We have been trying to build a classification model for credit default prediction using two different models one being Random forest and another being the Logistic regression based scorecard model. Once we have the samples in place, the number of bad events are too low for the time window we had chosen (<1%).
Willing to know whether additionally bringing in only the bad observations from the time period before the considered time window will help. By this way, we can be able to increase the number of bad events/rate for the model to learn from.
We have the following thoughts for doing so;
- “The goal is to classify the ‘good’ and ‘bad’ characteristics. We have enough good characteristics but not enough bad. So can we bring some bad observations which brings in bad characteristics giving the model enough bads to learn from? These bads are of the same clients type/population from the past years”.
Thanks in advance.