Comparing Linear Regression, Random Forest, and LSTM for Antam Gold Price Forecasting

Abstract

Forecasting gold prices plays a crucial role in supporting investment decision-making and financial risk management, particularly in Indonesia, where Antam gold is widely used as an investment instrument. Although numerous forecasting techniques have been proposed, comparative studies evaluating statistical, machine learning, and deep learning models for Antam gold price prediction using long-term historical data remain limited. A systematic comparison is therefore important to identify forecasting approaches that can provide reliable and interpretable predictions for the Indonesian gold market, especially when different lengths of historical data are considered. This study compares the forecasting performance of Linear Regression, Random Forest, and Long Short-Term Memory (LSTM) using historical Antam gold price data spanning January 2010 to June 2026, collected from Kaggle and the official Logam Mulia website. The preprocessing stage included date parsing, chronological sorting, missing value inspection, data normalization for LSTM, lag feature engineering, and sequence construction. Five experimental scenarios were designed to investigate the influence of historical data length on forecasting accuracy. Model performance was evaluated using Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and the Coefficient of Determination (R²), with RMSE selected as the primary evaluation metric. The experimental results show that Linear Regression consistently achieved the best forecasting performance across all scenarios. The highest prediction accuracy was obtained using the 10-year historical data scenario, resulting in an RMSE of 17,064.29 and an R² of 0.9978. Random Forest produced competitive results only when the training dataset contained the highest observed gold price, whereas LSTM did not outperform Linear Regression in any experimental scenario. These findings demonstrate that, for Antam gold price forecasting with predominantly linear historical trends, a simple and interpretable statistical model can outperform more complex machine learning and deep learning approaches. Practically, the findings can support investors, financial analysts, and policymakers in Indonesia in selecting appropriate forecasting methods and developing more evidence-based investment strategies and financial risk management policies. The results also highlight the importance of selecting an appropriate historical data length to achieve reliable forecasting performance in financial time-series applications.