Abstract:
The price of Indonesian crude oil (ICP) is highly volatile due to fluctuations in global demand, energy policies, and geopolitical tensions, making accurate forecasting challenging. This study compares three forecasting models: ARIMA, Random Forest, and Hybrid ARIMA–Random Forest. The models are evaluated using Time-Series Cross-Validation (TSCV) with MAPE, sMAPE, and RMSE as performance metrics. The results indicate that the Hybrid ARIMA–Random Forest model achieves the lowest MAPE and sMAPE, while Random Forest attains the lowest RMSE, and ARIMA exhibits the highest forecast errors. Diebold–Mariano (DM) tests confirm that ARIMA’s predictive accuracy is significantly lower than both machine-learning-based models, whereas no significant difference is found between Random Forest and the hybrid model. Out-of-sample forecasts for January–June 2026 show relatively stable price movements within 59–63 USD per barrel, with short-term fluctuations reflected in wide prediction intervals. These findings suggest that Indonesian crude oil prices contain both linear and non-linear components, which are effectively captured by the hybrid approach. Overall, the Hybrid ARIMA–Random Forest model provides the most accurate forecasts in percentage-based metrics, offering a robust and reliable tool for policymakers, investors, and market participants navigating volatile oil markets.
