Management of credit risk, one of the main bank activities, is currently a very important issue. This paper contains comparison of two instruments used in prediction of probability that consumer fails to fully repay a loan in agreed time: artificial neural networks and models for polychotomous ordered data. For the empirical research each client has been assigned to one of four categories reflecting his/her delay in payments. Estimation and validation of methods was performed on a 3000-item sample containing information about each loan agreement and repayment history originating from one of Polish banks, covering years 2000-2001. The dataset was repeatedly divided into train and validation sets. Multi-layer architecture of artificial neural network with logistic activation function was proposed. Ordered logit and probit models were estimated within maximum likelihood framework. Several alternative specifications were proposed differing in independent variable set (including their products and squares). Bank income was chosen as the main criterion of fitness. Problem of optimal decision and defining appropriate loss function was formulated on the basis of statistical decision theory. Furthermore, properties of estimated models related to inference about probability of repayment and credit risk factors were presented.