Understanding BITCOIN Market Mechanics Using Feature Engineering, Data Modeling, and Forecasting Methods

Bitcoin (BTC) has emerged as a groundbreaking and influential cryptocurrency, revolutionizing the financial landscape. Traders operating in the Bitcoin market encounter numerous challenges when it comes to making informed decisions due to the inherent volatility of the cryptocurrency market. Given the challenges posed by the volatile nature of the Bitcoin market, this thesis focuses on understanding the market mechanics (i.e., the underlying factors influencing price movements) to assist traders in making well-informed and profitable decisions in the unpredictable cryptocurrency market. This includes the development of prediction models that can utilize both structured (such as trading data) and unstructured data (such as social media posts) to anticipate the direction of Bitcoin's price movements and support decision-making, especially in unstable markets (e.g., the COVID-19 pandemic). The thesis represents a compendium of published papers. Article 1 provides a literature review and comparative analysis of state-of-the-art time series prediction models. In Article 2, the BTC market mechanics are simulated using a feature set of endogenous and exogenous variables. It is necessary to recognize patterns within images of time-series data charts using deep learning, as shown in Article 3. Existing forecasting models fall short of providing a robust model that handles unstructured data while providing accurate forecasting results. Thus, in Articles 4 and 5, an efficient forecasting model using ensemble and consensus learning, respectively, are proposed, which accurately analyzes the trend of BTC during the COVID-19 pandemic using Twitter posts using labeled and unlabeled data. Collectively, this thesis has contributed new insights into the BTC market. Future research could build on these findings to focus on three key areas: 1) Obtaining a greater understanding of other cryptocurrencies and stock data, 2) varying the adopted baseline models, and 3) including federated learning to handle the large size of the social datasets.