Forecasting inflation in Turkey: A comparison of time-series and machine learning models

Abstract

Purpose: This paper aims to test the accuracy of some Machine Learning (ML) models in forecasting inflation in the case of Turkey and to give a new and also complementary approach to time series models.  Methods: This paper forecasts inflation in Turkey by using time-series and machine learning (ML) models. The data is spanning from the period 2006:M1 to 2020:M12. Findings: According to our findings, although the linear-based Ridge and Lasso regression algorithms perform worse than the VAR model, the multilayer perceptron algorithm gives satisfactory results that are close to the results of the time series algorithm. In this direction, non-linear machine learning models are thought to be a reliable complementary method for estimating inflation in emerging economies. It is also predicted that it can be considered as an alternative method as the amount of data and computational power increase. Implication: The findings are expected to be useful as a guide for central banks and policy-makers in emerging economies with volatile inflation rates. Originality: We evaluate the forecasting performance of ML models against each other and a time series model, and investigate possible improvements upon the naive model. So, this is the first study in the field, which uses both linear and nonlinear ML methods to make a comparison with the time series inflation forecasts for Turkey.