Thank you Zindi for a good competition.

Below is a summary of my approach:

1. Feature engineering

2. Winsorization (Replacing extreme values with a quantile threshold instead of removing them.)

3. BorderlineSMOTE and Scale-pos-weight to combat class imbalance

4. Model ensembling (XGB, LGBM and CatBOOST)

5. Post-processing: The post processing is based on the fact that a loan id only has one class of target even if it has several lenders. Finally, Ghana datasets, type-3 was adjusted because the models performed poorly on that particular dataset.

Full details and code can be seen here:

https://www.kaggle.com/code/tobby18/african-credit-scoring-challenge-zindi-final#Model-Development