An Experimental Study For Transforming and Differencing Effects in Multiplicative Neuron Model Artificial Neural Network For Time Series Forecasting

Abstract

Forecasting problems play an important role in time series. In recent years, to solve these problems, many good alternative methods like artificial neural networks have been proposed in the literature. Although the most used artificial neural networks type is multilayer perceptron artificial neural networks, multiplicative neuron model artificial neural networks (MNM-ANNs) have been used to obtain forecasts for a few years. Many of previous studies were used to original series without any transformations such as differencing operation, Box-Cox transformations. Although stationary is an important assumption, previous studies have shown that forecasts obtained from ANNs were employed to non-stationary time series. Box-Cox transformations have been often used to time series because of heteroscedasticity. We used particle swarm optimization (PSO) and artificial bee colony (ABC) algorithms to train MNM-ANNs, and investigated differencing effects of original and transformed data which are obtained from Box-Cox transformations. Istanbul stock exchange (IEX) data sets which are made up of five time series for years between 2009 and 2013 are analyzed. The sets have comprised of first five months for these years. The results are interpreted and discussed. It is shown that transformation operations are useful for forecasting IEX as a result of statistical hypothesis tests.