| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 905.8 KB | Adobe PDF |
Orientador(es)
Resumo(s)
Nesta dissertação, iremos explorar um pouco do universo que são os algoritmos Deep Learning, para
series temporais. Em particular, através do algoritmo LSTM (Long Short-Term Memory), foi
desenvolvido um modelo capaz de prever o S&P500, com resultados significativos, no período de
análise desde 1 de janeiro 2020 até 31 de dezembro de 2022.
A linguagem de programação Python foi escolhida para implementar o código, dado que disponibiliza
uma grande quantidade algoritmos pré-definidos em bibliotecas que facilitam tanto a escrita, como a
leitura e interpretação do código.
Além disto, foram escolhidos vários indicadores fundamentais, macroeconómicos e técnicos, com
expectativa de que o modelo LSTM consiga captar relações entre eles.
Após serem abordados os conceitos chave para compreensão da análise, exploramos de forma detalhada
a metodologia, raciocino e técnicas usadas no tratamento de dados, desenvolvimento, treino e tuning
dos hiperparâmetros do modelo LSTM.
De seguida, passamos à análise de sensibilidade, onde é avaliada a relevância de cada indicador e
parâmetro para o resultado do modelo, realizando ajustes caso necessário e por fim comparamos o
modelo final com o modelo Naive.
Concluímos esta dissertação respondendo a algumas questões e interpretando os resultados do algoritmo
do ponto de vista económico.
We will explore in this dissertation, a small part of the universe of Deep Learning algorithms for time series data. Using the LSTM (Long Short-Term Memory) algorithm, we will create a model capable of predicting the S&P500 with significant results. The data period analysed spans from January 1, 2020, to December 31, 2022. The Python programming language was chosen to implement the code because it provides a vast number of predefined algorithms in libraries make it easier, both code writing, reading and interpretation. We have selected several fundamental, macroeconomic, and technical indicators, with the expectation that the LSTM model could capture relationships between them. Frist we will address the key concepts necessary for understanding the analysis and them explore in detail the methodology, reasoning, and techniques used in data preprocessing, model training, and tuning of LSTM model hyperparameters. We then proceed to the sensitivity analysis, where we assess the relevance of each indicator and parameter to the model's outcome, adjusting if necessary and finally, we compare our final model to the Naive model. Then we conclude this dissertation by answering some questions and interpreting the model results from an economic perspective.
We will explore in this dissertation, a small part of the universe of Deep Learning algorithms for time series data. Using the LSTM (Long Short-Term Memory) algorithm, we will create a model capable of predicting the S&P500 with significant results. The data period analysed spans from January 1, 2020, to December 31, 2022. The Python programming language was chosen to implement the code because it provides a vast number of predefined algorithms in libraries make it easier, both code writing, reading and interpretation. We have selected several fundamental, macroeconomic, and technical indicators, with the expectation that the LSTM model could capture relationships between them. Frist we will address the key concepts necessary for understanding the analysis and them explore in detail the methodology, reasoning, and techniques used in data preprocessing, model training, and tuning of LSTM model hyperparameters. We then proceed to the sensitivity analysis, where we assess the relevance of each indicator and parameter to the model's outcome, adjusting if necessary and finally, we compare our final model to the Naive model. Then we conclude this dissertation by answering some questions and interpreting the model results from an economic perspective.
Descrição
Tese de Mestrado, Matemática Financeira, 2023, Universidade de Lisboa, Faculdade de Ciências
Palavras-chave
Redes Neuronais LSTM Tuning Hidden layers RMSE Teses de mestrado - 2023
