Data pre-processing for neural network-based forecasting: does it really matter?

This study aims to analyze the effects of data pre-processing on the forecasting performance of neural network models. We use three different Artificial Neural Networks techniques to predict tourist demand: multi-layer perceptron, radial basis function and the Elman neural networks. The structure of...

Descripción completa

Detalles Bibliográficos
Autores: Claveria, Oscar, Monte Moreno, Enrique|||0000-0002-4907-0494, Torra Porras, Salvador
Tipo de recurso: artículo
Fecha de publicación:2015
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/81362
Acceso en línea:https://hdl.handle.net/2117/81362
https://dx.doi.org/10.3846/20294913.2015.1070772
Access Level:acceso abierto
Palabra clave:Forecasting
Neural networks (Computer science)
Artificial neural networks
Multiple-input multiple-output (MIMO)
Seasonality
Detrending
Tourism demand
Multilayer perceptron
Radial basis function
Elman
Previsió
Xarxes neuronals (Informàtica)
Àrees temàtiques de la UPC::Economia i organització d'empreses
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial
Descripción
Sumario:This study aims to analyze the effects of data pre-processing on the forecasting performance of neural network models. We use three different Artificial Neural Networks techniques to predict tourist demand: multi-layer perceptron, radial basis function and the Elman neural networks. The structure of the networks is based on a multiple-input multiple-output (MIMO) approach. We use official statistical data of inbound international tourism demand to Catalonia (Spain) and compare the forecasting accuracy of four processing methods for the input vector of the networks: levels, growth rates, seasonally adjusted levels and seasonally adjusted growth rates. When comparing the forecasting accuracy of the different inputs for each visitor market and for different forecasting horizons, we obtain significantly better forecasts with levels than with growth rates. We also find that seasonally adjusted series significantly improve the forecasting performance of the networks, which hints at the significance of deseasonalizing the time series when using neural networks with forecasting purposes. These results reveal that, when using seasonal data, neural networks performance can be significantly improved by working directly with seasonally adjusted levels.