A Comparative Analysis of the Long Short-Term Memory (LSTM) and Prophet Models in Gold Price Forecasting Using Time Series Data

Abstract

Gold price forecasting is crucial for investors in mitigating financial risk, particularly due to high market volatility. This study aims to evaluate and compare the performance of two distinct time series modeling approaches, representing additive statistical models. The research utilizes 4,733 daily closing price recods of PT Aneka Tambang Tbk (ANTAM) gols spanning from Januari 4, 2010, to November 20, 2025 partitioned into an 80:20 chronological ratio for training (3,786 days) and testing (947 days) sets. Unlike previos studies that included lag features, this paper evaluates both models in a pure univariate setup using only dates and prices to reflect their true intrinsic forecasting capabilities. Models were configured through systematic hyperparameter tuning: LSTM utilized a 90-day window size, 128 units, a single layer with 0,2 Dropout, and Adam optimizer with a learning rate of 0.001, evaluated via a one-step wal-forward scheme. Prophet was configured with a linear growth, multiplicative seasonality, and a changepoint prior scale of 0,8. Evaluation metrics demonstrate that LSTM significantly outperforms Prophet across all metrics, achieving a Mean Absolute Error (MAE) of Rp15,038, Root Mean Square Error (RMSE) of Rp23,637, and Mean Absolute Percentage Error (MAPE) of 0.89%, compared to Prophet’s MAE of Rp130,982, RMSE of Rp174,014, and MAPE of 8.62%. The findings indicate that LSTM’s dynamic memory cell mechanism is superior in adapting to non-linear market trends, offering valuable technological insights for automated commodity decision-support systems.