Forecasting Gold Prices (Global Ounce) Using Time Series Models and Deep Learning

The Autoregressive Integrated Moving Average (ARIMA) model is one of the most widely used methods for forecasting economic and financial time series. It is mostly has effect for short-term linear time series forecasting. With the necessary need of modern modeling techniques, especially in deep learning methods, which are capable of capturing advanced features in the time series data, researchers now have the opportunity to measure and compare the predictive power of traditional and deep learning approaches. This study aims to assess the accuracy of time series forecasting models by comparing the performance of the traditional ARIMA model with three deep learning models: Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU), using daily global gold prices (XAU/USD) as its dataset. The freshness of this research lies in three key aspects: first, the vast numbers of applications of three different recurrent neural network architectures on a single financial dataset; second, the use of data from January 2020 to October 2024, a period marked by high volatility and unusual economic conditions; and third, a comprehensive performance evaluation using six well-known statistical metrics: RMSE, MAE, MAPE, MSE, $\mathbf{R}^{\mathbf{2}}$, and RMSPE. The dataset includes 1,207 daily observations of global gold prices, which were normalized and been split into training (80 %) and testing (20 %) sets. The results show that the GRU model achieved the best predictive performance with $\text{RMSE}=54.46, \text{MAE}=49.5$, and $\mathrm{R}^{2}=0.77$. In contrast, the ARIMA model showed the weakest performance with RMSE $=163.82$, MAE $=120.72$, and $\mathbf{R}^{2}=-0.97$. These findings indicate that deep learning methods, particularly the GRU architecture, can significantly improve forecasting accuracy in the gold market and serve as a valuable tool for financial decision-making.