This thesis explores the ability of machine learning methods combined with the use of blockchain characteristics, a stable coin proxy and technical indicators to predict the price movement of Bitcoin correctly. The experimental work attempts to prove that machine learning models can outperform a Buy & Hold strategy with the selected features and compares the classification performance of Decision Tree, a Support Vector Machine and a Long Short-Term Memory Recurrent Neural Network. The results from the feature selection supported the findings of existing literature. Additionally, the results suggest that mining difficulty and blockchain characteristics that measure transaction activity can provide supplementary information about Bitcoin. In terms of the performance of the models, results showed that a Decision Tree model has the propensity to overfit due to its low complexity and the type of data. The results from the SVM showed that although it achieved the highest accuracy, it only managed to identify the overall trend and therefore it was not able to beat a Buy & Hold strategy. Even though the LSTM did not outperform the benchmarks, it was the model that showed the most promising results, and its performance would likely improve by additional hyperparameter tunning. Therefore, the inability to outperform the benchmarks was not conclusive. Finally, the fact that the Logistic regression model was able to outperform the Buy & Hold strategy in terms of returns and volatility leads us to conclude that machine learning methods might be effective in predicting bitcoin price movement with the selected features.
|Date of Award||1 Feb 2021|
- Universidade Católica Portuguesa
|Supervisor||Dan Tran (Supervisor)|
- Price movement prediction
- Machine learning
- Technical analysis
- Stable coin