Project Introduction

Project : French Motor Claims
Author : Bhishan Poudel, Ph.D Physics
Goal : Implement Frequency modelling, Severity modelling and Pure Premium Modelling
Tools : pandas, scikit-learn, xgboost,pygam

References: - https://www.kaggle.com/floser/french-motor-claims-datasets-fremtpl2freq - https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.TweedieRegressor.html

Project Notebooks

Notebook	Rendered	Description	Author
a01_data_cleaning.ipynb	ipynb, rendered	ohe, kbin, logscaling	Bhishan Poudel
b01_freq_modelling.ipynb	ipynb, rendered	Poisson	Bhishan Poudel
b02_severity_modelling.ipynb	ipynb, rendered	Gamma	Bhishan Poudel
b03_pure_premium_modelling.ipynb	ipynb, rendered	Poisson*Gamma and Tweedie	Bhishan Poudel
b04_tweedie_vs_freqSev.ipynb	ipynb, rendered	comparison	Bhishan Poudel
b05_lorentz_curves_comparison.ipynb	ipynb, rendered	Lorentz Curve	Bhishan Poudel
c01_xgboost_tweedie.ipynb	ipynb, rendered	'objective':'reg:tweedie'	Bhishan Poudel
d01_gam_linear.ipynb	ipynb, rendered	n_splies=10, grid_search	Bhishan Poudel

Data

openml french motor freq (Multiple features.)
openml french motor severity (Two features Id policy and Claim Amount.)

Data Cleaning

Some of the features are chosen for modelling.

one hot encoding = ["VehBrand", "VehPower", "VehGas", "Region", "Area"]
kbins discretizer = ["VehAge", "DrivAge"]
log and scaling = ["Density"]
pass through =  ["BonusMalus"]

Results

Module	Distribution	y_train	sample_weight	train D2	test D2	train MAE	test MSE	train MAE	test MSE
sklearn	Frequency Modelling (Poisson Distribution)	df_train['Frequency']	df_train['Exposure']	0.051384	0.048138	0.232085	0.224547	4.738399	2.407906
sklearn	Severity Modelling (Gamma Distribution)	df_train.loc[mask_train, 'AvgClaimAmount']	df_train.loc[mask_train, 'ClaimNb']	-	3.638157e-03	-4.747382e-04	1.859814e+03	1.856312e+03	4.959565e+06
sklearn	Pure Premium Modelling (TweedieRegressor)	df_train['PurePremium']	df_train['Exposure']	2.018645e-02	1.353285e-02	6.580440e+02	4.927505e+02	1.478259e+09	1.622053e+08
xgboost	Xgboost Tweedie Regression	dtrain.set_base_margin(np.log(df_train['Exposure'].to_numpy())	dtest.set_base_margin(np.log(df_test['Exposure'].to_numpy()))	-	-	1.760538e+03	1.588351e+03	1.481952e+09	1.659363e+08
pygam	GAM Linear Model	df_train["AvgClaimAmount"].values	N/A	-	-	1.686438e+02	1.655408e+02	1.785332e+06	1.647533e+06