Standard and Stressed Value at Risk Forecasting Using Dynamic Bayesian Networks

Battle Count: 86

The paper is primarily focused on regulatory risk management (Basel Accords compliance) rather than alpha generation or trading strategy development. However, the DBN framework for forward-looking return forecasting and the comparison of volatility models (GARCH, EGARCH, RiskMetrics) are relevant to quantitative risk management within trading desks. The causal inference approach using 41 macroeconomic/financial variables could inform systematic trading signals, though the paper does not explore this application.

High complexity due to: (1) training three DBN learning algorithms with 41 variables using rolling period methodology over 7,286+ trading days; (2) handling missing data and variable frequency mismatches (daily to quarterly); (3) implementing multiple traditional models with two distributional assumptions; (4) performing comprehensive backtesting (traffic light, Kupiec, Christoffersen) and forecasting error analysis; (5) constructing conservative stressed periods for SVaR. The dbnR package in R helps but the overall pipeline is substantial.

The paper uses R with the dbnR package for DBN training and forecasting. However, the primary data source is Bloomberg (proprietary), and the specific rolling period parameters, variable transformations, and network structures are described but not fully reproducible without access to the same data. The 41 variables are listed in Appendix A. No code repository is provided.