Loading...
Publication
Variabilidade e quebras de estruturas em séries temporais: comparação de métodos e aplicação a séries económico-financeiras
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 6.33 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Nesta dissertação, exploramos a análise de séries temporais financeiras, focando em métodos clássicos e mais recentes. Investigámos a deteção de pontos de mudança e a previsão de volatilidade em séries como a taxa de juro SELIC, os preços do ouro, os ETFs (fundos negociados em bolsa) e as criptomoedas.
Realizámos uma análise de pontos de mudança, utilizando os métodos PELT (Pruned Exact Linear Time) e SONDE (Self Organizing Neural Network for Detecting Novelties). Foram detetadas diversas rupturas na estrutura da série, consistentes com eventos de mercado conhecidos como a crise financeira de 2008 e a pandemia de COVID-19. Utilizámos modelos como SARIMA, Filtro de Kalman e GARCH, diretamente sobre as séries e sobre as suas componentes obtidas pela decomposição por modo empírico (EMD).
Observámos que as séries de volatilidade financeira apresentam diversos desafios na aplicação de um único método a toda a série, entre outros motivos pela quantidade de pontos de mudança, no entanto uma abordagem em janela móvel pode gerar resultados satisfatórios.
Chama-nos a atenção o mau desempenho do algoritmo GARCH em relação ao SARIMA e ao Filtro de Kalman na nossa abordagem.
Discutimos a aplicação destes resultados em séries temporais reais do mercado financeiro e as suas aplicações práticas neste mesmo contexto, tais como a gestão de carteiras e a gestão de relações com clientes. A investigação sugere que há muito a explorar nesta área dinâmica e desafiante.
In this dissertation, we explore the analysis of financial time series, focusing on both classical and more recent methods. We investigate the detection of changepoints and volatility forecasting in series such as the SELIC interest rate, gold prices, ETFs (exchange-traded funds), and cryptocurrencies. We performed a changepoint analysis using the PELT (Pruned Exact Linear Time) and SONDE (Self Organizing Neural Network for Detecting Novelties) methods. Several structural breaks in the series were detected, consistent with known market events such as the 2008 financial crisis and the COVID-19 pandemic. We employed models such as SARIMA, Kalman filter, and GARCH, directly on the series and on their components obtained by empirical mode decomposition (EMD). We observed that financial volatility series present several challenges in applying a single method to the entire series, due in part to the number of changepoints. However, a rolling window approach can yield satisfactory results. We were particularly struck by the poor performance of the GARCH algorithm compared to SARIMA and the Kalman filter in our approach. We discuss the application of these results to real-world financial time series and their practical applications in this context, such as portfolio management and customer relationship management. The research suggests that there is much to explore in this dynamic and challenging field.
In this dissertation, we explore the analysis of financial time series, focusing on both classical and more recent methods. We investigate the detection of changepoints and volatility forecasting in series such as the SELIC interest rate, gold prices, ETFs (exchange-traded funds), and cryptocurrencies. We performed a changepoint analysis using the PELT (Pruned Exact Linear Time) and SONDE (Self Organizing Neural Network for Detecting Novelties) methods. Several structural breaks in the series were detected, consistent with known market events such as the 2008 financial crisis and the COVID-19 pandemic. We employed models such as SARIMA, Kalman filter, and GARCH, directly on the series and on their components obtained by empirical mode decomposition (EMD). We observed that financial volatility series present several challenges in applying a single method to the entire series, due in part to the number of changepoints. However, a rolling window approach can yield satisfactory results. We were particularly struck by the poor performance of the GARCH algorithm compared to SARIMA and the Kalman filter in our approach. We discuss the application of these results to real-world financial time series and their practical applications in this context, such as portfolio management and customer relationship management. The research suggests that there is much to explore in this dynamic and challenging field.
Description
Dissertação de Mestrado em Estatística, Matemática e Computação, apresentada à Universidade Aberta
Keywords
Análise financeira Taxa de juro Análise estatística multivariada