A carregar...
Publicação
Cryptocurrency Price Prediction via Graph-Based Representations
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 2 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
This study investigates the application of Graph Neural Networks (GNNs) for cryptocurrency
price prediction by integrating structural relationships from both the digital asset market and
traditional finance. Motivated by the growing body of research on spatio-temporal graph
learning, this work explores how the quality and design of the input graph affect forecasting
performance. Specifically, three graph structures are constructed: one based on Granger
causality, another on the precision matrix via Graphical Lasso, and a third using mutual
information. These graphs capture different perspectives on inter-asset relationships and are
used as adjacency matrices in a hybrid deep learning framework combining 1D Convolutional
Neural Networks (CNNs), Graph Convolutional Networks (GCNs), and Multi-Layer Perceptrons
(MLPs). The model is trained on a dataset comprising daily closing prices of four
cryptocurrencies—Bitcoin, Ethereum, Solana, and Tether along with the Financial Stress Index
(FSI), spanning from October 1, 2019 to October 1, 2024 (1,534 days). Log returns are used
for cryptocurrencies, and first differences for the FSI to enhance stationarity. Forecasts are
generated for three time horizons: 1-day, 3-days, and 6-days ahead, using a 14-day input
window in a sliding prediction framework. The results show that all three graph-based models
perform well but exhibit different strengths across time horizons and assets. The mutual
information-based model achieves the lowest average error for 1-day predictions (MAPE
1.8%), while the Granger causality model excels at 3-day forecasts (MAPE 2.7%), and the
precision matrix model leads at 6-day horizons (MAPE 2.8%). Asset-wise, the Granger model
produces the most accurate predictions for Bitcoin and Ethereum. These results outperform
that of correlation based graphs from similar work by Patil (Patil et al., 2020) and it validates
prior literature suggesting that financial graphs encode valuable relational information that
can enhance predictive modelling. Although this work uses a small graph and shallow
architecture, it demonstrates the potential of graph-based deep learning in financial
forecasting. Future work could involve scaling to larger graphs with more assets and
indicators, employing deeper or attention-based STGNN architectures, and experimenting
with dynamic and weighted graphs to further exploit the rich structure of financial systems.
Descrição
Dissertation presented as the partial requirement for obtaining a Master's degree in Data Driven Marketing, specialization in Data Science for Marketing
Palavras-chave
Machine learning Deep learning Cryptocurrency Graphical Neural Network SDG 4 - Quality education SDG 8 - Decent work and economic growth SDG 9 - Industry, innovation and infrastructure