From Features to Decisions: Analysis of Transformer Frameworks for Bitcoin Price Prediction

Abstract

This research comprehensively evaluates the performance of eight Transformer-based frameworks in Bitcoin price prediction and backtest trading. A novel comprehensive scoring index is proposed, which combines nine key performance indicators. Each model's performance is quantified and ranked using this index. Furthermore, the impact of different feature combinations on model performance is analyzed by identifying the top 10 optimal feature combinations for each model. The Apriori algorithm is applied to mine frequent feature combinations and generate high-confidence association rules, providing valuable insights for feature selection. The results show that the Convolutional Transformer framework achieves the highest overall performance, outperforming other models in terms of annualized return, Sharpe ratio, and feature adaptability. The feature combination analysis reveals three features frequently appear in the top combinations of multiple models, highlighting their importance in Bitcoin price prediction. This research contributes to the understanding of Transformer-based models' performance and characteristics in cryptocurrency market forecasting, guiding researchers and practitioners in framework selection and feature engineering.