Prediksi Harga Bitcoin Menggunakan Model Hibrida LSTM–Transformer dengan Integrasi Indikator Teknikal dan Validasi Statistik

Fajar Rohmattulloh (Universitas Amikom Purwokerto)
Fandy Setyo Utomo (Universitas Amikom Purwokerto)
Taqwa Hariguna (Universitas Amikom Purwokerto)

Jurnal Algoritma
Vol 23 No 1 (2026): Jurnal Algoritma

This study aims to develop an accurate, stable, and adaptive Bitcoin price prediction model by integrating Long Short-Term Memory (LSTM) and Transformer Encoder architectures with technical indicators as additional features. Four deep learning architectures were comparatively evaluated: LSTM, Bidirectional LSTM (BiLSTM), Convolutional Neural Network–LSTM (CNN–LSTM), and a hybrid LSTM–Transformer model, using historical Bitcoin to US Dollar (BTC/USD) price data from 2014 to 2025 obtained from Yahoo Finance. The technical indicators incorporated include Moving Average (MA), Exponential Moving Average (EMA), Relative Strength Index (RSI), and Moving Average Convergence Divergence (MACD). Model performance was assessed using three primary metrics—Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination (R²)—along with paired t-tests to evaluate the statistical significance of performance differences among models. Experimental results indicate that the hybrid LSTM–Transformer model achieves the most competitive performance, with an RMSE of 0.0412, a MAPE of 4.36%, and an R² of 0.9617. The paired t-test results confirm that the performance differences among the models are statistically significant (p-value < 0.05), thereby providing empirical support for the superiority of the hybrid approach. The integration of technical indicators enhances the model’s ability to capture price trends and volatility patterns in Bitcoin markets. However, further analysis—such as ablation studies or explicit before-and-after comparisons—is required to isolate and quantify the individual contributions of these indicators. From a scientific perspective, this research reinforces the effectiveness of attention mechanisms in capturing long-term temporal dependencies and demonstrates that combining technical indicators with hybrid deep learning architectures can improve both the stability and validity of cryptocurrency price predictions. The main contribution of this study lies in proposing a cryptocurrency price prediction framework that emphasizes not only predictive accuracy but also reliability and statistical significance, making it a promising approach for digital financial market analytics.