Research topics on credit risk management

Abstract

Συνοψίζοντας παρακάτω τους κύριους στόχους, τα κύρια χαρακτηριστικά και τη συνεισφορά της παρούσας διατριβής: • Ο κύριος στόχος αυτής της διατριβής είναι η πρόταση τόσο για περιγραφικούς όσο και για προβλεπτικούς σκοπούς, μιας καινοτόμου ευέλικτης και αξιόπιστης προσέγγισης για τη μοντελοποίηση της πιστοληπτικής ικανότητας, η οποία έχει σημαντική σημασία στη χρηματοδότηση και την τραπεζική λόγω της άμεσης σύνδεσής της με την πιστοληπτική ικανότητα. • Η πρωτοτυπία και μία από τις κύριες συνεισφορές της προτεινόμενης μεθοδολογίας μοντελοποίησης έγκειται στο γεγονός ότι συνδυάζουμε αποτελεσματικά οικονομικά χαρακτηριστικά μαζί με χαρακτηριστικά πιστωτικής συμπεριφοράς αλλά και εναλλακτικά δεδομένα που δεν έχουν εξεταστεί ποτέ πριν καθώς οι περισσότερες χώρες και ιδρύματα χρησιμοποιούν μόνο οικονομικά δεδομένα για τη μοντελοποίηση της πιστωτικής βαθμολόγησης. • Πραγματοποιήθηκε μια συγκριτική μελέτη αξιολόγησης δώδεκα αλγορίθμων ταξινόμησης σε ένα πραγματικό σύνολο δεδομένων πιστοληπτικής ικανότητας για τη σύγκριση καινοτόμων και παραδοσιακών μεθόδων ταξινόμησης που προσφέρουν πολύτιμες γνώσεις τόσο στους επαγγελματίες όσο και στους μη επαγγελματίες. • Μια αποτελεσματική και φιλική προς τον χρήστη αλγοριθμική διαδικασία που έχει προταθεί και εφαρμοστεί στη διατριβή αποτελεί μία ακόμα συμβολή δεδομένου ότι ανταποκρίνεται στην ανάγκη μείωσης της διάστασης, ένα ζήτημα που συναντάται συχνά στην πράξη, ειδικά σε προβλήματα που ταξινομούνται στην περιοχή της Ανάλυσης Μεγάλης Κλίμακας Δεδομένων (Big Data). Από όσο γνωρίζουμε, αυτή είναι η πρώτη φορά που ο συνδυασμός των παραπάνω τεχνικών πολλαπλών επιπέδων μείωσης διάστασης χρησιμοποιείται και εφαρμόζεται αποτελεσματικά, στη μοντελοποίηση της πιστοληπτικής ικανότητας. • Τέλος, παρέχουμε μια αξιολόγηση των πρόσφατων μεθόδων πιστωτικής βαθμολόγησης για να συνδράμουμε τη μελλοντική έρευνα.

Credit risk is one of the biggest threats facing credit institutions. The gradual development of credit risk control leads to the need for continuous improvement of credit risk models in order to face or predict it. For this reason, this thesis focuses on the contribution to areas related to methods of forecasting and selection of explanatory / independent variables with the ultimate goal of enhancing the profitability of credit risk models. Initially, our interest is focused on the category of forecasting models, including the nature (type) of explanatory / independent variables that can be used in credit scoring models but also new classification algorithms for valid and reliable evaluation of the performance of the proposed models.We first explore the effectiveness of alternative data in credit rating models. As alternative data we consider data derived from non-traditional sources and can be used to supplement traditional data in order to provide better information that would otherwise not have been achievable and which was considered unique, unusual or expensive a few years ago.For this purpose, we created and introduced variables that come from alternative sources of information, in an already existing forecast model for Greek hotels that uses only credit behavior data.For this analysis, a real set of credit scoring data of 678 Greek hotels was used, which was provided by the private database of TEIRESIA SA. (a company founded by almost all banks in Greece). Comparing the "alternative" model with the existing one using the performance indicators K-S, Gini Index and accuracy, we came to the conclusion that the alternative data contribute to the performance of the model. Having seen the improvement in model performance for Greek hotels, we can easily conclude that it would be wise to explore the usefulness of alternative data in other industries as well.Then, a comparative evaluation study of 12 classification algorithms in the same data set was performed to compare novel with traditional classification methods. In pursuit of this goal, we compared these classification algorithms in terms of AUC and accuracy. Our results showed that there are small differences between the values ​​of the performance indicators in each classifier and this is probably because we are working on a homogeneous sample. In particular, we observed that logistic regression and neural networks performed better than other (new or not) classifiers and logistic regression had the highest AUC.The contribution of this analysis lies primarily in the use of alternative data in forecasting models that traditionally use only classical credit behavior data. In addition, it contributes to the relevant literature by defining and utilizing variables from alternative sources of information with application in the hotel sector. It also provides valuable information for professionals, as they can take advantage of new classification algorithms for predictive models. In addition, we provide an assessment of recent scoring methods to assist in future research.Continuing to aim for maximum performance of forecasting models and wishing to contribute to the business sector as well as to the wider industry (not just the hotel industry), we decided to explore a combination of data on the independent variables that will make up forecasting models for companies . Given that financial data is usually the only data used in modeling (both in Greece and in other countries) to assess a company's credit risk, we decided to use a combination of financial and credit behavior data.The proposed models (with the data combination) were then compared with the traditional models (containing only financial data) using three performance indicators, the accuracy, the K-S and the Gini Index.After comparing the models, we came to the conclusion that the new models contribute to the assessment of credit risk as shown by their performance.Finally, this thesis also contributes to the analysis of big data since it deals with the problem of selecting variables which in combination with the utilization of the dimension reduction technique achieves the construction of flexible and reliable classification models for Greek companies (one model for small and one for large companies), based on their credit behavior.For the modeling, we propose an algorithmic procedure of 3 (4) steps to reduce the dimensions with an initial stage of preliminary data processing (step 0) which was performed in the previous analyzes and is as follows:We used the Weight-of-Evidence (WOE) encoding to generate dummy variables in order to group all the independent variables.The main part of the algorithm is based on dimension reduction techniques taking into account the Akaike step-by-step information criterion and principal component analysis (PCA). The proposed procedure allows an optional 4th step based on Elastic Net Regularization to further reduce the dimension if the researcher believes this is useful.The findings of this analysis clearly show the importance of using credit behavior variables, as some of these variables have been found to play a key role in creating credit rating models for both small and large businesses. Indeed, in the final small business model, each PCA variable depends on 6 credit behavior variables (out of a total of 15 variables) while in the large enterprise final model, each PCA variable depends on 10 credit behavior variables (out of a total of 18 variables). The use of such combinations is one of the main contributions of the present thesis, as countries rely almost exclusively on economic variables. It is also worth noting that the proposed methodology responds to the need to reduce the dimensions for the construction of flexible but also reliable credit models not only for descriptive but also mainly for predictive purposes.In addition, the proposed methodology provides, among others, insurers, financial planners and lenders with an automated reliable financial instrument for assessing creditworthiness according to some statistically significant financial and credit variables and at the same time offers creditworthiness to borrowers increased lending opportunities.Also, the proposed dimension reduction model can be applied to the modeling of fiscal debt creditworthiness. The importance of forecasting arising from fiscal debt rating agencies is another area of ​​potential extensions. The significance of the provisions concerning both credit risk assessment models and fiscal debt assessment models can be tested using non-parametric methods.