Publicação
What is the best automatic forecasting method to deal with high frequency data and little human intervention?
| datacite.subject.fos | Ciências Sociais::Economia e Gestão | pt_PT |
| dc.contributor.advisor | Fernandes, Pedro Afonso | |
| dc.contributor.author | Meneses, Mário Pedro Sobrinho Alves Ruber de | |
| dc.date.accessioned | 2023-07-04T14:31:03Z | |
| dc.date.available | 2023-07-04T14:31:03Z | |
| dc.date.issued | 2023-01-26 | |
| dc.date.submitted | 2023-01 | |
| dc.description.abstract | The aim of this thesis is to study which automatic forecasting method, between the Theta and Prophet models, can return the most appropriate predictions about the future, by computing two different performance metrics and with the use of three different ways of an out-of-sample cross-validation to successfully train the models. Hence, it seems that Prophet has more difficulty to capture good predictions when working with certain nature of data when compared to the Theta model. Another relevant note is that the out-of-sample approach using the growing window and one-fold type of cross-validation seems to perform better, when looking to all series used is this research. In addition, the Prophet model also has a higher computational cost in terms of time to make its predictions. Further, it should also be noted that the forecasts, in this analysis, regarding markets prices and the market yield are computed to calculate their expected value, however it is very hard to estimate this. So, the justification for using this type of information is related to the fact that it is more common to find daily values in this context and the models work better with daily observations. This exercise was attempted based on easily available series and with merely illustrative purposes, in order to make a comparative evaluation of the two models. For these reasons, the Theta model is the best automatic forecasting method in terms of results and computational costs when compared to the Prophet model. | pt_PT |
| dc.description.abstract | O objetivo desta tese ´e analisar qual ´e o m´etodo de previs˜ao autom´atica, entre o modelo Theta e o Prophet, que tem a potencialidade de obter as melhores previs˜oes e fornecer as visualiza¸c˜oes mais ver´ıdicas em rela¸c˜ao ao futuro, com o aux´ılio de dois indicadores de performance, como tamb´em da utiliza¸c˜ao de diferentes processos de valida¸c˜ao cruzada para treinar os modelos com sucesso. Deste modo, o modelo Prophet aparenta ter dificuldades em gerar boas previs˜oes quando trabalha com dados de diferentes tipos de natureza. Adicionalmente, o tipo de valida¸c˜ao cruzada com o aux´ılio de uma janela crescente e com um fold tem um melhor desempenho, comparativamente com os outros tipos. Al´em disso, o Prophet possui um custo computacional maior em termos de tempo para realizar este processo. E relevante ressalvar que as previs˜oes, nesta an´alise, relativas a pre¸cos e yield ´ de mercado s˜ao calculados com o intituito de estimar os seus valores esperados, no entanto ´e bastante dif´ıcil realizar essa opera¸c˜ao. A justifica¸c˜ao para o uso deste tipo de informa¸c˜ao deve-se ao facto de ser mais comum encontrar valores di´arios no contexto do tipo de pesquisa elaborada, como tamb´em o melhor desempenho dos modelos com observa¸c˜oes di´arias. Deste modo, tentou-se realizar esse exerc´ıcio com base em s´eries facilmente dispon´ıveis e com prop´ositos meramente ilustrativos, de modo a se fazer uma avalia¸c˜ao comparada dos dois modelos. Assim, o modelo Theta ´e o melhor m´etodo de previs˜ao autom´atica em termos de resultados e custos computacionais quando comparado ao Prophet. | pt_PT |
| dc.identifier.tid | 203252837 | pt_PT |
| dc.identifier.uri | http://hdl.handle.net/10400.14/41544 | |
| dc.language.iso | eng | pt_PT |
| dc.subject | Time series | pt_PT |
| dc.subject | Forecasting at scale | pt_PT |
| dc.subject | Automatic forecasting | pt_PT |
| dc.subject | Theta model | pt_PT |
| dc.subject | Prophet model | pt_PT |
| dc.title | What is the best automatic forecasting method to deal with high frequency data and little human intervention? | pt_PT |
| dc.type | master thesis | |
| dspace.entity.type | Publication | |
| rcaap.rights | openAccess | pt_PT |
| rcaap.type | masterThesis | pt_PT |
| thesis.degree.name | Mestrado em Análise de Dados para Gestão | pt_PT |