Portfolio edition: Kith Climate cohort case, fictional client. Report date 16 October 2026.
Small grey tags are references to the project record: A- an assumption (assumption ledger, page 08), DR- a data request (page 11), D-5- a decision, DP- a data-plan item.
Physical climate risk for Northwood: method note (v1)
This note explains how the Week 5 physical-risk figures are built, so that Northwood can check, update or re-run them without Kith. It goes with the dashboard (my-work/week-5/fund-level/physical-risk-dashboard-v1.html), the assumption ledger (my-work/week-5/fund-level/leave-behind/assumption-ledger-v1.csv), the reference-data index (my-work/week-5/fund-level/leave-behind/reference-data-index-v1.csv), the result files in data/, and the re-run note (my-work/week-5/fund-level/leave-behind/re-run-note-v1.md). Every number on the dashboard is a row in one of those files; this note does not restate them.
1. What question each figure answers
The work answers Sarah Mitchell's three investment-committee questions for all ten portfolio companies:
- What could one bad event cost before exit? Gross loss in a named event, who pays it (insured, retained, next renewal), and what it does to covenant headroom.
- What does physical risk cost on average each year, and is that rising? Expected loss over each company's hold, and the recurring cost at exit on today's climate and on the buyer's horizon.
- Will the companies stay insurable at an acceptable price? The premium path above a 3% non-climate trend, in the current and a hardened market, and the retained loss a captive could pool.
For each material risk the lever library gives a dollar value protected, a cost, the timing against the exit date, and a named Northwood owner.
2. Dating convention
- Report date. The figures carry the deliverable's real date, held as the manifest setting
report_dateand never typed into a script. Until it is recorded, the engine uses the run date and marks every output provisional (D-5-42). - Hold. Each company's exit date comes from the Week 4 model: the Week 4 report date (9 October 2026) plus the company's
n_years(D-4-170, D-4-174). The hold is the exit date minus the report date, with a one-year floor. - Age of a figure. A figure's age runs from the end of the period it covers. Financials are the last twelve months to 31 January 2026, the data room's own date.
- Reference data. Public data carries its vintage (in the fetch logs) and is not flagged as stale. The cohort runs on the real calendar; the data room is fictional and its "today" is about 31 January 2026.
3. The chain: hazard, exposure, vulnerability, cost
The analysis starts from what each company depends on, not from hazard scores.
Vulnerability register. One row per dependency: each own site, each supplier, each client base, the workforce where heat matters, and shared infrastructure (Bowling Green's water). Each row carries:
- the revenue that depends on it (seeded where the data room has it, otherwise synthesised by rule);
- one of four archetypes assigned by written rule: can't produce, can't serve, can reroute, exposed through clients' assets;
- its region, hazards and flood status;
- its worst plausible 1-in-50 loss.
The data room's 1 to 5 hazard scores are the "before" picture only.
Hazards and regions. There are seven hazards: hurricane, flood, winter and ice, extreme heat, wildfire and smoke, drought, and severe convective storms. There are six regions: Gulf coast, Southeast Atlantic, Central Valley, Southwest, Mountain West, and Ohio Valley and Mid-Atlantic.
- A hazard and region pair is plausible only if its NRI annual frequency is above 0.01 in at least one county with a site or supplier there (A-01).
- Wildfire smoke is also allowed where a company's questionnaire names it (D-5-62).
- Severity has three levels: 1-in-10, 1-in-50 and 1-in-100. Base downtime is set at 1-in-50 and scaled by 0.4 and 1.6 (A-04).
Exposure rules.
- Hurricane hits every site in its region at once. This overstates the Gulf and Southeast Atlantic presets.
- A river flood hits each company's worst county or its single worst supplier, whichever loses more.
- A convective storm hits one site per company at 1-in-50 and the worst county at 1-in-100.
- Facility coordinates are the case's site locations. The FEMA flood zone at the coordinate sets flood status.
- Where the map and the company disagree (13 sites), or a point has only a city (62 points), the site is carried as a range from "not in zone" to "in zone" (D-5-56).
Vulnerability.
- Own-site downtime days come from priors by facility class and hazard (A-05), adjusted by event length.
- Field operations add client-driven days: Pinnacle's clients shut for 2 to 3 times the storm (A-06).
- Supplier outages last 1 to 26 weeks by severity (A-07). Stock, substitution by alternate status, and requalification time then apply (A-08, A-09, A-12).
- Physical damage uses FEMA Hazus depth-damage curves for flood and stated wind ratios (A-45).
Cost. Loss is contribution never recovered, plus the cost of recovery.
- Within the first four weeks of an outage (A-47), part of the lost revenue is recovered:
- Recovered revenue goes on a separate cash-timing line.
- Recovery cost is 10 to 25% of the recovered contribution (A-31).
- One company's lost revenue in one event is capped at its revenue times the longest outage in that event.
- Nothing is netted: upside, if evidenced, is a separate line.
4. Who pays
No insurance schedules are in the data room (data request DR-1), so each company has a programme synthesised by rule (A-17 to A-21):
- an all-risk deductible by size band;
- a 2 to 5% named-storm deductible on coastal sites;
- flood sublimits;
- a 72-hour waiting period for business interruption;
- contingent business-interruption cover on named sole suppliers.
A hardened-market toggle doubles deductibles, sets the named-storm deductible to 5%, halves flood sublimits and removes contingent cover. After a claim, the next renewal steps up by 10 to 25% (A-23).
Insured and retained each report their own range, because cover only triggers in the worse cases. Gross = insured + retained holds within each case.
Covenants.
- Sterling, Trident and NovaCare have seeded covenants (D-5-49).
- For the other seven, the covenant level is entry leverage divided by one minus a 25 to 35% cushion (A-32).
- A scenario flags a breach when retained loss plus the renewal step exceeds the headroom. It also reports Northwood's own trigger: a cushion under 15% (
_reference/risk-framework.md).
5. Recurring cost and exit value
Recurring cost a year at exit is the sum of four parts:
- premium growth above a 3% non-climate trend (seeded histories where they exist; A-24, A-25);
- heat productivity loss on outdoor and field labour (A-26, A-27, CMRA county hot-day projections);
- cooling and water uplift (A-28, A-46);
- in the hardened market only, higher retained deductibles.
It is shown on two horizons: the hold, to each exit, and the buyer's horizon (about 2035 to 2040). It is also multiplied by each company's exit multiple, as value a buyer may normalise away. One-off event cash and recurring cost times the multiple are always shown apart.
6. Presets, the do-nothing figures and the builder
Presets. Seven presets are fixed settings of the same engine:
- Gulf landfall
- Southeast Atlantic landfall
- Sterling superalloy supplier outage, 26 weeks (D-5-67)
- Central Valley drought, two seasons
- Ohio Valley and Mid-Atlantic winter storm
- Southwest heat summer
- Hold-period "do nothing"
The do-nothing preset shows two figures that are never added together:
- The severe-but-plausible stack. The rule takes the two multi-company presets with the highest gross loss at 1-in-50, from different hazard families, counting a row hit by both once, and adds the chronic baseline over each hold (D-5-31, D-5-47).
- The rule picks the Gulf landfall and the winter storm.
- NRI ranks flood and convective higher by expected annual loss. Both are local hazards that do not stack across companies at 1-in-50, so their weight is in the expected loss (D-5-65).
- Expected loss over the hold. A trapezoid over 1-in-10, 1-in-50 and 1-in-100 for every plausible hazard and region, times each hold (spec s8). It is expected loss from events at or rarer than 1-in-10 (the 1-in-10 to 1-in-100 band, with the tail held at the 1-in-100 loss), on today's climate, beyond the routine weather already in earnings; a lower bound.
Why events more frequent than 1-in-10 are left out. The hazard intensities, downtime and supplier-outage ranges are only defined at 1-in-10, 1-in-50 and 1-in-100, so anything more frequent would be extrapolation with nothing to calibrate it. Routine weather that recurs every year is already in the trailing-twelve-month EBITDA the model measures against, so counting it again would double-count. The companies' own reported events were replayed at 1-in-10 severity and the seven tested all passed, so the 1-in-10 point matches the size of the events they actually have. And it is the rarer events that drive covenants and value at exit. The result is a lower bound: the frequent band and the tail beyond 1-in-100 would both add to it.
The builder reads a precomputed grid. Hazard-led settings cover hazard × region × severity × length × market (768 plausible settings). Dependency-led settings cover each site and supplier × 2 to 26 weeks × market (2,020 settings). The page looks results up and does no arithmetic of its own (D-5-12).
7. Calibration
Nine events the companies reported for 2023 to 2026 were replayed through the engine with their actual sites, hazard, storm length and a severity fixed before the replay (1-in-10). Modelled lost-plus-delayed revenue was compared with the reported figure. A pass is 0.5 to 2 times (A-33). The results are in my-work/week-5/fund-level/leave-behind/data/calibration-groups-v1.csv.
- First run: five of nine passed.
- Two priors were narrowed by decision: food-plant winter downtime (D-5-69) and clinic hurricane downtime (D-5-70).
- Two events are reported but not tested, by decision:
- Keystone (D-5-71): its session count and its reported revenue contradict each other, and data request DR-13 asks Keystone to resolve it.
- Ridgeline (D-5-72): see limitations.
- Final run: all seven tested events pass.
An independent check (my-work/week-5/checks/chk_calibration_band.py) re-derives every replayed figure without using the engine.
8. Levers
Each lever in my-work/week-5/fund-level/leave-behind/data/levers-v1.csv names the input it changes and a cost range (A-35 to A-39, A-50; Akron's HVAC cost is seeded). The engine is re-run with each lever applied, for the lever's companies only (my-work/week-5/fund-level/leave-behind/data/lever-results-v1.csv).
For each lever:
- Value protected is expected loss avoided over the part of the hold after the lever is in place, plus recurring cost avoided times the exit multiple if it is in place before exit.
- Ranking is by value protected per dollar of cost over the hold.
- Tail protection, the retained loss avoided in a 1-in-50 event, is shown beside the value and never added to it.
- Timing: a lever that cannot be in place before exit is listed for the buyer's plan and earns no exit credit.
Two portfolio overlays sit beside the levers:
- Captive feasibility flag: retained expected loss across the ten against $2M a year (A-40).
- Pipeline screen: Clearview's Target B (Houston), run through the Gulf preset as a synthesised Clearview field site.
9. How far to trust it
- Ranges. Every output has a low, central and high. The wide-range priors (group 3 in D-5-48) are carried to every figure.
- Sensitivity.
my-work/week-5/fund-level/leave-behind/sensitivity-v1.csvmoves each group 3 prior from end to end with everything else at central. Drought pass-through (A-15) and supplier outage weeks (A-07) move the answers most. The Harvest Ridge covenant breach holds across the full covenant-cushion range (A-32). - Assumption ledger.
my-work/week-5/fund-level/leave-behind/assumption-ledger-v1.csvgives each assumption's treatment: calibrated, cited, a visible range with sensitivity, or accepted judgment. - Data requests.
my-work/week-5/fund-level/leave-behind/data/data-requests-v2.mdlists every gap filled by a synthesised or assumed input, in priority order. The insurance schedules (DR-1), Sterling's superalloy stock and requalification (DR-2) and street addresses (DR-10) would narrow the ranges most.
10. Limitations
- Hurricane hits every site in a region at once. This overstates the Gulf and Southeast Atlantic presets for companies spread along the coast.
- Project schedule slippage is not modelled. Ridgeline's January 2026 storms deferred fit-out work for tenants by weeks, with no margin impact. The engine models closure days, not project slippage, so that event is reported but not tested (D-5-72). The effect is cash timing, not loss.
- Insurance is synthesised. Insured and retained splits, and the contingent business-interruption cover behind Sterling's superalloy result, are rules, not schedules (DR-1). Sterling's cover is a question to ask, not a reassurance.
- Supplier layers for eight companies are synthesised from their cost lines by sector template (A-41). Sterling's and Harvest Ridge's supplier files are seeded.
- Flood status is a range for 75 points (D-5-56), so the central flood figure is a midpoint, not a probability.
- Each figure's low, central and high are its own lowest, middle and highest of the three cases. For local floods and convective storms, a different county can be worst in each case. A company's channels, insured and retained then need not add to its gross column by column. Within each case they always do.
- Keystone's clinic closure cost rests on the spine revenue. The data room's session count implies about a third of that (DR-13).
- Revenue per closure day uses calendar days. Clinics and offices open about 250 days a year, so a closure on a working day may cost more than shown.
- The buyer's horizon changes recurring heat cost only. It uses CMRA county projections of days at or above 95°F for the 2030s and 2040s. The range runs across five model and scenario pairs: LOCA2 under SSP2-4.5, SSP3-7.0 and SSP5-8.5, and STAR under SSP2-4.5 and SSP5-8.5. The low case is the lowest of the five, the central case the median, the high case the highest. The hold period uses the 2020s. Event losses use today's climate on both horizons (spec s3).
- Lever costs are stated judgment (D-5-74). Broker quotes should replace the insurance pricing (A-39, A-50) when DR-1 is answered.
11. Portfolio edition addendum: the arithmetic the dashboard v3 adds (November 2026)
Sections 1 to 10 are the method as frozen on 16 October 2026, and every figure they describe is unchanged. For the portfolio edition, Daniel Vetsko lifted the no-recalculation rule for Week 5 only (P-30) and approved new views on the dashboard (P-31 to P-33). Each view is a stated step on figures the engine already stored. None of them feeds back into a stored figure.
- The case and insurance switches. These only choose which stored value to show: low, central or high, and today's terms or the hardened market (section 4; pull-back terms A-17 to A-22). The do-nothing stack and expected loss are gross and were stored for today's market only, so the insurance switch does not move them.
- The map. Own sites use the data room's locations (138 of 139; one office is given only as a city). Of the 223 register rows, 174 sit at a mapped point; the other 49 have no single location (supplier groups, generic suppliers such as telecom and software, outdoor workforces). Suppliers and managed properties sit at their city centre (GeoNames), so a single point is indicative (
points-v3.csv); decision P-38 records this in place of the county centres first proposed, and actual addresses should be requested from the companies for a final delivery. No figure depends on a point's location: exposure comes from each dependency's county and region in the register. Each point is joined to the register by its site or supplier id. Its size is the sum of its rows' 1-in-50 gross loss, and its colour is the Northwood severity of its worst row. A point is in a scenario's footprint when one of its register rows names that scenario's region and hazard. The base map is the US Census state outlines (us-atlas 3.0.1). - The cash squeeze. Cash timing ÷ (revolver size − revolver drawn), using the data room's January 2026 balances. Cash on the balance sheet is not in the files, so this is a conservative test.
- Leverage after the event. Debt ÷ (EBITDA − retained loss − the renewal step), against the assumed covenant level (A-32). This is the engine's own covenant test (section 4), shown as a ratio. The breach flags are the engine's.
- Exit timing. The gross cost during each hold is expected loss a year × hold years, plus recurring cost a year at exit × hold years (hold at least one year, section 1). The buyer's figure is the same two measures on the buyer's horizon × 5 years. Five years is the midpoint of Northwood's own 4 to 6 year hold policy (data room investment thesis), assumed for a buyer (portfolio assumption PA-1); the page also shows the share after the sale at 4 and 6 years. The price at risk is the stored recurring cost × exit multiple (section 5): an upper bound, which assumes a buyer takes the whole recurring cost off the price. A switch on the page shows half and none (P-38): part of the recurring cost is already in exit-year earnings, and no source fixes how much a buyer would take off. Nothing is discounted.
- Expected loss by hazard. The section 6 trapezoid is applied to each hazard's stored 1-in-10, 1-in-50 and 1-in-100 losses at its default event length. The parts are then scaled so that each company's parts sum to its stored figure. Rebuilt without scaling, the page's grid reproduces eight companies exactly, and Pinnacle and Ridgeline within 2%, because a few wildfire cells are not stored at their default length.
- The climate outlook. CMRA county projections (section 10) for days at or above 90°F, 95°F and 100°F, dry days, the longest dry spell and annual rainfall, weighted by headcount over each company's own site counties. Only days at or above 95°F feed a priced cost (heat on outdoor work on the buyer's horizon, section 10); the other measures are shown for context and are not priced. The line is the median of the five model and scenario pairs and the band runs from the lowest to the highest. Decades before the 2020s are the models' own hindcast.
- The illustrative slider. The part of expected loss from the hazards the reader ticks (all of them to start) is multiplied by (1 + frequency uplift) × (1 + severity uplift), evenly at every return period. The reader sets the uplift. CMRA projects temperature and rainfall but not how often storms, floods or droughts strike a county, which is why events stay on today's climate and this is a what-if. It is a sensitivity of the stated answer, not a forecast, and no source stands behind it.
- The break-even. N = retained loss avoided in one 1-in-50 event ÷ (cost over the hold ÷ hold years). If an event of that size comes at least once every N years, the lever pays for itself on the tail alone. Hold years are the company's hold, or months to exit for portfolio levers, or else the longest hold.
- An average year against one event. The company's largest gross loss across the six single-event presets ÷ its expected loss a year.
- Reading return periods. The chance of at least one event of a given size in n years = 1 − (1 − 1/T)^n, for T = 10, 50 and 100 and n = 3 (a typical hold) and 5 (a buyer's years).
- The priority view of the grid. Likelihood within the hold = 1 − (1 − 1/T)^hold years, banded by Northwood's risk framework (Remote under 10%, Possible 10 to 35%, Likely 35 to 65%, Almost certain above), then combined with severity through the framework's priority matrix. An outage has no return period, so it has no priority.
- The stress case split. The recurring part = recurring cost a year (today's terms) × hold years, summed over the ten; the event part is the stored stress case minus that.
- Each company's worst event. For each company, the stored hazard and region cell with its largest 1-in-50 gross loss at the standard event length or a preset's length (section 6). It is a lookup that lets companies no preset reaches, and local hazards such as flood, run through the same stages. It shows the chosen company alone, because a local flood or convective storm sits on each company's worst county and is never added across companies (D-5-65).
To re-run all of it, run python3 my-work/portfolio/week-5/build-week5-portfolio-v2.py from the repository root. It reads the frozen Week 5 files and the data room, never edits them, and rebuilds delivery-v2/. The dashboard's Appendix B shows how the build fits together, the controls at each step and where there are none, and the flagged assumptions.
Checks and review. my-work/portfolio/week-5/checks/chk-v3-derived-v1.py recomputes every figure above independently in Python, from the frozen data and the data room, and compares it with the text the page renders (review/render-text-v1.py). The portfolio-edition judgments (the buyer's years, city-centre placement, the break-even hold fallback, headcount weighting) are listed in dashboard-v3/portfolio-assumptions-v1.csv. The November 2026 fact-check, client-voice and process reviews, with what was done about each finding, are in review/review-findings-v1.md.
Review round 1 additions (November 2026, decision P-37).
- Wider sensitivity.
sensitivity/sensitivity-wider-v1.pyimports the frozen engine read-only and moves 18 more assumptions one at a time to their low and high values, everything else central. It covers A-02, A-05, A-06, A-13, A-14, A-16, A-17 to A-21 as one group (A-20 has no range), A-23, A-25 to A-28, A-30, A-31, A-45 and A-49, plus Pinnacle's seeded premium growth across its broker's stated 10 to 15% (data room questionnaire 2025). Rows with no effect are kept, with a swing of 0. It recomputes six headline figures: the stress case, expected loss a year, recurring cost, price at risk, Gulf landfall kept plus renewal, and Harvest Ridge's drought loss. The base run must reproduce the stored figures, and the parameters must restore, before anything is written. The largest swings are the drought price rise (A-13, ±38% of Harvest Ridge's drought loss), heat productivity (A-27, ±19% of recurring cost and price at risk) and wind damage (A-45, ±15% of expected loss). An assumption is flagged "tested: moves the answer" at a swing of 5% or more in a headline figure. - Second-sourcing against the drought. The same script applies lever L-03's change (alternates for the two sole-source suppliers qualified) to the Central Valley drought and writes
sensitivity/lever-l03-drought-v1.csv. It avoids $0 in every case, because the drought loss is California input prices, not supplier outages. The lever's stored tail ($2.21M) is for its flood reference event only. - Price at risk split. Visible part = (premium above trend + retained deductibles) × exit multiple; modelled part = (heat productivity + cooling) × exit multiple; both × the pass-through share chosen.
- Harvest Ridge covenant trigger. The covenant breaks if the loss Harvest Ridge keeps is more than EBITDA − debt ÷ covenant level, using the same data-room debt and EBITDA and the assumed covenant level (A-32) as the engine's test. The assumed covenant is entry leverage (entry debt ÷ entry EBITDA, data room) ÷ (1 − the 25 to 35% cushion). The page compares the trigger with the drought scenario and with the low end of the drought price rise, and gives the drought's likelihood over the hold, 1 − (1 − 1/50)^hold years.
- Recommended programme. Sums of the stored cost and value protected for the levers whose verdict is "Do now". The cautious-end value per $ on each card is value protected at its low ÷ cost at its high.
my-work/week-5/fund-level/leave-behind/method-note-v1.md (frozen at G2/G3 on 16 October 2026; sha256 108627d5feb0), with the portfolio wording fixes listed in the portfolio changelog. Raw file: data/method-note-v1.md.