Forecasting Performance Of Time Series And Regression Models With Neural Network Models

Loading...
Thumbnail Image
Date
2016-10
Authors
Davou, Pwasong Augustine
Journal Title
Journal ISSN
Volume Title
Publisher
Universiti Sains Malaysia
Abstract
In this study, ten different sets of daily time series data of over three thousand observations for each set was modeled and forecasted using three linear techniques and four nonlinear techniques. The 3-LM in focused here are the quadratic regression (QRM) method,the random walk with drift method and the autoregressive integrated moving average (ARIMA) technique. The 4-NLM are the cascade forward backpropa- gation neural network (CFBNN) method, the layer recurrent neural network (LRNN) method; the hybrid quadratic regression method and cascade forward backpropagation neural network (QRM −CFBNN) method and the hybrid LRNN and ARIMA tech- nique. The proposed new hybrid QRM −CFBNN technique is possessed by three striking characteristics that differentiate it from any existing hybrid technique currently in the modeling and forecasting literature. Firstly, the quadratic regression model QRM is fitted using the time series data such that the residuals emerging after the fitting are considered as the target variable in the joint cascade forward backpropagation neu- ral network CFBNN. Secondly, the pooled final forecasting product is a weighted average facilitated through the application of the Bayesian model averaging (BMA) technique. Thirdly, the minimum mean square error (MMSE) and the minimum mean absolute error (MMAE) are objective functions, which are mutually castoff to train neural networks in parallel. The new hybrid QRM −CFBNN was trained and tested on the increment series of the ten different data sets of the univariate daily time series data. The increment series of the original time series data in this perspective refers to the log difference series of the data. The increment series is applied to conduct the Dickey-Fuller (DF) test to determine if the data series is stationary. The results obtained from the new hybrid QRM −CFBNN method were compared with the re- sults obtained from the hybrid LRNN−ARIMA technique, standalone cascade forward backpropagation neural network (CFBNN) technique, standalone layered recurrent neural network (LRNN) technique and standalone autoregressive integrated moving average (ARIMA) technique as well as standalone random walk with drift technique respectively. The comparison indicates that the results emerging from the new hybrid QRM−CFBNN method on the average, generally displayed a better forecasting per- formance when compared with the hybrid LRNN −ARIMA technique, the standalone CFBNN technique, standalone random walk with drift technique and the standalone LRNN technique. The RMSEs, MAEs, MPE and the MAPE were applied to ascer- tain the assertion that the new jointly integrated forecast has better forecasting perfor- mance greater than the forecasts of the standalone CFBNN method, standalone LRNN method, standalone autoregressive integrated moving average (ARIMA) method and standalone random walk with drift technique as well as the forecast of the combined ARIMA−RNN technique.
Description
Keywords
Forecasting performance of time Series , and regression models
Citation