1. Research Overview & Context

• Topic: Measuring systemic risk within the banking sector using ∆CoVaR.
• Primary Goals:
  • To explore a Neural Network Quantile Regression (NNQR) approach to calculate the systemic risk metric ∆CoVaR.
  • To build a predictive Forward ∆CoVaR model using lagged financial variables.
  • To provide practical, real-world interpretations of the findings for prospective banking/financial applications.

2. Key Concepts Defined

• Systemic Risk: A phenomenon of financial interconnectedness where a single initial shock spreads regionally or globally. It typically stems from a single "too big-to-fail" institution or a group of smaller institutions exhibiting highly correlated behaviors.
• Value at Risk (VaR): Quantifies the potential maximum loss for a financial institution within a specific confidence interval over a given timeframe.
• CoVaR & ∆CoVaR: Introduced by Adrian and Brunnermeier (2011), CoVaR is the VaR of the entire financial system conditional on a specific institution being in distress. ∆CoVaR measures the difference between the system's CoVaR when an institution is in distress versus when it is in its normal state, thereby isolating that specific institution's marginal contribution to systemic risk.

3. Core Problem & Methodology

The traditional approach relies on Linear Quantile Regression (LQR), which introduces limitations such as assuming linear relationships, requiring manual specification of non-linear terms, and oversimplifying complex financial interactions. The thesis proposes NNQR to overcome these limits.

The Dataset

• Scope: 1,284 banks (632 active, 622 inactive since 2006, 30 rest) across the United States and the Euro Area.
• Timeline: Weekly data stretching from December 26, 2003, to December 29, 2023 (1,045 weeks).
• Data Sources: Market/macro indicators (e.g., VIX, S&P500, US Treasury yields) and specific stock/firm metrics (e.g., market cap, leverage risk ratio, maturity mismatch, Tier 1 ratio) collected via Bloomberg and BankFocus.

4. Key Results & Model Comparison

•  ∆CoVaR Calculations: The NNQR approach shows slightly better performance than LQR when mapping out the historical  ∆CoVaR values, especially when evaluating commercial banks during the Great Financial Crisis (GFC).
• Forward  ∆CoVaR (Predictive Model): To combat procyclicality, a panel regression was used on lagged institutional characteristics.
  • Variables lagged by 1 year yielded the best predictive results.
  • Contradictory/Unexpected Outcome: For the forward-looking predictive framework of currently active banks, LQR unexpectedly outperformed NNQR in terms of accuracy. Also most firm-level variables did not fully capture the  ∆CoVaR values, which could be due to a "too small to care" effect or moral hazard among banks with exceptionally strong balance sheets.

5. Potential Future Improvements

To refine the predictive capacity of the models, the research highlights several areas for next steps:

• Integrating European market-specific variables, as the currently used ones are US based and might not be enough explanatory.
• Utilizing MICE (Multiple Imputation by Chained Equations) interpolation.
• Splitting the growth of market-valued total assets $X_t^i$ into separate positive and negative regimes (as proposed by López-Espinosa et al.)
• Swapping out panel regression in favor of advanced machine learning models like XGBoost.
• Incorporating direct data on financial interconnectedness although very difficult to obtain.