City Research Online

Abstract

The research for this thesis has been conducted at the junction of several intellectual disciplines. These include the theory of human information processing, economics, banking, accounting, and applied statistics, and the results that are reported here should be of theoretical and practical interest to practitioners in all of these fields. This thesis is rooted in the theory of human information processing, and its primary purpose is to test the hypothesis that formal multivariate statistical modelling techniques have a key contribution to make to human decisions. For this purpose, the chosen task-domain is that of creditworthiness assessments by international lenders to less developed countries.

The starting-point for the research is the recent history of lending to less developed countries, and the first original contribution is the derivation of formal multivariate statistical models of credit- worthiness. These are estimated on a set of panel data for a group of 55 LDCs, including all the major debtors. The techniques that are applied include principal components analysis, cluster and proximities analyses, discriminant and logit analyses, and the automatic interaction detector (AID). Using discriminant and logit analysis and AID, early-warning models are derived that perform well, and have a clear operational utility. Moreover, there is a broad consensus across the derived models, concerning variable-specification.

The second major result reported here relates to published and commercially produced ratings of country risk, including those of BERI, the EIU, the ICRG, and the Institutional Investor. The predictive performance of these ratings is shown to be poor. Paramorphic representations of the rating systems are obtained using multiple regression and AID, and these show that the problem is one of bias in the judgemental processes upon which the ratings are based, rather than inconsistency in the application of judgement. This conclusion is supported by an intertemporal statistical analysis of the Institutional Investor rating system.

The performance of the multivariate models is compared with that of the rating systems, and in broad terms the models are found to be superior. In certain circumstances the Institutional Investor system has a lower misclassification cost than the multivariate models, but the findings nevertheless support the case for man-model interaction.

Publication Type:	Thesis (Doctoral)
Subjects:	H Social Sciences > HA Statistics H Social Sciences > HG Finance
Departments:	Bayes Business School Bayes Business School > Bayes Business School Doctoral Theses Doctoral Theses

Export

Downloads