| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 729.81 KB | Adobe PDF | |||
| 262.94 KB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
A previsão de séries temporais relatvas a mercados fnanceiros é tda como uma tarefa desafante.
Prever o movimento e direção das séries pode ser mais lucratvo que o prever preço, mas de forma a
maximizar o retorno pode interessar prever movimentos acentuados das séries. Fazendo uso de 18
séries de ações provenientes de três mercados distntos e para um horizonte de previsão de 4
semanas, este estudo testa três hipóteses para avaliar a pertnência da previsão de movimentos
acentuados, considerando a existência, ou não, de dependência na dimensão tempo e também a
potencial alteração das relações entre as variáveis dependentes e independentes com a dimensão
tempo. Para cada série são testadas, as três hipóteses, com sete modelos de aprendizagem
automátca e selecionado o melhor para realizar as previsões através de uma metodologia de
otmização automátca, que seleciona de entre 103 variáveis possrveis para realizar as previsões, num
problema com classes desequilibradas. Este estudo atnge resultados compettvos com a literatura,
obtendo-se com a melhor metodologia uma probabilidade de acertar na direção do movimento da
série de 80% e num movimento acentuado de 65% das vezes.
Forecastng tme series on fnancial markets is regarded as a challenging task. Predictng the movement and directon of the series may be more proftable than the price forecast, but in order to maximize the return it may be interestng to predict the series sharp movements. Using 18 series of stocks from three diferent markets and for a forecast horizon of 4 weeks, this study tests three hypotheses to evaluate the pertnence of predicton of sharp movements, considering the existence or not of dependence in the tme dimension and also the potental change of the relatonships between the dependent and independent variables with the tme dimension. For each tested series, the three hypothesis, with seven machine learning models are tested and the best is selected to carry out the predictons through an automatc optmizaton methodology, which selects from 103 possible variables to do the predictons, in a problem with unbalanced classes. This study achieves compettve results with the literature, with the best methodology being likely to hit in the directon of the movement of the series in 80% of the cases and the sharp movements in 65% of tmes.
Forecastng tme series on fnancial markets is regarded as a challenging task. Predictng the movement and directon of the series may be more proftable than the price forecast, but in order to maximize the return it may be interestng to predict the series sharp movements. Using 18 series of stocks from three diferent markets and for a forecast horizon of 4 weeks, this study tests three hypotheses to evaluate the pertnence of predicton of sharp movements, considering the existence or not of dependence in the tme dimension and also the potental change of the relatonships between the dependent and independent variables with the tme dimension. For each tested series, the three hypothesis, with seven machine learning models are tested and the best is selected to carry out the predictons through an automatc optmizaton methodology, which selects from 103 possible variables to do the predictons, in a problem with unbalanced classes. This study achieves compettve results with the literature, with the best methodology being likely to hit in the directon of the movement of the series in 80% of the cases and the sharp movements in 65% of tmes.
Descrição
Project Work presented as the partial requirement for obtaining a Master's degree in Information Management, specialization in Knowledge Management and Business Intelligence
Palavras-chave
Previsão movimento ações Classes desequilibradas Seleção variáveis Aprendizagem automátca Forecast stock movement Unbalanced classes Feature selecton Machine learning Concept drif
