Machine Learning in Credit Risk Modelling

Santander Consumer Bank utvikler matematiske modeller for å kunne evaluere kundens kredittverdighet på en effektiv og nøyaktig måte. Disse modellene blir benyttet for kredittvurdering i søknadsprosessen. Prosjektet har fokusert på dyp læring, et underfelt av maskinlæring, som har oppnådd sterke resultater i klassifiseringsoppgaver. Vi har utviklet en teknikk som er i stand til å finne en datarepresentasjon, som viser veldefinerte grupperinger av bankens kunder. I tillegg benytter vi disse grupperingene for å utføre segment basert kredittscoring, som presterer bedre sammenlignet med en porteføljebasert tilnærming. I andre del av prosjektet, analyserte vi forskjellige teknikker for å utlede kredittverdigheten for avslåtte søknader, såkalt «reject inference». I denne analysen, ble «deep generative» modeller benyttet for å forbedre nøyaktigheten i kredittscoringsmodeller, ved å legge til avslåtte søknader. Eksperimentene viser at våre foreslåtte modeller oppnår bedre resultater enn klassiske og alternative maskinlæringsmodeller benyttet i «reject inference» for kredittscoring. Artiklene har blitt publisert og er tilgjengelig på følgende lenker: https://www.sciencedirect.com/science/article/abs/pii/S0950705120301660 - https://www.sciencedirect.com/science/article/abs/pii/S0957417420307910

The results of the project and the experience the institution gained from it is seen as a key contributor to further pursuing the usage and analyses of advanced techniques in the field of credit scoring over the currently available, traditional techniques.

The logistic regression is useful in calculating rankings of customer creditworthiness. However, there is large literature showing the advantages of machine learning in credit risk. Machine learning offers higher classification accuracy and solutions to deal with common problems faced in the classical credit risk models. In addition, machine learning can improve credit allocation and estimate credit risk relatively more accurate. Hence, the aggregated social welfare is improved. This is achieve in two ways 1)right customers have access to credit and the credit cost is relatively more accurate, and 2) bank's financial obligations are not threaten by customers' lack of payment. Hence, Santander Consumer Bank AS wants to learn from the best-in-class research institutes in the country, and to be able to build in-house competence and use this knowledge to improve bank's credit risk management by building machine learning credit risk models. Santander considers this project as an essential part of a long-term strategy for building in-house competence and expertise in machine learning for credit risk modelling. This competence is not limited to Mr. Mancisidor, the industrial PhD candidate. Over the duration of the project, and upon completion of the PhD, Santander aims for Mr. Mancisidor to actively transfer his acquired knowledge in this project to his fellow members of the Nordic Risk Models department.