MODELING AND FORECASTING HEADLINE INFLATION IN ETHIOPIA: A MACHINE LEARNING APPROACH

Abstract

Inflation is an important indicator of a nation's welfare and has become one of the major economic challenges globally, especially in Ethiopia. Several studies forecasted inflation in Ethiopia using traditional models, as accurate forecasts contributed to a more stable economic environment, even though forecasting with traditional models was challenging. Thus, this study aimed to model and forecast headline inflation in Ethiopia using a machine learning approach. The study was based on secondary data recorded on headline inflation and related factors from January 2000 to December 2023, obtained from various inflation-related organizations. The study utilized multivariable time series data for the past 24 years. The data were transformed, standardized and split into training and testing sets to enhance the forecast accuracy of both the machine learning and time series models. Cross-validation (CV) and grid search were used to tune the machine learning parameters such as the penalty parameter, learning rate, number of trees, maximum iterations, and the number of hidden nodes and layers and also select model with minimum MSE or RMSE. A stochastic gradient decent (SGD) was used to optimize the parameters. The selected models were evaluated based on performance evaluation criteria, including RMSE, MAE, and MAPE tests. The headline inflation in Ethiopia saw slight increases from 2000 to the second quarter of 2007, followed by a sudden shift in the fourth quarter of 2008, and then a rapid increase from the first quarter of 2015 to the fourth quarter of 2023. Food inflation, non-food inflation, export and import prices of goods and services, political stability index, exchange rate, numbers of vehicles, rainfall, world oil price, gross domestic fixed investment, unemployment rate, T-bill sales, agricultural production price were predictors significantly determine the headline inflation in Ethiopia. Furthermore, Food inflation, non-food inflation, export and import prices of goods and services, number of vehicles, gross domestic fixed investment were the most factors that determine the forecasting accuracy of the models. Among various forecasting methods, a specific ANN architecture called NNAR emerged victorious. It outperformed Ridge regression, LASSO, Elastic Net, Random Forest, and even the benchmark model in terms of accuracy for both in-sample and out-of-sample inflation forecasts in Ethiopia. NNAR achieved the lowest RMSE, MAE and MAPE, solidifying its position as the most effective model in this study. Finally, this study recommended that policymakers, financial analysts, investor and stakeholders should give attention to the identified drivers of headline inflation and consider using advanced machine learning models.