纽约联储-估算中央银行的HANK（英）-2023.8-37页_1mb

Estimating HANK for Central Banks: A Toolkit and Forecasting Accuracy Assessment

This paper introduces a Sequential Monte Carlo (SMC) method for efficient online estimation of HANK models, developed by Acharya et al. (2023). The SMC algorithm allows repeated estimation of HANK models over time, which is computationally intensive due to large state spaces and the inclusion of inequality factors.

Key Contributions:

• Methodology: Describes an SMC approach parallelizable in Monte Carlo steps, enabling "online" estimation without restarting from scratch when new data or model changes occur. It also proposes an adaptive schedule for incremental posterior approximation to maintain effective sample size.
• Empirical Comparison: Uses the SMC toolkit to compare the out-of-sample forecasting accuracy of the Bayer et al. (2022) HANK model to the Smets-Wouters (2007) representative agent (RA) New Keynesian model. The "online" estimation leverages previous posterior draws to bridge across data vintages.

Findings:

• HANK forecasts for real activity variables (e.g., GDP, consumption growth) were notably inferior to RA NK forecasts. Consumption growth showed particularly disappointing results despite HANK's focus on incorporating inequality through heterogeneous household consumption policies.
• Investment growth forecasts showed similar accuracy for short horizons but worse performance for medium horizons in HANK. Inflation forecasts, however, were comparable between models.
• Differences in forecasting performance are suspected to arise partly from calibrated parameters in HANK affecting steady-state dynamics, which are computationally costly to estimate repeatedly.

Policy Implications:

• While RA NK models have proven effective in forecasting, HANK models remain valuable for policy analysis involving inequality. Further research is needed to improve HANK forecasting accuracy and computational efficiency.