Scenario Analysis Under Deep Uncertainty
A utility is stress-testing its ERM framework against a plausible but severe scenario: a prolonged drought reduces hydroelectric output while simultaneously raising cooling-water constraints on a thermal plant, and a carbon price rises faster than planned. The risk team built a model with point estimates for each variable and produced a single 'expected loss' figure of $310 million. The board's risk committee notes that the single number hides the fact that the variables are correlated-drought and heat tend to co-occur-and that the carbon price path is deeply uncertain. A board member asks what the number would be under a worse but plausible combination, and the team cannot answer because the model was not built for that. The CRO wants a revised approach before the annual report.
1. Explain why a single expected-loss figure from point estimates is inadequate for this scenario, referencing correlation and deep uncertainty.
Read the worked answer
A single expected loss collapses several distinct things into one number and hides the structure that matters. First, correlation: drought and heat co-occur, so hydro loss and cooling constraints are positively dependent. Treating each variable's point estimate independently understates joint severity-the combined event is worse than the sum of separate averages. Second, deep uncertainty: the carbon price path is not a known distribution; it depends on policy, which is subject to discrete shifts. Assigning a point estimate or a tidy distribution implies knowledge the utility does not have and produces false precision. Third, tail focus: the board's question is about a worse-but-plausible combination, which is a tail question, not a mean question. An expected value is the wrong statistic for decisions about resilience and capital. The model should instead produce a range of outcomes under explicitly stated combinations, with the assumptions visible, so the committee can see how the loss changes as variables move together. The number should be accompanied by the combination that produces it.
2. Recommend a revised approach that addresses correlation and deep uncertainty without overstating precision.
Read the worked answer
Use scenario combination and stress framing rather than a single probabilistic output. Step one: define a small set of coherent scenarios (e.g., moderate, severe, severe-plus-accelerated-carbon-price) where variables move together in physically and politically plausible ways, rather than sampling independently. Step two: for each scenario, report a loss range with the key assumptions stated, and identify the drivers of the range. Step three: for the carbon price, use a few discrete paths (slow, planned, accelerated) rather than a fitted distribution, and show the loss under each-this respects deep uncertainty instead of disguising it. Step four: identify thresholds and nonlinearities-points where loss jumps (e.g., cooling constraints forcing curtailment)-because these matter more than averages. Step five: present a table of scenario × loss range with the worst plausible combination highlighted, plus the actions that reduce loss in each. This gives the board the tail view it asked for, keeps assumptions transparent, and avoids false precision. Complement with a qualitative discussion of what would have to be true for the worst case.
3. How should the CRO present this to the board so the revised approach is credible and useful for decisions?
Read the worked answer
Present the revision as a response to the committee's own question, not as a modeling upgrade for its own sake. Structure: (1) acknowledge the limitation of the single number and explain briefly why correlation and policy uncertainty make a point estimate misleading; (2) show the scenario table with loss ranges and the assumptions behind each; (3) highlight the worst plausible combination and the drivers; (4) translate into decisions-capital buffer, insurance, water-rights or supply diversification, and carbon-price hedging-and show which actions reduce loss in which scenarios; (5) state explicitly what the analysis cannot tell them and what would change the answer. Avoid presenting a single revised number; the value is in the range and the structure. Recommend a review cadence (annual, or triggered by material policy or climate signals) and an owner. This makes the analysis decision-useful, honest about uncertainty, and auditable, which is what a risk committee needs before signing the annual report.