Applications of time series analysis for forecasting fuel sales and change point detection

Abstract

This paper shows how time series models can be used to forecast motor vehicle fuel sales, and how to use those models to detect changes in the time series signals. The model used is a least squared regression model that considers seasonal trends, serial autocorrelation, day of week, holidays, and days since open. With these covariates, the models proved to be highly predictive with forecasts on gasoline and diesel fuel, as the maximum cumulative accuracy was at most 2.74% for gasoline and 1.33% for diesel fuel. It can also be shown that mean absolute error rates can be represented by a gamma distribution, and this can be used to detect changes in the time series signal.