A Hybrid Ensemble Framework Combining Transformer Networks, CNN-LSTM, and Prophet for Multi-Horizon Bitcoin Price Prediction Using 1-Minute Time Series Data

Siti Sarah Maidin (Faculty of Data Science and Information Technology (FDSIT), INTI International University, Nilai, Malaysia)
M. Hemalatha (PG and Research Department of Computer Science. Sri Ramakrishna College of Arts and Science, Coimbatore, India)
Jing Sun (Faculty of Liberal Arts, Shinawatra University, Thailand)

Journal of Current Research in Blockchain
Vol. 3 No. 1 (2026): Regular Issue March 2026

Bitcoin price forecasting at one-minute frequency presents significant challenges due to rapid volatility and noise in high-frequency markets. This study proposes a hybrid ensemble framework integrating a CNN-LSTM model, a Transformer architecture, and a Prophet-based component to perform multi-horizon prediction using 500,000 one-minute BTC/USD observations. The model is evaluated across 5-minute, 15-minute, and 30-minute horizons. The results show that the ensemble achieves the best performance for the 5-minute horizon with MAE = 41.565 USD, RMSE = 60.722 USD, and MAPE = 0.156. This outperforms the CNN-LSTM model (MAE = 47.838 USD) and the Transformer model (MAE = 53.733 USD). Performance decreases at the 15-minute horizon due to Transformer instability, where the ensemble reaches MAE = 269.347 USD and the Transformer reaches MAE = 530.429 USD. At the 30-minute horizon, performance stabilizes, with the ensemble producing MAE = 84.481 USD, close to the CNN-LSTM result (MAE = 84.186 USD) and better than the Transformer (MAE = 153.887 USD). These findings indicate that the hybrid ensemble is highly effective for ultra-short-term forecasting but requires horizon-specific tuning to remain stable for medium-range intervals.