Time series analysis and predictive modeling of prices: an empirical study on bitcoin and ethereum

Abstract

Over the last few years, cryptocurrency markets have been under the radar of people because of their decentralized nature, highly volatile nature in terms of price and large amounts of money that can be generated. It is the volatility that is making it hard for traders, investors and financial analysts to construct sound forecasting models. The classical financial theories do not take into account the special properties of cryptocurrency markets, therefore it is necessary to use sophisticated statistical and machine learning methods to predict prices correctly. This thesis explores the effectiveness of the performance of time series analysis and machine learning algorithms to predict cryptocurrency price movements with both short and long term trends considered. A statistical model for analyzing trends of past price history is time series analysis. Because such tools as the Autoregressive Integrated Moving Average (ARIMA) have been able to effectively extract trends and seasonality, they have been extensively used in financial markets. Therefore, ARIMA models are suitable to investigate and forecast stationary time series data. They are so efficient and yet not suitable to use them for non-linear and very volatile financial data, therefore we are using these machine learning approaches. However, data driven solutions to cryptocurrency price prediction are offered by the machine learning methods, in particular, deep learning models such as Long Short Term Memory (LSTM) networks. Specifically, recurrent neural networks (RNN) of the form known as LSTM networks are specifically developed to learn dependencies over a time interval that are long enough to be useful for predicting a time series. Predictive performance of the models can be increased with a wide array of input features, including past price, volume, social media sentiment analysis, and macroeconomic variables, and trained with such data. The effect of policy shifts and economic conditions on digital asset prices is also assessed with the help of which cryptocurrency prices moved under different U.S. presidential regimes, for example during the Biden and Trump regimes. An extensive methodology is used in this study in which statistical and machine learning models are evaluated based on their predictive accuracy. While this data is historical cryptocurrency price data (Bitcoin, Ethereum, other top digital assets), it was obtained from prominent financial databases, usually exchanges. Input data is made reliable and of good quality by means of preprocessing techniques such as normalization, feature engineering, etc. As to the performance of the model, I measured it using important metrics such as Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and R-squared values. This study draws conclusions on the strength and weakness of each modeling technique and bases it on relative strengths and limitations. Forecasting is interpretable with ARIMA models but they are not good at handling sudden market shocks or nonlinear relationships. On the other hand, LSTM networks keep better performance in postulating the long term trends and volatility patterns, however at the cost of higher computational complexity and training time duration. By adding these outside variables pertaining to political and economic spheres to these models, the explanatory power of these models is enhanced. The implications of this study are quite broad and important to stakeholders like financial analysts, institutional investors and policymakers. Knowledge of the forecasting ability of various models facilitates well-informed decision-making in cryptocurrency investment portfolios. Moreover, the results contribute to the domain of financial modeling by demonstrating that hybrid methods combining classical statistical methods with state of the art machine learning algorithms are indeed beneficial. Finally, in the thesis, a serious comparative analysis of a time series and machine learning approaches of the cryptocurrency price prediction is presented. Actionable information provided through the synthesis of statistical and computational matter sought to help working through the subtleties of the digital asset marketplace. The next stage of research could be the application of reinforcement learning techniques together with other data sources beyond classical historical price data, including blockchain network indicators, geopolitical events etc. to achieve even greater prediction accuracy and market sophistication.