Comparative performance of stochastic and deep learning models in forecasting crude oil prices under market shocks

Abstract

This study provides a comprehensive evaluation of crude oil price forecasting by systematically comparing stochastic models (GBM, FBM, BB), machine learning and time-series approaches (LSTM, ARIMA, VAR), and two hybrid strategies LSTM-GBM (time-varying drift estimation) and Bridge-ARIMA (terminal-level anchoring) across short-, medium-, and long-term horizons. Using daily price data spanning 34 years (1990–2024), the analysis encompasses major global crises, including the Asian Financial Crisis, Dot-com Bubble, Global Commodity Boom, Global Financial Crisis, Arab Spring, Oil Price Collapse, COVID-19 pandemic, and the Russia–Ukraine conflict. Stochastic models exhibited sensitivity to the number of simulation runs, with FBM yielding the most robust forecasts, BB demonstrating stability through mean-reversion, and GBM showing larger errors over extended horizons. LSTM consistently outperformed alternative methods, capturing nonlinear dynamics and long-term dependencies with high accuracy across all horizons. ARIMA performed adequately for short-term forecasts but declined over longer periods, whereas VAR consistently underperformed, reflecting its limited capacity to capture complex market dynamics. The hybrid methods further enhance performance: Bridge-ARIMA reduces bias by aligning simulated paths with an ARIMA-implied terminal, improving crisis-period tracking, whereas LSTM-GBM leverages learned drift to refine path forecasts and risk metrics (e.g., VaR). These results underscore the complementary advantages of stochastic and deep learning approaches and suggest that hybrid frameworks can enhance both volatility assessment and trend prediction. The findings offer practical insights for energy market planning, risk management, and strategic investment, particularly under crisis-driven volatility.