Introduction

The U.S. residential real estate market likely has not fully accounted for the financial costs of flooding. Milliman estimates that only 4% of homeowners in the United States have a flood insurance policy today, and homeowners often believe that, if they are not forced to buy flood insurance, then they do not have flood risk. During Hurricane Harvey, it is estimated that approximately 80% of homeowners in the areas that experienced the most damage did not have flood insurance.¹ Homeowners’ misconceptions of their exposure to flood lead us to conclude that the costs of flooding today are likely not fully considered in property values. This leaves the potential for most homeowners’ greatest asset, their homes, to lose significant value if the costs of flooding become realized in the real estate market. In turn, mortgage markets could be adversely impacted by a potentially abrupt repricing event to account for flood risk.

While property values have been relatively resilient to flood events to date, the past may not be representative of the future. Potential home buyers, investors, and lenders could all become less willing to look past the costs of flood risk as climate change increases the frequency and severity of future flood events.

Executive summary

This paper indicates that flooding, and increases in flood risk due to climate change, represent a major potential exposure to homeowners and lenders supporting a currently robust mortgage market.

Milliman has estimated “unpriced flood costs” by determining the difference between the present value of flood costs (either from flood insurance premiums or uninsured flood damages) and the actual flood costs priced into properties today. Unpriced flood costs represent an actual cost to consumers as they pay out of pocket for flood damages and flood insurance premiums that are likely not fully considered when purchasing a home; these flood costs are estimated both in current conditions and under future climate states with consideration for both sea-level rise and precipitation changes.

Unpriced flood costs could result in abrupt property value declines if future buyers consider the full financial impact on the purchase price of a property. Milliman measured the effect property value declines could have on mortgage loan investors using its proprietary credit risk model, which predicts the resulting increase in mortgage defaults and the severity of loss to mortgage investors if these property value declines were recognized today.

The results of this paper show that climate change may have a large impact on future mortgage credit losses and home retention as, by definition, mortgage credit losses only occur when there is a default on the mortgage and the borrower loses their home. Therefore, not only does flood risk present a risk to the value of properties and mortgage investors, it may also present a risk to sustainable homeownership.

• This paper estimates that single-family residential properties in the United States may be overvalued by $520 billion today by not fully accounting for the potential costs of flooding.
• The $520 billion in unpriced flood costs today represents an average potential overvaluation of about 2% for the properties included in this study. Beyond the average, individual results are skewed and come close to matching the Pareto Principle, with only 20% of properties representing 77%, or $402 billion, of the total unpriced flood costs today. The impact to these markets will be greater than to the average market.
• Under the definition of a “High” climate change scenario, unpriced flood costs could grow to as much as $643 billion in 2050. Much of the increased climate risk is in coastal areas, where a High climate change scenario could increase unpriced flood costs by an average of approximately 50%.
• This study estimates that almost 3.5 million homeowners are exposed to major repricing due to unpriced flood costs, where "major" is defined as a decrease in property value greater than 10%. Under a High climate change scenario, about 4.2 million homeowners will be exposed to a major repricing due to unpriced flood costs.
• There is substantial disparity in unpriced flood costs by income level. The average unpriced flood cost as a percentage of property value for households below the poverty line is a relative increase of 17% compared to median-income households (2.3% vs. 2.0%), and a relative increase of 68% compared to high-income households² (2.3% vs. 1.4%).
• This study explores the exposure of mortgage defaults to a repricing of flood risk resulting in mortgage defaults that could increase by as much as 40% in a High climate change scenario; with severity of loss to the mortgage investor of each default increasing as well. As described later in this report, this study sequences natural catastrophe and credit risk models to find credit losses to mortgage investors of government-sponsored enterprise (GSE) loans could be as high as $72.1 billion for mortgage loans held by the GSEs as of September 30, 2020. This compares to $16.4 billion in credit losses projected by Milliman for the same portfolio of loans under the Moody’s Baseline economic scenario.

Flood loss estimates

The Milliman Market Basket is a portfolio representing a cross-section of approximately 10% of all homes in the U.S. single-family home market. This portfolio was modeled in the KatRisk SpatialKat Flood and Storm Surge Modeling suite to obtain estimates of future flood damages, in particular the average annual loss (AAL). AALs represent the expected damages to a home in an average year due to flooding.

In addition to modeling AALs based on current conditions (the Standard scenario), KatRisk provided AALs under several climate change scenarios:

1. A Medium scenario based on the Representative Concentration Pathway (RCP) 4.5³ scenario, which is a reasonable estimation of what will happen if, on average, global economies are able to stabilize and begin to reduce greenhouse gases in the near future.
2. A High scenario based on the RCP 8.5⁴ scenario, which is generally considered to be close to a worst-case scenario, wherein global economies progress with a "business as usual" approach.

AALs were estimated as of conditions in 2050 for the Medium and High scenarios. Each scenario also assumes that local flood defense infrastructure improvements keep pace with climate change at rates similar to today's. In other words, a community mitigating to protect from a 1-in-100-year flood standard will continue to assess and protect to that level, even if a 1-in-100-year flood is a larger flood event in the future. This means that without future infrastructure investment to defend against any increased risk of flooding in the future, estimated flood losses for future climate scenarios would be higher than those presented in this report.

All scenarios contemplate losses from riverine flooding, precipitation-based “flash” flooding, and storm surge from tropical cyclone events. The climate change scenarios also include the impacts of changes in precipitation patterns on riverine and flash flooding, and changes in sea levels that can increase storm surge during tropical cyclones.

More information on the KatRisk SpatialKat model and the Milliman Market Basket can be found in Section 2 of the Society of Actuaries Report coauthored by KatRisk and Milliman, “Residential Flood Risk in the United States.”⁵

Unpriced flood cost estimation

Using the Milliman Market Basket appended with KatRisk-modeled AALs and April 2021 National Flood Insurance Program (NFIP) rates, Milliman calculated the cost of flooding relative to property values to estimate what the potential exposure to homeowners and their property values are with respect to flood risk. The remainder of this section provides detailed calculation steps, and concludes with the table in Figure 1 showing example calculations for three hypothetical properties.

The cost of flooding relative to property values was calculated by separating homeowners into two buckets—those likely to purchase insurance and those unlikely to purchase insurance. Homeowners likely to purchase insurance included all residences inside the Special Flood Hazard Area (SFHA), where flood insurance is mandatory for homeowners with a federally backed mortgage. For homeowners outside the SFHA, where flood insurance purchase is voluntary, Milliman estimated which homeowners were most likely to purchase insurance by developing a predictive model to estimate the probability a home had an NFIP policy based on factors relevant to flood risk such as distance to coast, distance to river, elevation, and AAL based on the KatRisk results.

For homeowners inside the SFHA, or outside and likely to purchase insurance, the NFIP’s April 2021 rates were used to calculate an insurance premium under the Standard scenario. For the Medium and High scenarios, the insurance premium was increased consistent with the change in AAL between each scenario and the Standard scenario, to account for the increase in premium due to an increase in annual losses.⁶ Each insurance premium was then converted to a present value by assuming an interest rate of 2.5% applied in perpetuity. The interest rate was selected based on Moody’s inflation forecasts. Finally, the present value of the insurance premium perpetuity was converted to a percentage of property value by dividing by the estimated property values for each Market Basket location. The resulting estimate is the cost of flooding relative to property values.⁷

Only about 2% of homes outside of the SFHA are likely to purchase flood insurance today. For the remaining homes that were unlikely to purchase flood insurance today, the cost of flooding relative to property values is calculated similarly. The present value of the AALs, rather than an insurance premium, is used for each property.

After calculating the cost of flooding relative to property values, we evaluated how much of this cost might be priced into current property values. Existing research on the impact of flood risk and flood events to property values is limited, often with variable results.⁸ One of the most comprehensive papers on this subject (Hino and Burke, 2020⁹) estimates that, all else equal, homes in the floodplain are valued 2.1% less than those outside the floodplain (the “floodplain discount”). To obtain an estimate of the cost of flooding relative to property values that is not currently priced into homes today, 2.1% was subtracted from the cost of flooding relative to property values for all homes in the SFHA.¹⁰ For example, if the estimated cost of flooding is 10% of the property value, the unpriced cost of flooding is 10% less 2.1%, or 7.9% of the property value. This study assumes that flood risk is not currently priced into homes outside the SFHA, which is a reasonable assumption as very few homeowners outside the SFHA purchase flood insurance, and many incorrectly believe that their homeowners insurance already covers flood damage.¹¹

Note: Figure 1 shows the calculation of unpriced flood costs as a percentage of property value for three hypothetical properties. Example 1 has flood insurance and is inside the SFHA. Example 2 differs from Example 1 only because it is outside the SFHA. This means that the floodplain discount does not apply. All else equal, this increases the unpriced flood costs because we assume that all flood risk is unpriced today. Example 3 differs from Example 2 only because it does have flood insurance. In this case, we use AALs from the KatRisk model to estimate the uninsured flood damages. As the uninsured flood damages were assumed to be equal to the premium for this example, there is no difference in the unpriced flood costs between Examples 2 and 3.

After accounting for the cost of flooding priced into homes today in the SFHA, the unpriced flood cost as a percentage of property value is computed. This metric shows the potential exposure of homeowners to price decreases of their often-largest asset due to flood costs under current and future climate scenarios. It’s important to note that it is not expected that these costs will be realized instantly as there is substantial uncertainty at local levels around how much flood risk is priced into properties today. However, the gap between the estimated costs of flooding and what existing research shows is accounted for in property values today indicates that there is a substantial amount of flood risk not being considered in the residential real estate markets overall.

Unpriced flood cost results

The table in Figure 2 shows the overall results of the unpriced flood costs. As a percentage of property value, it is estimated that countrywide flood costs equal to 2.0% of property values are not accounted for today, and are thus unpriced flood costs relative to property values. This represents a total value of $520 billion countrywide. Under a scenario of high sea-level rise, the average would increase by approximately 20%, netting an increase of $123 billion of unpriced flood costs (for a total value of $643 billion) relative to property value.

Coastal homes, defined in this report as homes within 25 miles of the Gulf of Mexico or Atlantic coast, tend to be most exposed to climate change. Under the High scenario, unpriced flood costs are approximately 50% greater for coastal homes relative to noncoastal homes.

In addition to reviewing the regional costs of flooding relative to property values, this study examined whether particular demographic groups were more exposed to flooding and climate change risk. By weighting the unpriced flood costs by census estimates of income distribution and poverty status, the unpriced flood costs across each group may be inferred. Significant correlation was found between higher unpriced flood costs and lower income/poverty status,¹² which is shown in Figure 3. The average unpriced flood cost as a percentage of property value for households below the poverty line is a relative increase of 17% compared to median-income households (2.3% vs. 2.0%), and a relative increase of 68% compared to high-income households¹³ (2.3% vs. 1.4%).

The above estimates are an average of all homes, which can mask a key trait of flood risk in that it is a high-resolution hazard that results in highly skewed effects on a home-by-home basis. Many homes may have unpriced flood costs relative to property value at or below the average, and some have unpriced flood costs that can be several multiples of the average. For example, in our Standard scenario, 20% of properties represent 77%, or $402 billion, of the total $520 billion in unpriced flood costs today. Further, the 5% of properties with the highest unpriced flood costs represent over 40% of the total unpriced flood costs.

The table in Figure 4 summarizes the number of homeowners who have unpriced flood costs that are more than 10% of the current property value. Under a High climate change scenario, compared to today’s conditions reflected in the Standard scenario, it is estimated that approximately 750,000 more homeowners will exceed the 10% threshold relative to today’s conditions as reflected in the Standard scenario.

Mortgage performance results

Given the large amount of flood risk that is not currently capitalized into property values, there is a risk that a realization of those costs could occur, possibly triggering an increase in mortgage defaults and losses to mortgage security investors.

Milliman used its proprietary econometric credit risk model coupled with the climate change scenarios of Standard, Medium, and High, as defined in this report, to estimate future credit losses for loans acquired by the GSEs, as shown in Figure 5. One key driver of credit losses for mortgages are home prices. Historically, if home prices decline the frequency and severity of mortgage defaults increases. This study estimated credit losses assuming property values experience an immediate reduction in value to account for underpriced flood risk (i.e., a parallel shock to the future home price forecast).

In the Standard scenario which reflects current conditions, credit losses are projected to increase by $35.8 billion for GSE loans in-force as of September 30, 2020 relative to expected credit losses in the baseline scenario.¹⁴ Credit losses increase by $43.5 billion and $55.7 billion for the Medium and High scenarios, respectively.

Figure 5 illustrates the possible exposure of mortgages to a repricing of flood risk in property values. In other words, if property values decline because of home buyers pricing future flood risk into the purchase price of new homes, then losses consistent with those reported in Figure 5 would be anticipated.

The next few sections of this paper outline the data and approach used to develop the above estimates.

Data sets

Fannie Mae and Freddie Mac, collectively referred to as the GSEs, make up a large portion of the mortgage market today. As of June 2021, agency mortgage-backed securities outstanding stood at $7.9 trillion, with Fannie Mae making up $3.4 trillion and Freddie Mac $2.5 trillion.¹⁵ Additionally, the GSEs’ share of first lien origination volume was 58.2% for 2020.¹⁶ Milliman relied upon Fannie Mae’s Single-Family Loan Performance Data Set (Fannie Mae Data), with loan acquisition data from 1999 through September 30, 2020, for use in this study. The Fannie Mae Data was combined with Freddie Mac’s Single-Family Loan-Level Data Set (Freddie Mac Data). The Freddie Mac data contains origination and performance data for single-family mortgage loans acquired by Freddie Mac starting in January 1999 with performance through September 30, 2020. The Fannie Mae Data and Freddie Mac Data were aggregated to serve as the exposure base for analyzing climate change and flood impacts on credit risk for this study (GSE Data). As of September 30, 2020, there were about 22 million GSE loans in-force in the GSE Data.¹⁷

Loan sampling

In order to forecast credit risk losses for the GSE Data, Milliman needed to map the property-level losses for each of the scenarios to the in-force¹⁸ GSE loans. Loan location data is scrubbed of personally identifiable information before being distributed by Fannie and Freddie for external use, and a loan can be identified by a three-digit ZIP Code at the most granular level. As an example, a mortgage on a house located in ZIP Code 53217 would show 532 as the ZIP Code in the loan data. This presents a challenge in estimating the effects of climate change—by way of property value—to mortgage credit risk. This lack of granularity is addressed by randomly mapping property locations to loans that reside in the same three-digit ZIP Code. The mapping is carried across all scenarios for consistency.

It is accepted that this mapping is not a one-to-one property-to-mortgage solution and that inconsistencies in the mapping should “wash out” across the large sample of loan and property value combinations. More exhaustive approaches to mapping are possible and matching similarities between property value from the property data and Home Price Index (HPI)-adjusted home value from the loan data may enhance the mapping process; however, an exact mapping may not exist. A location in the property data may not match a mortgage in the loan data. As examples, the property could have been purchased in cash, the mortgage may have already matured, or the loan may not be classified as a single-family fixed-rate mortgage. This study posits that this approach is therefore reasonable as a more comprehensive method but may not yield more accurate results.

Mortgage performance model

Milliman developed a web-based mortgage analytics tool, the Mortgage Platform for Investments and Reinsurance (M-PIRe), which is a turnkey solution for analyzing mortgage investment opportunities, including credit risk transfer securities.¹⁹ The platform includes the data, models (loan-level performance models and cash flow waterfalls), business intelligence dashboards, and other reporting tools to holistically manage and value a portfolio of mortgage exposures. The mortgage performance model combines loan-level data, loan-level econometric models, and economic scenario forecasts to produce deterministic cash flow estimates for each loan (and ultimately each exposure). Specifically, M-PIRe estimates rates of monthly prepayment, default, and loss severity (also referred to as “Loss Given Default”) for each active loan included in the GSE Data. These estimates are generated using the Moody’s baseline economic scenario described below and our econometric mortgage performance model.

Economic scenario

The economic environment serves as a key driver of future mortgage collateral performance and thus M-PIRe utilizes historical and current economic forecasts to evaluate mortgage transactions. For this study, Milliman used the Moody’s Baseline economic scenario of home prices, unemployment, and interest rates released in May 2021 as the expected mortgage credit loss or baseline expectation of credit losses absent any climate change impacts. The graph in Figure 6 details the forecasted credit losses by state for the Moody’s Baseline scenario.

Translation to credit loss

As described in the Unpriced Flood Costs sections above, Milliman estimated credit losses as a result of a repricing of property values to quantify the potential downstream financial impacts on the GSE in-force loans as of September 30, 2020. The additional credit losses estimated under these scenarios are ultimately absorbed by the mortgage loan investors. Milliman used the following procedure to translate property repricing to credit losses for the GSE in-force loans:

1. Through the random sampling technique mentioned above, attach the property value decline from unpriced flood costs to each loan.
2. Adjust the assumed future house price appreciation for each loan to reflect the property value decline based on unpriced flood costs.²⁰
3. The loan-level simulated decrease in property value is assumed to occur at time 0 and is treated as a “shock” to the home price. For example, if the original property value is $100,000 at time 0 and the assumed home price appreciation is 3% per year for the next five years, the model estimates the property value to be $100,000 at time 0, $103,000 at time 1, $106,090 at time 2, etc. If the impact to the value of property as a result of the unpriced flood cost is estimated at 20% at time 0, then the value of the property is estimated to be $80,000 at time 0, $82,400 at time 1, $84,872 at time 2, etc.
4. Estimate loan-level performance vectors for underlying mortgage pools using adjusted loan-level price appreciation vectors. Loan-level performance vectors were estimated as a function of each loan’s underwriting characteristics (e.g., original FICO score, original loan-to-value ratio, debt-to-income ratio, and others) and underlying property value, as measured by metropolitan statistical area (MSA)-level home price indices. In order to estimate the impact of the house price shock due to the unpriced flood cost, each loan was run through two separate iterations of the mortgage performance model in M-PIRe. The first run was the unadjusted baseline run for all loans in the pool to provide the model’s baseline projected performance. There were no shocks to house prices applied in this iteration, and each loan received a forecasted house price appreciation path equal to Moody’s Baseline forecast for the property’s MSA (or state if a property does not fall in an MSA). The second iteration of the model output was performed with the unpriced flood cost shock applied to the loan-level house price appreciation. After the shocks to house prices were applied, each loan reverted back to receiving a forecasted house price appreciation path equal to Moody’s Baseline forecast for the property’s MSA or state. The estimated drop in the value of the home has an immediate negative impact to borrower equity. This results in increased delinquencies, foreclosure rates, and severity rates for impacted loans.
5. Summarize the output for each scenario to estimate the potential impact of climate change on mortgage performance (e.g., default frequency increase, loss severity increase, etc.).

Acknowledgment

Milliman would like to thank KatRisk LLC for their contributions to the paper, including providing the flood modeling supporting this analysis. Founded in 2012, KatRisk LLC (KatRisk) is a leading provider of catastrophe risk models. For more information, visit their website at www.katrisk.com.

Milliman and KatRisk appreciate the review and engagement from Ceres Accelerator, ceres.org/accelerator, in developing the analysis and results for this paper.

1 Long, H. (August 29, 2017). Where Harvey is hitting hardest, 80 percent lack flood insurance. Washington Post. Retrieved January 27, 2022, from https://www.washingtonpost.com/news/wonk/wp/2017/08/29/where-harvey-is-hitting-hardest-four-out-of-five-homeowners-lack-flood-insurance/.

² Defined as the 90th percentile of household income.

³ For more information, see the Intergovernmental Panel on Climate Change (IPCC) glossary at https://www.ipcc-data.org/guidelines/pages/glossary/glossary_r.html.

⁴ Ibid.

⁵ See https://www.soa.org/globalassets/assets/files/resources/research-report/2020/soa-flood-report.pdf.

⁶ While this is a simplistic way to calculate the insurance premium under future climate states, limited data is available on the current premium structure of the NFIP’s premiums to perform a more detailed calculation. Risk Rating 2.0 is currently under implementation and may allow for more nuanced estimates of premium changes under future climate states. For more information on Risk Rating 2.0, see https://www.fema.gov/flood-insurance/risk-rating.

⁷ For more information on this model, please see https://www.milliman.com/en/insight/Insights-into-consumer-demand-for-flood-insurance-Trends-in-take-up.

⁸ Beltran, A., Maddison, D., & Elliot, R. J. (January 2018). Is flood risk capitalised into property values? Ecological Economics 146., 668-685.

⁹ Hino, M., & Burke, M. (February 2020). Does information about flood risk affect property values? National Bureau of Economic Research Working Paper 26807. Retrieved January 27, 2022, from https://www.nber.org/papers/w26807.

¹⁰ This is a simplistic calculation as local costs of flooding priced into property values likely vary around this 2.1% estimate. As noted above, existing research is not conclusive on how it varies and further research would be beneficial to understanding which local areas currently account for flood risk more than others.

¹¹ As discussed on page 17 of the report at https://www.soa.org/globalassets/assets/files/resources/research-report/2020/soa-flood-report.pdf, Milliman finds it likely that less than 5% of homeowners have flood insurance, while industry surveys find as many as 17% believe they have flood insurance.

¹² Data retrieved from American Community Survey (ACS).

¹³ Defined as the 90th percentile of household income.

¹⁴ Milliman’s baseline scenario reflects expected credit risk assumptions and Moody’s baseline economic forecast.

¹⁵ Source: Urban Institute July 2021 Housing Finance at a Glance A Monthly Chartbook.

¹⁶ Ibid.

¹⁷ The GSE Data is a subset of the GSEs’ 30-year and less, fully amortizing, full documentation, single-family, conventional fixed-rate mortgages. Please refer to Fannie's and Freddie’s descriptions of their data sets for a full list of loan exclusions.

¹⁸ In-force is defined as loans that have not terminated as of September 30, 2020.

¹⁹ For more information, see https://www.milliman.com/en/products/milliman-m-pire.

²⁰ Historical data and mortgage performance models indicate that mortgage default events are heavily correlated with changes in home prices. Specifically, declines in home values result in increased mortgage defaults and severity rates.