Forecasting of BBNI Stock Prices under Middle East Geopolitical Uncertainty Using Multivariate Time Series: A Comparative Study
Abstract
Global geopolitical uncertainties, such as the Middle East crisis, inherently trigger high volatility and systemic risks in capital markets, potentially threatening the stability of banking sector stocks. This study aims to analyze the structural shifts in trends and stock price volatility of PT Bank Negara Indonesia (Persero) Tbk (BBNI), while concurrently conducting a comparative evaluation of various multivariate-based short-term forecasting models amidst the crisis. The methodology compares five computational model architectures integrating exogenous macroeconomic variables (BI Rate, Inflation, and USD/IDR Exchange Rate), comprising the statistical approach of Autoregressive Integrated Moving Average with Exogenous Inputs (ARIMAX), machine learning techniques (Random Forest and XGBoost), and deep learning frameworks (Long Short-Term Memory [LSTM] and Gated Recurrent Unit [GRU]). Daily data are divided chronologically into a pre-geopolitical period (January 2024–May 2025) and a geopolitical crisis period (June 2025–May 2026).Trend analysis results reveal a unique market adaptation phenomenon, wherein the geopolitical crisis period paradoxically exhibits a positive trend recovery characterized by lower and more stable market risk within a volatility range of 0.010 to 0.026, compared to the pre-geopolitical period which peaked at 0.041. Predictive performance evaluation demonstrates that the tree-based ensemble model, specifically XGBoost, achieves the highest accuracy and efficiency, registering the minimum error metrics (MAE: 22.19; RMSE: 30.99; MAPE: 0.56%). The ARIMAX(2,0,1) model ranks second (MAPE: 0.65%) and successfully confirms a strong, negative linear influence of the BI Rate variable on stock prices. Conversely, the LSTM and GRU architectures underperform due to data scarcity constraints inherent to parameter-dense models, as well as a smoothing effect that hinders the capture of extreme price fluctuations. This study confirms that ensemble learning approaches, such as XGBoost, are superior, adaptive, and robust for modeling moderate-scale financial data amidst dynamic macroeconomic volatility.