Abstract

Three Essays on Cryptocurrency Analytics and Forecasting by Using Deep Learning Models

Bahareh Amirshahi, Ph.D.
Concordia University, 2024

Since the emergence of Bitcoin in 2008 as the first cryptocurrency, digital assets have become favored investment options worldwide. Understanding and predicting the behavior of
cryptocurrency markets are essential for effective risk management and investment decisions.
However, the rapid fluctuations in these markets make accurate predictions a challenging task. In
this thesis, key aspects of cryptocurrency market analytics are explored from three different angles.
The recurring theme in each study involves proposing hybrid prediction models by combining a
feature extractor component with deep learning models to enhance prediction performance.
The first study focuses on predicting cryptocurrency volatility, an underexplored area
despite extensive studies in financial markets. By combining traditional econometrics methods
with deep learning models, we forecast daily volatility with improved accuracies. The findings
revealed that deep learning models not only enhance the accuracy of traditional models but also
exhibit superior forecasting when combined with such models in a hybrid approach.
In the second study, we address the challenge of predicting cryptocurrency price values.
Recognizing the impracticality of a universal model due to unique cryptocurrency characteristics,
we propose a flexible architecture tailored for each cryptocurrency. Additionally, we explore the
impact of sentiment data from Twitter posts on prediction accuracy, employing state-of-the-art pretrained
language models in an ensemble manner for more robust sentiment analysis. We show that
sentiment data improves the prediction results for more than 70% of the cryptocurrencies studied.
This indicates that social media posts, in particular tweets play a significant role in the
cryptocurrency markets behavior.
iv
The third study focuses on predicting the direction of cryptocurrency prices. We propose
an innovative approach by combining data denoising techniques with machine learning methods,
that generates high-quality data for prediction models and achieves significantly higher accuracies.
Notably, we assess the proposed approach across distinct periods: before, during, and after the
COVID-19 pandemic, filling a critical gap in research regarding predictive models during crisis
periods. We show that the predictive performance of the proposed method is not affected by high
values of volatility in challenging periods like the COVID-19 pandemic.
This research helps in developing highly accurate prediction models that can aid investors
seize profitable opportunities and avoid potential losses. By analyzing over 25 cryptocurrencies,
collectively representing 75% of the total market capitalization, this study offers a comprehensive
perspective beyond the conventional Bitcoin-centric approach.

Keywords: Cryptocurrency, Forecasting, Machine Learning, Deep Learning, COVID-19