Country Risk Classification and Multicriteria Decision-Aid

Abstract

Country risk is an important concern in international business. Country risk classification refers to determining the risk level at which a country will not repay its international debt. Traditionally, country risk classification resorts to statistics methods such as discriminant analysis. In the past two decades, the so-called multicriteria decision aid (MCDA) methods have been proved to enjoy better performance than the standard statistics methods. Nevertheless, the performance of the MCDA methods is still far away from satisfactory and can be improved significantly. The better performance of several MCDA methods, such as UTADIS (UTilités Additives DIScriminantes) and MHDIS (Multigroup Hierarchical Discrimination), is achieved by exploiting the rater’s background knowledge. In the standard MCDA model, we assume that the criterion function for every factor is monotone and all the factors are independent. Then, we approximate the impact of every factor and use the sum of the corresponding criterion functions to determine the risk level of a country. By discretizing the feasible domain of the factor, the MCDA method solves a linear program to find a classifier for country risk classification. This thesis tries to enhance the capability of MCDA methods by allowing a class of non-monotone criteria: the unimodal ones. For this purpose, we developed an integer quadratic (non-convex) program for general unimodal criteria. Further, if we restrict ourselves to convex or concave unimodal criteria, then we can still use a linear program to find a classifier. For the case where all the factors are correlated, a simple quadratic form of aggregation is proposed to deal with it. Compared with the original UTADIS model, our generalized model is more flexible and can deal with more complex scenarios. Finally, our generalized model is tested based on cross-validation and our experiment is carried out under the AMPL+sovers environment. Promising numeric results indicate that except for its theoretical advantages, our generalized model exhibits practical efficiency and robustness as well.