A Study of Comparative Methods for Closed-Price Cryptocurrency Prediction

Abstract

This work aims to present a comparative study method of closed-price prediction for cryptocurrencies namely machine learning-based and deep learning-based methods. Three data normalization techniques in the data pre-processing stage were also compared. They are log scaling, min-max, and z-score normalization. In the machine learning-based method, support vector regression (SVR) was used to develop the predictive model, whereas long-short term memory (LSTM) was used in the deep learning-based method to develop the predictive model. In addition, three datasets are used in this study namely Bitcoin (BTC), Ethereum (ETH), and Litecoin (LTC). The results of evaluating the predictive models using RMSE and MAPE revealed that SRV with RBF kernel produced slightly better results than LSTM. Compared to other data normalization methods, log scaling normalization produced outcomes that are more satisfactory.