| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 7.05 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Num mundo tecnológico cada vez mais avançado a inflação é um problema sério que afeta todos. Existem algumas soluções para contornar este problema, sendo uma delas o investimento em mercados de ações. O mercado de ações permite a empresas angariarem capital e fornece aos investidores a possibilidade de gerar lucros. Estes mercados tendem a ser algo complexo de analisar e prever, pelo que, têm sido alvo de vários estudos. Nesta dissertação foram analisados vários métodos de previsão de séries temporais financeiras,
desde métodos estatísticos a métodos de aprendizagem automática. Foram analisados e implementados modelos da família ARMA e GARCH, assim como modelos de aprendizagem automática e modelos híbridos, para diferentes valores temporais de previsão. Estes modelos foram avaliados com quatro séries temporais distintas, CPIAUCSL, G17MVSFAUTOS, SHEL e LYG. Para CPIAUCSL, para um horizonte de 3 e 10 ARIMA é o melhor modelo, mas para um horizonte de 20 ML (SVR) é a melhor solução. Para a série temporal G17MVSFAUTOS o modelo com maior desempenho para os horizontes de 3 e 10 foi ML (SVR), enquanto que para o horizonte de 20 foi ML (KNR). Para a SHEL com um horizonte de 3 dias, o modelo ML (SVR) com uma janela de 10 apresenta desempenho superior, mas para os horizontes 10 e 20 o modelo ML (SVR) com
uma janela de 5 apresenta melhores resultados. Para a série temporal LYG com um horizonte de 3, o modelo ML (SVR) com uma janela de 5 apresenta bons resultados, para o horizonte de 10, ML (RFR) de janela 5 foi o melhor e para o horizonte de 20, os resultados são mais ambíguos, sendo que ML (KNR) de janela 5, ML (RFR) de janela 5 e ML (SVR) de janela 10 apresentam bons resultados.
In an increasingly advanced technological world, inflation is a serious problem that affects everyone. There are some solutions to overcome this problem, one of which is investing in stock markets. The stock market allows companies to raise capital and provides investors with the possibility of generating profits. These markets tend to be complex to analyze and predict, which is why they have been the subject of several studies. In this dissertation, various financial time series forecasting methods were analyzed, from statistical methods to machine learning methods. Models from the ARMA and GARCH family were analyzed and implemented, as well as machine learning models and hybrid models, for different temporal forecast values. These models were evaluated with four different time series, CPIAUCSL, G17MVSFAUTOS, SHEL and LYG. For CPIAUCSL, for a horizon of 3 and 10 ARIMA is the best model, but for a horizon of 20, ML (SVR) is the best solution. For the G17MVSFAUTOS time series, the model with the highest performance for horizons 3 and 10 was ML (SVR), while for horizon 20 it was ML (KNR). For SHEL with a horizon of 3 days, the ML model (SVR) with a window of 10 presents superior performance, but for horizons 10 and 20 the ML model (SVR) with a window of 5 presents better results. For the LYG time series with a horizon of 3, the ML (SVR) model with a window of 5 shows good results, for the horizon of 10, ML (RFR) with window 5 was the best and for the horizon of 20, the results are more ambiguous, with ML (KNR) with window 5, ML (RFR) with window 5 and ML (SVR) with window 10 showing good results.
In an increasingly advanced technological world, inflation is a serious problem that affects everyone. There are some solutions to overcome this problem, one of which is investing in stock markets. The stock market allows companies to raise capital and provides investors with the possibility of generating profits. These markets tend to be complex to analyze and predict, which is why they have been the subject of several studies. In this dissertation, various financial time series forecasting methods were analyzed, from statistical methods to machine learning methods. Models from the ARMA and GARCH family were analyzed and implemented, as well as machine learning models and hybrid models, for different temporal forecast values. These models were evaluated with four different time series, CPIAUCSL, G17MVSFAUTOS, SHEL and LYG. For CPIAUCSL, for a horizon of 3 and 10 ARIMA is the best model, but for a horizon of 20, ML (SVR) is the best solution. For the G17MVSFAUTOS time series, the model with the highest performance for horizons 3 and 10 was ML (SVR), while for horizon 20 it was ML (KNR). For SHEL with a horizon of 3 days, the ML model (SVR) with a window of 10 presents superior performance, but for horizons 10 and 20 the ML model (SVR) with a window of 5 presents better results. For the LYG time series with a horizon of 3, the ML (SVR) model with a window of 5 shows good results, for the horizon of 10, ML (RFR) with window 5 was the best and for the horizon of 20, the results are more ambiguous, with ML (KNR) with window 5, ML (RFR) with window 5 and ML (SVR) with window 10 showing good results.
Description
Keywords
Machine learning Financial time series Forecasting Aprendizagem automática Séries temporais financeiras Previsão