Global Gold Price Forecast Based on VAR Model

Abstract

 This research paper presents the out-of-sample forecasts of the global gold price based on the Vector Autoregression (VAR) approach using macro-financial variables. This paper uses monthly data from January 2006 to June 2024 for the following series: the S&P 500 Index, the WTI crude oil price, the US dollar index and the US 10-year government bond yield. The first-difference transformations are used to render all non-stationary series stationary. The estimation of the VAR(1) model and its validation with a 12-month holdout period (July 2024-June 2025) yields the final model with Mean Absolute Percentage Error (MAPE) of 15.92% and Root Mean Square Error‌ (RMSE) of 561.41. The results of analysis suggest that the dynamics of gold prices are almost completely self-determined and have almost no short-run reaction to other variables in the model. The results of impulse response function and forecast error variance decomposition analyses are consistent with the above conclusion and confirm the inertia of gold prices as the most likely explanation of the above-mentioned results. As a result, the gold behaves like a safe-haven asset. Finally, it is concluded that the linear VAR model cannot capture the dynamics in the sample well and, more importantly, is not a good forecasting model. It is necessary to use more complex non-linear models.