To populate Scholarly, sign in here .


Title On Hybridization of Time Series and Bayesian Regularization Neural Network Models
Posted by Bernadette Tubo
Authors Payla, Jayson N. and Tubo, Bernadette F.
Publication date 2019
Conference 14th National Convention on Statistics (NCS)
Abstract This study introduces the Bayesian Regularization Neural Network (BRNN) and the Multilayer Perceptron Neural Network (MLPNN) as hybridization methodologies to do modelling Time Series (TS) data. BRNN is an excellent approximation to an ideal Artificial Neural Network in a way that it modifies the standard back-propagation neural network by a regularization step incorporating Bayesian statistics. The procedure is applied to model the Gross Domestic Product (GDP) of the Philippines which is a linear data for hybridization process via BRNN and MLPNN to unveil the comparability in the forecasting performance for short term and long term scenarios. The final model of the GDP is ARIMA (1,1,1) x (1,1,1)4 – BRNN (4-2-1) with Mean Absolute Percentage Error (MAPE) of 0.9941 for long term forecasting. In general, the result of hybridization for linear models indicates that SARIMA and ARIMA are good when we want short term forecast while the hybrid MLPNN and hybrid BRNN are better if we want long term forecast.
Index terms / Keywords ARIMA, SARIMA, hybrid models, artificial neural networks, Bayesian Inference, multilayer perceptron, time series forecasting, combined forecast