PURE PREMIUM CALCULATION OF MOTOR VEHICLE INSURANCE USING SPATIOTEMPORAL GENERALIZED GAUSSIAN PROCESS MODEL ON CLAIM FREQUENCY WITH TAYLOR APPROXIMATION

• Main Author: AGAM ISLAMI AL MUTAQIN, MUHAMMAD
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/64970
• Institution: Institut Teknologi Bandung

The risk of vehicle accidents is one type of risk that is high enough to occur every day. Losses caused by accidents can cause a very large nominal economic loss. Auto insurance is a solution to reduce losses caused by accidents. The risk of accidents must be properly quantified by the insurance company so that any losses due to accidents can be covered. The characteristics of an area can be a big influence in increasing the risk of accidents. Accidents can be affected by weather, traffic density, or road conditions in certain areas. By studying the area and time of past accidents, it is hoped that information can be obtained that for a similar condition in the future, the risk of accidents can be predicted. The measure used as risk quantification in this research is the claim rate modeled by Gamma-Generalized Gaussian Process Model (Gamma-GGPM). The Gaussian Process is used as the basic model in this research because it is very good at capturing similarity information using the kernel. There are 4 models that measure the similarity of different variables, namely spatial, temporal, spatiotemporal, and spatiotemporal with other factors (Population, Temperature, Wind, and Wind Gust). The inference used by GGPM uses an approximation of the result of the integration constraint on the posterior and marginal likelihood gains. Taylor approximation is used because it is a non-iterative approximation process that can significantly lighten the computational load. The best model selection is based on the best model prediction performance which is quantified using the Weighted Mean Squared Prediction Error (WMSPE) measure. From the simulation data, it is found that the model that considers spatiotemporal and other factors is the best model. The addition of information provides better model performance as a result of the many similar conditions recorded by the model.