Performance evaluation of forecasting models based on time series and machine learning techniques: an application to light fuel consumption in Brazil

Abstract

PurposeFuel demand forecast is a fundamental tool to guide private planning actions and public policies aim to guarantee energy supply. This paper aims to evaluate different forecasting methods to project the consumption of light fuels in Brazil (fuel used by vehicles with internal combustion engine).Design/methodology/approachEight different methods were implemented, besides of ensemble learning technics that combine the different models. The evaluation was carried out based on the forecast error for a forecast horizon of 3, 6 and 12 months.FindingsThe statistical tests performed indicated the superiority of the evaluated models compared to a naive forecasting method. As the forecast horizon increase, the heterogeneity between the accuracy of the models becomes evident and the classification by performance becomes easier. Furthermore, for 12 months forecast, it was found methods that outperform, with statistical significance, the SARIMA method, that is widely used. Even with an unprecedented event, such as the COVID-19 crisis, the results proved to be robust.Practical implicationsSome regulation instruments in Brazilian fuel market requires the forecast of light fuel consumption to better deal with supply and environment issues. In that context, the level of accuracy reached allows the use of these models as tools to assist public and private agents that operate in this market.Originality/valueThe study seeks to fill a gap in the literature on the Brazilian light fuel market. In addition, the methodological strategy adopted assesses projection models from different areas of knowledge using a robust evaluation procedure.