Abstract:

Cryptocurrencies have become increasingly popular since their emergence in 2009. Particularly after the release of Bitcoin on January 3, 2009, the introduction of other cryptocurrencies gained momentum. Following this surge in popularity, the prediction of cryptocurrency returns has become a significant topic. The main objective of this study is to compare the forecast performances obtained using traditional time series methods, specifically ARIMA-GARCH, and the Long Short-Term Memory (LSTM) deep learning approach to predict returns for Bitcoin (BTC), Ethereum (ETH), and Binance (BNB) cryptocurrencies. In this context, the study aims to contribute novel insights to the existing literature. Daily data for each cryptocurrency from different time intervals were utilized, with the dataset being divided into 90% training and 10% testing data. The methods were compared using evaluation criteria such as RMSE and MSE. Overall, for the BTC series, the ARIMA-GARCH method demonstrated superior results in the training data, while the LSTM method proved to be more effective for the testing data. In the case of the BNB series, LSTM outperformed in both training and testing datasets. For the ETH series, the ARIMA-GARCH method yielded better results for both datasets. This study emphasizes that both methods can exhibit significant performance in predicting financial data.

Keywords: