Application of explainable AI in Machine Learning models to identify the main determinants of Bitcoin price

• Autor(a) principal: Morais, Ana Sofia Rosa
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.14/41429
• topic: Bitcoin; Determinants; LSTM; GLSAR; DNN; Complexity; Performance; Interpretability; AI; Decision-making; Determinantes; Complexidade; Desempenho; Interpretação; IA; Tomada de decisão; Domínio/Área Científica::Ciências Sociais::Economia e Gestão

Since 2019, Bitcoin has become one of the most popular assets in the world. However, this decentralised cryptocurrency is typically characterised by high volatility and, in that sense, creates some concerns mainly to regulatory authorities and other decision-makers, such as governments and legislators. Furthermore, there are multiple approaches and results in the literature regarding the most relevant determinants to predict the Bitcoin price, the complexity of the Machine Learning (ML) model used to predict the Bitcoin price, and the trade-off between interpretability and the model’s performance. As a starting point, the simple model called Generalized Least Squares with Autocorrelation covariance structure (GLSAR) was found to be unrealistic to predict something as complex as the Bitcoin price. Alternatively, two more complex black box models were tested: a Long Short Term Memory neural network (LSTM) and a simple Deep Neural Network (DNN). LSTM achieved the highest 2 score of 81.63% with DNN obtaining a 2 score of 81.27%. Explainability techniques were applied on DNN and the results indicate that 71% of the twenty-one most significant variables are transaction-based, although future analysis can be done for occasional events. Moreover, the three most important features are the S&P500, the Bitcoin price in the previous day and how difficult it is to mine a Bitcoin block.