A carregar...
Publicação
Volatility forecasting with machine learning and application to volatility swaps
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 1.9 MB | Adobe PDF |
Orientador(es)
Resumo(s)
Esta dissertação oferece novos contributos obre o uso de modelos de Machine Learning, especificamente o algoritmo XGBoost, para previsão de volatilidade realizada e sua aplicação em estratégias de investimento. O estudo concentra-se no índice SP500, de 1990 a 2024, e simula uma estratégia de investimento com swaps de volatilidade entre 2018 e 2024 com base nas previsões obtidas. Realiza-se uma análise para testar qual combinação de modelo de Machine Learning e variáveis apresenta melhor desempenho na tarefa. Os modelos avaliados incluem XGBoost, Florestas Aleatórias (Random Forests) e Redes Neuronais (Artificial Neural Networks), enquanto as variáveis testadas são as Funções Modais Intrínsecas (IMFs - Intrissic Mode Functions) resultantes da decomposição CEEMDAN e variáveis de mercado dos Estados Unidos. O algoritmo XGBoost, combinando IMFs e variáveis de mercado, destaca-se como o modelo de ML mais eficiente no estudo, tendo também melhor performance que dois “naïve models” diferentes usados como referência. A aplicabilidade das previsões é testada por meio da simulação de um investimento em swaps de volatilidade, que, com suposições realistas e diversos cenários, demonstra resultados encorajadores em termos de retornos, no entanto apresentado uma volatilidade extremamente alta.
This dissertation provides novel insights on the use of Machine Learning models, particularly the XGBoost algorithm, for realized volatility forecasting and its employment in investment strategies. The study focuses on the SP500 Index, from 1990 to 2024. A trading strategy using volatility swaps is simulated from 2018 to 2024 using the forecast results. We perform an analysis to test which combination of machine learning model and inputs perform better at the task at hand. The models studied are the XGBoost, Random Forests and Neural Networks, and the variable inputs the Intrinsic Mode Functions (IMFs) resulting from CEEMDAN decomposition, and market variables from the United States. The XGboost algorithm with both the IMFs and market variables is the best performing ML model out of the study made, while also outperforming two different naïve models used as benchmarks. We demonstrate the applicability of the forecasts through simulating a volatility swap investment that under realistic circumstances and various test scenarios is able to produce encouraging payouts, although with extremely high payoff volatility.
This dissertation provides novel insights on the use of Machine Learning models, particularly the XGBoost algorithm, for realized volatility forecasting and its employment in investment strategies. The study focuses on the SP500 Index, from 1990 to 2024. A trading strategy using volatility swaps is simulated from 2018 to 2024 using the forecast results. We perform an analysis to test which combination of machine learning model and inputs perform better at the task at hand. The models studied are the XGBoost, Random Forests and Neural Networks, and the variable inputs the Intrinsic Mode Functions (IMFs) resulting from CEEMDAN decomposition, and market variables from the United States. The XGboost algorithm with both the IMFs and market variables is the best performing ML model out of the study made, while also outperforming two different naïve models used as benchmarks. We demonstrate the applicability of the forecasts through simulating a volatility swap investment that under realistic circumstances and various test scenarios is able to produce encouraging payouts, although with extremely high payoff volatility.
Descrição
Trabalho Final de Mestrado, Finance, ISEG, 2025
Palavras-chave
Volatility Forecasting XGBoost Random Forests Neural Networks Volatility Swaps Previsão de Volatilidade XGBoost Florestas Aleatórias Redes Neurais Swaps de Volatilidade