Extending the Application of Dynamic Bayesian Networks in Calculating Market Risk: Standard and Stressed Expected Shortfall

Battle Count: 105

The paper is primarily focused on regulatory market risk measurement (Basel III/IV compliance) rather than trading strategy development. However, the findings on ES/SES forecasting accuracy, model performance comparisons, and the limitations of DBNs for tail risk are directly relevant to quantitative risk management within trading desks. The insights about distribution selection (normal vs. skewed Student's t) and model calibration are applicable to any quantitative trading operation that must manage tail risk. The paper's conclusion that no model achieves statistical accuracy at the 2.5% level is important for practitioners setting risk limits.

The DBN implementation requires structure learning over 41 variables with rolling training windows, using the dbnR R package. The traditional models (ARCH, GARCH, EGARCH, RiskMetrics) are standard but require careful calibration with both normal and skewed Student's t distributions. The stressed period construction methodology (amalgamation of most severe non-consecutive days) adds complexity. The backtesting framework involves four different statistical tests. Overall, the pipeline is moderately complex but uses well-established statistical methods and publicly available R packages.

The paper uses publicly available Bloomberg data (S&P 500 and 41 macroeconomic/financial variables), standard statistical models (ARCH, GARCH, EGARCH, RiskMetrics), and publicly available DBN algorithms (PC Stable, MMHC, SI-HITON-PC) via the dbnR R package. However, no code repository is provided, and the exact implementation details of the DBN weighting scheme and stressed period construction are described but not fully reproducible without the referenced Gross et al. (2025) paper. The rolling window methodology and AIC-based structure selection are described but specific parameter choices may need clarification.