Getting ready

The strategy that we'll adopt to predict default of a customer is as follows:

• Objective: Assign a high probability to the customers who are more likely to default.
• Measurement criterion: Maximize the number of customers who have actually defaulted when we consider only the top 10% of members by decreasing the default probability.

The strategy we will be adopting to assign a probability of default for each member will be as follows:

• Consider the historic data of all members.
• Understand the variables that can help us to identify a customer who is likely to default:
  • Income-to-debt ratio is a very good indicator of whether a member is likely to default.
  • We will be extracting a few other variables similar to that.
• In the previous step, we created the input variables; now, let's go ahead and create the dependent variable:
  • We will extract the members who have actually defaulted in the next 2 years by first going back in history and then looking at whether members defaulted in the next 2 years
  • It is important to have a time lag, as it might not give us any levers to change the outcome if we do not have a time gap between when a member is likely to default and the date of prediction.
• Given that the  outcome is binary, we will minimize the binary cross-entropy loss.
• The model shall have a hidden layer that connects the input layer and the output layer.
• We shall calculate the number of the top 10% probability members who have actually defaulted, in the test dataset.

Note that we assume that test data is representative here, as we are not in a position to assess the performance of a model on unseen dataset without productionalizing the model. We shall assume that the model's performance on an unseen dataset is a good indicator of how well the model will perform on future data.