Abstract:
The objective of this research is to study and compare the performance of
forecasting models using Machine Learning techniques. In this research, four time
series models are selected: 1) Recurrent Neural Network (RNN), 2) Support Vector
Regression (SVR), 3) Multi-Layer Perceptron (MLP), and 4) Seasonal Autoregressive
Integrated Moving Average with Exogenous Variable Model (SARIMAX). The research
also presents the development of three hybrid forecasting models: 1) RNN-SARIMAX,
2) SVR-SARIMAX, and 3) MLP-SARIMAX. These models are applied to monthly direct
premium income data from life insurance businesses in Thailand, spanning from
January 2003 to December 2022. For evaluating the performance of the models,
three criteria are used: Mean Absolute Percentage Error (MAPE), Root Mean Square
Error (RMSE), and Coefficient of Determination (R-Squared). The data is divided into
two parts: Part 1 includes data from January 2003 to December 2022 for model
training, while Part 2 includes data from January 2023 to December 2023 for model
comparison. The aim is to identify the most suitable and accurate model for
forecasting the growth of life insurance businesses in Thailand.