2023-03-16-IMF-替代数据模型_解释大规模机器学习危机预测模型（英）-_31页_2mb

Summary

Introduction

The paper addresses the challenge of predicting economic crises using machine learning (ML) models, which offer high predictive power but suffer from low interpretability due to high dimensionality and numerous features. Surrogate data models (SDMs) are proposed as a dimensionality reduction tool to enhance interpretability, facilitating policy guidance and scenario analysis.

Problem

Economic crises cause severe economic damage, and predictive models are crucial for early warning. However, many ML crisis prediction models use a large number of features, complicating interpretation and limiting their use in crisis prevention and mitigation policies.

Proposed Solution

SDMs simplify ML models by using fewer, interpretable features, based on economic theory and country surveillance variables to reduce complexity. This approach aims to explain model outputs and support policymakers in understanding crisis drivers.

Methodology

• Feature Selection: Variables selected are based on economic relevance and monitorability, with fewer than those in full ML models (e.g., 12–20 vs. hundreds).
• Estimation: Models like country-group and global random forests are used. Hyperparameters are optimized, and SDM outputs are approximated using dimensionality reduction techniques.
• Interpretability: SHAP values are employed to quantify feature contributions to crisis risk indices.
• Application: Tested on IMF's ML crisis prediction models for fiscal, external, and real sectors, demonstrating consistent results with economic intuition.

Results

SDMs effectively captured crisis dynamics across sectors and country groups with reasonable accuracy. Good performance was noted in fiscal crises (high correlation), though other sectors had limitations due to omitted variables. SHAP-based analysis clarified key drivers, such as low growth and tighter financial conditions.

Conclusion

SDMs are effective and intuitive tools for interpreting large-scale ML crisis prediction models, enhancing policy relevance. Their widespread adoption depends on careful application on a case-by-case basis, considering data limitations and interpretability trade-offs in surveillance work.