Implementation of Moving Average Filter in SARIMA-ANN and SARIMA-SVR Methods for Forecasting Pneumonia Incidence in Jakarta

Abstract

In this study, we implemented a moving average filter in SARIMA-ANN and SARIMA-SVR to predict Pneumonia incidence in Jakarta. Pneumonia is one of the highest causes of death in children throughout the world. Forecasting pneumonia incidence in the future can help to reduce the spread of cases, so that the number of deaths due to pneumonia can be reduced. In general, time series data consists of linear and nonlinear patterns, which cannot be properly modeled by linear or nonlinear models alone. One way to solve this issue is to use a hybrid model that combines several models to overcome the limitations of each component model and improve predicting performance. SARIMA-ANN and SARIMA-SVR methods combine a linear seasonal autoregressive integrated moving average (SARIMA) model and a nonlinear artificial neural network (ANN) or support vector regression (SVR) model to capture the linear and nonlinear characteristics of the data. Parameter estimation in SARIMA uses Gaussian Maximum Likelihood Estimation. Initially, the time series will be transformed by a moving average (MA) filter, so SARIMA can model the data well. Meanwhile, the remaining components separated from the transformation will be modeled with a nonlinear model such as ANN in the SARIMA-ANN method, or SVR in the SARIMA-SVR method. The simulation results show that the SARIMA-ANN method is superior to the SARIMA-SVR method in predicting incidences in West Jakarta and East Jakarta, with a MAPE difference ranging from 0.6% to 0.75%. Meanwhile, in North, South, and Central Jakarta, the SARIMA-SVR method is superior to the SARIMA-ANN method, with MAPE differences ranging from 1.6% to 3.99%. The SARIMA-SVR model achieves better results across the majority of municipalities, indicating that the SARIMA-SVR model generally provides better result for predicting Pneumonia incidence in Jakarta.