Fuel Sales Forecasting with SARIMA-GARCH and Rolling Window

Abstract

This research article proposes an innovative strategy to improve long-term forecasting accuracy for gasoline sales in Canada. The SARIMA-GARCH model was used with the rolling window forecasting technique to successfully address varying seasons, changing patterns, and conditional variance on the historical data of gasoline sales in Canada (1993-01-01 to 2015-12-01) with the sample size of 276. The rolling window forecasting technique was used to forecast one-step-ahead value and update the model to fresh observations while minimizing look-back bias and attaining good long-term forecasting accuracy. The findings revealed considerable improvements in forecasting accuracy. The proposed SARIMA-GARCH model with rolling window forecasting produced a RMSE of 151026.28 and a Mean Absolute Percentage Error (MAPE) of 0.0340. This outperformed other baseline models, including simple SARIMA model which had a RMSE of 329,689.88 and a MAPE of 0.0786, and the GARCH model which had a RMSE of 316,168.33 and a MAPE of 0.0685. The data shows that the proposed approach is effective for accurate long-term forecasting of gasoline sales in Canada. The study provides significant data for politicians, industry professionals, and energy investors, assisting them in making informed decisions about resource allocation, strategic planning, and risk management.