Comparative Deep Learning Models for Indonesian Gold Price Forecasting

Albi Pernata Jomantara Putra (Telkom University)
Baginda Mi’raj Williamsyah (Telkom University)
Achmad Rizal (Telkom University)
Favian Dewanta (Telkom University)
Anggunmeka Luhur Prasasti (Telkom University)
Said Ziani (Mohammed V University)

Advance Sustainable Science, Engineering and Technology (ASSET)
Vol. 8 No. 3 (2026): May - July

This study evaluates LSTM, CNN-LSTM, LSTM-GRU, and CNN-LSTM-GRU architectures for forecasting Indonesian gold prices using 1,269 daily observations (2022–2025). Models utilized Bayesian-optimized hyperparameters and were benchmarked against ARIMA-GARCH and Random Forest baselines across 30-day and 365-day horizons. Performance was assessed via MAE, RMSE, R², and MAPE, confirming deep learning’s superiority in capturing non-linear dynamics over classical methods. The LSTM-GRU achieved the best numerical results, with MAPEs of 1.21% (short-term) and 1.32% (long-term). However, qualitative evaluation revealed that the highest-scoring model produced unstable long-term predictions, indicating a critical trade-off between numerical accuracy and forecast realism. These findings suggest financial model selection must prioritize stability alongside statistical metrics. A key limitation is the exclusive use of univariate data, necessitating future multivariate validation with macroeconomic indicators.