Embedding neural network: Application to credit risk (default probability)
Files
Kembou_42261900_2023.pdf
Closed access - Adobe PDF
- 2.04 MB
Details
- Supervisors
- Faculty
- Degree label
- Abstract
- For several years, institutions have been concentrating their efforts in developing innovative modelling methods. The growing environment of Artificial Intelligence (AI) has enhanced the machine learning advancement, consequently impacting several business. As a result, the modern era of big data has presented some limitations in actuarial science to address complex relation in large sample size. This research main objective is to implement a machine learning technic to predict the default probability, particularly by applying embedding neural network on categorical variables during a logistic regression. Thus, executed using the binomial deviance loss function. Default probability is the likelihood that over a specified period, a borrower will not be able to make their scheduled repayments on a particular debt. The capacity to perfectly predict a default probability is vital for the development of any institution. Hence, a more accurate default prediction is important for institutions in order to avoid loss. In addition, the variability of the machine learning model will be examine and how it reacts with more or less embedding layers. The prediction performance obtain was compared with that of a classical GLM logistic model. Data was obtained from the Kaggle website containing information about each applicant and was recorded between 2007-2011 by the Lending club. However, the machine learning model completely outperform. Finally, the overall results was outlined and the best model selected according to a defined criterion undertaken. Nevertheless, the Generalised Linear Model (GLM) was implemented as a benchmark of this research. The results of this work shows the positive contribution of machine learning in actuarial science. Due to the rapid changes in modern technics, better performances are expected.