Project 02 / Quantitative finance
Modelling Systemic Risk with ΔCoVaR
Comparing linear and neural-network quantile regression.
PythonNeural networksQuantile regressionRisk modelling
01
Overview
A quantitative-finance research project comparing linear quantile regression and neural-network quantile regression for estimating ΔCoVaR, a measure of an institution’s contribution to systemic financial risk.
02
Focus areas
- Compared linear and neural-network approaches to ΔCoVaR estimation
- Worked with weekly data spanning 2003–2023 across US and euro-area banks
- Examined trade-offs between historical fit and forward-looking prediction
03
Research overview
Topic. Measuring systemic risk within the banking sector using ΔCoVaR.
Primary goals
- Explore a Neural Network Quantile Regression (NNQR) approach to calculate the systemic-risk metric ΔCoVaR.
- Build a predictive Forward ΔCoVaR model using lagged financial variables.
- Provide practical, real-world interpretations for prospective banking and financial applications.
04
Key concepts
- Systemic risk
- Financial interconnectedness where a single initial shock spreads regionally or globally, often from a too-big-to-fail institution or a group of highly correlated smaller institutions.
- Value at Risk (VaR)
- The potential maximum loss for a financial institution within a specified confidence interval over a given timeframe.
- CoVaR & ΔCoVaR
- Introduced by Adrian and Brunnermeier (2011), CoVaR is the VaR of the entire financial system conditional on an institution being in distress. ΔCoVaR measures the difference between the system’s CoVaR in distress and in its normal state, isolating an institution’s marginal contribution to systemic risk.
05
Core problem & methodology
The traditional approach relies on Linear Quantile Regression (LQR). It assumes linear relationships, requires manual specification of non-linear terms, and can oversimplify complex financial interactions. This research proposes NNQR to overcome those limitations.
The dataset
- Scope. 1,284 banks: 632 active, 622 inactive since 2006, and 30 rest across the United States and the Euro Area.
- Timeline. Weekly data from December 26, 2003 to December 29, 2023, covering 1,045 weeks.
- Data sources. Market and macro indicators such as VIX, S&P 500, and US Treasury yields, plus firm metrics including market capitalisation, leverage risk ratio, maturity mismatch, and Tier 1 ratio, collected via Bloomberg and BankFocus.
06
Results & comparison
- ΔCoVaR calculations. NNQR shows slightly better performance than LQR when mapping historical ΔCoVaR values, especially for commercial banks during the Great Financial Crisis.
- Forward ΔCoVaR. A panel regression on lagged institutional characteristics was used to reduce procyclicality.
- Best lag. Variables lagged by one year yielded the strongest predictive results.
- Contradictory outcome. For currently active banks, LQR unexpectedly outperformed NNQR in predictive accuracy. Firm-level variables also did not fully capture ΔCoVaR, potentially reflecting a too-small-to-care effect or moral hazard among exceptionally strong balance sheets.
07
Future improvements
Several directions could improve the predictive capacity of the models:
- Integrate European market-specific variables, since the current variables are US based.
- Use MICE (Multiple Imputation by Chained Equations) interpolation.
- Split the growth of market-valued total assets Xti into separate positive and negative regimes, as proposed by López-Espinosa et al.
- Replace panel regression with advanced machine-learning models such as XGBoost.
- Incorporate direct data on financial interconnectedness, although it is difficult to obtain.
Continue exploring