| Abstract [eng] |
The master’s thesis aims to introduce and study neural network prediction models, which could make it possible to predict the volatility of cryptocurrency prices. The research opens with a short introduction of time series, their definition, and usage scenario. Then the work presents the problem of predicting cryptocurrency prices and their volatility. Also, it presents how neural network models such as KNN, RF, SVM, LSTM, and GRU could help to predict those values better. As a result of the analysis of related works, not only the models that will be used for forecasting in the experimental part but also cryptocurrency volatility proxies were selected. The aim of the work is to study and compare each of the selected models. All of them were trained and tested in different ways: Litecoin and Bitcoin cryptocurrencies daily and hourly time series were taken. Data was divided into two parts: training and testing so that models could be trained and then tested, so predicted values would be compared with the real ones. MDA, RMSE, MAE, and R2 metrics were chosen for the effectiveness evaluation of the selected models. The effectiveness of the models is compared at the end of the experimental study. At the end of the work, the findings and future plans are described. |
