Unit:
Department of Informatics and TelecommunicationsΠληροφορική
Supervisors info:
Ιωάννης Εμίρης, Καθηγητής, Τμήμα Πληροφορικής και Τηλεπικοινωνιών, Εθνικό και Καποδιστριακό Πανεπιστήμιο Αθηνών
Original Title:
Deep Learning for cryptocurrency assets: Employing series forecasting models for price prediction and uncertainty quantification
Translated title:
Deep Learning for cryptocurrency assets: Employing series forecasting models for price prediction and uncertainty quantification
Summary:
Price prediction is one of the main challenges of quantitative finance. Within the realm of time series, cryptocurrencies are some of the most volatile and speculative, which makes the task of predicting them even more troublesome.
This thesis presents a series of Neural Networks to provide a deep machine learning solution to the price prediction problem.
Despite the difficulties, this research is concerned with two separate applications. The first focuses on predicting the next day’s Opening, High, Low and Closing prices for Bitcoin and Litecoin; whilst the second one estimates prediction intervals for Bitcoin’s Closing price. For the first application, a wide range of networks were tested and among them the one that gave the best results overall, proved to be a hyperparameter Bayesian optimized Long Short Term Memory (LSTM) network. For the second task, an approximate Bayesian Neural Network was created to provide an uncertainty estimation for Bitcoin. Uncertainty in this case is represented using prediction intervals, estimating the range within which future outcomes will fall. The research focuses specifically on Bitcoin for both tasks and Litecoin for the first, but could easily be extended for any other cryptocurrency or stocks using the methodologies that are presented.
Main subject category:
Technology - Computer science
Keywords:
Deep Learning, Neural Networks, Bayesian Neural Networks, LSTM, Bayesian optimization, Cryptocurrencies
