💰 quantitative finance

Forward-Looking Stress Testing Under Macro Scenarios: Stable SVaR Estimation Using a Hybrid GPR-HS Framework with SACS

This paper proposes a stable and regulator-aligned framework for forward-looking Stressed Value-at-Risk (SVaR) estimation under macroeconomic stress scenarios by extending the Hybrid GPR-HS model with a Scenario-Averaged Covariance Stabilization (SACS) approach to ensure numerical robustness and coherent capital projections for CCAR and ICAAP applications.

Original authors: Ujjwala Vadrevu

Published 2026-06-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ujjwala Vadrevu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the captain of a massive ship (a bank) sailing through the ocean of the global economy. Your job is to make sure you have enough lifeboats and emergency supplies (capital) to survive if a giant storm hits.

This paper is about building a super-reliable weather forecast and survival calculator for those storms. Here is how the author, Ujjwala Vadrevu, explains the new system using simple ideas:

1. The Problem: Old Maps Break in Storms

Usually, banks try to predict storm damage using old math formulas (like GARCH models). The author says these old formulas are like paper maps. They work fine on a calm day, but if you try to use them during a hurricane, they tear apart, give you nonsense directions, or simply stop working. This is dangerous because if your map breaks, you might think you're safe when you're actually sinking.

2. The Solution: A "Smart Compass" (The Hybrid GPR-HS Framework)

The author introduces a new tool called a Hybrid GPR-HS Framework. Think of this as a smart, self-correcting compass that doesn't rely on rigid rules.

How it works: Instead of forcing the data into a straight line, it uses a flexible "rubber sheet" (Gaussian Process Regression) that bends to fit the shape of the storm.
The Secret Sauce (ANI): The author adds a special trick called "Aggressive Noise Initialization." Imagine setting your compass to assume the storm is already violent before you even see the first wave. This forces the computer to be conservative. It's better to overestimate the danger and have too many lifeboats than to underestimate it and run out of supplies.

3. The Three Storms We Are Testing

The paper tests this new compass against three specific, scary "what-if" scenarios happening right now (as of 2026):

The West Asia War: A big fight that disrupts oil and trade. It's like a sudden, heavy fog and rough waves hitting the European and Asian parts of the ship hardest.
Climate Risk: Massive floods and heatwaves. This is like a slow, grinding damage to the ship's hull, affecting different parts of the world differently (like flooding in Asia but less in the US).
The AI Bubble: A tech crash followed by strict new laws. This is like a sudden explosion in the engine room (the tech sector), causing the whole ship to shake, especially the US part of the fleet.

4. The "Stabilizer" (SACS)

When a storm hits, the relationship between different parts of the ship changes. Usually, if one part sinks, others might stay afloat. But in a crisis, everything tends to sink together.

The author uses a tool called SACS (Scenario-Averaged Covariance Stabilization).
The Analogy: Instead of guessing how the ship parts will react right now (which is chaotic and confusing), the tool looks at how the ship reacted during three past major shipwrecks (like the 2008 crisis). It averages those past reactions to create a stable, reliable rule for how the ship behaves during a new storm. This prevents the calculator from panicking and giving wild, unstable numbers.

5. The Results: The Ship is Ready

The author ran the simulation for a full year (252 days) under these three storms.

Did the compass break? No. The old paper maps would have torn, but this new smart compass stayed steady and gave clear numbers for every single scenario.
How much damage? The worst-case scenario (West Asia War) showed the ship could lose about 2.2% of its value in a single day of extreme stress. The Climate Risk scenario was slightly less severe (about 2.1%).
Is it safe? Yes. The system proved that if you prepare for the worst-case loss (SVaR), you will automatically have enough to cover even the deeper, more catastrophic losses (SES).

The Bottom Line

This paper proves that you can build a digital stress-test engine that doesn't break when the world gets scary. By using a "conservative-first" approach (assuming the worst) and learning from past shipwrecks to predict future ones, banks can calculate exactly how much money they need to stay afloat, satisfying strict government rules (like CCAR and ICAAP) without the math collapsing under pressure.

Technical Summary: Forward-Looking Stress Testing Under Macro Scenarios

Problem Statement
Regulatory frameworks such as the Comprehensive Capital Analysis and Review (CCAR) and the Internal Capital Adequacy Assessment Process (ICAAP) mandate the estimation of Stressed Value-at-Risk (SVaR) and Stressed Expected Shortfall (SES) under forward-looking macroeconomic scenarios. A critical technical challenge exists in the computational production of these estimates: traditional parametric models (e.g., GARCH-family, EWMA) frequently exhibit numerical instability or model breakdown when conditioned on extreme negative return paths. This instability undermines the reliability and regulatory defensibility of capital projections. Furthermore, dynamically estimated correlation structures often become unstable under extreme conditions, failing to provide a robust representation of cross-asset dependencies during stress.

Methodology
This paper extends the Hybrid Gaussian Process Regression–Historical Simulation (GPR-HS) framework, previously introduced in Vadrevu (2026), to the domain of forward-looking stress testing. The methodology integrates three core components:

Stress-Specific GPR with Aggressive Noise Initialization (ANI):
For each of the seven global equity indices, a dedicated Univariate GPR model is fitted to a 252-day forward-looking stress path. To ensure numerical stability and conservative volatility estimation, the paper adapts the ANI strategy. Unlike base-case fitting, the White Noise kernel variance ( $\sigma_n^2$ ) is initialized to the full empirical variance of the stressed return series. This forces the Bayesian optimizer to begin in a region reflecting heightened uncertainty, preventing the model from interpreting extreme shocks as transient noise and reverting to low-volatility priors.
Scenario Construction:
Three distinct systemic risk scenarios are operationalized as ICAAP-style cumulative return shocks over a one-year horizon:
- West Asia War: A geopolitical conflict causing supply chain disruption and stagflation.
- Climate Risk: Physical climate damage (flooding, heatwaves) affecting agriculture and infrastructure.
- AI Bubble/Regulatory Burden: A technology sector bubble burst followed by stringent global regulation.
  Daily stressed return paths are generated using a deterministic linear drift component (calibrated to reach specific cumulative shocks by day 252) combined with historical stochastic noise to preserve realistic intraday volatility dynamics.
Scenario-Averaged Covariance Stabilization (SACS):
To address the instability of dynamically estimated correlations under stress, the paper introduces SACS. Instead of estimating correlations from the stressed path alone, the stress-period covariance matrix ( $\Sigma_{SACS}$ ) is constructed as a weighted aggregation of covariance structures observed across historical crisis regimes (e.g., 2008 GFC, COVID-19). The diagonal elements utilize the GPR-predicted conditional variances, while off-diagonal elements utilize the SACS-derived covariances.
Risk Calculation and Stability Assessment:
Terminal day (T=252) SVaR and SES are calculated analytically from the resulting conditional distribution under a Student's t-distribution assumption ( $\nu=5$ ). Since forward-looking scenarios lack realized data for traditional backtesting, model stability is assessed via the Log-Marginal Likelihood (LML). Consistent LML convergence across all fits serves as the primary diagnostic for numerical integrity.

Key Contributions
The paper identifies four primary contributions:

Stress Extension of Hybrid GPR-HS: It demonstrates that the ANI-stabilized GPR architecture maintains numerical stability under three distinct extreme macro scenarios, producing regulatory-compliant SVaR and SES projections.
Scenario Design Methodology: It develops an ICAAP-consistent framework for constructing forward-looking scenarios that combine deterministic drift with stochastic noise, ensuring paths are plausible rather than discontinuous.
Scenario-Averaged Covariance Stabilization (SACS): It introduces a stable framework for stress-period covariance construction by aggregating empirically observed crisis regimes, addressing the instability of dynamic correlation estimates.
Real-World Validation: It highlights the contemporaneous alignment between the modeled scenarios and ongoing global developments (specifically the West Asia conflict and AI regulatory enforcement), providing external context for the plausibility of the stress calibration.

Results
The framework was applied to a portfolio of seven global equity indices across the three scenarios:

Numerical Stability: All 21 stressed fits (7 indices $\times$ 3 scenarios) achieved successful LML convergence. No divergent or extreme LML scores were observed, confirming the GPR framework's stability under extreme inputs. The Hang Seng Index, noted for complex dynamics, achieved stable fits across all scenarios.
SVaR and SES Projections: The terminal day SVaR estimates ranged from -2.1020% (Climate Risk) to -2.2231% (West Asia War). The West Asia War scenario produced the highest capital impact, reflecting the depth of applied shocks.
Coherence: The fundamental coherent risk measure property $|SES| > |SVaR|$ was preserved across all scenarios, confirming compliance with Artzner et al. (1999).
Model Behavior: The results exhibited monotonic risk ranking consistent with shock magnitudes and maintained a conservative floor, preventing the model from collapsing to unrealistically low risk estimates.

Significance and Claims
The paper claims that the Hybrid GPR-HS framework, augmented with SACS and ANI, serves as a stable, interpretable, and regulator-aligned engine for forward-looking SVaR and SES estimation. Its primary significance lies in its ability to satisfy the three simultaneous requirements of stress testing: numerical stability, conservatism, and transparency.

The authors assert that the ANI strategy is a critical regulatory design choice, preventing model divergence in production environments (such as CCAR or ICAAP) where model failure is operationally unacceptable. By anchoring correlation estimates to historical crisis regimes via SACS, the framework provides a robust approximation of stressed dependencies that prioritizes stability over potentially unstable dynamic refinement. The paper concludes that this architecture is suitable for the 9-quarter projections required by CCAR and longer-horizon ICAAP assessments, offering a viable alternative to traditional parametric models that fail under extreme conditioning.