Household, Bank, and Insurer Exposure to Miami Hurricanes: a flow-of-risk analysis

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Household, Bank, and Insurer Exposure to Miami Hurricanes: a flow-of-risk analysis Benjamin N. Dennis 2023-013 Please cite this paper as: Dennis, Benjamin N. (2023). “Household, Bank, and Insurer Exposure to Miami Hurricanes: a flow-of-risk analysis,” Finance and Economics Discussion Series 2023-013. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2023.013. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Household, Bank, and Insurer Exposure to Miami Hurricanes: a flow-of-risk analysis Benjamin Dennis∗ Federal Reserve Board February 7, 2023 Abstract Weanalyzepossiblefuturefinanciallossesintheeventofhurricane damage to Miami residential real estate, where the hurricane’s destructiveness reflects climate-change. We focus on three scenarios: (i) a business-as-usual scenario, (ii) a Hurricane-Ian-spillovers scenario, and (iii) a cautious-markets scenario. We quantify bank exposures and loss rates, where exposures are proportional to the size of real estate markets and loss rates depend on post-hurricane devaluations and insurance coverage. This quantitative methodology could complement modeling of local economy impacts, stress on public finances, asset market losses, and other financial developments that will also affect banks. ∗The views expressed in this paper should not be attributed to the Federal Reserve Board and are the sole responsibility of the author. The scenario analysis in this paper is a research effort and is unrelated to the Federal Reserve Board’s recently announced Pilot Climate Scenario Analysis. This paper builds on contributions by my colleagues in S&R Policy Planning and Strategy to a review of Miami climate risks, including Joseph Cox, Jonathan Loritz, Jacy Su, Nick Tabor, Justin Warner, and Aurite Werman. Other participants in the Miami study included Brian Bailey, Kyle Binder, Saba Haq, Nick Klagge, Andy Polacek, John Schindler Jr., Solomon Tarlin, Lauren Terschan, and James Wang. I thank Liz Marshall and Roisin McCord for their review of the code. Jake Clark was instrumental in applying the Hazus tool. I also thank Benjamin Kay for many helpful comments. While the insights of these individuals were instrumental in shaping this analysis, all errors and omissions remain my own. 1

Contents 1 Introduction 3 2 Analytical approach 6 2.1 Loss absorption by insurers . . . . . . . . . . . . . . . . . . . 7 2.2 Loss absorption by borrowers . . . . . . . . . . . . . . . . . . 8 2.3 Loss absorption by creditors . . . . . . . . . . . . . . . . . . . 9 3 Scenarios 10 3.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 3.2 Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 3.3 The Business-as-Usual (BAU) Scenario . . . . . . . . . . . . . 14 3.4 The Hurricane Ian Spillover Effects Scenario . . . . . . . . . . 16 3.5 The Cautious Markets Scenario . . . . . . . . . . . . . . . . . 21 4 Conclusion 25 A Model 28 A.1 Homeowner types . . . . . . . . . . . . . . . . . . . . . . . . . 28 A.2 Residential real estate dynamics . . . . . . . . . . . . . . . . . 28 A.3 Equity held by non-homeowners . . . . . . . . . . . . . . . . . 29 A.4 Ensuring consistency between stocks and flows . . . . . . . . . 30 A.5 Average size of mortgage by cohort . . . . . . . . . . . . . . . 31 A.6 Number of homes in each mortgage cohort . . . . . . . . . . . 33 A.7 Solving for mortgage repayments . . . . . . . . . . . . . . . . 34 A.8 Determining equity holdings by banks, securities purchasers, and GSEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 B Climate Change Damage Generation Process 36 B.1 Apportioning flood and wind damage . . . . . . . . . . . . . . 36 C Insurance as the first loss-absorbing layer 36 C.1 The extensive vs the intensive insurance margin . . . . . . . . 38 D Homeowners as the second loss-absorbing layer 38 D.1 Insurance coverage as a fraction of replacement value vs total property value . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 D.2 Keeping track of loss cushions by cohort . . . . . . . . . . . . 40 2

D.3 Default risk by segment and cohort . . . . . . . . . . . . . . . 41 D.3.1 Cohort equity and strategic default . . . . . . . . . . . 41 E Non-homeownerassetholdersasthefinalloss-absorbinglayer 44 F Data addendum 46 G Basis for scenario assumptions 48 G.1 Hurricane shock devaluation . . . . . . . . . . . . . . . . . . . 49 G.2 Home price depreciation . . . . . . . . . . . . . . . . . . . . . 49 G.3 Housing unit growth . . . . . . . . . . . . . . . . . . . . . . . 50 G.4 Turnover rates . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 1 Introduction The accelerating impact of climate change has added urgency to efforts to understand how severe weather events might affect the safety and soundness of the financial system.1 This paper takes a bottom up approach by investigating the impact of sea level rise on the vulnerability of banks to losses on their residential real estate portfolios in Miami. The defining feature of this analysis is that it traces what we define as the flow-of-risk across entities. Pozsar (2014) [31] introduced the term flow-of-risk as a means of taking derivatives and other risk shifting mechanisms into account when quantifying exposure to specific financial risks in the shadow banking sector. We similarly use the term as a means of identifying true exposure to climate risks taking into account all known loss allocation arrangements. In the event of a hurricane in Miami, insurance companies take the first loss (net of deductables). When insurance coverage does not exist or is insufficient, losses spill over to homeowners. If homeowners default for whatever reason, losses accrue to mortgage originators or purchasers depending on their exposure. Thus, the model shares a kinship with waterfall models developed to understand loss distributions across asset managers in the wake of the Global Financial Crisis. This kinship extends to the recognition that institutional and contractual details matter in the passthrough of losses from one party 1See, for example, the most recent Sixth Assessment of the Intergovernmental Panel on Climate Change (IPCC): https://www.ipcc.ch 3

to another, and the degree to which losses are contingent on complementary factors. The focus on residential real estate is an attractive starting point given the significant role of banks in mortgage lending and the availability of bank balance sheet data that can be matched against hurricane and flood risks. While exposure to residential real estates in Miami is admittedly a small fraction of any large national bank’s portfolio, this exercise provides a template that can be extended to other assets and regions. With a suitable methodology for aggregating across these assets and regions, taking into account correlations between climate events and spillovers across regions, a given bank’s total exposure to climate risk can be evaluated. We choose Miami given that its unique exposure to sea level rise has received a lot of interest. The high risk to Miami from sea level rise combined with its susceptibility to hurricanes imply that homeowners might have a markedly higher incentive to strategically default on their mortgages in the event of climate-related losses. There are two recent books that cover real estate markets in Miami in the context of sea level rise (Ariza 2020 [1], and Goodell2017[20]). Eachcity-regionisdistinguishedbyuniquesetofphysical and transitional climate risks. Miami is wedged between two large bodies of water that will overwhelm the city if sea levels rise. Complicating decisionmaking, local climate officials must coordinate across 34 municipalities in Miami-Dade County alone. In addition, Miami is built on porous limestone through which water can infiltrate. Not only does this allow water to bypass seawalls, but saltwater infiltration is also degrading Miami’s freshwater aquifers. In the flow-of-risk framework, a climate-linked natural disaster leads to insurable claims for covered homeowners, generating losses for insurers.2 To theextentthatinsurancepolicypremiumsaccuratelyreflectclimaterisksand homeowners are appropriately insured against those risks, insurance should be sufficient to prevent any losses from passing through to banks. Yet not all households have insurance, especially if they are not required to have it, and those that do may not be fully insured. This applies both to flood and wind insurance which are sold by different insurance companies in separate markets (whether wind is bundled into general homeowner’s insurance 2The flow-of-risk considerations examined here are a subset of the effects described by Batten et al. (2016) [5], who provide a comprehensive mapping of direct and indirect impacts of a climate-linked natural disaster on the financial system. 4

automatically or is required as a separate rider depends on the insurers assessment of wind vulnerability within the policy-holders vicinity). Potential homeowners looking to take out a mortgage to purchase property located in FEMA-designated flood plains are often formally required to purchase flood insurance by lenders. However, there are concerns that lenders may be exempting as many as half of these borrowers from this requirement.3 For those with insurance, coverage may not be sufficient to replace the home, especially if it is now necessary to elevate the home on pilings, construct private water drainage infrastructure, or build to higher local construction codes. Policies written to cover the actual depreciated value of structures may fall far short of the necessary reconstruction funds, and homeowners may often insure themselves to the minimum required level. As discussed in Section 3.1, property-level data on insurance coverage is limited or unavailable. Insurance losses only pass through to banks if homeowners do not in turn absorb uninsured losses. Losses to homeowners include two distinct components: losses due to damage to structures, and losses due to devaluation of the land parcel. For residential mortgages, the historical record on price devaluation is limited. Many homeowners have benefitted from some form of insurance or disaster relief and strategic default rates have been low. In many cases, property values have rebounded from isolated natural disasters within three to seven years. When authorities have determined that a land parcel is no longer suitable for habitation, the government has often acted to buy out homeowners and make them whole.4 However, if the number of at-risk properties becomes too large, governments may not be able to absorb these losses. A scenario that produces the trifecta of uninsurability, large and unexpected adaptation costs (such as for condominium owners in the wake of the Surfside tragedy), and losses too large to be absorbed by the government could produce a large impact on prices. Although some research has found an impact of climate change on home prices, it does not seem likely that prices fully reflect climate risks (see, e.g., Dennis, 2022, [9]). Section (2) describes the flow-of-funds analytical approach, with a description of how losses are calculated for each loss-absorption layer. Section (3) describes the data, assumptions, and design of each of our three scenar- 3Thispoint,initiallyraisedininformaldiscussionswithinsuranceexperts,issupported by coverage data from NFIP as described in Section 3.1. 4See, e.g., NYT, “U.S. Flood Strategy Shifts to ‘Unavoidable’ Relocation of EntireNeighborhoods”: https://www.nytimes.com/2020/08/26/climate/flooding-relocationmanaged-retreat.html 5

ios, and presents results. Section (4) concludes. Addition details on model calculations and other information are provided in the appendix. 2 Analytical approach WemodelhurricanesateachboundaryoftheSaffir-SimpsonHurricaneWind Scale – Categories 1 through 5 – that follow a predetermined path through the heart of Miami-Dade County using the HAZUS FEMA physical modeling tool.[13]5 This tool uses the physical characteristics of both a defined hurricane system and Miami to determine damage rates on a census tract basis. These damage rates are then applied to properties. The HAZUS software takes into account various factors such as windspeed, track speed, tidal elevation, width of hurricane, etc. to calculate wind and flood damage in a combined model. The relative balance of flood and wind damage will depend on many factors, including storm characteristics in addition to the characteristics of built structures in the storms path (e.g., frame vs. masonry, age, height, purpose, etc.). For example, a wide, slowmoving storm with slower winds will create a high ratio of flood to wind damage. A narrow, fast-moving Category 5 hurricane will primarily cause wind damage (although flooding due to storm surge will depend on the topography of the shoreline). These many features are captured by the HAZUS model. HAZUS provides a series of wind and flood loss rates as a fraction of replacement value for single family homes, multifamily dwellings, and mobile homes by census tract. Replacement values are also available from HAZUS, allowing us to calculate losses. Actual property values are equal to the replacement value plus the value of the land parcel. Although it is generally the case that the land parcel adds value to the structure, this is not always the case. Local amenities may have a negative value that lowers the sales price of a property below the replacement value, especially in areas in which climate risks are rising sharply. We then map census tracts to flood zones using FEMA National Risk Index data and use U.S. Census data to identify areas of high-to-medium income owners, lower-income owners, and primarily rental properties. These 5Theboundariesaregivenbythefollowingmaximumsustainedwindspeeds: Category 1 = 74 mph, Category 2 = 96 mph, Category 3 = 111 mph, Category 4 = 130 mph, and Category 5 = 157 mph. 6

distinctions are used to impute the amount of insurance coverage likely to prevail by census tract. We assume that floodplain homes are more likely to carry flood insurance as a condition of securing a mortgage. Evidence suggests that this coverage is only around 50 percent, however, given that half of homeowners allow their flood coverage to subsequently lapse.6. We also assume that lower income households are less insured relative to highto-middle income households, again most likely due to lapses in renewing policies. Analysis of the aftermath of Hurricane Ian will provide much more information about how well insured Florida households are, and whether insurance will remain within the reach of most households. We derive the mortgage exposure of banks to each census tract, and therefore each of the six homeowner categories, using Home Mortgage Disclosure Act (HMDA) data. HMDA mortgage data provides information on the type of borrower, the location of the property, the loan-to-value ratio, and many other factors for a wide variety of financial institutions. However, it does not provide the stock of mortgage assets on these institutions? balance sheets. We get information on the stock of mortgages from Y-14M data. One benefit of Y-14M data is that it apportions bank-originated mortgages into mortgages held by banks, mortgages sold to the GSEs, mortgages that are securitized and sold, and mortgages sold in the interbank market. Bankoriginated mortgages account for only a third of all mortgages, with the majority of mortgages offered through non-bank financial institutions (NBFIs) such as Rocket Mortgage. Banks tend to keep around half of the mortgages they originate and sell the rest, and the choice of which mortgages to retain willreflectariskmanagementstrategy. Lackingdataonwhichmortgagesare retained and held on banks’ own books, we assume that mortgage retention and sales are proportional to the stock of mortgage loans in each of the six categories. This assumption will overstate the risk to banks (and understate the risk to purchasers) if banks retain the least risky categories of mortgages for their portfolios. 2.1 Loss absorption by insurers The first layer of loss absorption is insurance, net of deductible. Because homeowners insurance deductibles are so small, typically $500 to $2,000, we 6There are no comprehensive datasets on insurance coverage of households. We base our estimates of coverage rates on reporting on the impact of Hurricane Ian, see, e.g., Rozsa and Werner (2022) [33], and Flavelle (2022b) [15] 7

do not model them in our simulations. There are three types of homeowner policies in the model: (i) National Flood Insurance Program (NFIP), (ii) private flood insurance, and (iii) homeowners insurance. The first two types of insurance are specific to flooding, while the latter covers damage by wind. Intheeventthatwindandflooddamagecoincide,weassumethathomeowner insurers will insist that the damage be classified as flood damage. Most homeowners lack flood insurance. Recent reporting in the wake of Hurricane Ian suggests that only half of homes in floodplain areas in the counties evacuated for Hurricane Ian had flood insurance, while only 20 percent in non-floodplain areas had flood insurance (see, e.g., Rozsa and Werner (2022) [33], and Flavelle (2022b) [15]). It is unclear why so many mortgage borrowers lack flood insurance, which is a requirement of securing a mortgage in designated floodplain areas. Most likely flood policies are allowed to lapse after the original mortgage is secured. Homeowners insurance, which typically covers wind damage, is available for the replacement cost of the home’s structure. The combined value of the home’s structure and the land parcel on which the home sits is equal to the price of the home. If the land value does not change in the wake of a hurricane, it is possible for insurance to make the homeowner whole, which seems to be the historical norm. However, climate change may cause a climate event to lead to a devaluation of the land parcel in addition to structural damage to the home. This potential devaluation is most likely if insurers are led by climate change to withdraw coverage by declining to renew insurance policies (or to increase premiums beyond the reach of most homeowners). There is no practical method for insuring land value, so some homeowner losses are not covered by the insurance loss absorption layer. Our specific insurance assumptions are discussed further below. 2.2 Loss absorption by borrowers Borrowers are next in line to absorb losses not covered by insurers. Borrower equity will vary by many different factors including tenure in the home. We assume that all borrowers purchase their homes initially with a 30-year fixed rate mortgage and a 20 percent down payment. We assume that there is a constant rate at which homes are sold, which allows us to solve for the cohort distribution where cohorts are defined by the number of years since purchase. To the extent that uninsured structural damage and land devaluation caused by a hurricane reduces homeowner equity, homeowners may owe 8

more on their mortgages than their homes are worth. Recent cohorts, who have newly purchased their homes, will have a large loan-to-value ratio and will consequently be more likely to default. The literature on post-natural disaster home values tends to find that home values recover with three to seven years of the disaster implying that land values are typically durable. Homeowners who expect the eventual recovery of their home values will be less likely to default, viewing any devaluation as temporary. However, for Miami, climate change is likely to permanently devalue properties in harms way through at least three channels. The first is the carrying cost of modifying the property to withstand higher sea levels and enduring more frequent or damaging storms. The second is the cost in reduced amenities values caused by eroding public services, higher taxes, and other community effects of climate-induced changes. The third is the increased difficulty of insuring property, including reduced availability of insurance or increased premiums. This implies that studies that tie default rates to negative equity need to be modified for land value devaluation in addition to considerations such as whether the state has non-recourse laws in which the lender cannot go after more than the collateral of the home itself. Florida is a recourse state. We use an adaptation from Bhutta et al. (2010) [6], which focuses in part on Florida default rates, to impose a linear relationship between negative equity anddefaultprobabilitywithanadhocadjustmenttocapturetheexpectation of permanent devaluation. For each cohort of each type of borrower, we then apply the non-default rate times the homeowner’s loss (both uninsured damage and devaluation) to determine the amount of loss absorbed by homeowners. Remaining losses pass through to the next loss absorption layer. 2.3 Loss absorption by creditors Credit originators include both banks (which originate roughly one-third of mortgages) and non-bank financial institutions (NBFIs). However, banks only retain a portion of the loans that they originate on their balance sheets. The remainder are securitized, sold to government-sponsored entities such as Fannie Mae or Freddie Mac, or sold to other banks. Confidential Home Mortgage Disclosure Act (HMDA) data can track annual mortgage originations and sales by homeowner type and location, allowing us to calculate the exposure of different creditors and purchasers to homeowner type and 9

climate-vulnerable locations. We assume that the share breakdown of these mortgage flows is equivalent to the balance sheet composition for these institutions, with some institutions carrying a heavier exposure to climate risk. For each housing type (single, multi, mobile homes)/homeowner type (high/middle income, low income, investor)/locational vulnerability/cohort quad in each census tract (where census tracts differ in the mix of housing type – single family homes, multi-family homes, and mobile homes – and home prices), we apply the appropriate default rate (the portion of borrowers inthatcategorywhodefault)times80percentofthenetvalueofoutstanding principle minus the ex-post collateral value of the home. The percentage represents an assumption that there is a foreclosure friction for the bank of 20 percent of value of the home. Once we have the total quad losses, we allocate them across creditor institutions using the share exposures of each institution to each quad. For example, suppose net default losses for low-income homeowners in mobile homes in flood-prone census tract X total $100 across all cohorts, and Bank Y accounts for 50 percent of low-income, mobile home mortgages in this census tract. We would calculate that Bank Y experiences losses of $50. 3 Scenarios Scenario design is complicated due to the many different exogenous variables not calculated within the model. This iteration of the flow-of-risk model takes home price levels and growth rates, turnover rates, interest rates, insurance rates, sea level rise, home construction rates, and other factors as exogenous. These factors interact, however, and so they cannot be chosen completely independently. Ideally, consistency across exogenous variables could be enforced through an asset valuation model (or module) that would complement the flow-of-risk analysis. For example, if insurance premiums were to rise sharply (or if insurance is no longer available for some homes), we would expect a “syndrome” to follow in which insurance coverage falls, home construction slows, home prices decrease, and mortgage interest rates rise (if for no other reason than a rising risk premium). We also, critically, assume post-hurricane devaluation rates in which the amenities value (or the value of the land parcel of the property distinct from the value of the built structure) falls by a fixed amount. Historically, land parcel devaluation in the wake of a climate disaster has not played a significant role. However, 10

climate change impacts may lead to permanent revaluations of certain locations, and we allow for it. At the very least, the scenarios that we develop need to follow a consistent narrative. With this in mind, we construct four scenarios to explore how bank losses respond to the following conditions: • Business-as-Usual (BAU) • Hurricane Ian Spillover Effects • Cautious Markets 3.1 Data First, we discuss the key data sources for our analysis. From the Home Mortgage Disclosure Act (HMDA) dataset, we obtain the following data on Miami mortgages on an institution level basis by census-tract: combined loanto-value ratio, whether within limit for conforming loans, lien-status, loan amount, loan purpose, loan term, loan type, property type, property value, purchasertype,nameoflendinginstitution,whethersold/guaranteed/transfered to another institution, occupancy, construction method, dwelling category, debt-to-income ratio of borrower, business/commercial purpose, applicant income, median census-tract family income, ratio of average tract income to MSA income (as percentage), number of units in tract, number of owner occupied units in tract, action type, and applicant race. From Y14M data, we obtain the names of covered banks, their category (e.g., LISCC, RBO, etc.), total mortgage holdings, and custodial holdings for GSEs and securitiesholders. From Elliot et al. (2017) [10], we obtain by neighborhood: the number of housing units, the number of occupied units, the number of lowmiddle income (LMI) households, the number of very-low-income (VLI) plus (LMI) households, population, the number of households, the number of white/black/other race residents, the number of hispanic residents, the labor force unemployment rate, the poverty rate, the share of renter-occupied housing, the share of LMI renter-occupied housing, the share of VLI renteroccupied housing, the share of owner-occupied housing under $200 thousand, theshareofrentaloccupiedhousingwhererentislessthan$1,000permonth, the amount of single family housing, the median home value, and whether housing is flood prone. 11

We match the flow data from HMDA with the stock data from Y14M to generate the stock-flow figures from appendix section (A) in combination withthefollowingassumptions. Downpaymentsareassumedtobe20percent of the value of the loan in all cases. Low-income households are defined as having no more than $117 thousand in income.7. Unoccupied or renteroccupied homes are classified as other (i.e., “OX”). The number of “LX” homes is calculated as the share of sub-$200K owner-occupied homes (from Elliot et al., 2017, [10]) times the number of primary occupancy homes (from HMDA). The number of “HX” homes the therefore the total of primary occupancy homes minus “LX” homes. “OX” homes are all non-primary residence homes (from HMDA). Home prices are taken from the Zillow Home Value Indices (bottom-, middle-, and top-tier homes in the Ft. Lauderdale/Miami region) and the National Association of Realtors (median price of existing one-family homes for Miami-Ft. Lauderdale and West Palm Beach) as reported in HAVER. We use an average of top- and middle-tier home values for the price of “HX” homes, the value of low-tier homes for the price of “LX” homes, and the NAR median price for “OX” homes. Initial turnover rates are given by the number of home mortgages from HMDA divided by existing housing stock net of new home construction8 We also use data from the American Enterprise Institute on months supply of existing homes to calculate the turnover as follows: (12/months supply of housing) * MSA housing inventory * (Miami-Dade County home sales/MSA home sales).9 We use the best 30-year housing market growth rate (for the years 1976- 2005) to project a 2.2 percent annual growth rate for the business-as-usual (BAU) scenario described below. We use the methodology described in section (G.2) to generate price trends for floodplain homes. We assume flood plain homes grow at the rate of population growth (1.16 percent), and stop growing thereafter. We do not model the destruction of housing stock. However, as described above, we do assume that floodplain homes lose 80 percent of their value in the wake of a severe hurricane. 7Miami-Dade defines low-income households as $73.1 thousand in income for a family of four, or 80 percent of AMI as of April 2020) 8Whilethisneglectscash-onlysales,thisistheturnoverratethatisrelevantforfinancial institutions’ financial health. 9AvailableonHaveratUSRegional⇒SelectedRegionalIndicators⇒HousingMarket Statistics. 12

We use newspaper reporting on flood insurance coverage in the wake of HurricaneIan tocondition our scenarioinsurance assumptions (e.g., Flavelle, 2022b, [15]). Home replacement cost values are taken from the Hazus model. See appendix section F for alternative calculations for both the NFIP flood coverage rate and replacement cost values. 3.2 Assumptions Each scenario will share some common parameter values as shown in Table 1. We assume that both investors and owner-occupiers purchase their homes with mortgages with a 20 percent downpayment.10 Banks’ share of mortgage originations is 34 percent.11 If a home is foreclosed, the unrecoverable costs offoreclosureare20percent.12 Lastly, thenominalinterestrateis1.5percent throughout the period of analysis. Table 1: Universal parameter values Investor downpayment rate 0.2 Owner-occupier downpayment rate 0.2 Bank share of mortgage originations 0.34 Foreclosure recovery share 0.8 Nominal interest rate 1.5 Wenowdiscusseachscenariointurn. Formoreinformationonthechoices for scenario parameters, see appendix section G. 10Although a downpayment rate of 20 percent is ideal, the typical downpayment for first-time homebuyers was 7 percent in 2021 and 17 percent for repeat buyers according to the National Association of Realtors. https://www.nar.realtor/blogs/economistsoutlook/tackling-home-financing-and-down-payment-misconceptions . Our assumption of 20 percent is therefore conservative. 11According to 2021 Home Mortgage Disclosure Act (HMDA) data, non-depository, independent mortgage companies accounted for 63.9 percent of first lien, 1-4 family, sitebuilt, owner-occupied, closed-end home-purchase loans, an increase from 60.7 percent in 2020. https://www.consumerfinance.gov/data-research/hmda/summary-of-2021-data-onmortgage-lending/ 12Frame, 2010, [17], e.g., surveys foreclosure discount estimates that range from 22 to 50 percent. Pennington-Cross, 2006, [29] links the discount to loan size, time in realestate-owned (REO) status, local house price movements, and being located in a judicial foreclosure state. 13

3.3 The Business-as-Usual (BAU) Scenario We simply extrapolate current growth rates, turnover rates, home construction rates, interest rates, etc. We also set post-hurricane devaluation rates at their highest levels given that major hurricane damage was not expected. While this may represent the expectations of many (most?) home buyers in Florida, the BAU assumption is also unrealistic in light of climate change. This scenario may plausibly represent losses in distant future years out to 2050 only if Florida is enormously lucky in avoiding hurricanes between now and then. The BAU scenario assumes rapid annual growth in both prices and home units at growth rates of 2.2 and 1.6 percent, respectively. Properties in both flood and above-flood plain areas experience similar price and construction trajectories. In addition, historical turnover rates continue unabated. We use the estimated share of homeowners within and outside of flood plains that had flood insurance to proxy for well- and poorly-insured homes (See Rozsa and Werner, 2022, [33], Flavelle, 2022b, [15].) Table 2: Scenario – Business as Usual Floodplain Homes Above-floodplain Homes Devaluation Cat 3 0.4 0.3 Cat 4 0.6 0.4 Cat 5 0.8 0.5 Price Growth Through 2035 0.22 0.22 2040 plus 0.22 0.22 Unit Growth Through 2035 0.16 0.16 2040 plus 0.16 0.16 Turnover Rate Change Through 2035 0 0 2040 plus 0 0 Well-insured share 0.5 0.2 Additive propensity to default 0.4 14

We develop insurance profiles for well- and poorly insured homes for each of our six different homeowner types, given in Table 3. Beyond the higher insurance coverage of well-insured homeowners, who have NFIP flood insurance, high-income and investor properties in floodplains are assumed to have additional private flood insurance given that NFIP policies are capped at $250,000. Well insured homes’ wind coverage is 80 percent of replacement cost of the structure of the home (where the cost of the structure is provided in HAZUS data), while that of poorly-insured homes covers only 55 percent of the replacement cost. Table 3: BAU-Scenario Insurance Assumptions Type Well-insured? NFIP Private Flood Private Wind (y/n) ($K) ($K) (% structure val.) HA y 233 0 0.80 HF y 233 40 0.80 LA y 233 0 0.80 LF y 233 0 0.80 OA y 233 0 0.80 OF y 233 40 0.80 HA n 0 0 0.55 HF n 0 0 0.55 LA n 0 0 0.55 LF n 0 0 0.55 OA n 0 0 0.55 OF n 0 0 0.55 Notes: Homeowner types are: HA = high/mid income, above-floodplain, HF = high/mid income, floodplain, LA = low income, above-floodplain, LF = low income, floodplain, OA = other, above-floodplain, and OF = other, floodplain. Simulation results are given in Table 4. We calculate losses in billions of dollars resulting from a single incidence of the relevant simulated for the years 2025, 2030, 2035, 2040, 2045, and 2050. The local economy evolves over time according to each scenario’s assumptions in ways that may make it more or less vulnerable, depending on the scenario. These losses are reported for each loss-absorbing agent for each category on the Saffir-Simpson scale 15

from Cat 1 to Cat 5. Although we can disaggregate the data extensively, we report category totals as well as the share of the banking sector’s Miami- Dade mortgage portfolio that is lost. It is important to note that these damage estimates are for a specifically-parameterized hurricane and should not be viewed as representative of all possible hurricanes. For example, a moderately fast and narrow Cat 1 hurricane will cause far less damage than a wide and slow-moving Cat 1 hurricane. All parties suffer significant loss from Cat 3 or higher hurricanes. For a Cat 5 hurricane that strikes in 2050, the losses for insurers could rise to as much as $89 billion, with losses to banks on mortgages held on their balance sheet of $ 6.3 billion (or 54.8 percent of their Miami portfolio).13 Given that Miami-area mortgages constitute a small fraction of the overall asset portfolio of large banks, these losses in isolation are not likely to threaten bank solvency. But for smaller banks that are more heavily concentrated in the area, there might be significant distress. The exposure of the various creditors to our six different types of homeowners differs, and HMDA data allow us to draw distinctions between creditors based on relative exposure to the most risky households. Loss rates for other holders of bank-originated mortgages are 28.7 percent for GSE’s, 19.7 percent for securitized mortgage purchasers, and 19.3 percent for interbank purchasers.14 3.4 The Hurricane Ian Spillover Effects Scenario The next scenario postulates a strong reaction to Hurricane Ian that dramatically alters real estate and insurance markets in Florida.15 The continued existence of insurance as an initial loss buffer (after deductables) is question- 13RMS Moody’s best estimate of private insurance losses from Hurricane Ian is $67 billion, with an additional $10 billion loss to NFIP for a total of $77 billion. https://www.rms.com/newsroom/press-releases/press-detail/2022-10-07/rmsestimates-us67-billion-in-insured-losses-from-hurricane-ian . 14We ignore the possibility of ‘put-back’ risk, or the potential that purchasers of securitized mortgages might have a contractual right to return mortgages that fall below a performance standard to the banks that sold the mortgage. We also assume that the homeowner-type shares of the mortgages sold by banks matches the distribution of these shares held on the banks’ own balance sheets. In other words, we do not allow for the adverse selection or moral hazard that has been investigated by [21] (for securities) and [27] for GSE purchases. 15See, e.g., Flavelle (2022a) [14] which describes potential consequences of out-of-reach insurance for Florida’s housing market. 16

Table 4: Scenario – Business as Usual 2025 2030 2035 2040 2045 2050 Cat 1 Insurers 1,422,105 1,540,548 1,668,856 1,807,850 1,958,420 2,121,532 Homeowners 71,806 78,293 90,530 107,396 126,787 150,556 Bankheld 24,326 29,734 36,311 43,856 53,261 64,442 %ofmortageportfolio 0.6% 0.6% 0.6% 0.6% 0.6% 0.6% Bank-originatedbutnotheld 37,661 46,035 56,241 67,904 82,477 99,792 Bank-originatedother 25,564 31,249 38,177 46,094 55,986 67,741 Non-bankoriginated 169,951 207,741 253,769 306,423 372,170 450,304 Total 1,751,413 1,933,600 2,143,884 2,379,524 2,649,102 2,954,366 Cat 2 Insurers 3,707,877 4,016,694 4,351,233 4,713,635 5,106,219 5,531,502 Homeowners 226,653 249,923 275,385 308,828 342,464 389,847 Bankheld 24,326 29,734 36,311 43,856 53,261 64,442 %ofmortageportfolio 0.6% 0.6% 0.6% 0.6% 0.6% 0.6% Bank-originatedbutnotheld 37,661 46,035 56,241 67,904 82,477 99,792 Bank-originatedother 25,564 31,249 38,177 46,094 55,986 67,741 Non-bankoriginated 169,951 207,741 253,769 306,424 372,171 450,304 Total 4,192,032 4,581,376 5,011,115 5,486,742 6,012,578 6,603,627 Cat 3 Insurers 7,871,837 8,527,459 9,237,687 10,007,066 10,840,525 11,705,182 Homeowners 19,300,000 23,300,000 28,100,000 33,900,000 40,900,000 49,300,000 Bankheld 151,784 181,464 217,225 260,555 312,840 384,027 %ofmortageportfolio 3.5% 3.4% 3.4% 3.3% 3.3% 3.4% Bank-originatedbutnotheld 127,474 152,663 183,073 219,409 263,595 332,123 Bank-originatedother 95,795 114,684 137,482 164,778 197,927 247,924 Non-bankoriginated 728,044 871,220 1,043,926 1,251,558 1,503,173 1,871,438 Total 28,274,934 33,147,490 38,919,393 45,803,366 54,018,060 63,840,694 Cat 4 Insurers 20,037,181 21,715,237 23,499,792 25,391,391 27,590,618 29,898,111 Homeowners 28,200,000 33,900,000 40,800,000 49,100,000 59,100,000 71,200,000 Bankheld 871,819 1,057,617 1,258,331 1,497,623 1,785,934 2,132,540 %ofmortageportfolio 20.2% 20.0% 19.5% 19.2% 18.9% 18.6% Bank-originatedbutnotheld 628,287 756,040 897,038 1,064,928 1,267,365 1,510,590 Bank-originatedother 489,774 590,572 701,090 832,728 991,412 1,182,083 Non-bankoriginated 3,862,709 4,667,033 5,544,892 6,590,835 7,851,498 9,366,589 Total 54,089,771 62,686,499 72,701,144 84,477,505 98,586,827 115,289,912 Cat 5 Insurers 59,675,523 64,673,429 69,979,489 75,794,383 82,118,846 88,953,676 Homeowners 38,700,000 46,200,000 55,300,000 66,200,000 79,400,000 95,200,000 Bankheld 2,658,525 3,129,476 3,753,645 4,454,419 5,281,918 6,271,536 %ofmortageportfolio 61.5% 59.2% 58.2% 57.2% 55.8% 54.8% Bank-originatedbutnotheld 2,001,947 2,341,945 2,769,197 3,252,756 3,824,587 4,507,738 Bank-originatedother 1,548,189 1,813,848 2,151,897 2,533,719 2,984,721 3,523,673 Non-bankoriginated 12,100,000 14,100,000 16,800,000 19,900,000 23,500,000 27,800,000 Total 116,684,185 132,258,697 150,754,229 172,135,277 197,110,073 226,256,623 17

able. In the wake of Hurricane Andrew, some insurers went bankrupt, and several withdrew from the Florida market altogether (see, e.g., MGI 2020 [25]). The state set up a Florida taxpayer-backed supplemental insurance vehicle to fill the gap, and private captive re-insurance companies emerged in the Cayman Islands to backstop Florida insurers after more well-known re-insurers increased rates or left the market. It is possible that private insurance options may cease to exist or that new insurance company entrants are unreliable. Moreover, the patchwork regulation of insurance by local, state, and federal regulators implies different outcomes by region that challenge a one-size-fits-all model to metro region flow-of-risk analysis. We therefore model a stark reaction to Hurricane Ian in which insurers flee the state causing insurance coverage to decline sharply. The lack of insurability causes home price trends to turn negative and brings new home construction to a halt for floodplain housing. Table 5 displays our assumptions. We lower the ex-post devaluation for floodplain homes from the BAU scenario given that homeowners are already aware of the potential for hurricane damage. Prices for floodplain homes begin to fall at half the rate of recent increases through 2035, after which climate change leads them to fall even more sharply. For above floodplain homes, prices rise at half their previous rate through 2035, after which they plateau. Home construction in floodplain zones ceases (beyond replacement) while construction in above-floodplain zones proceeds at half its previous rate. Greater difficulty in selling floodplain homes leads to a decrease in the turnover rate that intensifies slightly in 2040 and beyond. The well-insured share of both homeowner types drops severely, and homeowners are more likely to default for a given loss of equity. We also make adjustments to insurance coverage for well- and poorlyinsured homeowners. Well-insured now means the homeowner in all six categoriesiscoveredbyNFIPfloodinsuranceupto$233,000,andwinddamageof up to 60 percent of the value of the structure. For poorly-insured homeowners, there is no flood insurance coverage (NFIP or private) and homeowners insurance only covers 30 percent of wind damage. Simulation results are given in Table 7. In this scenario, mortgage portfolios are much smaller given lower turnover rates and smaller mortgages (due to reductions in home prices over time). So the loss rates are applied to smaller balances. In this way, the reaction to Hurricane Ian can be seen as corrective,helpingtoright-sizetherealestatemarketrelativetoclimaterisks. For example, bank-held mortgages in this scenario reach only $7.4 billion by 18

Table 5: Scenario – Hurricane Ian Spillovers Floodplain Homes Above-floodplain Homes Devaluation Cat 3 0.2 0.10 Cat 4 0.3 0.15 Cat 5 0.4 0.20 Price Growth Through 2035 -0.014 0.11 2040 plus -0.080 0.00 Unit Growth Through 2035 0 0.008 2040 plus 0 0.008 Turnover Rate Change Through 2035 -0.2 0 2040 plus -0.3 0 Well-insured share 0.25 0.10 Additive propensity to default 0.55 19

Table 6: Hurricane Ian Spillovers Insurance Assumptions Type Well-insured? NFIP Private Flood Private Wind (y/n) ($K) ($K) (% structure val.) HA y 233 0 0.6 HF y 233 0 0.6 LA y 233 0 0.6 LF y 233 0 0.6 OA y 233 0 0.6 OF y 233 0 0.6 HA n 0 0 0.3 HF n 0 0 0.3 LA n 0 0 0.3 LF n 0 0 0.3 OA n 0 0 0.3 OF n 0 0 0.3 Notes: Homeowner types are: HA = high/mid income, above-floodplain, HF = high/mid income, floodplain, LA = low income, above-floodplain, LF = low income, floodplain, OA = other, above-floodplain, and OF = other, floodplain. 20

2050 compared with $11.5 billion in the BAU scenario. The distribution of those mortgages is also more skewed towards non-floodplain properties in the Hurricane Ian spillover scenario. Consequently, instantaneous losses are smaller for all parties, despite the relative high rates of default and lack of insurance. Again, focusing on a Cat 5 hurricane that hits in 2050, total losses are $63.3 billion in this scenario compared with $226.3 billion in the BAU scenario. Insurer losses fall from $95.2 billion to $31.8 billion, and bank-held mortgage losses fall from $6.3 billion (54.8 percent of the Miami mortgage portfolio) to $2.2 billion (or 29.5 percent of the portfolio). Milder hurricane scenarios due lead to higher losses under this scenario than the BAU scenario, however. The assumed deterioration in price trends and turnover rates beginning after 2035 lead to higher percentage losses on banks’ mortgage portfolios under the Hurricane Ian Spillover scenario for Cat 1 through 3 hurricanes from 2040 onwards. In a sinking real estate market, moderate shocks will lead to higher rates of default. However, once shocks become sufficiently large, the sinking-marketfragility effect is overwhelmed by the generally poor level of resilience of the entire market. 3.5 The Cautious Markets Scenario Our final scenario envisions real estate markets turning cautious while insurance coverage rates rise significantly. Agents take maximum action to anticipate and prepare for climate risks under this scenario, with specific assumptions given in Table 8. In many aspects, the parameter assumptions are similar to those of the Hurricane Ian Spillovers scenario. The main differences are that almost all households, regardless of floodplain status, are well-insured; that even prices for above-floodplain homes eventually begin to decline, and that propensity to default is lower given that expectations are better calibrated towards climate risks. Thedefinitionofwell-insurednowmeans90%coverageofstructuraldamage, full NFIP insurance, and an additional $40,000 of private flood insurance, as shown in Table 9. Poorly-insured is almost identical except for the absence of flood insurance. Simulation results are given in Table 10. Given that insurers now absorb the bulk of losses, loss rates fall for all other parties. Even so, insurer losses are roughly comparable to the BAU losses and are even lower in extreme cases, such as a Cat 5 hurricane in 2050, despite the far higher insurance 21

Table 7: Scenario – Hurricane Ian Spillovers Effect 2025 2030 2035 2040 2045 2050 Cat 1 Insurers 684,771 697,443 697,443 697,443 697,443 697,443 Homeowners 110,766 113,524 119,671 129,330 139,392 151,368 Bankheld 51,053 55,810 58,355 58,480 151,444 249,823 %ofmortageportfolio 0.8% 0.7% 0.7% 0.7% 2.0% 3.4% Bank-originatedbutnotheld 64,971 72,138 76,427 77,942 144,086 213,125 Bank-originatedother 42,141 46,886 49,760 50,894 98,952 149,601 Non-bankoriginated 307,027 339,385 358,229 363,613 765,759 1,189,066 Total 1,260,730 1,325,187 1,359,886 1,377,703 1,997,076 2,650,426 Cat 2 Insurers 1,867,320 1,903,494 1,903,494 1,903,494 1,903,494 1,903,494 Homeowners 363,299 375,052 380,256 391,229 399,174 416,154 Bankheld 51,053 55,811 58,355 60,202 175,490 293,667 %ofmortageportfolio 0.8% 0.7% 0.7% 0.8% 2.3% 4.0% Bank-originatedbutnotheld 64,972 72,139 76,428 79,073 160,323 242,720 Bank-originatedother 42,141 46,886 49,760 51,786 111,214 171,957 Non-bankoriginated 307,029 339,386 358,231 370,885 867,759 1,375,022 Total 2,695,814 2,792,768 2,826,524 2,856,671 3,617,454 4,403,015 Cat 3 Insurers 4,034,063 4,116,560 4,116,560 4,116,560 4,116,560 4,116,560 Homeowners 7,370,094 7,459,375 7,437,682 6,470,597 5,828,364 5,406,228 Bankheld 53,633 58,976 62,276 302,479 605,080 748,601 %ofmortageportfolio 0.8% 0.8% 0.8% 3.8% 7.8% 10.1% Bank-originatedbutnotheld 66,777 74,341 79,144 248,691 461,308 566,296 Bank-originatedother 43,646 48,720 51,987 175,705 331,205 408,183 Non-bankoriginated 318,463 353,368 375,438 1,410,992 2,712,975 3,344,802 Total 11,886,676 12,111,341 12,123,087 12,725,024 14,055,492 14,590,670 Cat 4 Insurers 10,362,955 10,623,717 10,623,717 10,623,717 10,623,717 10,623,717 Homeowners 12,400,000 12,600,000 12,600,000 11,100,000 10,100,000 9,500,216 Bankheld 217,369 244,060 251,394 820,514 1,198,659 1,266,701 %ofmortageportfolio 3.2% 3.3% 3.2% 10.4% 15.5% 17.1% Bank-originatedbutnotheld 186,410 208,434 215,371 618,625 885,641 933,088 Bank-originatedother 130,698 146,923 152,227 445,024 639,499 675,121 Non-bankoriginated 1,037,514 1,163,575 1,201,573 3,657,495 5,287,376 5,580,707 Total 24,334,946 24,986,709 25,044,282 27,265,375 28,734,892 28,579,549 Cat 5 Insurers 31,028,955 31,834,483 31,834,483 31,834,483 31,834,483 31,834,483 Homeowners 20,200,000 20,500,000 20,400,000 18,500,000 17,200,000 16,200,000 Bankheld 1,073,418 1,253,443 1,321,399 2,168,634 2,387,930 2,187,986 %ofmortageportfolio 15.8% 16.8% 16.8% 27.5% 30.9% 29.5% Bank-originatedbutnotheld 940,160 1,073,457 1,115,956 1,743,196 1,899,883 1,751,174 Bank-originatedother 652,266 747,185 778,090 1,224,745 1,341,413 1,241,419 Non-bankoriginated 5,174,873 5,967,342 6,241,746 9,970,999 10,900,000 10,100,000 Total 59,069,671 61,375,911 61,691,674 65,442,057 65,563,710 63,315,062 22

Table 8: Scenario – Cautious Markets Floodplain Homes Above-floodplain Homes Devaluation Cat 3 0.2 0.10 Cat 4 0.3 0.15 Cat 5 0.4 0.20 Price Growth Through 2035 -0.02 0.00 2040 plus -0.08 -0.02 Unit Growth Through 2035 0 0.008 2040 plus 0 0.008 Turnover Rate Change Through 2035 -0.3 0 2040 plus -0.3 0 Well-insured share 0.9 0.9 Additive propensity to default 0.4 23

Table 9: Cautious Markets Insurance Assumptions Type Well-insured? NFIP Private Flood Private Wind (y/n) ($K) ($K) (% structure val.) HA y 250 40 0.9 HF y 250 40 0.9 LA y 250 40 0.9 LF y 250 40 0.9 OA y 250 40 0.9 OF y 250 40 0.9 HA n 0 0 0.9 HF n 0 0 0.9 LA n 0 0 0.9 LF n 0 0 0.9 OA n 0 0 0.9 OF n 0 0 0.9 Notes: Homeowner types are: HA = high/mid income, above-floodplain, HF = high/mid income, floodplain, LA = low income, above-floodplain, LF = low income, floodplain, OA = other, above-floodplain, and OF = other, floodplain. 24

coverage rates. This is due to a smaller amount of home construction in floodplain areas over our study period. Banks hold far smaller real estate portfolios, with a combined total of $5.9 billion in mortgage loans held in 2050,comparedto$7.4billionand$11.5billionintheIANandBAUscenarios respectively. Loss rates for banks are much smaller, with maximum losses of 19.3% in the 2050 Cat 5 outcome, compared with 29.5% and 54.8% in the IAN and BAU scenarios respectively. Total losses under the Cautious Markets scenario amount to $98.9 billion, compared with $63.3 billion and $226.3 billion in the IAN and BAU scenarios respectively. 4 Conclusion This paper simulates a flow-of-risk approach to a specific climate event that affectsstrategicmortgagedefaults. Itsvalueliesinthevariousconsiderations that affect insurance, homeowner, and creditor decisions in the presence of diversityinborrowercharactisticsandclimaterisk. However, eveninthenarrow confines of the exercise, this flow-of-risk model does not consider market, counterparty, or operational risks, nor does it endogenously model real economy impacts. Moreover, even though hurricanes of different categories are considered under conditions of a rising sea level, there are in principle an infinite number of, say, Category 5 hurricanes that could be designed that differ by track speed, width, shear, and other characteristics. Each of these theoretical hurricanes could lead to different levels of damages and loss. For these reasons, the flow-of-risk model should be thought of as a module in a larger suite of models that could help evaluate climate risk to financial institutions. We now turn to several possible paths forward. As mentioned, Miami mortgages are likely to be a limited portion of any large bank’s balance sheet implying that even momentous losses in Miami are manageable in isolation. However, the correlation of climate events across both time and space is rising significantly.16 If a given bank faces repeated climate disasters affecting a portion of its portfolio, or simultaneous climate events across all regions of its business footprint or asset types, losses might add up sufficient to threaten distress. One approach might be to repeatedly conduct joint flow-of-risk-type analyses across a bank’s major business 16See, for example, the increase in both the number and the joint occurrences of large climate disasters in NOAA’s Billion Dollar Disasters data. https://www.ncei.noaa.gov/access/billions/mapping 25

Table 10: Scenario – Cautious Markets 2025 2030 2035 2040 2045 2050 Cat 1 Cat1 Insurers 3,054,115 3,128,427 3,128,427 3,128,427 3,128,427 3,128,427 Homeowners 16,189 16,604 17,365 18,606 19,905 21,474 Bankheld 50,973 54,387 55,017 52,522 115,595 179,885 %ofmortageportfolio 0.7% 0.7% 0.7% 0.7% 1.7% 3.0% Bank-originatedbutnotheld 63,943 69,217 70,812 68,686 107,556 146,145 Bank-originatedother 41,463 44,966 46,068 44,776 73,083 101,529 Non-bankoriginated 303,559 327,226 333,682 322,204 575,042 829,967 Total 3,530,242 3,640,828 3,651,371 3,635,221 4,019,607 4,407,428 Cat 2 Insurers 6,188,516 6,328,672 6,328,672 6,328,672 6,328,672 6,328,672 Homeowners 54,480 56,093 56,760 58,178 59,236 61,449 Bankheld 50,973 54,387 55,017 52,611 118,437 185,168 %ofmortageportfolio 0.7% 0.7% 0.7% 0.7% 1.8% 3.1% Bank-originatedbutnotheld 63,943 69,217 70,812 68,738 109,278 149,338 Bank-originatedother 41,463 44,966 46,068 44,814 74,369 103,918 Non-bankoriginated 303,559 327,226 333,683 322,551 586,398 851,058 Total 6,702,934 6,880,561 6,891,011 6,875,563 7,276,388 7,679,603 Cat 3 Insurers 11,785,069 12,051,615 12,051,615 12,051,615 12,051,615 12,051,615 Homeowners 6,288,613 6,092,540 5,808,625 4,636,502 3,785,778 3,157,120 Bankheld 50,992 54,420 55,076 191,105 374,898 456,171 %ofmortageportfolio 0.7% 0.7% 0.7% 2.7% 5.6% 7.7% Bank-originatedbutnotheld 63,954 69,236 70,847 157,592 273,523 326,250 Bank-originatedother 41,471 44,980 46,093 108,466 192,494 231,269 Non-bankoriginated 303,633 327,353 333,914 887,434 1,632,365 1,967,752 Total 18,533,732 18,640,144 18,366,170 18,032,714 18,310,673 18,190,177 Cat 4 Insurers 28,232,298 28,888,386 28,888,386 28,888,386 28,888,386 28,888,386 Homeowners 9,676,562 9,388,947 8,962,466 7,200,111 5,919,813 4,972,503 Bankheld 116,142 120,848 116,936 421,167 656,459 729,795 %ofmortageportfolio 1.7% 1.6% 1.6% 5.9% 9.9% 12.3% Bank-originatedbutnotheld 105,926 111,977 110,612 305,499 454,178 497,932 Bank-originatedother 71,438 75,567 74,582 214,570 322,714 355,708 Non-bankoriginated 569,747 598,642 586,487 1,827,106 2,782,389 3,073,727 Total 38,772,112 39,184,366 38,739,468 38,856,839 39,023,939 38,518,051 Cat 5 Insurers 82,381,354 84,440,942 84,440,942 84,440,942 84,440,942 84,440,942 Homeowners 13,500,000 13,100,000 12,500,000 10,200,000 8,453,299 7,176,015 Bankheld 336,820 394,137 388,143 885,749 1,173,052 1,144,842 %ofmortageportfolio 4.9% 5.3% 5.2% 12.3% 17.7% 19.3% Bank-originatedbutnotheld 250,641 291,023 288,556 611,340 794,921 774,158 Bank-originatedother 173,863 202,510 200,789 431,884 564,813 551,489 Non-bankoriginated 1,477,863 1,723,123 1,703,357 3,744,476 4,916,584 4,795,656 Total 98,120,540 100,151,734 99,521,786 100,314,390 100,343,610 98,883,102 26

regions for different climate event severities and correlations supported by climate modeling. The output could be used to develop a probability distribution of losses. Similar to stress-testing methodologies, a focus on a pre-specified level of tail risk might be used to judge safety and soundness. In addition, while the scenario inputs are chosen to be plausible, a better approach would be to derive them from a companion model that integrates regional economic outcomes with home prices, incomes, and other key variables. For example, we have modeled the default decision as strategic (based on the willingness of borrowers to repay their debts). However, even willing borrowerswillnotbeabletorepayiftheylosetheirjobsandareunablemake their payments. A companion model that provides estimates of the impact on incomes can factor in how changes in the ability to pay might change default rates. It would also be useful to include expected public support, which is currently absent from the model. A more holistic approach would also address market, counterparty, and operational risks in addition to the credit-risk outcomes addressed by the flow-of-risk model. 27

Appendix A Model To set up the modeling framework, we first establish the flow and stock relationships for the adding, financing, and distributing of real estate equity. Prices and price trends are taken as exogenous to the model. Two important distributional concerns are tackled here. First, homeowners are separated into six categories reflecting income, purpose of homeownership, and exposure to flood risk. Second, homeowners are divided into cohorts that reflect the amount still owed on mortgages relative to the value of the original loan. Both of these factors will influence the decision to default in the event of hurricane damage. A.1 Homeowner types We will exploit the homeowner type to distinguish between high- and lowincome residents, and outside investors. We will also distinguish between homes built inside and outside high-risk flood plains. This in turn gives six separate types of homeowners. Let the first character denote identity (H=high-income, L=low-income, O=outside investor) and the second character denote location (F=floodplain, A=above floodplain). Thus, we have: j ∈ {HF,LF,OF,HA,LA,OA}. We do not count low income households who rent as part of the “LF, LA” category, as they do not hold mortgages. Properties that are rented to low-income households are assumed to be part of the high-income, “OF, OA”, category, whereby the owner will presumably have engaged in the same investor-motivated behavior that characterizes outside owners. A.2 Residential real estate dynamics Homeowner exposure to real estate is determined by the rate at which the homeowners’ home equity grows. Home equity, E, grows with: 1. The degree of price appreciation realized by (all) homeowners of type j, which is γj·Hj ·pj , where γj is price appreciation between time t−1 t t−1 t−1 t 28

and t of the average home of homeowner type j, Hj is the number of t−1 properties of type j, and pj is the t−1 price of the average home of t−1 type j; 2. The routine payment of mortgage principal based on the book value of the home at the time of purchase, which is πj · Hj · pj , where t t−1 t−1 πj reflects a steady-state relationship calibrated to represent principal t payments relative to the overall level of the housing stock at time t−1, exclusive of prepayment; 3. Thedeacquisition of homes bysellingexistingproperties, whichis−ωj· t Hj ·pj , where ωj is the rate of turnover of homes of type j between t−1 t−1 t times t−1 and t; 4. Theprepayment of mortgage debt,whichwewillassociatewithturnover of properties, given by πˆj·ωj·Hj ·pj , where πˆj is a steady-state adt t t−1 t−1 t justment factor that calibrates prepayments to the home’s book value; 5. Home acquisition as homeowners of type j purchase existing housing stock, which is ψj·ωj·Hj ·pj , where ψj is the rate of downpayment t t t−1 t−1 t as a fraction of the home’s value; 6. Purchase of new housing stock, which is ψj · ∆Hj · pj, where ∆Hj t t t t represents the increase in the housing stock between times t−1 and t. Putting this together, home equity changes according to: J ∆E = (cid:88)(cid:8)(cid:2) γj +πj −(1−πˆj −ψj)ωj(cid:3) Hj pj +ψjpj∆Hj(cid:9) . t t t t t t t−1 t−1 t t t j=1 where E represents housing equity owned by all homeowners at time t. t A.3 Equity held by non-homeowners Changes in home equity not held by homeowners, Q , is equal to: t J ∆Q = (cid:88) ∆ (cid:0) Hjpj(cid:1) −∆E . t t t t j=1 The amount of this home equity initially held by banks is given by the following elements: 29

1. Reductions based on the payment of mortgage principal based on the book value of the home at the time of purchase as described above, which is −πj ·Hj ·pj ; t t−1 t−1 2. Reductions based on the prepayment of mortgage debt, as described above, which is −πˆj ·ωj ·Hj ·pj ; t t t−1 t−1 3. The financing of home purchases, both existing and new, equal to: (1−ψj)·(ωjHj ·pj +∆Hj ·pj), where ωj is the rate of turnover of t t t−1 t−1 t t t homes of type j between times t−1 and t; Thus:     (cid:88) J     ∆Q = (1−ψj)(ωjHj pj +∆Hjpj)−(πj +πˆjωj)Hj pj , t t t t−1 t−1 t t t t t t−1 t−1 (cid:124) (cid:123)(cid:122) (cid:125) (cid:124) (cid:123)(cid:122) (cid:125) j=1   New mortgages Repayments   J K (cid:88)(cid:88)(cid:16) (cid:17) = mjk −ojk . t t j=1 k=1 where the k superscript refers to lender type, mjk is new mortgages, and ojk t t is repayments. A.4 Ensuring consistency between stocks and flows Inordertomaketheparameterizationastractableaspossible,wewillassume that the shares of financing across homeowner types and lender types are stable over time. Thus: Qjk = (cid:36)jkQ . t t where j is homeowner type and k is lender type, and (cid:36) is the appropriate fixed share value and (cid:80) (cid:36)jk = 1, ∀ jk ∈ J×K. jk We will also assume that the distribution of originations remains fixed such that: mjk = χjkmj, t t ojk = χjkoj. t t where χ is another fixed share such that (cid:80) χjk = 1, ∀ k ∈ K. k 30

This implies that repayments are given by: χjkoj = χjkmj −(cid:36)jk∆Q . t t t Summing over k: oj = mj −(cid:36)j∆Q . t t t where (cid:36)j = (cid:80) (cid:36)jk. k We can then solve for: mj −(cid:36)j∆Q (πj +πˆjωj) = t t . (A.1) (cid:124) t (cid:123)(cid:122) t t (cid:125) Hj pj t−1 t−1 Unknown (cid:124) (cid:123)(cid:122) (cid:125) Known The right hand side of this equation is composed of known variables. In order to address the left hand side of the equation, we take the following approach. We determine the amount of mortgage prepayment (the second term on the RHS), by making use of the turnover rate, the average length of a mortgage, the historical interest rate, and the average historical value of a mortgage. More specifically, we determine the average amount of mortgage remaining for each cohort of each homeowner type, the number of homes for each cohort-homeowner dyad, and finally the amount of prepayment due to the selling of existing homes by homeowner type. The method described below does not include prepayment for other motives, such as refinancing at a lower interest rate, which could in principal be included in equation (A.1). A.5 Average size of mortgage by cohort Let Mj be the amount remaining on mortage j, with Mj equal to the original t 0 loan amount and the subscript 0 referring to the first year of the mortgage. For simplicity, we assume that interest is compounded annually. At the end of the first year, the amount owed will be: Mj = Mj(1+rj)−Fj. 1 0 0 where Fj is the fixed (annual) payment (interest and principal) on the mortgage, and rj is the contractual interest rate on the mortgage. 0 31

Likewise, in the second period: Mj = Mj(1+rj)−Fj, 2 1 0 = Mj(1+rj)2 −Fj(1+rj)−Fj. 0 0 0 and so on. In general, if Mj is the original value of the mortgage, than at 0 any time t: (cid:34) (cid:35) t−1 Mj = Mj(cid:0) 1+rj(cid:1)t −Fj (cid:88) (1+rj)i . (A.2) t 0 0 0 i=0 where T is the total number of years of the original mortgage (e.g., 30 years). We can determine Fj as a function of the initial mortgage amount and interest rate by noting that at the end of the life of the mortgage (at time T), the principal has to be equal to zero, i.e., Mj = 0. Thus: T Mj(1+rj)T Fj = 0 0 (cid:104) (cid:105) (cid:80)T−1(1+rj)i i=0 0 The amount of principal at any given time 1 can be calculated as: Pj = Pj −(Fj −Mjrj). 1 0 0 0 where Pj is principal remaining at time t. At time 2, we have: t Pj = Pj −(Fj −Mjrj), 2 1 1 0 = Pj −2Fj +rj(Mj +Mj). 0 0 0 1 In general, we will have: (cid:32) (cid:33) t−1 (cid:88) Pj = Pj −t·Fj +rj Mj . (A.3) t 0 0 i i=0 The average mortgage for each homeowner type j is calculated from HMDA data. We impose this mortgage on all homeowners of type j. We assume that there is a constant probability (equal to the turnover rate for homeowner of type j) that any homeowner cohort sells their home in any given time t. This will then drive the size distribution of cohorts of homes with mortgages and those that have been paid off. We then take the value of 32

average outstanding mortgage principal for each cohort times the amount of homesremainingineachcohortandmultiplyittimestheturnoverratetoget the overall principal repayment for that cohort in a given year t. Summing ˆ over cohorts in j gives us πjωj. This allows us to solve for the one remaining t t free variable in equation (A.1), which is πj. t A.6 Number of homes in each mortgage cohort Let us define Hj,s as the number of homes owned by homeowners of type j t at time t who took out mortgages s number of years ago. From the HMDA data, we know the average loan term for homeowners of type j and assume this is constant. Furthermore, we assume that the turnover rate of home ownership, ωj is the same across all cohorts. Homeownership will then be t distributed across cohorts as: S t (cid:88) (cid:88) Hj = Hj,s(1−ωj)t−r +Hj,0 (1−ωj). (A.4) t r t−1 t s=1 r=t−s where S is the loan term of the typical mortgage for homeowner type j (e.g., S = 30ifthetypicalmortgagewasa30-yearloan),Hj,s istheoriginalamount of households taking out mortgages at time t = s, the number of cohort s households of type j at time t will be equal to Hj,s = Hj,s(1 − ωj)t−s, and t t Hj,0 are homes owned by homeowners of type j that are fully owned. The t amount of homes funded by mortgages taken out s years ago is given by total loanstohomeownersoftypej foundintheHMDAdata. Weimputeturnover rates from HAVER data provided by Zillow, although we cannot determine these rates by year. Finally, we know the total amount of homes held by homeowers of type j at time t. Using this information, we can calculate Ej,0 . t−1 Let Aj,s be equal to the average mortgage taken out by a homeowner of t type j who bought a home t−s years ago, which we assume to be equal to (1 − ψj )pj . Using equations (A.2) and (A.3), the amount of mortgage t−s t−s principal remaining at any given time t will be equal to: Pj,s Pj,s t = t , Aj,s Pj,s t 0 (cid:40) (cid:34) (cid:34) (cid:35)(cid:35)(cid:41) = 1−(t−s) F 0 j,s +rj,s t− (cid:88) s−1 (cid:0) 1+rj,s(cid:1)i − F 0 j,s (cid:88) i−1 (cid:0) 1+rj,s(cid:1)k . Pj,s 0 0 Pj,s 0 0 i=0 0 k=0 33

where Fj,s/Pj,s is purely a function of the contractual interest rate: 0 0 Fj,s (cid:0) 1+rj,s(cid:1)T 0 = 0 . (cid:104) (cid:105) Pj,s (cid:80)T−1(cid:0) 1+rj,s(cid:1)i 0 i=0 0 Let: Pj,s Υj,s = t = F (cid:0) rj,s(cid:1) . t Aj,s 0 t A.7 Solving for mortgage repayments We can now write our expression for prepayments due to home sales. Somefractionofeachcohortofeachhomeownertype,equaltotheturnover rate at time t, will sell their home and retire (prepay) their mortgage. To determine the amount of total repayments by homeowner type, we need to: (i) determine the contemporary number of borrowers in each homeowner type-cohort dyad, (ii) apply the appropriate (time-sensitive) turnover rate to determine the number of homeowners retiring their mortgages, and (iii) calculate and aggregage the amount of principal left on the mortgages by homeowner type. Consider the case of a single homeowner type and set aside the j superscript. Designate the amount of homeowners in the present cohort s = t = 0 as H0. The one-period-earlier cohort s = −1 will have a total size equal to H−1(1 − ω ) at time t, where ω is the turnover rate of the prior pe- −1 −1 riod. Likewise, the remaining number of cohort s = −2 households will be given by H−2(1 − ω )(1 − ω ). In general, the amount of cohort s −2 −1 remaining at the beginning of time t will be given by: HsΠt−1(1 − ω ). r=s r The share of this cohort that sells their home will be given by the present turnover rate and will equal: ω HsΠt−1(1−ω ). The value of the mortgages t r=s r that they prepay will be equal to the share of principal left to repay times the value of the original mortgage times the remaining size of the cohort, ΥsAsω HsΠt−1(1−ω ). Summing across cohorts gives us the total amount t t r=s r of prepayment: (cid:80)S ΥsAsω HsΠt−1(1−ω ). s=1 t t r=s r Finally, we acknowledge the different homeowner types j to get: (cid:80)S Υj,sAj,sω Hj,sΠt−1(1−ωj) πˆjωj = s=1 t t r=s r . (A.5) t t Hj pj t−1 t−1 34

where the prepayment rate is calibrated to the current value of type-j homeowner housing stock. We can solve for repayments in two ways. We can calculate repayments directly using the mortgage rates and principal owed by each homeowner type and cohort, or we can use the following relationship:17 mj −(cid:36)j∆Q πj = t t −πˆjωj. (A.6) t Hj pj t t t−1 t−1 There is a continuum of choices for the turnover rate ωj that lead to t corresponding repayment rates πj in equation (A.6). In principle, a unique t combination of repayment rate and turnover rate can be calibrated to mortgage income reported on the income statement, but this is beyond the scope of the paper. Rather, we use information on historical turnover rates, and consider the future path of turnover rates one of the key scenario choices of the modeler. A.8 Determining equity holdings by banks, securities purchasers, and GSEs Although these mortgages initially sit with banks, banks will securitize and sell mortgages on to other parties. These shares are obtained for flows from HMDA data and Y14M data provide custodial holdings of GSE and securitized mortgages by LISCC institutions for stocks. We can therefore represent the change in home equity held by investment funds and other parties as: ∆Q = ∆B +∆G +∆F +∆NBFI , t t t t t Q = B +G +F +NBFI . t t t t t where B is bank holdings of home equity, G is GSE holdings of bankt t originated mortages, F is investment fund holdings of bank-originated mortt gates, and NBFI is holdings of all non-bank financial institution (NBFI) t originated mortgages. Note that NBFI-originated mortgages account for the majority of new mortgages (as high as two-thirds).18 17Note: Detailed historical data on home sales are provided by Miami-Dade Office of Appraisal - See bbs.miamidade.gov. 18https://www.wsj.com/articles/nonbank-lenders-are-dominating-the-mortgagemarket-11624367460 35

B Climate Change Damage Generation Process For each scenario, we model five hurricanes at the boundary of each Saffir- SimpsoncategoryusingtheFEMAHazustool. Wechooseatrackthatcarries the hurricanes through the main business district using the “near wave surge model” approach. B.1 Apportioning flood and wind damage The total amount of structural damage that a homeowner can experience is limited to the replacement cost value of the unit structure. In many cases the sum of flood and wind damage exceeds the replacement value. It is common tohearstoriesofhomeownersstrugglingtogetfloodandwindinsurerstopay claims because each has determined that the primary damage was inflicted by the condition that they do not insure (e.g., flood insurers insist that the damage was caused by wind, whereas homeowners insurers insist that the damage was caused by flooding).19 We assume that flood insurance stands first in line, such that if flood damage alone is equal to the replacement cost of the property, wind damage is equal to zero. In general, wind damage will be limited to the difference between the replacement value and flood damage if wind and flood damage together exceed the replacement cost value. C Insurance as the first loss-absorbing layer Thedegreetowhichinsuranceabsorbsriskdependsonthenatureofthedamage (flood or wind, as described above) and the number of households with coverage (the extensive margin) and the extent to which those households are insured (the intensive margin). These two margins will vary significantly across our six different household categories. We therefore adopt a two-step procedure in which the first step is to model the specific type of loss (either actualized or anticipated), followed by determining the size of the loss and the amount that would be covered by insurance. 19See, e.g., https://www.nytimes.com/2021/09/10/your-money/ida-flood-damageinsurance-policy.html. 36

In general, the analysis will determine each entity’s exposure, and then apply loss rates. These loss rates will be dependent on an ordering of priority in claims on the underlying asset. The first losses will be borne by insurers up to the limits of their obligations (or their resources). Any losses above these amounts will spill over to homeowners. Whether homeowners will completely absorb remaining losses will depend on the share of ownership of the home’s equity, and their ability and willingness to continue to honor mortgage obligations. Under circumstances in which they cannot or do not honor these obligations and default, losses will spill over to banks, investment funds, and GSEs in proportion to their share of the mortgage pool. This pari passu assumption may be incorrect if, say, securitization contracts allocate first losses to different tranches of blended assets (circa 2008 CDOs) or require the securitizer to take first losses. Insurance coverage for a shock occurring between times t−1 and t will equal: Ij = (cid:88) (cid:0) λj,r ·Rj,r(cid:1) (C.1) t I,Z,t I r∈N,F,W whereλj,r isType r insurers’lossrateasashareoftotalexposureforashock I,Z,t of Type Z at time t of coverage of homeowners of Type j (covered in detail in section (E)); and Rj,r is exposure of insurers of Type r to homeowners of I Type j.20 In matrix notation: I = L ×R . t I,Z,t I,t where I is a j ×1 vector given by: t   IHF t  ILF  I =  t . . t  .  .   IOA t L is a j×3·j matrix of insurance sector loss rates applicable to each I,Z,t 20Congressional Budget Office. (2019) [8] 37

of our 3·j combinations:  λHF,N, λHF,F, λHF,W 0, 0, 0 ··· 0, 0, 0  I,Z,t I,Z,t I,Z,t  0 ··· 0 λLF,N, λLF,F, λLF,W, ··· 0 ··· 0  L I,Z,t =    . . . . . . . . . I,Z,t . . . I,Z . . . ,t . . . I,Z,t ... . . . . . . . . .    . 0 ··· 0 0 ··· 0 ··· λOA,N, λOA,F, λOA,W, I,Z,t I,Z,t I,Z,t and R is a 3·j ×1 vector given by: I,t   RHF,N I,t  RHF,F   I,t   RHF,W   I,t   .  R I,t =  . . .    ROA,N   I,t   ROA,F   I,t  ROA,W I,t We discuss the determination of the λ terms below. C.1 The extensive vs the intensive insurance margin For the extensive insurance margin, or the number of households that have flood insurance, we assume differing NFIP flood insurance coverage rates for occupied floodzone residential units and above-floodplain units (see scenario assumptions). This essentially doubles our homeowner categories as we now have well- and poorly-insured versions of each of our six homeowner classifications. In what follows, the subscript j therefore refers to 12 different homeowner types: the original six as NFIP-insured, and the original six as non-NFIP-insured. D Homeowners as the second loss-absorbing layer Anylossesthatarenotcoveredbyinsurancespillovertohomeowners. Homeowners must decide whether to absorb these losses in full, or to default on their mortgages. The default choice depends on both the ability of homeowners to service their mortgages (or to reschedule), which has not yet been 38

added to the model, and the willingness to continue servicing their debt (the strategic default motive). D.1 Insurance coverage as a fraction of replacement value vs total property value Insurance coverage is calibrated to the cost of replacing the structure of the home, which can be either greater or less than the value of the property itself. Inexpensiverealestatemarkets, thecostofthestructureisoftensmall relative to the cost of the land parcel. On the other hand, in declining cities, thecostofrebuildingastructuremightbemanytimesgreaterthantheZillow price of the home. If we differentiate between the land parcel cost and the replacement cost of the structure, it becomes apparent that climate change can damage either or both of these categories. A climate event might reduce the desirability of a parcel of land, leading to a permanent reduction in value evenifthestructureonthatparcelhasnotexperienceddamagedirectly. This kind of loss is uninsurable through homeowner insurance riders for flood and wind damage. Alternatively, a climate event might damage the structure of the home without reducing the desirability of the land parcel itself, leading to a loss in value that is insurable. Of course, a climate event may cause both effects simultaneously. We therefore confine insurance coverage to the replacement value of the home even as we allow climate “damage” to pass through to home prices via a reduction in the desirability of a property. Denote the damage accruing to each household by s˜j ·vj ·Hj, where vj t t t t is the replacement value of the home, and s˜j is the damage rate as a share t of home replacement value. Clearly, homeowners will experience no loss on the damage to built property so long as: Ij = s˜j ·vj ·Hj. t t t t This equation holds with equality because insurance payouts will never exceed the amount of damage. However, it is possible for s˜j to exceed unity t if, for example, remediation to better protect the property against future climate events is required (by, say, raising the property up on pilings). Homeowner losses due to property damages are therefore equal to: (cid:26) s˜j ·vj ·Hj −Ij, for s˜j ·vj ·Hj > Ij Losses to homeowners of type j = t t t t t t t t . 0, for s˜j ·vj ·Hj = Ij t t t t 39

To these losses must be added any reductions in the market value of the land parcel: Losses due to land parcel devaluation = ∆pj = sˆjpj = f(Sj). t t t t where Sj is a climate forcing process described below. It is clear that a t good proxy for how much residential real estate prices might decline due to unanticipated local developments is needed. D.2 Keeping track of loss cushions by cohort Because each cohort will have different amounts of equity at risk, it is likely that different cohorts of homeowners of the same type j will reach the threshold that triggers default at different levels of sustained damage. In general, we would expect default rates to be a function of: (i) home price appreciation since purchasing the home, (ii) the turnover rate, and (iii) the contractual mortgage interest rate, inter alia.21 For any given cohort s of homeowner type j, the amount of home equity held will equal the value of the home minus the principal owed: Ej,s hj,s = t = pj −Υj,sAj,s (D.1) t Hj,s t t t t We can therefore include hj,s (or its distribution) as an argument in det termining the default rate in the wake of a shock. Equation(D.1)isextremelyuseful, however, inshowingthatfinancialsector vulnerability to mortgage default will depend in part on the composition of cohorts within a given homeowner type j, the size of the original mortgage relative the the value of the home (which implicates both price appreciation since purchase as well as the prevalance of refinancing), and the mortgage interest rate (with consideration that high interest rates can be refinanced but also that adjustable rate mortgages may trap less wary borrowers into higher interest rate payments). Now that we have established the amount of losses passing through to homeowners, we focus on the amount of this residual loss that homeowners 21Thereisalargeliteratureondefaultsandnegativeequitythatcaninformthisdiscussion, including: Bhutta et al. (2010) [6], Scharlemann and Shore (2016) [34], and Foote et al. (2008) [16]. 40

are willing and able to absorb. The ability to continue to service mortgage debt will depend on a homeowner’s income and their ability to reschedule their mortgage. These factors are beyond the homeowner’s control. However, the homeowner’s willingness to default, that is, the strategic default motive, is more complicated. To institutional factors that impose penalties for defaulting (such as whether the state is non-recourse, and the impact on credit scores) must be added homeowner expectations about whether sharp (hurricane-induced) home price devaluations will be reversed. The historical experience is that home prices tend to rebound to their pre-disaster levels after a few years. Under these circumstances, it makes sense for homeowners to hang onto their homes until prices recover. However, this historical tendency may not be a good guide for climate change in that sea level rise or perpetual wildfire risk may make such price recoveries impossible. For example, if the land on which a house is built is permanently submerged due to sea level rise, home prices will not recover. Based on the previous section, total damages for the homeowner equal: sjp = s˜jvj(cid:37)j +sˆjp . (D.2) t t t t t t t where (cid:37)j is the ratio of home price to replacement cost for homeowner of t type j. D.3 Default risk by segment and cohort We assume that an increasing fraction χj of homeowners of type j will walk t away from their mortgages as their losses rise. Let losses per household net of insurance be defined as: lj = (cid:0) s˜jvj −ij(cid:1) − sˆjpj . (D.3) t t t t t t (cid:124) (cid:123)(cid:122) (cid:125) (cid:124)(cid:123)(cid:122)(cid:125) Structural damage Devaluation where ij = Ij/(Hj). t t t D.3.1 Cohort equity and strategic default Recalling our discussion of cohort equity and equation (D.1), we can write net (post-event) equity holdings by cohort and segment as: ej,s = hj,s −lj. t t t 41

We follow Bhutta et al. (2010) [6] in estimating the share of homeowners who walk away from their mortgages. Bhutta et al. find support for a dual trigger in which both reductions in the value of the home and reduction in income inform the decision to default. Climate induced damage will likely affect both of these variables, but for now we focus solely on the home price reduction. Florida is a recourse state, meaning that homeowners are still theoretically liable for mortgage debt even if the home is foreclosed. Fortunately, Bhutta et al. [6] focus on Florida as one of the four states in their analysis. According to results presented in their Table 5, it takes a reduction of equity equal to 46 percent of their home’s value for 25 percent of Florida homeowners to strategically default (ceteris paribus), a reduction equal to 79 percent of their home’s value for 50 percent to strategically default, and a decrease in home value equal to 128 percent for all homeowners to strategically default. However, there is an external validity concern that we must also address. In the case in which a negative event leaves the property intact, an individual homeowner might have some expectation that any decrease in value is temporary. However, a climate event may leave the property unsuitable for reconstruction, in which case it is not a question of ‘walking away from your mortgage,’ butyour‘homewalking(floating)awayfromyou.’ Inotherwords, strategic default rates are likely to be much higher than those estimated in Bhutta et al. (2010) [6] After exploring different functional forms, we estimate the percentage of homeowners who walk away from their mortgages by extracting a linear relationship based on the California trend from Figure 6 in Bhutta et al. [6]: (cid:12) (cid:12) χj,s = 0.6× (cid:12) (cid:12) (cid:12)1− (cid:32) P t j,s (cid:33)j,s(cid:12) (cid:12) (cid:12)+Wj, t (cid:12) pj (cid:12) t (cid:12) t (cid:12)   (cid:32) (cid:33)j,s Pj,s ∀1− t  < 0. pj t where as a reminder Pj,s is the remaining loan balance and pj is the (postt t shock) value of the home. The term Wj is a potential adjustment factor to t take into account the unsuitability of reconstruction. Theamountoflossesabsorbedbyhomeownersoftypej willthenbeequal to the share of homeowners who hang onto their homes and fully absorb 42

losses, plus the share of homeowners who walk away times the amount of positive equity that was destroyed. Thus, for Hj > 0: t−1 S S (cid:88) (cid:88) λj Ej = (cid:15)j,s(1−χj,s)ljHj + (cid:15)j,sχj,sEj H,Z,t t−1 t t t t t t t−1 s=1 s=1 (cid:34) (cid:35) (cid:88) S ljHj (cid:88) S λj = (cid:15)j,s(1−χj,s) t t + (cid:15)j,sχj,s H,Z,t t t Ej t t s=1 t−1 s=1 (cid:124) (cid:123)(cid:122) (cid:125) (cid:124) (cid:123)(cid:122) (cid:125) Stay W.A. where (cid:15)j,s is the share of cohort s households in homeowner type j at time t t, and the abbreviation “W.A.” stands for “walk away.” Note that in the case that all homeowners walk away, χj = 1, and the loss rate is also unity, t λj = 1. In other words, homeowners of Type j lose all of their equity, but H,Z,t nothing more. In the case in which no homeowners walk away, homeowners may absorb losses greater than the amount of their equity, depending on whether lj > hj. t t In the case in which there is no positive equity, this approach must be modified. Homeowners who walk away from negative equity lose nothing (note that we address the costs of defaulting separately in the determination of χj,s). The status of those who stay is more complicated. If we adopt a t mark-to-market approach, these homeowners lose an amount equal to the devaluation of their homes. However, it is not possible to express this as a ‘loss rate’ if equity holdings are negative. As an alternative, we think of the choice to stay as maintaining an option to benefit from an upside if home prices were to rise (in addition to the flow of shelter services). The value of this real option is sufficiently low as to change very little with the scale of home devaluation. Consequently, we treat the loss rate λj as effectively H,Z,t zero when Ej < 0. t−1 Define the following net exposure vector (inclusive of unrealized equity 43

gains) for homeowners:  EHFI  t ELFI  t   .   . .     EOAI  R =  t . E,t  EHFU   t   ELFU   t   . .   .  EOAU t And define the diagonal matrix of homeowner loss rates as:   λHFI 0 ··· 0 E,Z,t  0 λLFI ··· 0  L E,Z,t =   . . . E . . . ,Z,t ... . . .   .   0 0 ··· λOAU E,Z,t where λj = 0 if Ej < 0. E,Z,t t−1 Losses faced by each homeowner group can thus be represented by the column vector: N = L ×R . (D.4) t E,Z,t E,t E Non-homeowner asset holders as the final loss-absorbing layer Non-homeowner asset holders experience losses when homeowners default, and risk-mitigation fails to cover losses.22 In this section, we combine our loss estimates into a framework that apportions losses between banks, GSEs, and investment funds. Define Sj as the following vector of dollar damages by homeowner type t 22The model does not consider losses stemming from mark-to-market concerns leading, for example, to collateral calls or other disruptive events. 44

j:   SHFI t  SLFI  Sj =  t .  t  .  .   SOAU t Define Vj as the following (pre-climate event) vector of total home values t pjNj for homeowner type j: t t   VHFI t  VLFI  Vj =  t .  t  .  .   VOAU t The net value of mortgages left to banks, funds, and the government after deducting unabsorbed losses will then be given by the j ×1 vector: J S (cid:88)(cid:88) T = (V −R )− (S −N ) + χj,sVj,s , (E.1) t t−1 h,t−1 t−1 t−1 t t−1 (cid:124) (cid:123)(cid:122) (cid:125) (cid:124) (cid:123)(cid:122) (cid:125) j=1 s=1 Orig. net value Unabsorb. loss (cid:124) (cid:123)(cid:122) (cid:125) Collateral val. J S (cid:88)(cid:88) = P − χj,s(Pj,s −Vj,s ) . (E.2) t t t t−1 (cid:124)(cid:123)(cid:122)(cid:125) Outstanding mort. j=1 s=1 (cid:124) (cid:123)(cid:122) (cid:125) Mort. defaults net collateral where the first underbraced term is the value of home equity held by parties other than the homeowner, and the second term is the spillover loss not absorbed by homeowners (inclusive of insurance coverage). Weassumethatlossesaredistributedacrosscreditortypesonapari passu basis with loss shares equal to: (cid:18) (cid:19) B Banks : αB = t (E.3) t Q t (cid:18) (cid:19) G Government : αG = t (E.4) t Q t (cid:18) (cid:19) F Funds : αF = t (E.5) t Q t (cid:18) (cid:19) B +G +F NBFI : αNBFI = 1−αB −αG −αF = 1− t t t (E..6) t t t t Q t 45

We calculate losses using equation (E.2). We can directly calculate the principle owed by each homeowner type and cohort, Pj,s, using the methods t given above and apply our calculations of χj,s to implement the following for t each institution: J S (cid:88)(cid:88) L = χj,s(Pj,s −Vj,s ). (E.7) Z,t t t t−1 j=1 s=1 F Data addendum We use NFIP data to generate alternative estimates. The number of NFIP policies per state can be used as a proxy for the number of NFIP policies per municipality. NFIP policies per state (by floodzone designation) are provided by NFIP at: https://nfipservices.floodsmart.gov. The number of Floridaoccupiedresidenceslocatedin100-yearandcombined100-to-500year floodplains are provided by FloodZoneData.us [32]. We take the statewide ratio of NFIP policies written on high risk properties (zones A, AE, AH, and AO) per residences located in 100-year floodzones to proxy for the take up rate of NFIP policies by floodzone properties at the municipal level. We likewisetakethestatewideratioofnon-floodzoneNFIPpoliciestonon-floodzone residences to proxy the take up of above-floodplain residences. Data on total occupied residences in Florida is taken from the American Community Survey (https://data.census.gov), as described in Table 11. Based on these numbers, 84 percent of Miami’s NFIP policies should be allocated to floodzone units, or 292,979 policies. Total occupied floodzone units in Miami equal 661,242, for a coverage percentage of 44 percent. The remaining policies, equaling 53,802 cover a total of 222,931 above-floodplain units, for a coverage percentage of 24 percent. We subtract the number of floodzone properties from the total number of occupied units to arrive at above-floodzone property numbers. We take the ratio of non-floodzone NFIP policies to above-floodzone properties to determine the coverage rate for above-floodzone policies in Miami. We use this ratio in combination with the number of floodzone residences in Miami to estimate the number of NFIP policies going to floodzone residences. We assume identical takeup rates by HF, LF, and OF homeowners. The excess of NFIP policies above this estimated figure is apportioned equally between above-floodplain residences. 46

Table 11: State-of-Florida NFIP Coverage Data 100-year Combined # of occupied floodzone units 1,893,920 2,611,010 # of NFIP floodzone policies 1,041,842 1,041,842 % of floodzone units NFIP-insured 55.0 39.9 # of occupied above-floodzone units 6,011,912 5,294,822 # of NFIP above-floodzone policies 605,767 605,767 % above-floodzone units NFIP-insured 10.1 11.4 Miami-Dade County # of NFIP Policies 346,781 346,781 Avg. policy coverage ($ thous.) 233 233 To perform a consistency check on structual replacement values, we used replacementcostvaluesfromHomeConstructionProMatcher(https://homebuilders.promatcher.com/cost/miami-fl-home-builders-costs-prices.aspx). According to the surveyed construction firms in the Miami area, the cost of custom home building in Miami ranges from $111.35 to $165.33 per square foot. We assume that the average lower-income home is 1,800 square feet (22 percent of new single-family homes completed in the South region of the United States were 1,800 square feet or less according to the US Census 2020 Annual Characteristics of New Housing, which reports data collected by the US Department of Housing and Urban Development (HUD)) which implies a replacement cost value of $111.35 × 1,800 ≈ $200,000 for LX homes. Around 65 percent of the total number of homes lie between 1,800 and 3,999 square feet, so we set HX square footage at 2,600 and take a construction cost of $134.61 per square foot to arrive at a replacement cost of $350,000. For OX homes, we take the middle of the cost range, $120, multiplied by the median home square footage (2,261 in 2020) to get approximately $270,000. 47

G Basis for scenario assumptions It is widely recognized that: (i) the underlying stationarity required for the application of standard quantitative risk assessments – including the determination of a ‘fundamental’ price or the assumption of normal statistical moments – is not present with climate change, and (ii) the adjustment of coastal home prices is highly contingent on beliefs about the reality of climate change (see, e.g., Pindyck, 2021, [30], Weitzman, 2011, [37], Bakkensen and Barrage, 2021, [3], and Baldauf et al., 2020, [4]). These factors motivate us to use a scenario analysis approach to our simulations. There are four main dynamic housing market variables that differ across the two scenarios presented here: 1. Thedegreetowhichhomessufferdevaluationinthewakeofahurricane of a given strength. 2. The trend growth of home prices in the presence of chronic sea level rise. 3. The growth in the housing stock for different homeowner categories. 4. The rate at which homes turnover. While each of these variables could in principal be obtained through a dynamic programming solution technique, the stationarity and statistic moment concerns described above make it exceedingly challenging to solve for thesevariables. Rather, wecreatescenariosbasedonknowledgeoflocalcharacteristics. For example, in localities with ample fiscal resources, the modeler might reasonably expect that the government might build the infrastructure necessary to support continued home price appreciation. Alternatively, if the local economy is highly vulnerable to climate shocks, the degree of local home price devaluation might be much higher as jobs disappear and individuals are no longer able to pay their mortgages. Yet another case is one in which migration from vulnerable to non-vulnerable areas within the same locality is desirable leads to falling home prices in the former and rising home prices in the latter. Whether this last effect dominates is an open question. We therefore use scenarios to illustrate how the model can be used to process scenarios, where the specific scenario assumptions are given in the text. We describe the basis for some of these assumptions below. 48

G.1 Hurricane shock devaluation For our scenarios, we assume that instantaneous devaluations due to a hurricaneshockonlyoccurforhurricanecategories3andabove, withthedegreeof devaluation rising in the strength of the hurricane. The highest devaluation of 80 percent is based on a study by the McKinsey Global Institute.23 G.2 Home price depreciation We follow MGI’s projections and set the price trend for homes in floodplains at a pace to bring them to a 30 percent devaluation relative to above-floodplain homes by 2030, and an 80 percent devaluation relative to above-floodplain homes by 2050.24 Sustained home devaluation implies that negative equity will set in at some point depending on the difference between the rate of home price depreciation and the rate of mortgage interest. More generally, if it is clear that prices will face sustained downward pressure, prices should jump to the foreseen lower level immediately consistent with rational pricing models. This is a difficult issue to handle in light of the large empirical literature attempting to explain coastal real estate pricing anomalies. We leave proper consideration of coastal home pricing for future work. For above-flood-plain homes, let: PIXAegA·10 = PIXA. 2020 2030 where PIy X e A ar is a price index for a given year for above-flood-plain homes of type XA, and g is the annual growth rate of those prices. Likewise: A PIXF egF·10 = PIXF . 2020 2030 If flood plain homes are to depreciate 30 percent relative to above-flood-plain homes, it must be the case that: PIXF egF·10 = 0.7×PIXA. 2020 2030 Set PIXA = PIXF = 1. Combining this with the above equations, we 2020 2020 23McKinsey Global Institute (2020) [25], p. 20. 24McKinsey Global Institute (2020) [25], p. 20, maximum devaluation projections. 49

have the following growth rates for the period 2020-2030: ln(0.7) g = +g , F A 10 = g −0.036, A = 0.022−0.036 = −0.014. where g is set equal to 2.2 percent, equal to the best 30 year average growth A in the US housing market (1976-2005). Performing the same exercise for the period 2030-2050, we have: ln(0.7) g = +g , F A 10 = g −0.080, A = 0.00−0.080 = −0.080. where we assume prices for above-flood-plain homes in Miami are flat. One consideration that sustained home devaluation introduces is that the speed with which negative equity sets in depends on the difference between the rate of home price depreciation and the rate of mortgage interest. More generally, if it is clear that prices will face sustained downward pressure, rational pricing models would generate an immediate jump to a lower price consistent with the fundamentals. However, the empirical evidence on the impact of climate change (particularly sea level rise) on home prices is mixed, with some evidence that greater exposure to climate risk lowers home prices but also evidence that risks are not fully capitalized. Moreover, homeowner beliefs tend to affect the degree to which climate risks are incorporated into prices. Empirical studies of natural disasters suggest that prices tend to rebound, andthathomeownerswhoholdonlongenoughtendtoberewarded with the recovery of any lost equity. There is always a real options value to waiting to see how uncertainty is resolved before taking an irreversible action. In light of these conflicting factors, we assume that homeowners do not default outside of a hurricane event. G.3 Housing unit growth In some scenarios, the housing stock grows at the Miami population growth rate during the 2020s (projected to be 1.16 percent according to the planning horizon figures used by Miami-Dade County).25 25https://www.miamidade.gov/water/library/reports/reuse-feasibility-iii.pdf 50

G.4 Turnover rates Bank vulnerability to climate shocks will depend significantly on the cohort structure of home ownership that in turn depends upon the rate of real estate turnover. Consider the polar case in which turnover is zero, new home construction is zero, and all home equity is fully owned by the homeowners. Under such a scenario, banks and other non-homeowners do not hold risk. It is reasonable to assume that turnover rates will fall in riskier areas as it becomes difficult to see in a sinking market. Add to this the possibility that government programs buy out homeowners living in high risk areas, and the probability of additional lending in these areas declines. There are no clear empirical examples of which we are aware to guide us in our choice of the path of turnover rates. 51

References [1] Ariza, Mario Alejandro, 2020. Disposable City: Miami’s Future on the Shores of Climate Catastrophe. New York: Bold Type Books. [2] Auerbach, Alan, Yuriy Gorodnichenko, and Daniel Murphy, 2019. Local Fiscal Multipliers and Fiscal Spillovers in the United States, NBER Working Paper No. 25457. http://www.nber.org/papers/w25457. [3] Bakkensen, Laura, and Lint Barrage. (2021) ”Flood Risk Belief Heterogeneity and Coastal Home Price Dynamics: Going Under Water?” NBER Working Paper No. 23854. [4] Baldauf, Markus, Lorenzo Garlappi, and Constantine Yannelis. (2020) ”Does Climate Change Affect Real Estate Prices? Only If You Believe In It,” The Review of Financial Studies, 33(3): 1256–95. [5] Batten, Sandra, Rhiannon Sowerbutts, and Misa Tanaka, 2016. Let’s talk about the weather: The impact of climate change on central banks, Bank of England Working Paper No. 603. [6] Bhutta, Neil, Jane Dokko, and Hui Shan, 2010. The Depth of Negative Equity and Mortgage Default Decisions, Federal Reserve Board Finance and Economics Discussion Series 2010–35. [7] Bjarnadottir, Sigridur, Yue Li, and Mark G. Stewart, 2014. Regional loss estimation due to hurricane wind and hurricane-induced surge considering climate variability, Structure and Infrastructure Engineering. 10 (11): 1369–1384. doi:10.1080/15732479.2013.816973. [8] Congressional Budget Office, 2019. Expected Costs of Damage From Hurricane Winds and Storm-Related Flooding. Available at: www.cbo.gov/publication/55019. [9] Dennis, Benjamin, 2022. Climate Change and Financial Policy: A LiteratureReview.Federal Reserve Board Finance and Economics Discussion Series 2022–48. [10] Elliot, Diana, TanayaSrini, ShivaKooragayala, andCarlHedman, 2017. Miami and the State of Low- and Middle-Income Housing: Strategies to 52

Preserve Affordibility and Opportunties for the Future, Urban Institute Research Report. [11] Elliott, Diana, Tanaya Srini, Shiva Kooragayala, and Carl Hedman, 2017. Miami and the State of Low- and Middle-Income Housing: Strategies to Preserve Affordability and Opportunities for the Future, Urban Institute Research Report. [12] Emanuel, Kerry, and Thomas Jagger, 2010. On Estimating Hurricane Return Periods, Journal of Applied Meteorology and Climatology. 49: 837–44. doi:10.1175/2009JAMC2236.1. [13] Federal Emergency Management Agency. FEMA’s Hazus Program. https//www.fema.gove/flood-maps/products-tools/hazus [14] Flavelle, Christopher. (2022a) “Why Ian May Push Florida Real Estate Out of Reach for All but the Super Rich.” New York Times. October 13, 2022. [15] Flavelle, Christopher. (2022b) “Hurricane Ian’s Toll Is Severe. Lack of Insurance Will Make It Worse.” New York Times. September 29, 2022. [16] Foote, Christopher, Kristopher Gerardi, and Paul Willen, 2008. Negative Equity and Foreclosure: Theory and Evidence, Journal of Urban Economics. 64 (2): 234–45. [17] Frame, W. Scott, 2010. Estimating the Effect of Mortgage Foreclosures on Nearby Property Values: A Critical Review of the Literature, Economic Review, Federal Reserve Bank of Atlanta, No. 3. [18] Genovese, Elisabetta, and Chloe Green, 2014. Assessment of Storm Surge Damage to Coastal Settlements in Southeast Florida, Journal of Risk Research. 18 (4): 407–27. [19] Goldsmith-Pinkham, Paul, Matthew Gustafson, and Ryan Lewis, 2019. Sea level rise and municipal bond yields, Jacobs Levy Equity Management Center for Quantitative Financial Research Paper. Available at: http://dx.doi.org/10.2139/ssrn.3478364. [20] Goodell, Jeff, 2017. The Water Will Come: Rising Seas, Sinking Cities, and the Remaking of the Civilized World. New York: Little, Brown and Company. 53

[21] Keenan, Jesse, and Jacob T. Bradt, 2020. Underwaterwriting: from theory to empiricism in regional mortgage markets in the U.S., Climatic Change, 162: 2043–67. [22] Keenan, Jesse, Thomas Hill, and Anurag Gumber, 2018. Climate Gentrification: from Theory to Empiricism in Miami-Dade County, Florida, Environmental Research Letters. 13. doi:10.1088/1748-9326/aabb32. [23] Huang, Z., D.V. Rosowsky, and P.R. Sparks, 2001. Long-term Hurricane Risk Assessment and Expected Damage to Residential Structures. Reliability Engineering and System Safety. 74: 239–249. [24] McAlpine, Steven, Jeremy Porter, 2018. Estimating Recent Local Impacts of Sea-Level Rise on Current Real-Estate Losses: A Housing MarketCaseStudyinMiami-Dade, Florida, Population Research and Policy Review. 37: 871–95. doi:10.1007/s11113-018-9473-5. [25] McKinsey Global Insitute, 2020. Will mortgages and markets stay afloat in Florida?. MGI Climate risk and response: Physical hazards and socioeconomic impacts Case Study. Available at:https://www.mckinsey.com/ /media/mckinsey/business%20functions/ sustainability/our%20insights/will%20mortgages%20and%20markets%20stay %20afloat%20in%20florida/mgi-will-mortgages-and-markets-stayafloat-in-florida.pdf. [26] Montero Kuscevic, Casto Mart´ın, 2014. Okun’s law and urban spillovers in US unemployment, The Annals of Regional Science. 53 (3): 719–730. [27] Ouazad, Amine, and Matthew Kahn, 2021. Mortgage finance and climate change: Securitization dynamics in the aftermath of natural disasters, NBER Working Paper 26322, Doi: 10.3386/w26322. [28] Painter, M., 2020. An inconvenient cost: The effects of climate change on municipal bonds, Journal of Financial Economics. 135 (2): 468–82. doi:10.1016/j.jfineco.2019.06.006. [29] Pennington-Cross, Anthony, 2006. The Value of Foreclosed Property. Journal of Real Estate Research. 28 (2): 193–214. 54

[30] Pindyck, Robert. 2021. What We Know and Don’t Know about Climate Change, and the Implications for Policy, in Matthew Kotchen, James H. Stock & Catherine Wolfram (eds.) Environmental and Energy Policy and the Economy, Volume 2, Chicago: University of Chicago Press. [31] Pozsar, Zoltan. 2014. Shadow banking: The money view, OFR Working Paper 14-04. Office of Financial Research. [32] Rosoff, Stephanie, and Jessica Yager, 2017. Housing in the U.S. Floodplains, NYU Furman Center Data Brief. [33] Rozsa, Lori, and Erica Werner. (2022) “Florida’s insurance woes could make Ian’s economic wrath even worse,” Washington Post, September 30, 2022. https://www.washingtonpost.com/climateenvironment/2022/09/30/ian-florida-economy-insurance/ [34] Scharlemann, Therese, and Stephen Shore, 2016. The Effect of Negative Equity on Mortgage Default: Evidence from HAMP’s Principal Reduction Alternative, The Review of Financial Studies. 29 (10): 2850–83. doi:10.1093/rfs/hhw034. [35] Sullivan Sealey, Kathleen, Ray King Burch, P.-M. Binder, 2018. Will Miami Survive?: The Dynamic Interplay Between Floods and Finance. Springer Briefs in Geography, doi:10.1007/978-3-319-79020-6. [36] Union of Concerned Scientists, 2016. Encroaching Tides in Miami-Dade County, Florida: Investing in Preparedness to Manage the Impacts of Rising Seas. Available at: www.ucsusa.org/EncroachingTidesMiamiDade. [37] Weitzman, Martin. 2011. Fat-tailed uncertainty in the economics of catastrophic climate change, Review of Environmental Economics and Policy, 5(2): 275–92. Available at: http://reep.oxfordjournals.org/content/5/2/275 55