Nurul Mukhlisah Abdal, Asmaul Husnah Nasrullah, Dewi Fatmarani Surianto, Wirawan Setialaksana
This study compares classical, deep learning, and hybrid approaches for cryptocurrency price forecasting across Bitcoin (BTC), Ethereum (ETH), Binance Coin (BNB), and Cardano (ADA) using daily data from 15 September 2022 to 15 September 2025. We implement an ARIMA baseline, a stacked LSTM network, and an inverse-error-weighted ARIMA-LSTM ensemble. Feature engineering includes trend, momentum, volatility, and volume indicators; models are evaluated with expanding walk-forward validation and multiple metrics (R2, RMSE, MAE, MAPE, sMAPE, MASE). Statistical significance is assessed via Diebold-Mariano tests. Results indicate that ARIMA consistently outperforms LSTM across all assets, with average performance of R2 = 0.924 ± 0.051, MAPE = 2.19 ± 0.81%, and MASE = 0.95 ± 0.30, compared with LSTM’s R2 = 0.527 ± 0.709, MAPE = 4.36 ± 0.26%, and MASE = 1.91 ±1.25. The ensemble attains R2 = 0.902 ±0.084 and MAPE = 2.36 ± 0.58%, retaining ∼97% of ARIMA’s explanatory power while reducing volatility relative to LSTM. Asset-specific analyses show strong ARIMA performance for ETH (R2 = 0.972 ± 0.038) and BNB (R2 = 0.964 ± 0.045), and LSTM failure on BTC (R2 = −0.534 ± 0.214). Cross-asset dispersion in R2 is markedly lower for ARIMA than for LSTM, indicating superior generalization. DM tests confirm ARIMA’s advantage over LSTM (p < 0.001) across assets, with ARIMA versus ensemble differences nonsignificant for BNB. Under rigorous out-of-sample evaluation, the parsimonious ARIMA model provides the most accurate and stable forecasts, while the ensemble offers a robust alternative when cross-asset stability is prioritized. © 2025 IEEE.
Department of Informatics and Computer Engineering, Universitas Negeri Makassar, South Sulawesi, Indonesia
