A New Bootstrapped Hybrid Artificial Neural Network Approach for Time Series Forecasting
- PDF / 1,498,342 Bytes
- 29 Pages / 439.37 x 666.142 pts Page_size
- 67 Downloads / 178 Views
A New Bootstrapped Hybrid Artificial Neural Network Approach for Time Series Forecasting Erol Eğrioğlu1,2 · Robert Fildes2 Accepted: 2 November 2020 © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract In this study, a new bootstrapped hybrid artificial neural network is proposed for forecasting. This new neural network provides input significance, linearity and nonlinearity hypothesis tests in a unique network structure via a residual bootstrap approach. The network has three parts: linear, non-linear and a combination with associated weights and biases. These weights are used to test the input significance, linearity and nonlinearity hypotheses with this new method providing empirical distributions for forecasts and weights. The proposed method employs a bagging approach to obtain forecasts. It is then applied to real-time series including the M4 Competition data set and stock exchange time series where its performance is compared with appropriate benchmark methods including other popular neural networks. The proposed method results are less affected than other neural networks by initial random weights, which means that the results of the proposed method are more stable and precise. The new method provides improvements in forecasting accuracy over the established benchmarks. Keywords Artificial neural networks · Deep learning · Forecasting · Input significance · Interval forecast · Bootstrap
* Erol Eğrioğlu [email protected] Robert Fildes [email protected] 1
Department of Statistics, Faculty of Arts and Science, Giresun University, 28200 Giresun, Turkey
2
Department of Management Science, Management Science School, Marketing Analytics and Forecasting Research Center, Lancaster University, Lancaster, UK
13
Vol.:(0123456789)
E. Eğrioğlu, R. Fildes
1 Introduction Artificial neural networks (ANNs) can be used to obtain forecasts for linear or nonlinear time series. Many types of artificial neural networks have been proposed in the literature. The findings of studies about the performance of ANNs for forecasting purpose vary from study to study. There is no consensus about the reasons behind the success or failure of ANNs performance on forecasting problem. In early research, Gorr et al. (1994) stated that ANN can (1) automatically transform and represent complex and highly non-linear relationships and (2) automatically detect different states of phenomena through independently variable data patterns and switch on/off model components as appropriate. Besides these good properties, Gorr et al. (1994) emphasised that ANNs have several limitations, mostly noticeable in ‘explanation research’ (causal modelling and hypothesis testing) but not when used in forecasting. This occurs because ANN models are non-linear in the model coefficients and the normal probability models are not applicable. As a result of this, they do not have parametric statistical properties based on the t and F distributions. In this study, these problems are focused on and a new method proposed to s
Data Loading...