Machine Learning-Based Predictions of Customers’ Decisions in Car Insurance

Figures & data

Table 1. Distribution of target attribute.

Value	number of records	% of dataset
N/A	32471	72.89
CASH	1227	2.75
RENTAL_INS	3988	8.95
RENTAL_EXT	6857	15.39

Table 2. Features used in presented approach.

Attributes group	Attributes examples	Gap percentage
Person description (victim/perpetrator)	sex	0.1/0.02
	age	3.54/14.54
	postal code of place of living	0.16/0.04
Vehicle description (victim/perpetrator)	vehicle segment	0.20/0.75
	vehicle brand	0.02/0.00
	vehicle year of production	0/0
Perpetrator insurance history (TPO/Compr.)	accident history	3.20/1.77
	number of cars owned and insured	32.19/32.19
Accident circumstances	date	0
	place of accident	0
	time from nearest holiday	0
	number of days off in a week	0
	police presence	0
	condition of victim’s vehicle	50
	was the vehicle towed	35.90
Damage report circumstances	date	0
	how many days between the accident and report	0
	days off in a 10-day period	0
	will the claimant repair his car	0.01

Figure 1. ROC for 10-fold cross-validation for binary classification (cash or rental car).

Figure 2. Results of leave-one-out cross-validation for the gradient boosting method.

Table 3. Synthetic results of 10-fold stratified cross-validation for binary classification.

Classifier	AUC	EER
Random forest	0.81	26%
Logistic regression	0.79	28%
AdaBoost	0.82	26%
Gradient boosting	0.83	24%
Multilayer perceptron	0.79	28%

Figure 3. Histogram of probabilities of the ‘CASH’ class for the gradient boosting method.

Figure 4. Confusion matrix.