Reject inference is a method to exploit the information about the rejected accounts to improve the predictive performance and remove the selection bias. The methods used for the analysis are fuzzy augmentation, Heckman’s two stage correction, parceling, augmentation (reweighting) and reclassification. First, an application scorecard is developed using the known observations. The reject inference models are fitted on the through the door population. Results show that only Heckman’s two step correction method is giving satisfactory results and that a modest improvement in the scorecard is possible.