Abstract
The determination of the financial credibility of a person for a loan is a challenging task as many variables are taken into consideration. Recently, there has been a surge in the application of machine learning approaches in the design of robust and effective credit scoring models as part of the human social development agenda under the assumption that the variables will remain stable for a long time. However, in real-life, the behavior of customers changes over time and the variables used to quantify the financial credibility of a person for a loan such as past performances on debt obligations, profiling, main household, income and demographics tend to drift and evolve over time. This paper considers credit scoring as an ephemeral scenario as variables tend to drift over time and proposes the application of data stream learning techniques in credit scoring since they are tailored for incremental learning. This makes the scoring model to be able to detect and adapt to changes in the customer behavior. We propose the Adaptive and Dynamic Heterogeneous Ensemble (ADHE) approach that is capable of learning incrementally and adapting to drifting variables and consists of models derived from different learning algorithms to exploit diversity. The prediction performance of ADHE is evaluated using datasets that are publicly available and we compared the accuracy and computational cost of ADHE with existing state of the art models. Our proposed approach performs significantly well when compared to existing state of the art benchmark models on prediction accuracy according to the non-parametric test.