TFT-ACB-XML: Decision-Level Integration of Customized Temporal Fusion Transformer and Attention-BiLSTM with XGBoost Meta-Learner for BTC Price Forecasting

Researchers have developed a novel hybrid framework called TFT-ACB-XML for Bitcoin price forecasting, presented in a new arXiv paper published on February 12, 2026, to address the persistent challenges of predicting cryptocurrency prices in non-linear, highly volatile, and temporally irregular markets. The research introduces an innovative stacked-generalization methodology that combines multiple specialized models to enhance prediction accuracy while maintaining interpretability across diverse market conditions. The framework specifically targets closing price prediction, a crucial metric for traders and investors in the cryptocurrency market. The TFT-ACB-XML architecture integrates two parallel base learners—Customized Temporal Fusion Transformer and Attention-BiLSTM—with their outputs combined through an XGBoost meta-learner at the decision level. This multi-stage approach allows the system to capture various temporal patterns and market dynamics more effectively than single-model approaches by leveraging the unique strengths of each component. The Temporal Fusion Transformer excels at capturing long-term dependencies, while the Attention-BiLSTM focuses on short-term temporal patterns, with the XGBoost meta-learner effectively weighting these diverse predictions. This advancement represents a significant contribution to the field of financial forecasting, particularly for cryptocurrencies like Bitcoin that exhibit unique market characteristics. By combining the strengths of different neural network architectures with the interpretability of gradient boosting methods, the TFT-ACB-XML framework offers a potentially more robust solution for Bitcoin price prediction that could outperform existing models in accuracy and reliability. The research addresses a critical need in the cryptocurrency ecosystem where accurate price predictions remain elusive despite advances in machine learning techniques.