Forecasting Inflation Spikes with Machine Learning

Abstract

This study aims to forecast spikes in the inflation rate for the U.S. The data set includes inflation and 35 other relevant variables spanning the period from 2000:1 to 2022:12 in monthly frequency. These variables are fed to three different machine learning methodologies, Support Vector Machines (SVM), Decision Trees (DT) and Random Forests (RF) with an Elastic-Net logit model used as a benchmark from the area of traditional econometrics. The optimal model is an SVM model coupled with the RBF kernel reaching a total out-of-sample accuracy of 87.27%, with 69.23% for spike detection and 7.14% false alarms.