Coastal Mortgage Book, Scenario Choice, and the 'Unmodellable' Physical Risk Problem
A US regional bank holds $6.8bn of residential mortgages, with $2.4bn concentrated in three coastal counties. Downscaled projections for those counties indicate 0.45–0.65m of sea-level rise by 2050 under SSP3-7.0, and about 30% of the exposure sits in FEMA A/E zones currently being remapped. Flood insurance premiums on the book have risen 41% in two years after risk-rating changes, and two carriers have stopped writing in the state. The mortgages are 30-year fixed, average remaining term 24 years, average current LTV 78%, and many were underwritten on underinsured values. The CRO proposes running the next climate stress test with a single orderly 1.5°C scenario 'because that is what supervisors use for peer comparison', and wants physical risk kept as a qualitative overlay because the bank has no coastal default history to calibrate a model. The board wants a defensible position before the next supervisory cycle.
1. Evaluate the CRO's two proposals: single 1.5°C scenario, and physical risk as qualitative overlay only. What scenario set would you recommend and why?
Read the worked answer
Both proposals fail, but for different reasons.
Scenario choice. An orderly 1.5°C transition scenario is useful for transition risk and is the least severe for physical hazard. Using it alone on a book whose dominant risk is coastal flood is a category error: it answers a question the board did not ask. Peer comparability is a legitimate secondary objective, not a primary basis for selection; supervisory exercises run multiple scenarios precisely because no single one is adequate.
A defensible set for this book is three: (i) an orderly 1.5°C/NZ2050 transition, for transition and policy risk; (ii) a current-policies, high-warming pathway in the SSP3-7.0 range, which most climate-financial practitioners treat as the credible severe end for planning; (iii) a short, sharp hazard event - a landfall across the concentrated counties - to test liquidity and capital, not solvency.
I would resist making SSP5-8.5 the primary scenario. It is increasingly treated as a low-probability tail, and anchoring the board there invites the 'implausible scenario' dismissal that kills the exercise. Use it as a tail sensitivity only.
Qualitative overlay. Defensible as an interim step, not as the end state. Supervisors have moved from asking whether banks can model physical risk to asking what they did with the answer. The bank does not need default histories to quantify: it can model the transmission channels with documented assumptions and backtest the loss-given-default component on observed events. Declaring the risk unquantifiable is itself a finding the board would have to disclose.
2. Given thin default data, how would you make physical risk a quantitative credit input for this book? Set out the modelling choices and the key assumptions you would document.
Read the worked answer
The transmission channel is not default-today - it is insurance, value, and refinance. Build the model around those three.
1. Insurance-adjusted affordability. Model premium as a share of borrower income at each loan's next repricing year, not at 2050. Where modelled premium plus principal and interest exceeds roughly 45% of income, flag for higher PD. This is the fastest-moving channel and the one already visible in the 41% increase and the carrier withdrawals.
2. Hazard-conditioned LGD. Take the hazard at the loan's expected resolution date - refinance, maturity, or default - not at origination. Apply a forced-sale discount of 10–20% in the highest-hazard zones, calibrated on observed post-event sales in comparable markets. A 78% current LTV becomes a materially higher stressed LTV once that discount lands.
3. Refinance cliff. A 30-year amortising loan effectively reprices when it is sold or refinanced. Build a rollover schedule showing how much of the $2.4bn reaches a decision point before 2050, and in which hazard zones. That converts a long-horizon climate question into a three-to-seven-year capital planning question the board can act on.
4. Insurance retreat as an explicit variable. Model a scenario in which private carriers withdraw from the two highest-risk zips and borrowers fall back to a residual market at two to three times the premium. That is a credit event, not a footnote.
5. Calibrate with events, not with climate models. Use historical hurricane and flood losses to calibrate LGD and the correlation between hazard and house prices; use climate models only to shift the frequency and severity distribution.
Document every assumption, run sensitivities, and report a range rather than a point estimate.
3. What governance and disclosure would you put in place, given that the stress horizon (2050) far exceeds the bank's three-year capital planning cycle?
Read the worked answer
Governance. The scenario set and its assumptions should be owned by the board risk committee, not the modelling team, with a written rationale for why each scenario is included and what decision it informs. Set a standing annual review plus event-driven reviews after major hazard or insurance-market shifts, and name an owner for physical-risk data lineage.
Disclosure. Whatever the fate of any single national rule, the direction of travel is ISSB S2-style disclosure: governance, strategy, risk management, metrics and targets, with scenario assumptions described. Four things matter more than the template:
- Say which scenarios were used and why, including why the severe one was or was not treated as plausible for planning.
- Disclose the concentration: $2.4bn across three counties, share in A/E zones, insurance status and underinsurance.
- Disclose the transmission assumptions - premium pass-through and valuation discount - so users can disagree with them.
- Report the refinance schedule, which is the part with near-term capital consequences.
On the horizon mismatch. Do not resolve it by shortening the climate horizon. Resolve it by showing both: a three-year capital path under each scenario, and a longer asset-life view that explains why today's underwriting and servicing standards matter. The cheapest risk mitigation available here is tighter origination standards on new coastal lending, which bites immediately, not a better 2050 model.
Supervisory engagement. Bring the multi-scenario approach and the modelling limitations together. Presenting one benign scenario reads as avoidance; presenting three scenarios plus a candid account of data gaps reads as competence. Flag residual-market exposure to state insurance regulators as well.