Build combined scenarios and use probabilities only for the claim they support

Select scenarios from the ways the corporation’s outcome can change. Begin with individual drivers where they clarify the mechanism, then combine changes that can interact: rates and debt resets, exchange rates and collection, commodity prices and quantities, collateral values and drawable finance. Recalculate the account under each combination. Do not sum separately calculated “worst losses” as if their assumptions necessarily coexist, or rely on historical diversification after the scenario removes its operating cause.

A scenario is a conditional account, not a forecast merely because it has precise numbers. Separate an illustrative stress, a plausible planning case and a probability-weighted estimate. For a historical replay, apply the selected past changes to today’s positions and terms; yesterday’s portfolio loss is not today’s exposure. For a hypothetical stress, explain the changed drivers and why the combination is useful for this decision. To find a failure threshold, work backward from the unacceptable cash, value or permission result and solve for changes that would reach it; then examine their plausibility and available responses.

If the use requires a loss distribution, name its baseline, horizon, units and model. Generate losses by applying each modeled factor state to the same positions, including the nonlinear and performance conditions that matter, and attach supported probabilities. A historical sample uses an explicit observation window; a parameter model or simulation needs its distribution, dependence and calibration grounds. More simulated observations reduce sampling noise within the model; they do not validate its missing events or its dependence assumptions.

An expected loss averages those losses. A chosen percentile locates a tail boundary. A tail average describes losses within a specified tail. None is the maximum possible loss, the cash needed at every earlier date or a decision rule without an associated tolerance. Where probabilities are poorly supported, retain conditional scenarios and thresholds instead of assigning invented confidence. A richer statistical model is useful only when its additional grounds improve the receiving decision.

Check whether the measure could miss a consequential failure outside its selected dimensions. Low market volatility can coexist with a single-customer default, inaccessible group cash or an untested settlement route. A market-value model generally needs a separate dated-cash return before it can support a funding conclusion. FIN.2 supplies that return without requiring the exposure model to become the corporation’s entire cash forecast.