Generating density nowcasts for U.S. GDP growth with deep learning: Bayes by Backprop and Monte Carlo dropout

Kristóf Németh, Dániel Hadházi·May 24, 2024

Summary

This study compares artificial neural networks (ANNs) with the dynamic factor model (DFM) in generating US GDP growth nowcasts. ANNs, equipped with Bayes by Backprop and Monte Carlo dropout, produce density nowcasts, including uncertainty measures, outperforming the DFM during 2012-2022. The 1D CNN-based ANNs excel, especially during economic turbulence, offering a competitive alternative to traditional methods. The algorithms dynamically adjust predictive distributions, making them suitable for policy decisions. The study employs Bayesian neural networks to model uncertainty, addressing the DFM's linear structure and scalability limitations. Monte Carlo dropout, by averaging predictions with dropout layers, provides a measure of uncertainty in nowcasts. The analysis uses FRED-MD data and evaluates the algorithms' performance in real-time economic assessment. The findings suggest that ANNs with these adaptations are a valuable tool for policymakers.

Introduction

Background

Evolution of GDP growth forecasting methods

Importance of accurate nowcasts for policymakers

Objective

To assess the performance of ANNs vs DFM in GDP growth prediction

Evaluate the role of Bayesian techniques and CNNs in enhancing uncertainty quantification

Methodology

Data Collection

FRED-MD dataset: Source and time period (2012-2022)

Data preprocessing: Cleaning, normalization, and feature selection

Data Preprocessing

Handling missing values

Feature engineering for ANNs (if applicable)

Time series decomposition for understanding underlying trends

Artificial Neural Networks (ANNs)

Bayes by Backprop: Bayesian approach for parameter estimation and uncertainty quantification

1D Convolutional Neural Networks (1D CNNs):

Architecture and design

Real-time adaptability during economic turbulence

Monte Carlo Dropout:

Implementation for uncertainty estimation in nowcasts

Density Nowcasts:

Generation and evaluation of predictive distributions

Dynamic Factor Model (DFM)

Overview and limitations (linear structure, scalability)

Comparison with ANNs in nowcasting performance

Performance Evaluation

Real-time assessment: Rolling window analysis

Evaluation metrics: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and coverage of uncertainty intervals

Turbulence periods: Identification and analysis

Results

ANNs outperform DFM during 2012-2022

1D CNN-based ANNs' superiority in economic downturns

Comparison of uncertainty measures provided by ANNs and DFM

Discussion

Advantages of Bayesian ANNs for policy decisions

Scalability and adaptability of ANNs in dynamic economic environments

Implications for future forecasting models

Conclusion

Artificial neural networks, particularly with Bayesian techniques and CNN adaptations, are a valuable tool for US GDP growth nowcasting

Recommendations for policymakers and future research directions

Advanced features