A modified time series model using conditional and unconditional estimations with applications to a real dataset

Abstract

Modern statistical techniques offer a range of methodologies for modelling time series data, with conditional and unconditional approaches providing complementary insights that enhance overall model accuracy. This article introduced a modified ARIMA model employing conditional and unconditional parameter estimates. The methodology for the new model based on novel methods is provided. The prediction process, one and two steps ahead, is covered in detail, and a novel algorithm is presented. The best model is picked based on various measurement criteria, such as coefficient of determination (R²), root mean squared error (RMSE), and mean absolute scaled error (MASE). The suggested model is applied to a monthly petrol sales dataset (Jan 2014 to Dec 2023), where the real data in this article was taken from the U.S. Census. Eventually, the predicted petrol sales in the U.S. over the following four years are offered. As showed in the results the modified model fits the data better and improves forecast accuracy as measured by R², RMSE, and MASE. The enhanced performance demonstrates the effectiveness of the modified time series model, and it provides a valuable tool for practitioners and opens avenues for further research in advanced forecasting methodologies. All calculations and visualizations presented in this article were conducted using version 4.3.2 of the R programming language.