欧洲央行-打开局部投影的黑匣子（英）-2025_67页_5mb

Summary of ECB Working Paper Series No 3105

Local projections (LPs) are widely used in empirical macroeconomics to estimate impulse response functions (IRFs) for policy interventions. However, they often function as "black boxes," making it unclear what drives their estimates. This paper introduces a decomposition method that expresses LP estimates as a sum of contributions from historical events, revealing how past intervention episodes shape the results.

The decomposition shows that weights, which are a key component of LPs, can be interpreted in two ways:

• As standardized and purified shocks, accounting for control variables.
• As proximity scores reflecting similarity between the current intervention and past events in the sample.

This approach extends naturally to machine learning methods, allowing for interpretability in non-linear models like Random Forests. By visualizing weights and cumulative contributions, researchers can assess the external validity and robustness of IRFs.

Key empirical findings include:

• Monetary Policy: Cholesky VAR shocks misinterpret 1970s stagflation, leading to the "price puzzle," while other methods highlight political monetary interventions in the 1970s.
• Fiscal Policy: Fiscal multipliers are dominated by events like World War II, raising concerns about external validity.
• Climate Shocks: Global temperature impacts on GDP are fragile, primarily driven by the 1964 Mount Agung eruption, but medium-term effects are more robust.
• Financial Shocks: Non-linear models show sparser weights, aiding interpretation, and highlight size-dependent effects.

The framework enhances transparency by connecting IRFs to historical data points, enabling better diagnosis and refinement of economic models. Future research could explore extensions to non-linear operators and finer control variable analysis.

Conclusion: Decomposing LP estimates provides deeper insights into their historical foundations, diagnostic power, and application in machine learning, promoting transparency in economic analysis.