4.5.4 Reverse Stress Testing

Basic Prerequisites: 03 · Scenario Construction and Linear Algebra (quadratic forms, positive-definite matrices).

---

1. Intuition & Practical Objective

Ordinary stress testing is forward: choose a scenario, compute the loss. Reverse stress testing flips the arrow: fix the intolerable outcome - a capital breach, insolvency, a $30M loss that wipes the firm - and ask for the smallest market shock that produces it. You are not asking “what happens if equities fall 40%?” but “what combination of moves, as small as possible, is enough to break me?” (BIS 2009, citing CRMPG III; now a supervisory requirement for large banks.)

The intuition: every portfolio has an Achilles’ heel - a direction in factor space where it is maximally exposed. Reverse stress testing finds that direction directly. It answers: what is the nearest market state that kills us, and how far away is it (measured in standard deviations)? If the fatal state is 2σ away, you have almost no margin; if it is 6σ, you are more robust - but the composition of the fatal shock tells you what to hedge even more than its size.

Formally (for a linear factor portfolio): we want the smallest shock vector $z$ (in standardized, Mahalanobis units) whose P&L reaches the fatal loss $-C$:

$$\underset{z}{\min }{z}^T{\Sigma }^{-1}z\text{s.t.}{\beta }^{T}z\le -C.$$

The “smallest” is measured by $z^T{\Sigma }^{-1}z$ - the number of combined standard deviations (Mahalanobis distance) the scenario represents, respecting the correlations among factors. The optimizer will naturally push the shock toward the factors where the portfolio has the largest exposure per unit of their risk, and along the most correlated direction.

The one-sentence essence. “Forward stress tests ask ‘how bad is this scenario?’; reverse stress tests ask ‘what is the smallest scenario that breaks the firm?’ - a forward test tells you the damage, a reverse test tells you the distance to ruin and exactly which risk factors are doing the killing.”

---

2. Mathematical Ground Truth & Derivations

Setup. Let factors be standardized (each $\sim N(0,1)$, correlation matrix $\Sigma$), and $\beta =({\beta }_1,\dots ,{\beta }_K)$ be sensitivities in dollars per one-standard-deviation move of each factor. Portfolio P&L is linear:

$$\Delta V(z)={\beta }^{T}z.$$

The reverse-stress optimization. Minimize the Mahalanobis norm $z^T{\Sigma }^{-1}z$ subject to ${\beta }^{T}z\le -C$. The constraint is linear and the objective is a convex quadratic (positive-definite ${\Sigma }^{-1}$), so the optimum is unique and given by the KKT first-order condition. Writing the Lagrangian $L=z^T{\Sigma }^{-1}z+\lambda ({\beta }^{T}z+C)$ and setting ${\nabla }_{z}L=2{\Sigma }^{-1}z+\lambda \beta =0$ gives $z=-\frac{\lambda }{2}\Sigma \beta$. Substituting into the binding constraint ${\beta }^{T}z=-C$ yields $\frac{\lambda }{2}=\frac{C}{{\beta }^T\Sigma \beta }$, hence:

$$z^{\ast }=-\frac{C}{{\beta }^T\Sigma \beta }\Sigma \beta ,\sqrt{z^{\ast T}{\Sigma }^{-1}{z}^{\ast }}=\frac{C}{\sqrt{{\beta }^T\Sigma \beta }}.$$

So the fatal scenario is a scalar multiple of $\Sigma \beta$ - the “worst direction” is not the biggest beta, but the direction found by rotating the beta vector through the covariance: factors that are correlated with a large-exposure factor contribute even when their own beta is small (because they move with it). The distance to ruin is $C/\sqrt{{\beta }^T\Sigma \beta }$ combined standard deviations.

Why correlation matters - and why stress changes the answer. The objective uses $\Sigma$ as the metric for “small.” In normal times correlations are low, so the fatal direction is spread across factors and the distance $C/\sqrt{{\beta }^T\Sigma \beta }$ is large (diversification inflates ${\beta }^T\Sigma \beta$). Under stress, correlations rise toward 1 (05 · Failure Modes), which inflates ${\beta }^T\Sigma \beta$ and shrinks the distance to ruin: the same $C$ is reached by a smaller combined shock. The reverse stress test is itself a place to plug in a stressed correlation matrix.

Relationship to coherent measures. Reverse stress testing shares its philosophy with ES rather than VaR: it asks about the tail shape and worst direction, not a single quantile. It is also the complement to the Artzner subadditivity argument - VaR underweights exactly the correlated-bad-states that reverse stress makes the target.

---

3. Computational Implementation - analytic reverse stress test

Stdlib only. Three factors (equity, credit, rates) with sensitivities in $M per 1σ and a $30M capital line. The optimal fatal shock and its distance to ruin are computed in closed form - first under normal correlations, then under stressed correlations (→0.8).

Read the output two ways. Direction: the fatal shock is dominated by equity ($-1.44\sigma$) and credit ($-1.46\sigma$), with rates barely moving ($-0.21\sigma$) - under normal correlations the fund’s Achilles’ heel is the equity+credit combination, not rates, even though rates has a beta. Distance: under normal correlations the fatal state is $1.87\sigma$ away; under stressed correlations it collapses to $1.28\sigma$. Stressed correlations make ruin a smaller shock away - the diversification that the “distance to ruin” depended on has evaporated. That shrinkage is a core message: measure reverse-stress distance with the stressed covariance, or you will believe you have more room than you do.

---

4. Failure Modes & First-Principles Breakdowns

1. Measuring distance with normal correlations. The most dangerous error. If $\Sigma$ in the objective is the low-correlation normal matrix, $C/\sqrt{{\beta }^T\Sigma \beta }$ overstates the distance to ruin - the exact regime you are stress-testing for is the one whose correlations make ruin nearer. Always recompute reverse-stress distance under a stressed $\Sigma$.
2. Linearity hides convex tails. The closed form assumes $\Delta V={\beta }^{T}z$. For option/convex books, the true P&L curve is quadratic/convex, and the fatal region may be reachable by smaller shocks than the linear distance suggests (gamma accelerates losses). Reverse stress on convex books needs the full revaluation, not the delta-only proxy.
3. One fatal scenario ≠ one fatal mechanism. The optimizer returns the nearest point; but “nearest in Mahalanobis distance” is a modeling choice. A slightly larger but differently-shaped shock (e.g. a pure liquidity-freeze with no factor move at all) may be the real killer even if it is not the metric-nearest. Reverse stress identifies the direction; judgment must confirm it is a plausible mechanism, not just a mathematical optimum.
4. The no-probability trap. “Fatal at 1.87σ” sounds comfortable - but σ here is a model standard deviation, not a guarantee. Fat tails mean the real probability of a 1.87σ joint event is far higher than the normal assumption implies (see EVT & Fat Tails). The distance is a metric, not a probability.

---

5. References

• BCBS, Principles for Sound Stress Testing Practices and Supervision (2009, CN14)
• Schuermann, Stress Testing Banks, IJCB (2014)
• Hull, OFOD
• McNeil, Frey & Embrechts, QRM

---

6. Connected Graph Bridges

• Back: 03 · Scenario Construction · Index Hub
• Continue: 05 · Failure Modes & Practice
• Sibling: Linear Algebra (quadratic forms) · Multivariable Calculus & Optimization (KKT)