Cryptocurrency Price Prediction Improvements Through Artificial Intelligence and Model Features Refinement

Time series forecasting has always been one of the most investigated scientific areas and this is probably due to the large amount of fields involving a time component including, meteorology, finance, econometric, astronomy, earthquake prediction, just to mention a few. The present work studies the field of financial time series forecasting, mainly focusing on the Cryptocurrency market. This field poses very sophisticated problems for several reasons. Due to its recent birth, this market has a dynamic nature continuously giving rise to new cryptocurrencies leading to frequent and sudden variations in their prices. Cryptocurrencies are virtual currencies based on an innovative technology known as Blockchain. As a consequence, their economy, along with traditional macroeconomic variables, depends on technology variables that can be directly measured from the Blockchain platform. All this features lead to high volatility in cryptocurrency prices. At the same time, the analysis of a market whose price behaviour is still largely unexplored has a fundamental impact not only in the scientific field but also within economic and financial fields, serving as a source of information for speculators and investors. The two main approaches used to study this research topic are traditional econometric methods and artificial intelligence algorithms. The purpose of the present work is to study an innovative approach, based on the techniques described above, to gain relevant insights in order to enrich the state-of-the-art in the field of cryptocurrency time series forecasting. This study started with a regression forecasting problem of cryptocurrency time series prices where the main goal was to compare the results obtained through traditional econometrics models with those obtained using state-of-the-art artificial intelligence algorithms. Consequently, classification problems were conducted with the aim of predicting cryptocurrency price changes and verifying that the addition of technical analysis features to the dataset can lead to an effective improvement in the prediction of cryptocurrency price movements. Cryptocurrencies arouse keen interest not only in the scientific and financial fields but also within social media communities, making the analysis of their price beahviours one of the most discussed topics of the last fiew years. Several are also the studies that tried to use online information, including social media topics discussions, to predict cryptocurrencies price changes, proving the existence of possible cause-effect relationships between the cryptocurrency price changes and online information. These considerations led to the development of a study based on identifying and modeling relationships between cryptocurrencies market price changes and topic discussion occurrences on social media. This analysis is a further confirmation on how the addition of social variables to the dataset lead to an effective improvement on cryptocurrency price prediction. For this reason we decided to conduct a deeper investigation on whether developers emotions can effectively provide insights that can improve the prediction of cryptocurrency prices. This approach proved how the progressive addition of variables of different sources allows the development of more accurate cryptocurrencies prediction models.