Nowcasting U.S. inflation using mixed frequency real-time data

Abstract: Different models were developed with the aim of nowcasting inflation at a daily basis with high frequency variables, while using real-time data to avoid look ahead bias. Both popular machine learning models such as Random Forest and XGBoost, and more traditional models such as UMIDAS and Almon distributed lag models were used to make the nowcasts. The MIDAS framework was utilized as a way of handling predictors sampled at mixed frequencies and variations of LASSO were used to select the best variables and features for the model. The main analysis considers the performance of the models compared to each other and a simple benchmark AR(1) model. It can be concluded that the ML models outperform all other models, with XGboost at the top. UMIDAS and Almon were slightly outperformed by the AR(1) model which could probably be explained by overparametrization and that the LASSO did not do a good enough job to remove enough features. Further, other topics related to the nowcasting of inflation was investigated. It was concluded that inflation has become harder to nowcast during the recent years. However the variables used have stayed relatively constant throughout time. The inclusion of higher frequency variables, such as daily, improved the nowcasts compared to the more traditional approach of only using monthly released macro variables.