Comunicação em evento científico
Forecasting financial time series: a comparative study
Anabela Costa (Costa, A.); Vivaldo Mendes (Mendes, V.); Filipe Ramos (Ramos, F.R.); Vivaldo Mendes (Mendes, V.);
Título Evento
IO2018 - XIX Congresso da APDIO 2018
Ano (publicação definitiva)
2018
Língua
Inglês
País
Portugal
Mais Informação
Web of Science®

Esta publicação não está indexada na Web of Science®

Scopus

Esta publicação não está indexada na Scopus

Google Scholar

N.º de citações: 6

(Última verificação: 2024-03-25 13:18)

Ver o registo no Google Scholar

Abstract/Resumo
The main purpose of this paper it is to show that machine learning methods (neural networks and k-nearest neighbours) can be used to uncover the non-linearity that exists in financial time series and provide high quality forecast. First, we analyse the linearity (BDS test) and stationarity (ADF, PP unit rot test) of the Portuguese stock market index, PSI20, and also some typical features are studied (descriptive statistics, Hurst exponents, among others). The first forecast it is provided by traditional linear ARMA models. Secondly, we train several types of neural networks for the PSI20 index and use the models to make 1 and 5-day forecasts. The artificial neural networks are obtained by using a three-layer feed-forward topology and the back-propagation learning algorithm. Thirdly, k-nearest neighbours chartist method it is used. Finally, we compare the out-of-sample forecast error (MAE) for the several models, in order to conclude about the forecasting performance.
Agradecimentos/Acknowledgements
--
Palavras-chave
Time-series,ARMA,Artificial Neural Network,Forecasting