Hybrid MODWT-FFNN model for time series data forecasting

Abstract

In this research, we propose a hybrid MODWT-FFNN model for non-stationary and Long-Range Dependence (LRD) of time series data. The hybrid MODWT-FFNN model is a combination of the Maximal Overlap Discrete Wavelet Transform (MODWT) and the Feed-Forward Neural Network (FFNN) models. The decomposition of time series data using the MODWT model will produce wavelet (detail) and scale (smooth) coefficients. The detail and smooth coefficients are then estimated using the FFNN model. The final result of time series data forecasting obtained from the combined forecast value of the detail and smooth coefficients. In the case study of daily Rainfall data in Aceh, the Root Mean Squared Error (RMSE) and Median Absolute Deviation (MAD) values obtained by our model are smaller than those of the ARIMA, exponential smoothing, and MODWT-ARMA models. The second case study of the Jakarta Stock Exchange Composite (JKSE) daily data, obtained the smallest RMSE and MAD values in the hybrid MODWT-ARMA model, the model we propose is in the second number. This indicates that the hybrid MODWT-FFNN model is useful for adding forecasting accuracy to seasonal data patterns.