A carregar...
Publicação
Time series forecasting by combining LSTM RNN with ARIMA method
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 872.53 KB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
This study aims to analyze the performance of the ensemble model – the combination of Long ShortTerm Memory Recurring Neural Network model with the ARIMA model. We developed these models
separately to perform the best on their own in predicting prices 1, 7, and 14 observations ahead while
taking into account the last 30, 60 and 90 observations and checking if the combination of them
outperforms the standalone models. We evaluated the models based on RMSE and their ability to
predict the turning points. Models were developed and tested on two different types of securities –
index S&P 500 and cryptocurrency Bitcoin (BTC). The combined methods demonstrated strong
performance on BTC data set and gave at least 90% turning point prediction accuracy when predicting
the price for one observation ahead. For the S&P 500 data set, the performance of the stacked model
was poor – it outperformed the standalone models only in one test out of eighteen – while predicting
prices one observation ahead, looking back at the past 30 observations.
Descrição
Dissertation presented as the partial requirement for obtaining a Master's degree in Information Management, specialization in Knowledge Management and Business Intelligence
Palavras-chave
Reoccurring Neural Networks Deep Learning Long Short-Term Memory Financial Time Series ARIMA