INFLATION AND OUTPUT GROWTH FORECASTING IN A DATA-RICH ENVIRONMENT USING EXTREME GRADIENT BOOSTED TREES WITH ONLINE LEARNING

Abstract

Forecasting of inflation and output growth is notoriously difficult with na‹ve benchmarks being hard to beat. This paper examines a new tool in data-rich macroeconomic forecasting using eXtreme Gradient Boosting (XGB) and introduces a novel online learning of optimal boosting iterations algorithm that exploits the persistence present in macroeconomic data. Compared to traditional benchmarks, factor models, and random forest, our XGB implementation had substantially lower RMSE for longer horizons forecasts with comparable performance for shorter horizons. While past literature generally reported a decline in predictive ability for models after the Great Moderation, XGB models were more temporally robust to the structural changes. Using interpretable machine learning tools, an analysis using SHAP values uncovered new stylized facts where predictive content of predictors across the board did not actually fall after the Great Moderation and that reduced form inflation persistence were largely derived from inherited rather than intrinsic persistence.