Home > Research > Publications & Outputs > Feature selection of autoregressive Neural Netw...

Text available via DOI:

View graph of relations

Feature selection of autoregressive Neural Network inputs for trend Time Series Forecasting

Research output: Contribution in Book/Report/Proceedings - With ISBN/ISSNConference contribution/Paperpeer-review

Published
Publication date31/10/2016
Host publication2016 International Joint Conference on Neural Networks, IJCNN 2016
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages1515-1522
Number of pages8
ISBN (electronic)9781509006199
<mark>Original language</mark>English
Externally publishedYes
Event2016 International Joint Conference on Neural Networks, IJCNN 2016 - Vancouver, Canada
Duration: 24/07/201629/07/2016

Conference

Conference2016 International Joint Conference on Neural Networks, IJCNN 2016
Country/TerritoryCanada
CityVancouver
Period24/07/1629/07/16

Publication series

NameProceedings of the International Joint Conference on Neural Networks
Volume2016-October

Conference

Conference2016 International Joint Conference on Neural Networks, IJCNN 2016
Country/TerritoryCanada
CityVancouver
Period24/07/1629/07/16

Abstract

The capability of artificial Neural Networks to forecast time series with trends has been a topic of dispute. While selected research following Zhang and Qi has indicated that prior removal of trends is required for a Multilayer Perceptron (MLP), others provide evidence that Neural Networks are capable of forecasting trends without data preprocessing, either by choosing input-nodes employing an adequate autoregressive lag-structure of lagged realisations or by adding explanatory variables with trends. This paper proposes a novel variable selection methodology of autoregressive lags for trended time series with and without seasonality, and assesses its efficacy using the dataset of the International Time Series Forecasting Competition conducted at WCCI 2016. Our experiments indicate that MLPs are capable of forecasting different trend forms, but that more than a single lag-structure is required to do so, making the use of multiple input-lag variants and a robust model selection strategy necessary to achieve robust forecast accuracy.

Bibliographic note

Publisher Copyright: © 2016 IEEE. Copyright: Copyright 2017 Elsevier B.V., All rights reserved.