The client was a leading P2P financial services organization based out of Indonesia and the objective was to develop a credit scoring model for the loan applicants, which could be used to design custom plans for the applicants based on their credit scores.
Client was capturing different data points from a customer, like form data, relevant ID, demographic information etc.
In addition to that, since the login was facilitated though a mobile app, we also had access to the customer’s device/app usage behavior.
Initially loans were disbursed for a period of 10 months to customers who had valid documents and the default patterns were marked. This data gathered over 10 months was then used to create a credit scoring model.
Customers who had defaulted in the past 10 months were analyzed carefully and that was correlated with the available data points.
Features were generated based on the analysis. A gradient boosted classification tree was identified as the ML algorithm since the number of data points available was less.
SMOTE was implemented as the dataset was highly imbalanced, i.e. 1 in 100 cases of defaulters. The model was deployed in production which was used to monitor new applications on an on-going basis.
The model gave an initial AUC of 0.75, but as the volume of data increased, the model reached AUC close to 0.82. The probabilities obtained from the ML algorithm were converted to credit scores in the range 300-850.