A Hybrid Artificial Neural Network Model for Forecasting Short Time Series

Forecasting has long been the domain of traditional statistical models. Recent research has shown that novel and complex forecasting models do not necessarily outperform simpler models. These include in particular Artificial Neural Networks (ANNs). Even though claims of superior forecasting performa...

Full description

Bibliographic Details
Main Author: Mohan, Anil
Format: Dissertation (University of Nottingham only)
Language:English
Published: 2012
Online Access:https://eprints.nottingham.ac.uk/25828/
Description
Summary:Forecasting has long been the domain of traditional statistical models. Recent research has shown that novel and complex forecasting models do not necessarily outperform simpler models. These include in particular Artificial Neural Networks (ANNs). Even though claims of superior forecasting performance were made by Neural Network researchers, these claims were often unsubstantiated. Artificial neural networks are information processing paradigms motivated by the information processing functions of the human brain. ANNs are widely recognized as universal function approximators and are capable of exploiting nonlinear relationships between variables. Given these strengths, we believed it was possible to design a neural network that would provide excellent forecasting ability over a wide variety of data. Inspired by recent research into deep learning nets, we were able to model a new Hybrid ANN model and compared its performance to other forecasting models used in the M3 Time Series Competition. The results show that on average the Hybrid model outperforms the other methods investigated and