Machine Learning-Enhanced Time Series Forecasting for Macroeconomic Variables: A Comparative Evaluation of Deep Learning, Gradient Boosting, and Traditional Econometric Approaches

Abstract

This paper presents a comprehensive comparative evaluation of machine learning and traditional econometric approaches to short- and medium-term forecasting of key U.S. macroeconomic variables including GDP growth, headline and core inflation, unemployment rate, and federal funds rate. The evaluation combines monthly data from 1970 through 2024 with an expanded set of high-frequency indicators and text-derived sentiment features. Six forecasting approaches are compared: traditional autoregressive integrated moving average (ARIMA) models, dynamic factor models, Bayesian vector autoregression (BVAR), gradient boosted trees (XGBoost, LightGBM), long short-term memory (LSTM) recurrent networks, and Temporal Fusion Transformer (TFT) architectures. Evaluation uses out-of-sample rolling window forecasts across horizons of 1, 3, 6, and 12 months, with root mean squared error (RMSE), mean absolute error (MAE), and directional accuracy as principal metrics. Gradient boosted trees achieved the lowest RMSE for 1- and 3-month horizons on inflation and unemployment forecasting (11.4 per cent and 9.8 per cent improvement over ARIMA respectively). Temporal Fusion Transformer models achieved the strongest performance for longer horizons on GDP growth. Bayesian VAR remained competitive across all horizons and provided superior uncertainty quantification for policy applications. The paper contributes empirical evidence supporting the practical role of machine learning in macroeconomic forecasting while identifying contexts where traditional econometric methods retain advantages.