PARAMETERS ESTIMATON OF ROBUST VARIOGRAM ON PROPERTY INSURANCE IN BANDUNG USING GAUSS- NEWTON METHOD AND GENETIC ALGORITHM

• Main Author: Fardiaz Kuswanda, Giraldi
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/46573
• Institution: Institut Teknologi Bandung

Variogram analysis is a method that widely used in various fields to describe the nature of spatial correlation of data between locations. One of interesting observation is the distribution of claim values of property insurance based on the location. In general, the claim values of property insurance is a highly skewed data, so natural logarithmic transformations need to be done to reduce the skewness and variance of the data. Besides that, a robust variogram approach is needed in the modeling. One important step in modeling a variogram is parameter estimation. Unfortunately, simultaneous parameter estimation is not easy to do because of the nonlinearity of variogram model functions . In this thesis the Gauss-Newton method and genetic algorithm are used as a method of estimating parameters that can estimate variogram parameters simultaneously. The Gauss-Newton method provides a parameter estimate quickly, a maximum of 51 iterations, but the results are possible to get out of bond of variogram parameters. For example, there is negative number for estimated nugget value and this is out of bond because nugget is a measure of variability that is not naturally negative. In genetic algorithms we can limit the value of estimated parameters so this method is more reliable. However, one of the difficulties of genetic algorithms is to determine the stopping criteria based on the mean square error (MSE) between the variogram model and the experimental variogram. Both methods can estimate variogram parameters well, based on MSE values that are not significantly different, with a level of accuracy of 10?3 . Then validation of the best model is using the Jackknife Kriging method, and the spherical model that estimated by Gauss-Newton method was selected as the best model with the root mean square error (RMSE) of 58.4 million rupiah.