EBA欧洲银行-Session-2-A-new-approach-to-Early-Warning-Systems-for-smaller-European-banks-D.-Malikkidou2C20M.-BrC3A4uning2C20S.-Scalone_6页_308kb

Summary of "A New Approach to Early Warning Systems for Smaller European Banks"

Core Content

This paper introduces a new early warning system (EWS) specifically designed for smaller European banks, known as Less Significant Institutions (LSIs). The system leverages decision tree modeling to predict bank distress events, offering a more transparent and interpretable alternative to traditional statistical models. The approach is based on the Bank Recovery and Resolution Directive (BRRD), which allows for a broader and more accurate definition of distress, thereby enhancing the model's predictive power.

The EWS is built using three types of explanatory variables: bank-specific, banking-sector, and country-level macro-financial indicators. These variables are used to assess a bank's financial health and its exposure to systemic risks. The model is calibrated using a dataset of over 3,000 LSIs from the euro area, covering the period 2014Q4–2016Q1.

Main Viewpoints

• Expanded Distress Sample: Traditional models use a limited number of distress events, but this paper expands the sample by using the BRRD's criteria, including early interventions, capital breaches, and special administration, leading to a more robust model.
• Machine Learning Application: The paper employs decision tree techniques, particularly the C5.0 algorithm, which is known for its accuracy, efficiency, and ability to handle missing and noisy data.
• Interpretability: Decision trees provide a clear and interpretable structure, allowing supervisors to understand the logic behind the model's predictions and avoid "black box" issues.
• Asymmetric Misclassification Costs: The model assigns higher cost to Type I errors (missing a distress event) than Type II errors (false alarms), reflecting the importance of early detection in financial supervision.
• Model Performance: The EWS demonstrates strong in-sample and out-of-sample predictive performance, with an AUC of 0.95 and 0.92, respectively, and a Cohen’s Kappa of 0.89 and 0.80. The model also outperforms logistic regression in terms of robustness to missing data.

Key Information

Data and Variables

• Data Source: Over 3,000 LSIs in the euro area from 2014Q4 to 2016Q1.
• Explanatory Variables:
  • Bank-specific: Capital adequacy, profitability, credit risk, liquidity, and operational risk indicators.
  • Banking-sector: Sector structure, lending, leverage, and asset quality.
  • Country-level: Macro-financial indicators such as GDP growth, deficit-to-GDP ratio, and market risk measures.
• Preprocessing:
  • Data cleaning to remove incomplete or non-informative data.
  • Normalization and aggregation to ensure consistency across different accounting standards.
  • Correlation threshold of 0.9 used to remove redundant variables.
  • Use of boosting techniques to rank variable importance and select the top 20 variables for the final model, supplemented by expert judgment.

Methodology

• The model is developed using the CRISP-DM methodology, ensuring a structured and robust process.
• C5.0 Decision Tree Algorithm: Chosen for its accuracy and ability to handle missing data.
• Model Evaluation:
  • In-sample: AUC = 0.95, Cohen’s Kappa = 0.89.
  • Out-of-sample: AUC = 0.92, Cohen’s Kappa = 0.80.
  • The model is compared with logistic regression, showing better robustness to missing values.

Results

• The final decision tree has 19 nodes and 12 explanatory variables.
• Profitability is the primary indicator used to split the data into profitable and loss-making banks.
• For profitable banks, non-performing loans ratio and coverage ratio are key indicators.
• For loss-making banks, the model considers GDP growth, leverage ratio, equity exposures, liquidity coverage ratio (LCR), and membership in institutional protection schemes (IPS).
• The model successfully identifies distress events with low false positive and false negative rates, indicating high reliability.

Future Work

• Data Enrichment: Plans to expand the dataset and extend the prediction horizon to up to six months.
• Severity Classification: Development of a multi-class EWS to distinguish between mild, moderate, and severe distress, enabling more nuanced supervision.
• Back-testing: Further back-testing will be conducted to assess the model's robustness over time.

Conclusion

The proposed LSI-EWS provides a promising tool for financial supervisors to detect distress in smaller European banks early. By expanding the definition of distress and using decision tree models, the system offers improved accuracy, transparency, and interpretability, making it a valuable addition to the regulatory toolkit.