Support Vector Machines for Credit Scoring

Quantitative methods to assess the creditworthiness of the loan applicants are vital for the profitability and the transparency of the lending business. With the total loan volumes typical for traditional financial institutions, even the slightest improvement in credit scoring models can translate into substantial additional profit. Yet for the regulatory reasons and due to the potential model risk, banks tend to be reluctant to replace the logistic regression as an industrial standard with the new algorithms. This does not stop researchers from examining such new approaches, though. This thesis discusses the potential of the support vector machines, to become an alternative to logistic regression in credit scoring. Using the real-life credit data set obtained from the P2P lending platform Bondora, the scoring models were built to compare the discrimination power of support vector machines against the traditional approach. The results of the comparison were ambiguous. The linear support vector machines performed worse than logistic regression and their training consumed much more time. On the other hand, support vector machines with non-linear kernel performed better than logistic regression and the difference was statistically significant at 95% level. Despite this success, several factors prevent SVM from the widespread applications in credit scoring, higher training times and lower robustness of the method being two of the major drawbacks. Considering the alternative algorithms which became available in the last 10 years, support vector machines cannot be recommended as a standalone method for credit risk models.