纽约联储-基于组件的动态因子Nowcast模型（英）-2025.4_37页_3mb

Component-Based Dynamic Factor Nowcast Model

Authors: Hannah O’Keeffe, Katerina Petrova (Federal Reserve Bank of New York)
Publication Info: Staff Report no. 1152, April 2025
Link: doi.org/10.59576/sr.1152

Abstract

This paper proposes a component-based dynamic factor (CBDF) model for GDP growth nowcasting. It combines "bottom-up" approaches (using national income accounting identity) with dynamic factor (DF) models (handling mixed-frequency data). Key advantages:

1. Respects GDP accounting identity (unlike standard DF models).
2. Models all GDP components jointly (unlike bottom-up approaches).
3. Generates nowcast densities and impact decompositions for each GDP component.

The model uses market indicators (e.g., consumer sentiment, labor market data, financial conditions) and GDP components to produce real-time GDP nowcasts with uncertainty quantification.

Key Features

• Component-Based Approach: Combines DF models with national accounting identity to close the theoretical gap.
• Impact Analysis: Decomposes nowcast revisions into data and parameter-based impacts for transparency.
• Performance: Improves point nowcast RMSE by 15% and density forecast log-scores by 20% compared to benchmark models (Almuzara et al. 2023, Higgins 2014) using real-time data from 2006:Q2 to 2024:Q4.

Methodology

Dynamic Factor Model Structure

• State-space model with stochastic volatility and variance outliers to capture time-varying uncertainty.
• Four common factors and one temporary Covid-specific factor.
• GDP components are combined via Laspeyres weighting through a linear approximation of quarterly GDP growth.

GDP Nowcast Construction

• Uses national accounting identity to derive GDP growth from component-specific nowcasts, incorporating nominal shares and prior period weights.
• Delivers full probability density forecasts through Bayesian estimation and Gibbs sampling.

Applications and Examples

• 2025:Q1 Example: CBDF had a GDP nowcast (-3.16) compared to 2.58 for the Fed’s current model.
• Component Impact: High volatility in inventories contributes disproportionately to forecast errors despite its small nominal share of GDP.

Evaluation

• Real-Time Testing 2006:Q2–2024:Q4:
  • Point forecast RMSE improved by ~15%.
  • Density forecast log-predictive scores improved by ~20%.
• Baselines: Outperforms the New York Fed’s DF model (15% RMSE improvement) and Atlanta Fed’s GDPNow model (point accuracy comparable to professional forecasts like SPF/BlueChip).

Conclusion

The CBDF model offers significant improvements in accuracy and transparency by integrating accounting constraints and joint component modeling. It provides tools for early quarterly GDP estimates and uncertainty quantification, addressing limitations of both DF and bottom-up approaches.

JEL Classification: C32, C38, C53
Keywords: dynamic factor models, GDP nowcasting, national accounting identity.