Thesis (MSc) - Cyprus International University. Institute of Graduate Studies and Research Computer Engineerinig

Digital money has become a popular investment choice due to its decentralized structure and the potential for large returns. However, the cryptocurrency market’s extremely unpredictable and volatile character makes it difficult for investors to foresee price fluctuations and make lucrative purchases. One of the most popular and effective price forecasting approaches is time series analysis, which detects trends and patterns in previous pricing data to estimate future price movements. The combination of machine learning (ML) methods and time series analysis can greatly improve forecasting accuracy.
Predicting future Bitcoin prices is crucial for customers to maximize their profits and minimize their losses. However, this task is challenging because of the complex temporal relationships between Bitcoin-related features. Moreover, external factors can influence cryptocurrency movement, resulting in unpredictable price fluctuations. To address this problem, deep recurrent neural network (DRNN)-based sequence learner models have been used to learn complex sequential features. In this study, multiple bidirectional versions of LSTM, GRU, and RNN recurrent layers were designed on DRNN models, and their performances were compared for a one-day-ahead Bitcoin price prediction task. Algorithms are trained and evaluated using data generated from Yahoo Finance API historical price data from May 2017 to May 2023. After this model is tested using various criteria, At the end of the day we will have a complete comparison table with all data to show the best model in each scenario.
The results show that using a convolutional layer with three bidirectional GRU layer-based DRNN models achieves superior performance, with an average deviation of 3.81% from the actual Bitcoin price.