FORECASTING BITCOIN PRICES USING DEEP LEARNING AND MACHINE LEARNING

Abstract

This paper deals with financial time-series forecasting. It aims to predict daily Bitcoin prices using a multitude of forecasting models. In total, nine models have been implemented. These are three traditional forecasting models – Random Walk, ARIMA and ARMAX, two machine learning models – Support Vector Regressor and Random Forest Regressor and, three deep learning models – Multi-Layer Perceptron, Recurrent Neural Network (RNN) and Long Short-Term Memory Network (LSTM). Finally, a combination forecast for RNN and LSTM models is constructed. 32 predictor variables are considered and 5 are selected for modelling using Mutual Information Regression. The models are evaluated using the root mean squared error (RMSE). Training time for models is also included in the results. Deep learning models outperform traditional and machine learning models. The top-performing model is LSTM with RMSE of 3.9377%.