Bitcoin Time Series Forecasting with Exogenous Factors: Deep Leaning Approach

Over the last years, major developments have been made in cryptocurrency technology, resulting in their increased presence in worldwide financial transactions, providing increasing value for society by serving as a means of exchange and investment asset class. Modeling cryptocurrency prices is relevant for institutional and retail investors, contributing to informed decision-making in a highly volatile asset class. Despite the importance of forecasting the steep fluctuations observed in this asset class, this task is extremely complex and relies on multiple exogenous factors such as the blockchain network, market trends, and macroeconomic data. As simple statistical methods are unable to capture the complexity of temporal dependencies, researchers are turning to advanced machine learning and deep learning algorithms to tackle this non-stationary time series problem. This work is the result from an internship carried out at Klever in collaboration with ISEG within the scope of the Master of Quantitative Methods for Economic and Business Decision. We present a methodology for building sequence-to-vector deep learning models to predict the price, return, and directional state of Bitcoin. Leveraging a comprehensive feature engineering system, this research achieves an accuracy up to 90.9%, a MAPE up to 1.74% for price prediction, and up to 0.11 MAE for return prediction, surpassing each task’s respective baseline model.