Hybrid ARIMA-LSTM Model for Gold Price Forecasting at Pegadaian

Abstract

Accurate forecasting of digital gold prices at PT Pegadaian is essential for managing volatility driven by macroeconomic factors, including exchange rates, inflation, and global gold prices. Conventional models present limitations: ARIMA effectively captures linear trends but fails to model non-linear patterns, whereas LSTM handles non-linearity but is prone to overfitting and poor generalization. This study proposes a hybrid ARIMA-LSTM model based on a quantitative time series approach. The analysis uses secondary data comprising daily digital gold prices from PT Pegadaian (2024–2025) and related macroeconomic indicators obtained from BPS and Bank Indonesia. Data are preprocessed to ensure stationarity and quality prior to modeling. The hybrid model combines linear forecasts from ARIMA with LSTM modeling of the resulting non-linear residuals. The hybrid model achieved MSE = 54,294.23, MAE = 113.56, and RMSE = 233.01 on the test set, representing reductions of approximately 75% in MSE, 54% in MAE, and 50% in RMSE relative to the standalone LSTM on testing data. The hybrid model outperforms both individual ARIMA and LSTM models in terms of generalization and accuracy. A primary limitation is the use of manual hyperparameter tuning; implementation of automated methods, such as grid search or Bayesian optimization, could further improve performance and robustness.