BIS国际清算银行-Answering-the-Queen_-Machine-learning-and-financial-crises_66页_2mb

Summary of "Answering the Queen: Machine Learning and Financial Crises"

Core Content

This paper explores the use of machine learning techniques, specifically online or sequential prediction frameworks, to forecast systemic financial crises. The authors argue that traditional models and methods are insufficient for predicting financial crises due to the complexity and non-linear nature of such events, and that macroprudential policies need better predictive tools to be effective in mitigating their impact.

The paper introduces a meta-statistic approach that combines multiple predictive models ("experts") to improve forecasting accuracy. It emphasizes the importance of early warning indicators and the need for flexible, dynamic models that can adapt over time, as financial crises are often unpredictable and their mechanisms are not fully understood.

Main Viewpoints

• Financial crises have severe economic, social, and political consequences.
• Macroprudential policies are increasingly used to enhance financial system resilience but are under-researched compared to monetary and fiscal policies.
• Traditional forecasting methods such as panel data econometrics and event studies are limited in their ability to predict financial crises early and accurately.
• Machine learning offers a more flexible and robust approach, especially in the context of out-of-sample forecasting.
• The paper advocates for the use of online machine learning, which allows for time-varying weights on different models and does not assume a specific data generation process.

Key Information

1. Methodology

• The authors use sequential prediction, also known as online machine learning, to forecast systemic financial crises.
• This method allows for dynamic model aggregation and can incorporate a wide range of predictive models.
• The framework is "meta-statistic" in that it does not assume a particular model but instead combines the predictions of multiple models to improve accuracy.

2. Crisis Definition and Data

• The systemic crisis is defined using data from the European Central Bank and the European Systemic Risk Board, which combines quantitative indicators and expert judgment.
• The paper focuses on predicting pre-crisis periods (12 quarters before a crisis occurs).
• A total of seven countries are analyzed: France, Germany, Italy, Spain, Sweden, the United Kingdom, and the United States.

3. Variables Used

• A large set of macroeconomic and financial variables is used to build the predictive models, including:
  • Debt, GDP, unemployment, investment, credit, interest rates, monetary aggregates, asset prices, sentiment proxies, commodity prices, housing prices, and external imbalances.
  • These variables are selected based on classical financial crisis literature, such as the works of Kindleberger (1978), Minsky (1986), and Diaz-Alejandro (1985).

4. Real-Time vs. Quasi-Real-Time Data

• Quasi-real-time data includes a wide range of variables, including those that are revised.
• Real-time data is more limited, using only vintage data and excluding some variables that are not available in real-time.
• The paper presents real-time forecasts for France and the United Kingdom, using a subset of the variables.

5. Limitations of the Approach

• The online learning framework has limitations, particularly in its inability to predict new types of crises that have not occurred historically.
• It relies on historical data, which may not capture novel or unprecedented events like cyber-attacks or pandemics unless their correlates with past crises are known.

Conclusion

• The authors conclude that the online machine learning framework is a promising tool for forecasting financial crises.
• It allows for dynamic model selection and flexible forecasting, which is essential given the unpredictable and complex nature of financial crises.
• The methodology is robust, transparent, and informative, offering insights into which models perform best at different times and which variables are most predictive.

Key Contributions

• The first application of online machine learning to financial crisis forecasting.
• A comprehensive dataset of systemic crisis episodes and macroeconomic indicators.
• A flexible and adaptive framework for early warning and policy anticipation.
• A meta-statistic approach that can incorporate a variety of models and variables.
• Diagnostics and model comparisons that highlight the time-varying nature of crisis prediction.

References to Classic Literature

• The paper draws heavily on classical financial crisis theories, including:
  • Kindleberger's theory of manias, panics, and crashes.
  • Minsky's financial instability hypothesis.
  • Fisher's debt-deflation theory.
• These theories emphasize the role of credit booms, debt, and financial overvaluation in the buildup of systemic risk.

Methodology and Models

• The paper uses model averaging with time-varying weights.
• It includes several types of models, such as:
  • Bayesian averaging models.
  • Machine learning models (e.g., random forests, support vector machines, general additive models).
  • Elastic-net logits.
• The Shapley values are used to provide economic interpretations of the models' performance.

Conclusion

The paper concludes that the online learning framework is a valuable tool for forecasting systemic financial crises, offering greater flexibility and adaptability compared to traditional methods. It highlights the importance of early warning indicators and the need for a more dynamic and comprehensive approach to financial stability analysis.