2013年-IMF国际货币组织全球_Understanding_DSGE_Filters_in_Forecasting_and_Policy_Analysis_23页_737kb

Summary of "Understanding DSGE Filters in Forecasting and Policy Analysis"

Core Content

This working paper by Michal Andrle (2013) introduces methods for decomposing unobserved quantities in Dynamic Stochastic General Equilibrium (DSGE) models into contributions from observed data. These techniques are particularly useful for understanding the role of observed variables in estimating structural shocks and unobserved variables such as the output gap. The paper emphasizes the importance of using filters, especially the Kalman filter, in the context of state-space models and explores their application in forecasting and policy analysis.

Main Points

• Observables Decomposition: The paper presents a method to decompose unobserved variables (e.g., output gap, technology shocks) into contributions from observed data (e.g., output, inflation, interest rates). This is referred to as observables decomposition and complements the traditional shock decomposition.

• Filter Representation: The state-space model is represented through a linear filter, which can be two-sided (non-causal) and time-invariant. The filter weights are determined by the Wiener-Kolmogorov formula, and the filter is expressed as:
  $$
  \mathbf{X}_{t|\infty} = \mathbf{O}(L)\mathbf{Y}_t
  $$
  where $\mathbf{O}(L)$ is the filter function.

• Finite Sample Implementation: In practice, the filter weights are time-varying for finite samples. The weights can be calculated using the Kalman filter and smoother, and the paper suggests that the finite sample version can be implemented with a small number of Kalman smoother runs, making it computationally efficient.

• Data Revisions and News Effects: The paper discusses how data revisions affect the estimates of structural shocks and unobserved variables. It provides a formula to quantify the impact of data revisions on the forecast:
  $$
  \mathbb{R}{t|T} = \sum{\tau = t_0}^{T} \mathbf{O}{\tau|(t|T)}(\mathbf{Y}\tau^A - \mathbf{Y}\tau^B)
  $$
  where $\mathbb{R}{t|T}$ is the revision vector. The paper also introduces the concept of news effects, which represent the impact of new data on the estimates of structural shocks:
  $$
  \mathbb{N}{t|T+1} = \sum{\tau = t_0}^{T+1} \widehat{\mathbf{O}}{\tau|T+1}(\mathbf{Y}\tau^T - \mathbf{Y}_\tau^{T+1})
  $$

• Handling Missing Observations: The paper addresses how missing data can be incorporated into the filter framework. It suggests that the Kalman filter can be adapted to handle missing observations by using a time-varying selection matrix $\mathbf{W}_t$ that maps the full data vector to the observed data. The model can then be augmented with these missing observations, and the decomposition can be carried out accordingly.

• Imposing Expert Judgment: The paper introduces a method to incorporate expert judgment into the filtering process. This is done by adding stochastic linear restrictions to the model, effectively treating them as dummy observations. The modified model includes both observed and restricted data, and the Kalman filter is applied to the augmented dataset. This allows for the imposition of subjective constraints on the estimates of unobserved variables.

Key Information

• Filter Types: The paper discusses both time-invariant and time-varying filters. Time-invariant filters are derived from the Wiener-Kolmogorov formula and are used for doubly-infinite samples, while time-varying filters are used for finite samples.

• Practical Implementation: The implementation of the filter and its decomposition is computationally feasible using the Kalman filter and smoother. The method requires a limited number of runs, depending on the number of data points or groups being analyzed.

• Applications: The paper provides examples using the Smets and Wouters (2007) DSGE model, specifically focusing on the flexible-price output gap and real marginal costs. These examples demonstrate how the decomposition can be used to better understand the impact of observed data on unobserved variables.

• Revisions and News Effects: The paper highlights the importance of analyzing how data revisions and new data releases affect forecast accuracy and structural shock identification. It shows that the effects of these revisions can be quantified and decomposed using the filter weights.

Structure of the Paper

1. Introduction: Introduces the concept of observables decomposition and its importance in DSGE models.
2. Theoretical Background: Explains the state-space representation of DSGE models and the role of filters in decomposing unobserved variables.
3. Missing Observations and Imposing Judgement: Discusses the handling of missing data and how expert judgment can be incorporated into the filtering process.
4. Examples and Applications: Applies the decomposition method to the Smets and Wouters (2007) model, demonstrating its practical utility.
5. Conclusion: Summarizes the key findings and implications of the proposed methods for forecasting and policy analysis.

Conclusion

The paper presents a comprehensive approach to analyzing DSGE models through observables decomposition, emphasizing the role of filters in capturing the contributions of observed data to unobserved variables. It provides practical implementation techniques, discusses the effects of data revisions and news, and illustrates how missing data and expert judgment can be integrated into the filter framework. The methods are applicable to both stationary and non-stationary models and are useful for policy analysis and forecasting.