Paper: arXiv 2512.12334
Authors: Eden Gross, Ryan Kruger, Francois Toerien
Abstract
In the last five years, expected shortfall (ES) and stressed ES (SES) have become key required regulatory measures of market risk in the banking sector, especially following events such as the global financial crisis. Thus, finding ways to optimize their estimation is of great importance. We extend the application of dynamic Bayesian networks (DBNs) to the estimation of 10-day 97.5% ES and stressed ES, building on prior work applying DBNs to value at risk. Using the S&P 500 index as a proxy for the equities trading desk of a US bank, we compare the performance of three DBN structure-learning algorithms with several traditional market risk models, using either the normal or the skewed Student’s t return distributions. Backtesting shows that all models fail to produce statistically accurate ES and SES forecasts at the 2.5% level, reflecting the difficulty of modeling extreme tail behavior. For ES, the EGARCH(1,1) model (normal) produces the most accurate forecasts, while, for SES, the GARCH(1,1) model (normal) performs best. All distribution-dependent models deteriorate substantially when using the skewed Student’s t distribution. The DBNs perform comparably to the historical simulation model, but their contribution to tail prediction is limited by the small weight assigned to their one-day-ahead forecasts within the return distribution. Future research should examine weighting schemes that enhance the influence of forward-looking DBN forecasts on tail risk estimation.
Complexity vs Empirical Score
- Math Complexity: 8.5/10
- Empirical Rigor: 7.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced stochastic processes (GARCH/EGARCH models) and Bayesian network structure learning, requiring significant mathematical sophistication. However, it demonstrates high empirical rigor through rigorous backtesting of multiple models on real market data (S&P 500) over three decades, evaluating regulatory compliance with Basel standards.
Research Flowchart
flowchart TD
A["Research Goal:<br>Extend Dynamic Bayesian Networks<br>to estimate Stressed ES & ES"] --> B{"Data & Models"}
B --> C["S&P 500 Index Returns"]
B --> D["Models:<br>3 DBN Algorithms, GARCH(1,1),<br>Historical Simulation"]
B --> E["Distributions:<br>Normal vs. Skewed Student's t"]
C & D & E --> F["Computational Process:<br>10-day 97.5% ES & SES Estimation<br>& Backtesting"]
F --> G["Key Findings/Outcomes"]
G --> H["ES: EGARCH(1,1) Normal<br>most accurate"]
G --> I["SES: GARCH(1,1) Normal<br>performs best"]
G --> J["All models fail ES/SES<br>backtesting at 2.5% level"]
G --> K["Skewed Student's t<br>deteriorates performance"]
G --> L["DBNs: Limited tail impact<br>due to low forecast weighting"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper extends dynamic Bayesian networks to estimate 10‑day 97.5% expected shortfall and stressed expected shortfall.
Source sentence
We extend the application of dynamic Bayesian networks (DBNs) to the estimation of 10-day 97.5% ES and stressed ES
Assumes 6
- The S&P 500 index proxies the equities trading desk of a US bank. market
Source sentence
The S&P 500 index is used as a proxy for the equities trading desk of a US bank
- Return distributions are either normal or skewed Student’s t. model
Source sentence
using either the normal or the skewed Student’s t distribution as the underlying return distribution
- DBN structure is selected by minimizing AIC on each rolling training window. model
Source sentence
each training iteration was the one that minimized the Akaike information criterion (AIC)
- Non‑daily variables are converted to daily by carrying forward the last known value. data
Source sentence
non-daily data were converted to daily by carrying forward the last known value
Show 2 more
- Conditional backtest assumes ES forecasts are produced after corresponding VaR forecasts. statistical
Source sentence
The conditional backtest (Acerbi & Székely, 2014) rests on the assumption that ES forecasts are produced and tested after their corresponding VaR forecasts have been produced and tested.
- Stressed period is defined by a rolling methodology aggregating the most severe S&P 500 return days. other
Source sentence
the stressed period is determined using a rolling period methodology as the amalgamation of the most severe return days of the S&P 500 index
Data
1991-03 – 2020-02 · S&P 500 index · equities · daily
- S&P 500 daily closing values market-prices commercial
Source sentence
We use daily closing values of the S&P 500 index from 15 March 1991 to 14 February 2020, obtained from the Bloomberg database
Costs
Not applicable
Finds 5
- For ES, the EGARCH(1,1) model with normal distribution yields the most accurate forecasts. predictive-accuracy
Source sentence
For ES, the EGARCH(1,1) model (normal) produces the most accurate forecasts
- For SES, the GARCH(1,1) model with normal distribution performs best. predictive-accuracy
Source sentence
for SES, the GARCH(1,1) model (normal) performs best
- All models, including DBNs, fail statistical backtests at the 2.5% level, indicating inaccurate ES and SES forecasts. statistical
Source sentence
... forecasts were statistically inaccurate at the 97.5% confidence level
- DBNs perform comparably to the historical simulation model but have limited tail‑prediction contribution. performance
Source sentence
The DBNs perform comparably to the historical simulation model, but their contribution to tail prediction is limited
- Models using the skewed Student’s t distribution deteriorate substantially compared to normal distribution. performance
Source sentence
All distribution-dependent models deteriorate substantially when using the skewed Student's t distribution
Checked by others
Contradicted by
- Forecasting Probability Distributions of Financial Returns with Deep Neural Networks
Compare the findings
This paper: Models using the skewed Student’s t distribution deteriorate substantially compared to normal distribution.
That paper: The LSTM with skewed Student’s t distribution performs best across multiple evaluation criteria, capturing heavy tails and asymmetry.
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.