Caravan-Insurance-Challenge

Identify potential purchasers of caravan insurance policies

Problem setting & dataset

Problem setting

Given the customer and sales information, the task is to identify customers that could be interested in a caravan insurance and especially why.

Dataset

The dataset of the competition can be downloaded from the Kaggle website. The dataset contains 5822 observations with 85 features and a target variable (information of whether a customer has (1) or does not have (0) a caravan insurance policy). Each observation corresponds to a postal code. A test set contains 4000 observations without the information on the caravan insurance policy.

Models

All models were trained using the nested cross validation procedure with hyperparameter tuning.

Decision Tree classifier

AUC (Area under the curve): 0.7466
with the averaged hyper parameters: {'criterion': 'gini', 'max_depth': 3}

Random Forest classifier

AUC: 0.7560
with the averaged hyper parameters: {'criterion': 'log_loss', 'max_depth': 5, 'n_estimators': 325}

Feature selection

Motivation for the feature selection:

1. Prevent overfitting
2. Explainable models
3. Variance reduction

Eliminate features which importance is smaller than a given threshold. The threshold here is the mean value of importance of all features.

Top 10 most important features:

• 47 PPERSAUT Contribution car policies
• 68 APERSAUT Number of car policies
• 59 PBRAND Contribution fire policies
• 1 MOSTYPE Customer Subtype see L0
• 37 MINKM30 Income < 30.000
• 18 MOPLLAAG Lower level education
• 42 MINKGEM Average income
• 22 MBERMIDD Middle management
• 43 MKOOPKLA Purchasing power class
• 10 MRELGE Married

Re-train the Random Forest classifier with the useful features only

AUC of the Random Forest classifier with 26 most important features: 0.7669 (vs. 0.7560 of Random Forest with all features)

with the averaged hyper parameters: {'criterion': 'log_loss', 'max_depth': 5, 'n_estimators': 775}