Current Credit Risk Management

Literature Review

This chapter looks at some of the prior study that has been done on the issue. It examines the history of credit risk models as well as the development of new models in this sector. A study of the existing theoretical and empirical literature will be used to explore the need for effective credit risk modeling.

The Current Credit Risk Management Framework

The CBK recommends banks must have a good enough assessment procedure to ensure they are able to assess the risk profile of the borrower and they advocate looking at the credit rating report obtained from any licensed credit bureau (CBK, 2013).Non-banking financial institutions are also in the lending business and as such, their risk also needs to be managed especially with the rise of mobile loans and digital lending. The growing use of the digital loans is due to the quick access of funds and the zero requirement for collateral and the use of alternative credit scoring models which use information on mobile money transactions when determining eligibility of a borrower. The use of such alternate data and data from the Credit Reference Bureau (CRB) can really fit well with the deep learning or machine learning model which are robust in examining the relationships among variables. This increased competition means that the industry has become a margins ‘industry such that a small increase in efficiency in their processes will improve the chances of success. If they are to compete with banks, then the credit scoring mechanism must be efficient and automatic thus reducing the cost of credit analysis. Rhyne and Christen (1999) emphasizes the need of efficient credit scoring models and Schreiner (2004) affirmed that the study done in Bolivia and Columbia reduced costs by up to $75,000 per year in those countries. This is something that can easily translate to the Kenyan market as it is of similar characteristics with these markets. According to the Fin-Access (2016) report, 27% of adults are borrowers of digital loans with the male/female ratio being 55/45%. This is clearly a big market and being able to assess the likelihood of default will prove important to the success of the institutions.

Empirical literature

Jensen (1992) used neural networks to perform credit scoring and discovered a true rate of classification of 76-80 percent, with a 16 percent false positive rate and a 4 percent false negative rate. This was done on a single partition of data with a test on 50 cases. They based the study on a data set that had as much as 181 labeled features that were obtained from the financial statements of companies for the year 2016/2017. They tried finding the most efficient of 7 models including random forests, gradient boosting, ANN’s and logistic regression. They found that tree based algorithms outperformed the rest based on AUC. Charpignon et al. (2014) from Stanford published their paper on credit risk modelling using a data set of around 100, 000 customers that was given in a Kaggle challenge. In their findings, they found that gradient boosting technique had a really good predictive accuracy with an AUC of around 85%. Cao and Yu (2018) also used the “Give me some Credit” Kaggle data set but used 8 models including gradient boosting, neural networks, MLP and SVM and assessed the models based on 3 factors; accuracy, AUC and the logistic loss. The study found that X gradient boosting model had the best performance compared to the rest. The study however did recommend using a bigger dataset so as to improve the accuracy of the predictions.


Leave a Comment

Your email address will not be published. Required fields are marked *

+1 587-331-9072
We will write your work from scratch and ensure that it is plagiarism FREE, you just submit the completed work.