Stress Testing: Work Backward From the Loss You Cannot Absorb

Stress testing asks what a specified adverse market state would do to the portfolio. Reverse stress testing asks which market states could breach a defined loss or viability limit. Both are useful because a statistical risk measure fitted to recent observations may say little about concentrations that become dangerous only when several assumptions fail together.

REFERENCE · NO CALL

This report combines historical-scenario and reverse-stress teaching material from September 2026. The desk, shocks and losses are an illustrative case. Historical episode labels describe scenario shapes; the displayed shocks should not be presented as independently verified returns over specific historical dates. The calculations are conditional examples, not market forecasts.

Define the failure being investigated

An economic stress programme begins with a question. It might test whether a trading desk can absorb a mark-to-market loss, whether margin calls exceed available liquidity or whether a funding model remains viable after a market closure. A large loss alone is not a complete failure definition.

The distinction affects the model. A solvent portfolio may still fail through liquidity exhaustion if collateral calls arrive before assets can be sold. A profitable long-term position can generate a short-term capital breach. A stress test that models only final valuation changes cannot answer a question about the path of cash requirements without additional machinery.

For the worked market-risk case, the initial objective is current-book valuation under ten-day factor shocks. The reverse exercise then searches for states producing a ₹60 crore loss, defined as an illustrative allocated-capital threshold. That threshold is a teaching input rather than an identified institution’s risk appetite.

Apply scenario moves to the current book

Historical stress should take a historical or historically shaped joint factor move and apply it to the portfolio being assessed. It does not simply report what a different portfolio lost during the original event.

The example book contains a ₹300 crore equity basket with beta 1.28, a short index-put position and ₹120 crore of government bonds. The source uses the following scenario inputs:

Retained joint shocks for the illustrative scenario shapes — Scenario shape: Global financial crisis style

Retained joint shocks for the illustrative scenario shapes — Scenario shape: Pandemic-dislocation style

Retained joint shocks for the illustrative scenario shapes — Scenario shape: Inflation and rates repricing style

Volatility points and basis points are different units. A rise from 14% to 36% volatility is 22 volatility points, not a 22% relative increase. The distinction is essential when applying vega or reconstructing a volatility surface.

The scenario should preserve an economically coherent combination of factors, while recognising that historical dependence is not a physical law. Deliberately testing a breakdown in historical correlation can be valuable if the rationale is explicit. The objective is not to forbid unfamiliar combinations; it is to avoid arbitrary combinations masquerading as a sourced historical episode.

The composition matters as much as the total

The original full-revaluation exercise reports:

Source full-revaluation components and total losses — Scenario shape: Financial crisis style

Source full-revaluation components and total losses — Scenario shape: Pandemic style

Source full-revaluation components and total losses — Scenario shape: Rates repricing style

The numbers are retained source-example outputs. In the crisis-style case, the bond gains as yields fall. In the rates scenario, the same bond loses alongside the equity positions. A claim that the bond hedges the equity book is therefore incomplete without specifying the scenario and risk horizon.

The short-put losses also reveal why premium received is a poor risk-size measure. A position can collect a relatively small initial premium while carrying a much larger loss in an adverse state. The relevant exposure depends on the payoff and the path to that state, not just on the cash received at inception.

Revaluation error grows where precision matters

The teaching case compares full revaluation with a delta-gamma-vega approximation:

Full-revaluation and approximation losses in the source example — Scenario shape: Financial crisis style

Full-revaluation and approximation losses in the source example — Scenario shape: Pandemic style

Full-revaluation and approximation losses in the source example — Scenario shape: Rates repricing style

Those differences are about 11.7%, 10.7% and 5.2% of the full-revaluation losses. They demonstrate the error in this source example; they do not establish that every sensitivity approximation always understates losses. The sign depends on the payoff and shock.

Full revaluation is a stronger benchmark only when its model, market inputs and contractual treatment are themselves appropriate. Calling a computation full revaluation does not make it an economic truth. Validation should examine exercise features, barriers, surface extrapolation, cash-flow changes and model behaviour outside the calibration region.

A practical control is to measure approximation error across a ladder of increasing shocks. The output identifies where a fast sensitivity engine remains usable and where a different valuation route or explicit overlay is required.

Do not interpret a stress-to-VaR multiple as a probability

The pandemic-style loss is about 20.7 times the case’s one-day 99% VaR of ₹8.37 crore. That ratio highlights a difference in scale, but it compares a ten-day adverse scenario with a one-day percentile forecast. It is not a direct backtest, a return period or proof that the VaR is wrong by a factor of 20.7.

VaR and stress answer different questions. Their joint use is valuable precisely because neither replaces the other. A risk committee should see the horizon and conditioning assumptions alongside any headline multiple.

Reverse the direction of the exercise

Let $\mathbf{x}$ be a vector of market shocks and $L(\mathbf{x})$ the resulting portfolio loss. A basic reverse-stress problem searches for a scenario satisfying

$$L(\mathbf{x})\geq L^{*},$$

where $L^{*}$ is the failure threshold. Many scenarios may satisfy the inequality, so the search needs a criterion for comparing them.

The source uses Mahalanobis distance:

$$d^2(\mathbf{x})=(\mathbf{x}-\boldsymbol\mu)^{\mathsf T} \Sigma^{-1}(\mathbf{x}-\boldsymbol\mu),$$

In words: Measure the joint shock relative to its reference mean, accounting for each factor’s scale and how the factors move together.

where $\boldsymbol\mu$ is the reference mean shock and $\Sigma$ its covariance matrix over the chosen horizon. The method finds a low-distance point on the loss boundary. It accounts for factor scale and correlation, so several jointly consistent moves may require less distance than one isolated extreme shock.

The covariance matrix must be well conditioned and economically appropriate. Strongly collinear factors can make inversion unstable. Historical correlation can also be a poor guide in stress. Alternative covariance estimates, constraints and scenario narratives should therefore be tested rather than treating one optimiser result as definitive.

The retained reverse-stress result

For the ₹60 crore loss threshold, the source’s search produces:

Retained reverse-stress solution at the teaching threshold

The equity-only route requires an approximately 11.94% decline with a quoted joint-distance measure of 4.988 under the same construction. The lower-distance joint route combines falling equities with rising volatility and a small bond rally. It does not require every factor to move adversely in isolation.

Under a common multivariate-normal density assumption, the distance difference corresponds to a point-density ratio of approximately 1,407. The source describes this as a relative-likelihood comparison. It must not be reported as a ratio of event probabilities or as evidence that the scenario will occur. Points in a continuous distribution have zero probability; probability depends on regions, model assumptions and the full distribution.

Extend the exercise to actions and path dependence

A useful reverse-stress output identifies the exposure that makes the failure region accessible. That could be short volatility across several desks, a concentration in one funding channel or a hedge that depends on a correlation likely to weaken.

Management actions should be modelled with constraints.

Management actions retain execution and funding constraints

The analysis should distinguish actions available before the event from those assumed after it begins.

For nonlinear books, the optimiser may find a local solution. Multiple starting points, alternative shock parameterisations and independent repricing of candidate scenarios improve confidence that the result is not merely a numerical artefact. Failure modes outside the selected factors still remain possible.

What the report should change

The final stress report should connect each scenario to a decision: exposure limit, liquidity buffer, hedge redesign, contingency action or further investigation. It should show the loss decomposition, valuation method, assumptions and operational feasibility of proposed responses.

Historical scenarios challenge assumptions already associated with known episodes. Reverse stress challenges the analyst’s scenario selection itself. Together they make the portfolio’s failure mechanisms more visible without pretending to assign reliable probabilities to every extreme state.