Hybrid SARIMAX–SVR Forecasting of Retail Diesel Prices in Morocco

Abstract

Retail fuel prices are economically sensitive in Morocco, where diesel remains a key input for transport and logistics and domestic pump prices are highly exposed to external shocks. Forecasting in this setting is difficult because price dynamics combine structured persistence with nonlinear adjustments during volatile periods. This paper proposes a hybrid forecasting framework for Moroccan retail diesel prices that combines a Seasonal Autoregressive Integrated Moving Average model with exogenous variables (SARIMAX) and Support Vector Regression (SVR). SARIMAX captures the main linear timeseries structure while incorporating gasoline-based explanatory signals, and SVR is then used to learn a nonlinear correction from the remaining forecast residuals. Using bi-weekly national average diesel and gasoline prices from 2014 to 2022, out-ofsample evaluation indicates that the hybrid model increases the coefficient of determination by approximately 14.6% relative to SARIMAX and 7.6% relative to SVR, while reducing the mean squared error by about 39.7% compared to SARIMAX and 26.3% compared to SVR. Overall, the results show that combining statistical structure with machine learning flexibility leads to more accurate and stable forecasts, offering a practical and interpretable tool for fuel price analysis in an energy-importdependent economy.