Application of neural network autoregression (NNAR) and ARIMA-GARCH based on interpolation for forecasting direct economic losses of earthquake in Indonesia

Artificial Neural network are commonly used for time series forecasting, especially in financial forecasting, and when external information is useful. The aim of this study is to forecast direct economic losses of Earthquake in Indonesia using NNAR and ARIMA models and selects a model that produces a forecast with a minimum root mean square error (RMSE). The accurate prediction of the direct economic losses of earthquake is critical for allowing policy makers to take a decision to anticipate losses in the future. For the purpose of forecasting accurately, an additional method is used, namely GARCH method for resolving the white noise problem in residual, such as heteroscedasticity, autocorrelation and normality. In this paper, two interpolations are applied to expand the original small sample with virtual points. The data of direct economics losses in Indonesia during the period from 1989 to 2021 were collected from EM-DAT (The International Disaster Database). The data partitioned into training and validation periods, such that the last 5 years are the validation period. The software used for ARIMA-GARCH is EViews 10 and NNAR accomplished by forecast package in R software. Based on the analysis, obtained that the result of this study is the neural network autoregression in the training period has an RMSE of 2.412, compared to ARIMA-GARCH 9,197. The neural network autoregression is significantly worse in the validation period. Its RMSE is 24,66 versus ARIMA-GARCH 8,4. According to the RMSE value, we can notice that the ARIMA-GARCH method outperform the NNAR (4,1,20) [4] model for validation period and to forecast the direct economic losses of Earthquake in Indonesia.