Comparing Out-of-Sample Performance of Machine Learning Methods to Forecast U.S. GDP Growth.

Area of Science:

• Economics
• Data Science
• Forecasting

Background:

• Forecasting U.S. GDP growth is challenging due to data instability.
• Machine learning (ML) and deep learning (DL) are increasingly used for economic forecasting.

Purpose of the Study:

• To compare the performance of various ML and DL methods for U.S. GDP growth forecasting.
• To evaluate the impact of predictor set size and type on forecast accuracy.
• To identify optimal forecasting strategies for different time horizons.

Main Methods:

• A recursive forecasting strategy was employed to assess out-of-sample performance across multiple subperiods.
• Three predictor sets were used: a large set (224 predictors), a small set of strong predictors (9 predictors), and a combined set including a high-frequency index.
• Forecasting methods included density-based ML (bagging, boosting, neural networks), sparsity-based ML (Lasso), and deep learning (DL) models.

Main Results:

• Density-based ML methods showed slight advantages over sparsity-based methods for short-horizon forecasts with large predictor sets.
• Density-based ML methods performed better with larger predictor sets than smaller ones, particularly for shorter horizons.
• Parsimonious models incorporating a strong high-frequency predictor outperformed complex ML/DL models using numerous low-frequency predictors for long-horizon forecasts.
• Ensemble ML methods demonstrated superior performance compared to popular DL methods.

Conclusions:

• The choice of predictors, especially high-frequency data, is crucial for accurate economic forecasting.
• Parsimonious models can be highly effective, even outperforming sophisticated ML/DL approaches.
• Density-based ML methods show promise, particularly when utilizing extensive predictor datasets for short-term forecasting.