Am Willing to learn more so please share your ideas if you don't mind : (keep in mind am still a novice)

data cleaning:

1. mapping LGA_Name and State values to correct values using provided "NigerianStateNames.csv"

2. removing duplicates from train data using subset('Policy Start Date', 'Policy End Date', 'ProductName', 'Age', 'No_Pol','Gender', 'Car_Category', 'Subject_Car_Colour', 'Subject_Car_Make', 'LGA_Name', 'State')

3. other methods like filling nan values etc

NB: more could be done here

feature engneering:

1. binary variables (1, 0) 1 imputed value else 0

2. binary variables (1, 0) 1 rare value else 0

3. date features (year, month, week, day) and policy duration (year, month, week, day)

4. interaction features (level 2 combinations of (categorical and original numerical features(No_Pol, Age)))

4. original numerical interactions features for (No_Pol and Age) eg. addition, mulitplication and subtraction

5. Weight of Evidence Encoding for categorical features. (Target Encoding worked well too in my case)

6. sum, min , max, kurtosis, min, std, skew, median etc row wise

7. transforming the data using power transform

modeling:

1. weighted voting classifier (LGBMClassifier and CatBoostClassifier, each with class weights)

2. Probability threshold moving