Drawing on a dataset of over 200 million building permits across the United States, we identify 8.8 million climate and adaptation related permits filed between 2010 and 2024, representing an estimated $219 billion in private spending.

Subscribe now

The Permit Data: Identifying Adaptation Expenses

We analyze building permit records from Builty, a comprehensive database covering all 50 U.S. states, and reporting over 200 million permits. For each permit, we classify whether it was adaptation-related using two complementary methods:

• Boolean flags: The dataset includes a roof replacement indicator that directly identifies reroofing permits, the largest category of climate-related work.

• Description keywords: For the 81% of permits with a text description, we search for terms associated with hurricane protection (impact windows, storm shutters, wind mitigation), flood adaptation (elevation, flood-proofing), fire damage repair, and storm damage.

We group the permits in five categories, shown in the table below:

Roughly 44% of climate-related permits include a reported job value. For the remainder, we imputed values using the median job value for the same permit category, state, and year. The resulting $219 billion total should be read as an estimate, while the directly reported figure of $144 billion provides a conservative lower bound. Coverage is also imperfect and expands substantially over the sample period, so our estimates likely understate adaptation investment in the earlier years. Spending reaches roughly $25 billion per year in 2023 and 2024.

Roofing dominates the total, accounting for 67% of all spending. This is not surprising: roof replacement is both the most common home repair triggered by weather events and the most frequently permitted type of work.

We classify all roof replacements as weather-related, on the grounds that weather exposure (UV, hail, wind, and rain) is the primary determinant of roof lifespan. This convention is standard in the insurance industry, where roof age and storm exposure dominate claims data. It is nonetheless an aggressive classification: some roofs are replaced for cosmetic reasons or as part of broader renovations. If 20–30% of roofing turns out to be non-climate-driven, the true total would fall closer to $170-190 billion.

Where Adaptation Spending Is Highest

The figure below shows that climate adaptation spending is sharply unequal across states. To minimize biases in comparisons, we restrict this chart to 2020–2024, the years with the most reliable coverage, and to the 32 states in which at least 10% of permits report a job value and the average permit value falls within three times the national mean. States that fail these checks are excluded from the ranking. In most cases, these states have a handful of unusually expensive permits (likely commercial projects) that pull the average upward.

Florida is in a league of its own. Between 2020 and 2024, spending on climate-related building permits totaled $1,750 per person. This is more than triple the next-highest state. Florida’s exposure to Irma (2017), Michael (2018), Ian (2022), and Helene and Milton (2024), combined with one of the nation’s strictest wind-resistance building codes, means every roof replacement must now meet hurricane-rated standards. Each repair is therefore more expensive, but also more protective.

California ($580 per capita) ranks second, driven by wildfire damage repair and the state’s stringent fire-hardening requirements for construction in the wildland-urban interface. Rhode Island ($570) and Colorado ($550) follow, reflecting coastal storm exposure and the Front Range hail belt, respectively.

At the other end of the distribution, states like Virginia ($100 per capita) and New Mexico ($110) spend far less. Part of this gap reflects genuinely lower climate exposure in the built environment; part of it reflects thinner permit coverage in the dataset. The true gap is likely narrower than what appears here, but the broad pattern is robust: storm- and fire-exposed states have the highest spending.

The Hurricane Footprint

To show that the permit data captures the effects of individual disasters, we focus on Florida, the state with the largest adaptation spending. We compute climate-related permits as a share of all permits filed in the state, quarter by quarter. The share-based approach matters: if climate and non-climate permits grow in lockstep as new jurisdictions enter the dataset, the ratio stays flat. Thus, temporary fluctuations in the share that coincide with natural disasters represent genuine shifts in adaptation activity.

The figure below shows that the fingerprints of major hurricanes are clearly visible. Florida's climate permit share hovers around 8-10% through the early 2010s. Then Irma makes landfall in September 2017, and the share nearly doubles in a single quarter, from 12% to 19%. Ian repeats the pattern even more dramatically in late 2022, driving the share from 14% to 24%. Helene and Milton push it back to 19% in late 2024. Interestingly, between storms the share never fully retreats to pre-Irma levels, settling into a new baseline of 13-16%.

Conclusion

This post highlights how significant climate adaptation spending already is for homeowners and businesses, and how significant it is likely to remain.

The burden is also sharply concentrated. Spending in Florida amounts to more than $1,750 per capita. This is over triple the next-highest state and seventeen times what Virginians spend.

Building permits offer an underutilized lens on this landscape. While government agencies track disaster declarations and insurance claims, permit data captures the full spectrum of private adaptation spending, including the proactive efforts that take place outside the insurance system. As climate exposure broadens, this kind of bottom-up evidence will be essential for understanding who is paying, where, and how much.

Share

But I recently saw Refine promoting partnerships with journals. The stated goal: reduce reviewer workload in an era of rising submission volumes. If AI can help screen papers, summarize contributions, identify weaknesses, and streamline referee reports, many editors would understandably be interested.

Yet this raises a deeper question. What happens when both sides of the process rely on the same class of tools?

If authors use LLMs to preemptively polish papers, address likely criticisms, and improve exposition, then journals may receive submissions that are increasingly optimized for AI-based review criteria. In turn, review tools may be forced to search for smaller and smaller flaws in already polished manuscripts. The equilibrium could become strange: authors using AI to satisfy reviewers, reviewers using AI to uncover issues generated by authors anticipating AI reviewers. We may be entering a recursive loop in which machines are increasingly evaluating work prepared for machines.

Another possibility is that these tools eventually stop behaving like harsh gatekeepers and instead become validators. If a paper meets certain methodological, statistical, and presentation standards, the review system may simply certify it as technically sound. In that world, peer review shifts away from subjective judgments of novelty or style and toward verification.

This leads to an even more provocative implication. If journals can use agentic review tools, why can authors not use agentic response tools? Imagine submitting a paper alongside an expert AI agent trained on the manuscript, data, and code. Reviewers request robustness checks, additional tables, alternative specifications, or clarifications. The author’s agent runs the analysis, produces the output, and responds instantly. After several rounds of machine-to-machine interaction between review agents and author agents, the paper emerges revised and publication-ready.

Push the idea one step further, and the role of journals themselves becomes less obvious. Researchers could upload papers to open repositories such as arXiv or SSRN, accompanied by a transparent AI-generated review report evaluating correctness, assumptions, robustness, and contribution. Readers would then observe both the paper and the audit trail. Rather than waiting months for editorial gatekeeping, the market for ideas could operate in real time.

Under this scenario, the central scarcity is no longer correctness. AI systems may make it easier to detect coding errors, flawed identification strategies, missing citations, or weak robustness checks. Technical quality becomes cheaper to verify. The scarce resource instead becomes attention. If many papers are methodologically sound, then the key question is not whether a paper is “correct,” but whether it is important, useful, original, or worth reading.

That may be the real future of research publishing. Journals historically bundled multiple functions: quality control, certification, filtering, and distribution. AI may unbundle them. Verification can be automated. Distribution is already open. What remains hardest to automate is judgment about relevance.

If so, LLM review tools are not just changing peer review. They may be quietly redefining why peer review exists at all.

This post takes stock of where AI stands in residential real estate today. We focus on three areas: how consumers search for and discover homes to buy, how AI-driven valuation models hold up under stress, and the fact that automated decision-making in mortgage underwriting and tenant screening decisions may not be compliant with existing regulations. We draw on industry data, academic research, and recent developments to bring these dynamics into focus.

Subscribe now

Conversational AI Search: A New Interface for Home Discovery

Redfin launched its conversational search in November 2025, built in partnership with Sierra AI. Rather than selecting filters for bedrooms, price range, and zip code, users describe what they are looking for in their own words (and in a language of their choosing). The system interacts with the user and refines results through back-and-forth exchanges. The early results are notable: in Redfin’s initial testing, conversational search users viewed nearly twice as many listings as users using standard search. Users using conversational search were also 47% more likely to request home tours.

Zillow responded with “Zillow AI Mode” in March 2026. Beyond offering help with search, this conversational assistant provides guidance throughout the transaction. Users can ask questions like "How has this home's Zestimate changed over time?" or "What would a fair offer be?" The system remembers preferences across sessions, adapts to user behavior, and draws on Zillow's proprietary data and valuation models. Zillow's CEO has called generative AI a "bigger platform shift than mobile" for the company.

Both Zillow and Redfin also launched apps within OpenAI’s ChatGPT (Zillow in October 2025, Redfin in February 2026 ). These apps allow users to search for homes directly inside the chatbot. This degree of integration has raised legal questions about IDX licensing agreements, which govern how data from Multiple Listing Services (real estate brokers databases) can be displayed. Critics argue that transmitting listing data to a third-party AI platform may fall outside the scope of existing agreements.

An important backdrop to the developments discussed above is industry consolidation. Rocket Companies acquired Redfin for $1.75 billion in July 2025, creating a vertically integrated platform that spans home search, brokerage, and mortgage lending. The combined entity now controls over 14 petabytes (million gigabytes) of data across 100 million properties. Meanwhile, Zillow generated $2.6 billion in revenue in 2025, up 16% year over year, and is pursuing what it calls a “housing super app” strategy. The implication is clear: the AI arms race in real estate favors large, data-rich incumbents.

AI-Powered Home Valuations: The Limits of the Algorithm

Automated valuation models (AVMs) are now ubiquitous. Zillow’s Neural Zestimate covers over 100 million U.S. homes using deep learning, and similar models are embedded in mortgage underwriting, portfolio management, and consumer-facing apps. But AVMs remain subject to important limitations, particularly in markets with heterogeneous housing stock and exposure to catastrophic risk.

Zillow reports a median error rate of 1.9% for on-market homes and 7.1% for off-market properties. On a home priced at $1 million, the off-market error translates to approximately $71,000. These are median figures; the distribution has a long right tail, meaning a substantial share of estimates are considerably further off. Recent research has shown that these valuation errors have played a key role in the downfall of Zillow’s iBuyer business.

In a previous R.E.A.L. post, we documented that trust in algorithmic valuations has real market consequences. Following Zillow’s shutdown of its iBuyer program in November 2021, the absolute gap between listing prices and Zestimates increased by 26.8%, sellers more frequently priced above the Zestimate, and homes actually sold faster and at higher premiums. Diminished algorithmic trust did not reduce market activity; it shifted how participants used the information.

The LA Wildfire Stress Test

The January 2025 Los Angeles wildfires provided the most severe real-world test that consumer-facing AVMs have ever faced in a major U.S. metro. According to Zillow’s own analysis, approximately $46 billion in residential housing value was located within the fire perimeters, encompassing 19,605 units with a median value of nearly $1.95 million.

The impact on actual transaction prices was dramatic. In the Palisades fire burn zone, average sale prices dropped 33%, from approximately $3.6 million to $2.4 million. In the Eaton fire zone, the decline was 62%: from $1.8 million to roughly $684,000. As we documented in a recent R.E.A.L. post, this drop in valuations coincided with a wave of damaged homes coming to market.

AVMs are fundamentally unable to account for sudden physical destruction. They rely on comparable sales data, tax assessments, and historical transaction records. None of these can update instantaneously after a disaster. An algorithm cannot detect that a house has burned down. This is an inherent limitation, not a design flaw, but it underscores why algorithmic estimates should not be treated as substitutes for professional appraisals, especially in volatile or disaster-affected markets.

AI in Mortgage Underwriting: Faster, but for Whom?

AI-driven mortgage underwriting has moved from experimental to standard practice. Lenders using AI models report a 90% increase in processing speed. For standard approval cases, end-to-end origination (from application submission to fund disbursement) can be reduced from 3–5 days to under 60 minutes at some institutions. Approximately 85% of mortgage lenders now use AI for fraud detection, and industry data imply that AI has helped reduce mortgage application fraud by roughly half.

The emerging frontier is what the industry calls “agentic AI”. These are systems that autonomously retrieve documents, query data sources, run risk models, resolve exceptions, and generate underwriting memos without requiring human instruction at each step. This represents a shift from AI as a tool that assists underwriters to AI as an agent that manages routine workflows end-to-end, with human oversight reserved for complex cases and quality control.

The market for AI-powered lending was valued at $109.7 billion in 2024 and is projected to reach $2.01 trillion by 2037, growing at a 25.1% compound annual growth rate. These figures reflect an industry-wide bet that AI may become the default infrastructure for mortgage origination.

Speed and efficiency gains are real. But they raise two important questions: Do the benefits accrue equally across borrower populations? And are algorithmic decisions compliant with existing regulations? The Massachusetts Attorney General settled a fair lending action in July 2025 against a lender whose AI underwriting model produced disparate impact along racial and immigration-status lines. This case is unlikely to be the last. AI systems can likely infer restricted variables, which should be excluded from credit decisions (such as race and gender), from combinations of legally permissible variables.

Tenant Screening

Landlords increasingly rely on AI-powered tenant screening programs that evaluate applicants based on credit scores, eviction records, and criminal background checks. Evidence from multiple studies and lawsuits indicates that these programs disproportionately deny applications from Black and Latino renters. This is partly due to the usage of data that are incorrect or outdated.

In one notable case, Harbor Group Management was found to have deployed an AI leasing assistant that issued blanket denials to Housing Choice Voucher holders. A growing body of psychological research further suggests that bias introduced by an AI system can persist in human decision-making even after the AI is no longer being used.

Looking Ahead

The integration of AI into the housing market is proceeding rapidly and on multiple fronts. Conversational search is changing how buyers discover homes. Automated underwriting is accelerating mortgage approvals. AI-driven risk models are providing information that traditional assessments miss. Industry consolidation is concentrating data and market power in a small number of large platforms.

For researchers, the priority is clear: we need rigorous, independent measurement of how AI tools perform across demographic groups, property types, and market conditions, including in disaster-affected and historically underserved areas. For policymakers, the challenge is crafting regulations that preserves the genuine efficiency gains of AI while ensuring that automated systems meet the same standards of fairness and accountability that we expect of human decision-makers.

Share

Fuel price swings hit household budgets, particularly for families in car-dependent suburbs who face long daily commutes. While public transportation fares are also tied to fuel costs, they tend to be lower overall and adjust more gradually, offering commuters greater short-term stability. Car commuters, by contrast, feel every price spike immediately and directly.

We present here a census block group-level analysis of commuting fuel costs based on data from the American Community Survey and the Bureau of Labor Statistics. We focus on the broader commuting zones of the four most populous U.S. metropolitan areas: Los Angeles, Chicago, New York, and Dallas. The analysis is conducted at the census block group level because it is the finest geographic resolution at which the data are published. Each block group typically covers 600 to 3,000 people.

In recent weeks, conflict, geopolitical tensions, and disruptions to crude oil flows from the Middle East have led to higher prices at the pump and uncertainty about future price growth. Our calculations are based on fuel prices in February 2026 (the latest data available from the Bureau of Labor Statistics), before the beginning of the conflict in the Middle East, and likely represent a lower bound for projections of commuting fuel costs in the coming months.

Key takeaways from our study:

• New York stands apart from the other three major metropolitan areas. Its extensive transit network fundamentally reshapes commuting patterns, reducing reliance on private vehicles.

• In Dallas, Los Angeles, and Chicago, car is the main mode of commute in a large share of neighborhoods. Census block groups in which more than 80% of workers use a car as the main commute mode account for 60% of all block groups in Dallas, 47% in Los Angeles, and 43% in Chicago. New York still has a significant share of 20%, but much lower than the other three metros.

• Costs are significant, even though fuel prices were not at historically high levels in February 2026. Moreover, there are very large differences across metropolitan areas. Los Angeles has by far the highest car commuting cost. The median commuting fuel cost across block groups is at $26.24 per household per week, compared to slightly more than $16 in Chicago and Dallas, and just $10.23 in New York. Restricting the sample to car-owning households raises the Los Angeles median to $28.16 per week, while New York remains the lowest-cost city at $12.85 per week. This is striking because fuel prices in New York and Chicago are similar. However, New York benefits from shorter commutes and lower car usage.

• Interestingly, costs tend to follow an inverted U-shaped pattern in census block group-level household income within each metro. Thus, middle-income households experience the highest car commuting costs. The pattern is present both when considering all households and car-owning households only and highlights that households in middle-income blocks are, in general, the most reliant on their cars.

Subscribe now

Data Sources and Methodology

This section provides a detailed discussion of our data sources and assumptions. If you are only interested in the results, you can skip to the next section.

Our estimates are based on three primary data sources:

• American Community Survey (ACS) 2019–2023, 5-year estimates: We use published summary tables at the census block group level. Key tables include B08134 (travel time by mode of transportation), B08301 (means of transportation to work), and B25044 (tenure by vehicles available).

• Bureau of Labor Statistics (BLS) gasoline price series: We use the monthly average prices for regular unleaded gasoline by metro area (CPI Average Price data, series APU) for February 2026 (before the start of the war in the Middle East).

• EPA and FHWA fleet efficiency data: We use a national average fuel economy of approximately 25.4 miles per gallon, reflecting the current light-duty vehicle fleet.

For each block group, we observe the number of car commuters, the average one-way commute time (estimated from ACS travel-time bins specific to car/truck/van commuters), and a metro-specific average commute speed. From these, we derive the one-way commute distance, compute daily and weekly fuel consumption at the fleet average miles per gallon, and multiply by the local gasoline price. We calculate total commuting fuel cost per census block group per week, and divide by the number of households, or by the number of households who own a car.

An important assumption: we use binned travel-time estimates from table B08134, which provides commute-time distributions specifically for car, truck, and van commuters. This avoids the bias that would arise from using the all-modes travel time distribution (which includes transit riders with systematically different commute times). The trade-off is that binned data introduces a midpoint approximation, but coverage at the block group level is nearly complete.

The Geography of Car Commuting

The maps below show the share of workers who commute by car, truck, or van in the commuting zones of our four metropolitan areas. As expected, car dependence is highest in suburban and extra-urban areas and lowest in dense urban cores with transit access. Yet, especially in the commuting zones of Los Angeles, Chicago, and Dallas, the share of workers commuting by car is remarkably high.

In Dallas, car commute shares are uniformly high, exceeding 80 percent in the majority (nearly 60%) of block groups. Los Angeles and Chicago show more variation, with lower car shares in downtown cores where transit is viable. Still, 47% (in Los Angeles), and 43.3% (in Chicago) of block groups have more than 80% of workers commuting by car.

New York stands apart: the five boroughs of New York City, particularly Manhattan and Brooklyn, have car commute shares below 30 percent, while the surrounding suburbs in New Jersey, Long Island, and Connecticut are as car-dependent as the other three metros. Overall, only 20% of block groups have more than 80% of workers who mainly commute by car.

Commuting Fuel Costs: Distribution

The histogram below shows the distribution of estimated weekly household fuel costs across block groups in each metro. The dashed red line marks the median.

Los Angeles has the highest costs, with a median of $26.24 per week across block groups. This value is driven by long commutes and high fuel prices. Los Angeles has a price per gallon in February 2026 of $4.55, compared to Chicago’s $3.08, New York’s $2.96, and Dallas $2.52.

Chicago has a median car commuting cost of $16.14 across block groups. Dallas has long commutes, but low fuel prices reduce costs, and median commuting spending across block groups is $16.16, similar to Chicago. New York has the lowest costs, with a median of $10.21 per week, driven by the large share of households who have minimal expenses from commuting by car, as is also visible in the histogram.

The calculations above average fuel costs across all households in a census block group, including non-car commuters. The histogram below shows the distribution when we instead calculate averages using only car-owning households within each block group. We use this alternative measure because car ownership is a reasonable proxy for car commuting.

Average costs mechanically increase across the board. Los Angeles has the highest cost of $28.16 per week, followed by Chicago at $17.81, and Dallas at $16.96. New York, despite relatively high fuel prices, remains at the bottom of the list, significantly lower than other metros, with $12.85.

Annualized, these costs are substantial, and strikingly uneven across metropolitan areas, from roughly $670 in New York to nearly $1,460 in Los Angeles.

Subscribe now

Middle-Income Households Face the Highest Costs

A natural question is whether these commuting fuel costs fall disproportionately on particular households. The figure below plots average weekly fuel cost per household against median household income, with block groups binned into 20 income quantiles. The shaded area shows 95% confidence intervals. Estimates are weighted by the number of households in each block group.

The relationship is not linear and follows an inverted U shape. In the lowest-income block groups, household fuel costs are relatively low because fewer households own cars and fewer residents commute by driving. As income rises into the middle range ($50,000–$100,000), car ownership and car commuting rates increase, pushing fuel costs up. At the highest income levels, costs plateau or decline slightly, reflecting shorter commutes in affluent neighborhoods.

This inverted-U pattern is most pronounced in New York and Los Angeles. In Dallas, the curve is flatter: nearly everyone drives regardless of income, so the relationship reflects distance more than mode choice. Chicago falls in between.

As shown in the figure below, the pattern is also present when we compute average costs using only car-owning households. Thus, this evidence is driven not only by a larger share of car owners in middle-income block groups but also by longer commuting times.

The key policy implication is that middle-income suburban households bear the highest commuting fuel burden in absolute terms. However, while low-income households spend less on commuting fuel in total dollars, the share of their budget consumed by fuel might still be higher, an important distinction that this analysis does not directly capture.

Conclusion: Putting the Results in Perspective

Our estimates suggest that commuting fuel costs are significant in all four metros, but also that there are massive differences across metros, with the typical household spending $530 a year in New York, and $1,360 a year in Los Angeles. The typical car-owning household spends $670 in New York and $1,460 in Los Angeles.

How does this compare to total household car fuel spending? Data from the National Household Travel Survey shows that commuting accounts for approximately 30 percent of total vehicle miles traveled. If we take these estimates at face value, and do a back-of-the-envelope calculation, we find that total annual car fuel costs based on February 2026 prices would, in the median block group, equal approximately $4,530 in Los Angeles, $2,800 in Chicago and Dallas, and $1,800 in New York.

Should conflict or geopolitical tensions drive fuel prices in Los Angeles back to the 2022 peak of $6.5 per gallon, the annual median car commuting cost for households across Los Angeles would rise to approximately $1,940. The inverted U-shaped cost distribution documented above further implies that the related commuting cost shock would be largest in middle-class neighborhoods.

Overall, this post highlights that car fuel costs are significant for household budgets in the U.S., and that households are highly exposed to fuel price fluctuations.

Share

Since the early 2000s, much of the U.S. policy discussion around food access has been focused on “food deserts”, which are areas where residents lack convenient access to affordable, nutritious food. The USDA defines these at the census-tract level, combining low income with low proximity to supermarkets.

While the number of Americans living in designated food deserts has declined in recent years (see Karpyn et al. 2019), the landscape of grocery stores has grown more complex, with the rise of premium grocers and substantial reorganization in the industry. Nationally, retail analysts project thousands of store closures in 2026, and grocery closures have been a persistent concern particularly in lower-income neighborhoods.

In this post, we analyze the spatial distribution of grocery store locations and closures across California, and highlight the distinct patterns for premium and conventional (non-premium) grocers. Our results are based on establishment-level data from SafeGraph combined with demographic information at the census block group-level from the American Community Survey.

The following patterns stand out:

• There is massive demographic stratification in the locations of different types of stores. Premium stores target micro-locations with high-income renters. Conventional stores have a more even distribution, but lean, interestingly, toward lower-than-average income areas.

• Store closures over the 2019-2025 period have been a significant phenomenon. They have only weakly impacted premium stores, and rather have been concentrated in conventional stores.

• Store closures disproportionately affect areas with lower incomes, lower educational attainment, and higher Black and Hispanic/Latino population shares. Compared to state averages, these neighborhoods (census block groups) have 20% lower incomes, 6 percentage points (p.p.) lower share of bachelor-educated individuals, 10 p.p. higher share of Black and Hispanic/Latino residents, and 14 p.p. lower homeownership rates. A snapshot of closures in Los Angeles County shows that closures are also spatially concentrated.

Subscribe now

The Spatial Distribution of Premium and Conventional Grocers

As of the end of 2025, we identify 13,407 open grocery stores and supercenters (large retail stores that also sell groceries, such as Target) across California, of which 186 (just 1.4%) belong to seven chains that we classify as premium based on brand and pricing: Whole Foods Market (94 locations), Gelson’s (27), Nugget Markets (17), Bristol Farms (13), Erewhon (11), Lazy Acres (6), and Draeger’s (2). The remaining 13,221 stores encompass conventional supermarket chains, discount grocers, ethnic specialty stores, and independent operators.

The demographic profiles of the neighborhoods that host premium and conventional grocery stores differ considerably. For this analysis, we link store locations to census block group demographics from the American Community Survey (ACS) 5-Year Estimates (2019–2023). Block groups are the smallest geographic unit for which the ACS publishes demographic estimates. They typically contain 600-3,000 people. We restrict our analysis to block groups with at least 100 residents (25,471 of 25,586 block groups in California).

The figure below shows the distribution of demographic characteristics for census block groups containing at least one premium store. The dashed vertical lines represent average values across block groups.

Relative to state-wide averages across block groups, the locations with premium grocers have substantially higher shares of college-educated adults (on average, 59.4 p.p. vs 36.6 p.p), lower shares of Black and Hispanic/Latino residents (on average, 24.3 p.p. vs 43.1 p.p.), and elevated median household incomes (on average, $133.5k vs $109.5k). However, they also have substantially lower homeownership rates, about 43 p.p. on average, compared to the overall average of 57.4 p.p..

In summary, the locations of premium stores appear significantly skewed towards areas with high-education and high-income renters. Thus, premium stores’ micro-locations are a flag for locations with substantial demand for Class-A apartments and multifamily development.

We can see below that conventional stores, by contrast, are located in block groups with a distribution of demographic characteristics more similar to the overall population of block groups, but with on average a lower share of bachelor educated adults (31.1 p.p., on average), lower incomes ($94.2k), a higher share of Black or Hispanic/Latino residents (49.5 p.p.), and lower homeownership rates (48.8 p.p.).

These gaps are particularly significant when compared to the premium stores, with a 30% relative difference in average income across locations, and a 28 p.p. gap in the average share of bachelor-educated population.

Notably, 68.7% of California’s block groups contain no grocery store or supercenter at all. These neighborhoods in principle include lower-density areas at the outskirts of major population centers. However, in the figure below, we find that these block groups without grocery stores have a distribution of demographic characteristics similar to the overall population; if anything, they have slightly higher education and income, a lower Black and Hispanic/Latino share, and a higher homeownership rate. Thus, residential neighborhoods are likely to make up a substantial portion of these areas.

Closure Rates and Their Spatial Distribution

Between 2019 and 2025, SafeGraph recorded 3,427 grocery store and supercenter closures in California. The closure rate (closed stores divided by closed plus open stores as of 2025) diverges sharply by store type: 20.5% for non-high-end stores, compared to 5.6% for high-end stores. This is a nearly four-to-one ratio.

A comparison of block groups where conventional stores closed against those in which there were no closures reveals broadly similar demographic distributions. However, some differences are visible in average characteristics: compared to block groups with open stores, block groups experiencing closures tend to have slightly lower median incomes, somewhat higher shares of Black and Hispanic/Latino residents, and lower homeownership rates. These differences, while not dramatic, suggest a tilt toward more vulnerable communities.

In addition, block groups with conventional grocery stores already have lower incomes and educational attainment than the average block group in the state. Compared with the overall distribution of demographic characteristics across block groups in California, block groups affected by store closures tend to have incomes nearly 20% lower, a share of bachelor's degree holders more than 6 p.p. lower, a Black and Hispanic/Latino population share 10 p.p. higher, and homeownership rates 14 p.p. lower.

Geographic Concentration of Closures: Los Angeles County

Another important aspect of closures is that they are spatially concentrated. We illustrate this point below, with a figure focusing on the main metropolitan area of Los Angeles County.

We show areas with no conventional grocery stores in gray, areas with stores and no closures over the period between 2019 and 2025 in white. Then, block groups experiencing closures are color-coded, with a darker color corresponding to higher closure rates (closed stores divided by closed plus open stores as of 2025). We can see that many block groups experiencing closures lose all or more than half of their stores, and that closures cluster across neighboring block groups, especially close to downtown Los Angeles, in the East, and in the South.

Conclusion

Grocery stores are crucial neighborhood amenities that provide local residents with access to food and other products. While evidence of assortative matching between local socioeconomic characteristics and the likelihood of having access to premium stores is to be expected, it is still striking how strong this phenomenon is in the data. Moreover, it is interesting that conventional grocery stores tend to be located in neighborhoods with lower incomes and educational attainment. Finally, the scale and concentration of conventional store closures are concerning trends. Based on our results, store closures may, over time, worsen access to supermarkets for lower-income areas.

Share

However, Koijen et al., 2025 show that this approach typically fits commercial real estate data poorly, unlike in the more liquid residential market, and yields noisy predictions. The same paper also shows that modern machine learning methods can substantially outperform traditional hedonic regressions in this setting.

In this post, we run a horse race between OLS and machine learning for a major commercial real estate segment: multifamily properties in California. Following Koijen et al. 2025, we use LightGBM as our machine learning method.

LightGBM is a decision tree method. A decision tree makes predictions by sorting observations through a series of yes/no questions until they land in a final group. For example, if you are predicting property prices, the tree might first ask “has this property more than 100 units?”. Then it might ask “is it in a high income neighborhood?” and so on. Each observation follows a path down the tree based on its characteristics until it reaches an endpoint, called a leaf. The predicted outcome for the observation is the average of the target variable (in our case, the property price) within the leaf.

We evaluate the LightGBM model along three dimensions: cross-sectional predictive accuracy, interpretability of the key valuation drivers, and susceptibility to overfitting. We find that the model substantially outperforms OLS in accuracy, while performing well on the other two dimensions.

Subscribe now

Data and Estimation

We cover here the details of the data and estimation methodology. Readers interested primarily in seeing the results can skip ahead to the next section.

We collect a dataset of 12,177 arm’s-length sales of apartment buildings using data from Corelogic. We restrict the sample to buildings with at least 15,000 square feet of living area, spanning all 58 California counties from 2000 to 2025. We also collect neighborhood demographics from the American Community Survey at the Public Use Microdata Area level.

The dependent variable is the log of the sale price, demeaned by sale year to strip out aggregate time trends. Both the OLS and the LightGBM models compete purely on their ability to explain within-year price differences. The set of predictive features is identical for both: ten predictors covering building characteristics (log living area, log land area, log building age, and a “superstar county” indicator for Los Angeles, Orange, San Diego, San Francisco, San Mateo, and Santa Clara county), neighborhood demographics (population, elderly share, bachelor’s degree share, owner-occupied share, median income), and a CBSA market identifier.

The 12,177 observations are randomly divided into three non-overlapping sets: 60 percent for training (7,306 sales), 20 percent for validation (2,435), and 20 percent for the test set (2,436), which is used for the out-of-sample comparisons below. Note that the test set is never used in estimation. Year de-meaning is carried out using sale-year averages computed from the training set only, so that there are no information leaks from the validation or test sets.

The OLS model is estimated on the training set and its coefficients are applied to the test set covariates to generate predictions.

To explain how the LightGBM model is estimated, we begin by giving an intuition of how the model fits data. The LightGBM model starts with a naive guess, typically the unconditional mean of the variable it is trying to predict, which in our case is the log property price. It then looks at where the guess is wrong: for some observations it guesses too high, for others too low. A decision tree is built, splitting the data to try and predict these errors. The predictions of this tree are added to the original guess, scaled down by a factor (the learning rate) to avoid overcorrecting. The model then looks at the remaining errors, builds another tree to predict those, adds it in, and repeats. Each iteration of this process is called a boosting iteration.

LightGBM is estimated on the training set. Specifically, each boosting iteration runs estimates using a random subsampled of 70 percent of the training set, which decorrelates successive trees. At each boosting iteration, the Root Mean Squared Error is computed on the separate validation test, and the iteration with the lowest validation RMSE is selected for final prediction.

Readers experienced with non-parametric models may wonder how this validation-set approach compares to cross-validation. In a cross-validation approach, the model is estimated over multiple train-test partitions. LightGBM uses a fixed validation set for all boosting rounds, and an early-stopping routine (see below) to avoid overfitting. The downside relative to cross-validation is that performance estimates depend on a single random partition, but the benefit is computational efficiency and a clean separation between training set and evaluation set. Moreover, as mentioned, the model is re-estimated at every iteration on a different random subsample of the training set.

Both OLS and LightGBM forecasts are produced in the de-meaned space and then converted back to the original log price scale by adding the training set year means, so that RMSE and R-squared are directly comparable across models. Thus, the horse race between the models is based on their ability to predict prices across properties at a point in time.

All results below come exclusively from the held-out test set, and so are actual out-of-sample predictions.

LightGBM Wins on Out-of-Sample Predictions

On the held-out test set, LightGBM achieves a test RMSE of 1.332, which is roughly 14 percent lower than the OLS test RMSE of 1.553. In R-squared terms, LightGBM explains about 47 percent of out-of-sample price variation versus 28 percent for OLS. This is a gain of nearly 19 percentage points (close to a 70% relative improvement). Same features, same data, same year de-meaning. The improvement comes entirely from LightGBM’s ability to capture non-linear relationships and feature interactions.

A caveat: these absolute R-squared values are modest compared to single-family hedonic models, which often exceed 0.90. However, multifamily apartment buildings (and commercial real estate more broadly) are inherently harder to value due to fewer transactions and greater variation in asset characteristics. This challenge is compounded by the limited set of hedonic variables available for our analysis.

What Drives LightGBM Predictions?

A common knock-on of machine learning models is the “black box” problem. What drives predictions? Are the estimates sensible?

The analysis of SHAP values helps address this concern. For every test observation, SHAP decomposes the model’s prediction into additive contributions from each predictive feature. Averaging the absolute SHAP values across the test set gives us a global importance ranking, indicating which features are most important for prediction.

The top three features are log living area (mean |SHAP| = 0.282), bachelor’s degree share (0.247), and log land area (0.235). Building size is, unsurprisingly, the single strongest predictor. The prominence of the bachelor’s degree share is interesting. It proxies for neighborhood quality and tenant demand in an intuitive way. Areas with more educated residents tend to have higher rents and lower vacancy, both of which are capitalized into property values.

Log building age (0.171) and the superstar-city indicator (0.159) follow, capturing the age discount and the persistent price gap between California’s expensive coastal counties and the rest. Median income, elderly share, owner-occupied share, and population fill the middle ranks. The CBSA market identifier comes in last (0.079): most of the metro-level variation is already absorbed by the demographic features and the superstar indicator.

What about the direction of heterogeneity? What is the sign of the effect of each feature on prices? A Beeswarm plot shows this: each dot represents one observation; its horizontal position indicates the SHAP value, and its color encodes the actual feature value (blue = low, red = high). A feature that has predominantly red dots to the right and blue dots to the left has a positive effect on prices (higher values of the feature, higher prices).

Log living area shows a clean, positive gradient: larger buildings raise predicted prices, while smaller ones lower them. The spread is wide, confirming that building size drives the largest swings at the observation level.

Bachelor’s degree share displays similarly clean color separation. Neighborhoods with more college-educated residents receive positive SHAP contributions; those with low attainment receive negative ones. The effect is nearly monotonic, reinforcing that neighborhood human capital is a robust price driver.

Log building age is notably different. The relationship is noisy and non-linear. This is exactly the kind of complex pattern that tree-based models handle naturally but a log-linear OLS specification cannot.

The superstar-city indicator splits into two clusters, as expected from a binary variable. Properties outside the six superstar counties show a tight band of negative SHAP values, while those inside cluster around a positive contribution.

Across demographics, median income shows a clear positive gradient, whereas the elderly share shows the opposite: neighborhoods with large elderly populations are associated with lower apartment building prices.

Overfitting?

None of this matters if LightGBM is just memorizing training data. Several guardrails keep the model reliable. A conservative learning rate of 0.05 means each new tree contributes only five percent of its prediction to the running total. Trees are kept shallow (max 20 leaves, and a maximum of 8 levels of depth), and every leaf must contain at least 300 observations. Moreover, training is run at each iteration on a different subsample of the training set.

The most important safeguard is early stopping. At every boosting iteration, RMSE is evaluated on the validation set. If 100 consecutive rounds pass without improvement, training halts. The final predictions use the best-performing iteration, not the last one.

The best iteration in our exercise was round 2,208, with a moderate gap between training and validation error.

Conclusion

For multifamily property valuation, LightGBM delivers a meaningful improvement over OLS in out-of-sample performance, with 14 percent lower prediction error and 19 extra percentage points of explained variance, without sacrificing interpretability.

Does this mean we should throw out hedonic regressions? Not at all. OLS remains a useful, interpretable benchmark and, when combined with causal variation, can offer clear insights into what affects real estate prices.

However, when the goal is pure prediction (i.e., forecasting what a property will sell for), machine learning is a powerful tool that can improve on OLS.

Share

Over the past decade, a convergence of rising home prices, demographic shifts, and the historically unusual phenomenon of mortgage rate lock-in have reshaped who moves, when, and why.

We use nationally representative data from the U.S. Census Bureau’s Survey of Income and Program Participation (SIPP) to trace transitions within and between owning and renting. We track mobility at both the individual and household level across three panels covering 2013–2016, 2017–2020, and 2021–2023.

The picture that emerges is complex, with some expected results, some surprising facts, and several subtleties in the data that are essential for proper interpretation:

• There is no clear trend in owner-to-owner moves, although there is a sharp drop in 2021–2023, consistent with recent industry and academic research on mortgage rate lock-in.

• There is a strong, steady decline in renter-to-renter mobility over the entire period.

• Moves from owner-occupied homes to rentals are surprisingly frequent and have dropped sharply in 2021–2023, consistent with what suggested by recent research. However, these flows do not directly map into home sales or downsizing. Rather, they are significantly driven by young adults and other individuals breaking away from their current households. Thus, they need to be interpreted with care.

Subscribe now

Basic Facts from Individual-Level Statistics

To understand mobility trends, consider adults living in an owner-occupied home as facing one of three possible transitions each year: remaining in the same owner-occupied home, moving to a new owner-occupied home, or moving into a rental. The same framework applies to renters: they can remain in the same rental, move to another rental, or transition into an owner-occupied home.

The figure below summarizes the annual transition rates for individuals who do not remain in the same housing unit: owner-occupiers (top panel) and renters (bottom panel) who move to a new owned or rented home. We produce estimates for each of the three historical panels constructed from the SIPP data.

In 2013–2016, the share of owner-occupier individuals moving to another owner-occupied home is 5.5% per year. This increases to 6.4% in 2017–2020, before falling back to 5.6% in 2021–2023. The reversal aligns closely with the onset of mortgage rate lock-in. millions of homeowners who secured rates near historic lows in 2020–2021 now face a steep financial penalty for selling and repurchasing at much higher mortgage rates. The result is a housing market in which owners sit tight, reducing turnover, which is consistent with the findings of Fonseca and Liu (2024).

Interestingly, the data also reveal a large flow from owning to renting. In 2013–2016, 2.3% of owner-occupier individuals moved to a rental each year. This means that own-to-rent moves account for approximately 30%1 of all owner moves in that period. This is a somewhat surprising finding, but aligns with recent evidence discussed by Katz (2024).

This rate drops dramatically over time and is halved by the 2021–2023 panel, in which only 1.2% of owner-occupiers move to rentals. Katz (2024) argues that this empirical fact may have major implications for the effects of lock-in on house prices. However, the nature of these moves is somewhat unclear. Are they really capturing households selling their homes to become renters? As we show below, the answer is more nuanced than it first appears.

Another striking trend is that renter-to-renter mobility has steadily declined over time. In 2013–2016, roughly one in five renters moved to a different rental unit each year. By 2021–2023, that figure had dropped to approximately one in eight, a decline of nearly 40% in less than a decade.

This trend may be tied to several mechanisms. High rents and tight budgets may discourage renters from bearing the costs of relocating. Alternatively, renters’ preferences may be changing, making them less inclined to move.

Transitions from renting to homeownership have also declined, consistent with the progressive decrease in ownership affordability and the sustained growth in house prices. However, the magnitude of this decrease is more modest, with the annual transition rate falling from 7.3% in 2013–2016 to 5.5% in 2021–2023.

Digging Deeper into Owners’ Mobility

To better understand mobility patterns for owners, we now examine transition rates for household heads. Note that these data include single-person households.

Interestingly, the household-level statistics diverge from the individual-level results in important ways. Owner-to-owner moves are roughly a 0.5 to 1 percentage point smaller for household heads than for all individuals.

The larger discrepancy appears in own-to-rent transitions, which are much less frequent for household heads than for individuals overall. This suggests that many of the individuals moving from an owner-occupied home to a rental are forming a new household (perhaps leaving a family home), or joining a different household. Thus, a large share of these moves is unlikely to reflect existing households’ decisions to sell and rent.

Notably, there is also relatively little change in own-to-rent transitions for household heads when comparing 2017–2020 to 2021–2023: the rate drops only from 0.8% to 0.6%, rather than the much larger decline from 2.3% to 1.2% observed at the individual level.

To dig deeper into this pattern, we examine a breakdown by age across all panels. The figure below plots the share of individuals and households transitioning from owner-occupied homes to rentals across age groups.

Two interesting facts emerge. First, the share of individuals making own-to-rent moves is highest among younger adults (those under 35) and adults under 50. This pattern is present for both individual and household moves, but is much more pronounced for individual moves. Thus, these transitions are not mainly driven by downsizing among older homeowners, and, in the individual data, may have something to do with younger adults leaving the household to become renters.

Second, individual-level transition rates are two to three times larger than household-level rates across all age groups. This further reinforces the interpretation that a substantial portion of own-to-rent moves at the individual level reflects changes in household composition rather than homeowners choosing to sell and become renters.

Conclusion

The SIPP is a valuable tool for studying individual and household mobility in the United States. While its data are subject to limitations, it allows us to highlight several important patterns. We find that mobility is changing in meaningful ways, and that some less well-known aspects of the data, such as the non-negligible frequency of own-to-rent moves, deserve more attention. However, as our analysis shows, these facts require careful interpretation: what appears at first glance to be homeowners becoming renters is, in large part, individuals (especially young ones) leaving their current households to form new ones.

Share

For the past decade, my research agenda has been closely tied to Airbnb and the broader short-term rental market, with a focus on competition, housing outcomes, regulation, and market design. With that in mind, I am very excited to say that I have recently joined Airbnb as a Housing Scholar, a role that aligns closely with both my academic work and the mission of R.E.A.L.

This role fits naturally with my research agenda over the past decade, much of which has focused on Airbnb and short-term rentals more broadly. Over the years, I have studied how short-term rental platforms and their interactions with local housing markets, incumbents, regulations, and urban policy operate, using observational and quasi-experimental designs.

The goal of this appointment is straightforward: to conduct rigorous research on housing markets and housing policy and to help improve how evidence informs policy discussions on short-term rentals and local housing outcomes.

Housing markets are complex, reflecting the interplay of zoning, taxation, supply constraints, local demand shocks, and long- and short-term rental activity. As a result, simple narratives often overlook key mechanisms. My interest has always been in understanding these mechanisms using data, credible empirical designs, and transparent assumptions.

In this role, I will collaborate with Airbnb’s research and policy teams while continuing to serve as a professor at the University of Southern California. My views and conclusions will remain my own, and any research I conduct will adhere to standard academic norms and integrity.

This appointment also aligns with the work that Marco and I have been doing at R.E.A.L., an independent initiative that provides data-driven insights and research on real estate markets. Some of that work has already focused on short-term rentals, and more is coming. When possible, I will share on R.E.A.L. updates, results, and reflections as projects develop.

This is an opportunity to deepen research I have pursued for years, to gain a richer institutional context for how housing platforms operate, and to help move the debate away from anecdotes toward evidence.

As always, feedback is welcome.

Over the past two decades, U.S. states have enacted legislation offering tax incentives to attract data center development. The goal of these policies is to stimulate local economic and job growth. Many state and local officials view data centers as a catalyst for building a local tech industry.

This approach is not new. Local and federal tax incentive programs have long been used to stimulate investment and job creation in the United States. However, the effectiveness of these programs in attracting new investment activity and generating broader economic spillovers varies considerably. For many initiatives, evidence on their effectiveness remains inconclusive long after they are enacted.

Do state-level tax incentives for data centers meaningfully shape the geography of data center development? Are these programs substantial enough to alter developers’ decisions about where to invest? Do data centers act as catalysts for new tech hubs?

A recent working paper by Antonio Gargano and Marco Giacoletti provides evidence that incentives significantly increase data center development. However, the increase in investment is concentrated primarily in the largest-scale facilities. Tax incentives appear to be contributing to a shift in data center development toward hyperscale, single-tenant cloud facilities. Because of the scale of their operations, each of these projects receives millions of dollars in annual tax benefits that, when capitalized, represent a substantial share of total development costs.

Overall, these incentives meaningfully impact the data center industry and represent a substantial long-term commitment by U.S. states. Yet the paper finds that the effects of data center development on local tech job creation are limited or insignificant.

Given the scale of these policies, regulators should carefully assess their broader economic benefits. For the programs to be justified, their primary gains must extend well beyond local tech employment.

Subscribe now

State-Level Incentives Impact Development, but Scale Matters

The paper employs a staggered difference-in-differences (DiD) framework, using recent methodologies that correct for biases in traditional staggered designs. This approach allows the authors to identify the causal effect of incentive implementation on data center development.

Although the first incentive law was enacted in 2000, most programs were introduced during the 2010s. The newest provisions (in Kansas, Massachusetts, and West Virginia) took effect in 2025. These incentives primarily take the form of tax rebates. States do not offer direct subsidies that cover development costs. Instead, the benefits operate indirectly by reducing operating costs for new facilities, thus increasing their valuations, and the value created by new development.

The paper finds sizable effects, but also clear differences across categories of data centers. The analysis focuses on three categories: hyperscale, wholesale, and retail.

Hyperscale data centers are the largest category and began expanding rapidly in the 2010s. Designed for large-scale, high-speed computing, they have become the key infrastructure to modern LLM and AI workloads. In this industry, asset size is measured by the amount of power that can be reliably delivered to servers, which is called Uninterruptible Power Supply (UPS). Operating hyperscale facilities in the dataset used in the paper average about 33 megawatts of UPS. Wholesale data centers fall in the intermediate range and support either cloud computing or network infrastructure. Retail data centers are the smallest category, averaging roughly 4 megawatts of UPS, and focus on network connectivity and server colocation.

The event-study chart below shows the average effect of implementing a state-level incentive package on the development of new data centers across categories. On the x-axis, 0 denotes the year in which the incentive was enacted; positive values represent years after implementation, and negative values represent years before.

The results show very large effects on hyperscale development: within five years of implementation, the number of new hyperscale facilities increases by more than 50%, and their total combined UPS capacity increases by over 200%. Wholesale data centers also exhibit positive responses, though the estimated effects are smaller and substantially noisier. There is virtually no detectable impact on the development of retail data centers.

The analysis in the paper shows that this pattern is largely driven by the design of state incentive laws, which impose minimum size thresholds to qualify for incentives. These thresholds fall well within the operating scale of hyperscale, high–computing-power facilities but exceed the typical size of retail, network-oriented data centers. As a result, hyperscale projects are the most likely to be eligible for incentives, while most development projects for retail facilities are effectively excluded.

Incentives Create Substantial Value for Developers and Operators

Most states structure their incentives as sales and use tax exemptions, with a smaller number also offering property tax abatements. Sales and use tax exemptions typically cover electricity consumption as well as purchases of tangible equipment (servers, storage devices, and networking hardware), and key infrastructure components (electrical systems, UPS capacity, and cooling equipment). As noted earlier, these incentives indirectly stimulate data center investment by improving operating margins and increasing the value of new facilities.

How much value do these incentives deliver to newly developed facilities? The paper addresses this by estimating sales and use tax rebates for data centers operating in 2025. The figure below reports incentives per megawatt of UPS capacity. States shown in white have active incentive programs but no qualifying data centers in the paper’s dataset, while states in grey do not offer incentives. Alaska and Hawaii are omitted because neither state has an incentive program.

Although Virginia hosts the largest concentration of data centers in the country, it does not offer the highest subsidy per unit of power; its incentives amount to only about $22,000 per UPS megawatt, across qualifying operating facilities in 2025. Indiana, Minnesota, and Tennessee provide the most generous benefits. Texas, Iowa, and South Carolina also provide substantial incentives to qualifying facilities in 2025, each exceeding $80,000 per UPS megawatt.

The average operational hyperscale data center covered by the study has 33 megawatts of UPS capacity. With incentives of $80,000 per UPS megawatt, this translates into approximately $2.64 million in annual operating expense savings in 2025.

In commercial real estate, a common back-of-the-envelope valuation approach divides net operating income by prevailing market cap rates to find a rough estimate of property values. Using a current hyperscale cap rate of roughly 6.5%, these annual savings capitalize to about $40.6 million in additional asset value for a typical 33 megawatt facility.

Local Long-Term Effects on the Tech Job Market are Not Significant

Finally, the paper examines the effects of data center development on local labor markets, with a particular focus on the creation of new tech jobs. While the construction of data centers can generate short-term employment gains in the construction sector, and may temporarily stimulate local economic activity, these benefits are transient.

The incentives are structured as long-term tax rebates: they reduce operating costs for decades. This is capitalized into substantial increases in asset values. Thus, long-term creation of high-paying jobs appears to be a natural metric for evaluating the local economic benefits of these programs. Moreover, policymakers have argued that attracting large data center projects can serve as a catalyst for expanding the state’s broader tech sector.

Does data center development lead to an increase in new tech jobs? To answer this question, the paper uses new data from RevelioLab, which compiles individual job profiles from online sources and links them to specific employers and locations. Using these data, the authors construct three measures of annual tech job creation at the census place (city, town, or unincorporated area) level. The first measure identifies tech jobs based on profile text descriptions, the second one is based on standardized occupational classifications provided by RevelioLab, and the third one uses employer characteristics.

Because the location of data centers within a state is not random, the paper employs a matching estimator to ensure that census places exposed to new data center development are compared to similar places that are not. The matching procedure aligns locations on key demographic and economic characteristics.

The figure below reports the estimated effect of developing a new data center within 10 miles of a census place on the relative increase in the number of new tech jobs initiated in that same place three years later. Dots of different shapes correspond to point estimates for different measures of new tech jobs, while the solid lines represent confidence intervals.

Results are shown for Arizona, Iowa, Minnesota, North Carolina, Nevada, Ohio, South Carolina, Texas, and Virginia. These states have both substantial incentive programs and a significant amount of large-scale data center development over the past 15 years.

The results do not provide consistent evidence of positive effects. Instead, the estimates are either close to zero or extremely noisy. Similar patterns emerge across alternative time horizons and when varying the matching procedure. The only state showing systematically positive and significant impacts is Ohio, but even there the estimated effects are modest, on the order of a 1%–5% increase in new tech job starts. For census places that experience data center development, the average number of annual tech job starts in Ohio is about 150, so a 1%–5% increase translates into only 2 to 7 additional jobs per year.

Share

Platforms such as Airbnb have revolutionized the way people travel. With just a few clicks, travelers can access a vast and diverse array of lodging options (spare rooms, condos, or entire homes), often at prices well below those of traditional hotels. The broader supply of temporary accommodations lowers the overall cost of visiting a city, potentially attracting more visitors.

Share

Indeed, research shows that Airbnb attracts tourists who might not have visited otherwise and that its decentralized network of listings shifts tourism away from traditional hotel districts and into residential neighborhoods. This spatial redistribution of visitors can reshape urban spending patterns, increasing revenues for local restaurants, cafés, and neighborhood-based services.

However, critics argue that STR platforms also reduce the supply of long-term rental housing, inflate rents, and disrupt community life. In response, many U.S. cities have introduced increasingly stringent regulations, which require host registration, cap the number of rental nights, or prohibit the short-term rental of entire homes. While these policies are intended to stabilize housing affordability and preserve neighborhood character, they may also generate unintended ripple effects on local economic activity.

In September 2023, New York City enacted one of the strictest short-term rental regulations in the United States. The new rules required hosts to register with the city and prohibited most entire-home rentals unless the host was physically present during the stay. Within two months, the number of active Airbnb listings dropped from roughly 23,000 to just 4,600.

This effectively eliminated the majority of the city’s Airbnb market, and had an immediate effect on hotel prices. According to news reports, by December 2023, three months after the new regulations took effect, New York City’s average daily hotel rate had surged by 11%, reaching approximately $393 per night. Over the same period, revenue per available room (RevPAR) increased by 15.6%, compared to a national rise of only about 0.3%.

This drastic policy offers a unique opportunity to examine how restricting short-term rentals affects local economic activity, particularly in the restaurant sector, one of the neighborhood services that, as discussed above, tends to benefit from the decentralized flow of tourists facilitated by Airbnb.

Using detailed monthly credit and debit card transaction data from SafeGraph, covering thousands of restaurants across major U.S. cities between July 2022 and July 2024, we analyze how restaurant spending evolved before and after New York City’s 2023 regulation. Our empirical strategy combines propensity score matching with a difference-in-differences framework, comparing spending trends in New York City to those of a matched control group of restaurants in other major cities.

The results are striking: following the implementation of the regulation, restaurant spending in New York City declined by approximately 10% on average relative to the control group. The figure below plots the monthly changes in restaurant spending for New York City restaurants compared with their matched controls during the year preceding and following the policy’s implementation (the red dashed line indicates the month immediately preceding the policy's effect).

First, the figure shows that in the pre-regulation period (before September 2023), differences in spending between treated and control restaurants were close to zero, indicating that their spending trends were similar prior to the policy change. After the regulation took effect (September 2023 and after), restaurant spending in New York City declined steadily for approximately four months before stabilizing around five months after the policy's implementation, with a peak reduction of approximately 12% relative to the control group.

To put these results in perspective, we provide a back-of-the-envelope estimate of the implied revenue loss. The dataset used in the analysis includes 3,220 restaurants and debit and credit card transactions from roughly 9 million customers. However, New York City as a whole has far more restaurants and a much larger consumer base than what is represented in this sample.1 According to the Office of the New York City Comptroller, taxable sales at restaurants and other eating places totaled $26.904 billion between September 2023 and August 2024.2 Assuming that the observed decline in restaurant spending extends to the entire industry, and approximating the decline as a 10% reduction over the year, the counterfactual level of sales in the absence of the policy would have been $26.904 billion / (1 – 0.10) = $29.893 billion. The implied revenue loss is therefore the difference between the observed revenue ($26.904 billion) and the counterfactual revenue ($29.893 billion), amounting to approximately $2.989 billion.

Finally, we look at which restaurants are more affected by the STR policy. We find that the spending decline is particularly pronounced among higher-priced restaurants and establishments that cater to a larger share of non-local customers. These are the businesses most reliant on tourism demand.

In summary, New York’s 2023 short-term rental regulation had clear side effects on neighborhood economies. By cutting the supply of decentralized short-term accommodations and driving up hotel prices, it squeezed travelers’ budgets and dampened demand for restaurants and other local services. The broader lesson is that when crafting rules for platforms like Airbnb, policymakers need to consider not just housing affordability but also the local economic ecosystems that depend on tourism and short-term visitors.

When we think about neighborhoods, one of the first challenges we face is: how do we actually measure their profile and their change over time? We might rely on data about resident characteristics (income, jobs, or wealth) or on quality-of-life indicators such as amenities and crime rates. However, combining all of these into a single, coherent measure may require complicated weighting approaches that may lack interpretability.

There are, however, two basic facts that are intuitively true. First, the quality of the housing stock, especially single-family homes, is strongly correlated with the overall conditions of a neighborhood. Second, the development of new, high-quality single-family homes tends to occur in either already desirable neighborhoods or those that are actively undergoing gentrification.

These observations serve as the motivation for our approach. Instead of trying to measure every possible dimension of neighborhood life, we focus on the physical characteristics of the housing stock and on how these characteristics change over time. This offers an alternative method for classifying neighborhoods and identifying neighborhood change that is both practical and interpretable. The main challenge, of course, is that housing is highly complex and multi-dimensional. Anyone can walk down a street and get a sense of whether the homes are new, large, or in good condition, but so systematically with large datasets and statistical methods is much harder.

We tackle this challenge using a clustering algorithm called K-Means. This algorithm groups homes into buckets of similar properties. We then identify a manageable number of clusters that summarize the housing stock across the geographical area of interest.

For our exercise in this post, we work with data on residential parcel characteristics available from the Fitzgerald Institute of Real Estate at the University of Notre Dame and conduct a study of neighborhoods across the state of Indiana.

Share

Defining Home Types in Indiana

We focus on single-family parcels, and collect all the available physical characteristics of homes, such as effective year built, square footage, number of bedrooms and bathrooms, number of stories, the presence of amenities like fireplaces or pools, and roof quality. Effective year built is not the original construction year of the building, but rather a date that reflects the current structure and wear-and-tear of the home. We collect data on about 1.7 million single-family homes in Indiana as of the end of 2024.

We deliberately exclude location among the characteristics of interest because our goal is to use the physical features of local homes to classify neighborhoods.

After running the K-Means algorithm, we identify 9 distinct clusters,1 which we label as shown in the figure below. Note that these are names that we assign post-clustering, given the housing stock composition; the algorithm was not directed to bucket homes in this way. However, it turns out that clusters are strongly determined by vintage and size. You can also notice that we highlight some clusters in red and some clusters in green. This is to visualize the split between clusters of homes that have better quality (in green), and homes that do not have as good quality.

As mentioned, the definitions above are based on the prevailing characteristics within cluster. The figures below show, for each cluster, the median, top quartile, and bottom quartile of age and size. We can see that from left to right, the algorithm sorts homes into clusters belonging to different vintages, based on the effective year built. The differences in effective year built are substantially less pronounced when comparing the three clusters that we label as “Late 1980s”, “Small Early 1990s” and “Large Early 1990s”. However, these three clusters are substantially different in terms of home sizes, measured as living square feet. The median size for the “Late 1980s” cluster is 2000 square feet, the median size for the “Small Early 1990s” cluster is just 1,500 square feet, and the median size for the “Large Early 1990s” cluster is approximately 2,200 square feet.

The three clusters, “Late 1980s”, “Large Early 1990s”, and “1990s-2000s”, have the largest homes, with the third cluster having a median size of 4,000 square feet. These clusters also contain homes that tend to have a higher likelihood of having multiple stories and a higher likelihood of having amenities, such as fireplaces.

Do the clusters actually capture significant differences in the housing stock? We explore this by testing whether the clusters explain differences in house sales prices. Note that prices were not used in creating the clusters, and are an independent outcome. We look at sales of homes within the same census tract and year between 2010 and 2025, and thus compare prices across clusters while controlling for location and time of sale. The figure below reports estimates of the price premiums for all clusters with respect to the 1950s cluster. It is apparent that even within the same census tract and year, there are significant differences in prices for homes belonging to different clusters.

The “Late 1980s”, “Large Early 1990s”, and “Late 1990s-2000s” clusters have the largest premiums, equal to 70%, 60%, and 105%, respectively, which means that they are typically 60% to 105% more expensive than homes from the 1950s. This confirms that these three clusters combined capture the higher-quality homes in the state.

Classifying Neighborhoods Using Home Clusters

We then turn to classifying neighborhoods across Indiana using the local composition of single-family homes across clusters. We calculate for each census tract the share of “good quality” homes by pooling together the “Late 1980s”, “Large Early 1990s”, and “1990s–2000s” clusters.2 The map below shows how the share of good-quality homes varies across census tracts, and provides us with a perspective on the state of the single-family housing stock in Indiana as of the end of 2024 (the most recent update to our dataset).

We can see large heterogeneity, with the highest shares observed in the suburbs of Indianapolis and Fort Wayne, around Bloomington and south of Gary, in the north-west. We also observe large rural pockets where good-quality single-family housing is clearly scarce.

Zooming in on Indianapolis, we can see that the highest share of good-quality homes is located in the north of the city, particularly in areas to the north-west (such as Zionsville), and downtown. These neighborhoods have indeed experienced gentrification in recent years. On the other hand, the share of good quality stock is lowest in the areas surrounding downtown, especially to the south-west.

These patterns have important applications. Regulators, for instance, may want to identify pockets of underdeveloped or older housing stock to target redevelopment incentives. Developers may want to know which areas already have strong housing stock and where it is changing most rapidly, signaling opportunities for investment.

Analyzing Neighborhood Change in South Bend

Our measure can also serve as a useful tool for tracking neighborhood changes over time. For this exercise, we focus on South Bend, where the University of Notre Dame is located. This city has undergone several requalification efforts over the last decade. To track the effects of these efforts on the housing stock, we study how the share of good-quality single-family homes has changed across census tracts from 2010 to 2024.

The first figure below plots the 2010 shares, and the second one the 2024 shares. Despite the relatively small size of the city, there is substantial disparity in the housing stock, particularly when comparing the northeastern suburbs against the core metropolitan area.

However, most census tracts have not experienced meaningful change over the last 15 years. The only areas that saw an increase in good quality housing stock are part of the central area of the city (within the green square). Here, we can see several tracts turning substantially lighter, with the share of good-quality homes increasing from around 10-15% to close to 30%. These neighborhoods correspond to the immediate surroundings of the Notre Dame campus and to the area between campus and the Saint Joseph River. They have indeed been the target of substantial residential redevelopment. However, spillovers of these efforts to the housing stock in the rest of the city appear to be still limited.

Conclusion

In conclusion, our clustering approach offers a compact and interpretable method for classifying housing stock across neighborhoods. The classification is strongly predictive of individual house prices. More importantly, it provides a new, data-driven measure of neighborhood quality and changes over time. With this measure, we can classify neighborhoods, identify underdeveloped areas, and systematically study neighborhood change.

In this post, we propose an approach that leverages the characteristics of the local real estate market. Using detailed data on single-family parcels, we (1) cluster the existing housing stock into quality tiers based on structural attributes (age, size, etc.), and (2) analyze which clusters dominate in each neighborhood. This allows us to identify transforming areas, neighborhoods that have been overlooked by redevelopment, and those that have benefited from recent new construction.

We apply our method to Los Angeles County and demonstrate that it can be easily used to categorize local single-family homes into distinct types, and to develop a straightforward definition of transforming/gentrifying areas that is strongly correlated with recent price growth. Intuitively, transforming areas can be identified by a mix of homes belonging to the highest and lowest quality buckets within the metropolitan area.

Defining “Home Types“

We want to categorize single-family homes within a metropolitan area into distinct “home types”, based on their physical characteristics. We intentionally exclude location from the classification because our aim is to use the attributes of the housing stock to help differentiate neighborhoods within the city. Defining home types is a complex problem: How many categories should there be? Should homes be grouped by age, size, or some combination of features?

For our analysis, we adopt the following approach. First, we select four key physical characteristics that reflect the appearance and structure of a home:

• Effective Year Built: This is not the original construction year of the building, but rather a date that reflects the current structure and wear-and-tear of the home. For example, a house initially built in 1960 but heavily renovated in the early 2000s could have an effective year built of 2000, indicating that its current state aligns more closely with a newer property.

• Number of Bedrooms

• Living Square Feet: This is the size of the interior space of the home.

• Living Square Feet Divided by Lot Size: This is the ratio of the interior space of the home to the lot size. It reflects how intensively the lot has been developed.

Second, based on these characteristics, we cluster homes using the k-means algorithm, which is a widely used machine learning technique for unsupervised classification. Given a set of features, k-means assigns each observation to one of k clusters, grouping together observations that are similar across all dimensions. In this context, homes within the same cluster share comparable physical attributes. We apply this method to Los Angeles County, using approximately 1.39 million observations on single-family parcels as of the end of 2022.

Mechanically, increasing the number of clusters improves the fit of the k-means algorithm. To select a parsimonious number, we apply the Elbow rule, which recommends choosing the smallest number of clusters beyond which additional clusters yield diminishing improvements in fit. The figure below illustrates this using the within-cluster sum of squares, a standard measure of fit that captures the similarity of observations within each cluster. Lower values of the black line indicate a better fit. We choose to use four clusters, because the curve flattens noticeably after four clusters, suggesting that additional clusters offer only marginal gains.

Interestingly, the algorithm produces a highly intuitive grouping of properties, making it easy to interpret each cluster as a distinct type of home. The table below summarizes the characteristics of homes in each cluster. We define them as:

• Very Old Small Homes (296,226 observations, or 21.34% of homes): The first cluster mainly consists of homes that have an effective year built between the 1920s and the 1940s, 2 bedrooms, and living square feet between 880 and 1340. These houses also tend to occupy a smaller share of their lot, with living square feet equal to 15-20% of the lot size.

• Old Small Homes (99,681 observations, or 7.18% of homes): This is the smallest cluster, comprising homes that are slightly more recent, with an effective year built in the 1950s, and slightly larger, with 3 bedrooms and living square footage between 1,200 and 1,700.

• Intermediate Homes (644,179 observations, or 46.41% of homes): This is the largest cluster, consisting of homes built in the 1970s and 1980s, with 4 bedrooms and living areas ranging from approximately 1,000 to 2,000 living square feet. Interestingly, these homes utilize their lots more effectively, with living square footage equal to 20-30% of the lot size.

• New Large Homes (347,802 observations, or 25.06% of homes): This is the most modern group, comprising properties built in the 1990s and 2000s, and thus also including newly built properties. They are larger than homes in other clusters, with 4 or 5 bedrooms, living areas ranging from the high 2,000 to above 4,000 square feet, and they occupy a large share of their lots, between 40% and 60%.

Los Angeles Neighborhoods with a Prevailing Home Type

With the home type clusters defined, we can now use them to describe neighborhood composition across Los Angeles County. The figure below divides the county into Census tracts and highlights those in which a single cluster is prevalent, meaning it accounts for more than 50% of the homes in the tract, while no other cluster exceeds 20%. This allows us to identify neighborhoods dominated by a distinct home type.1

The dark red areas represent neighborhoods where Very Old Small Homes are prevalent. These areas are primarily located south and east of Downtown Los Angeles, and are neighborhoods where housing quality has lagged behind that of the broader metropolitan area. A similar prevalence of Very Old Small Homes is also present in parts of Long Beach. Old Small Homes are prevalent in several neighborhoods in the eastern portion of the county and throughout the San Fernando Valley. New Large Homes are prevalent only in a handful of neighborhoods, typically in the more affluent locations, such as Venice, Manhattan Beach, and Redondo Beach.

Transforming (Gentrifying) Neighborhoods

An even more insightful application of our classification is identifying neighborhoods that have been or are actively undergoing transformation and potentially gentrification.

Intuitively, the most striking cases of transformation should coincide with areas that currently or historically had a large share of Very Old Small Homes and Old Small Homes, but have since seen redevelopment, with many properties replaced by New Large Homes. We identify such neighborhoods as Census tracts in which New Large Homes is the most or second most frequent type, but the other most frequent type is either Very Old Small Homes or Old Small Homes.

The figure below highlights 64 Census tracts that fit these criteria, located within the core metropolitan area of the county.

Consistent with our hypothesis that the transforming neighborhoods identified by our approach are gentrified or gentrifying, we find that these areas have experienced significantly higher house price growth than the rest of Los Angeles County over the past decade. To quantify this, we calculate house price growth at the Census tract level using data on home values from the American Community Survey (ACS) for the years 2010 and 2022. We compare mean and median house price growth in the Transforming Census tracts identified by our approach to those in the rest of the county. While Census tracts outside these areas also experienced substantial appreciation, with average price growth exceeding 300% between 2010 and 2022, the Transforming tracts saw even faster increases, with both mean and median growth exceeding 400% (one-third higher than that of other tracts).

Subscribe now

Conclusion

It is crucial for investors and regulators to develop a bird’s-eye view of neighborhood disparities within metropolitan areas, as well as emerging trends in transformation and gentrification.

We demonstrate that a simple machine learning approach can be highly effective in identifying these patterns. It provides an interpretable classification of the local housing stock into distinct types and characterizes neighborhood composition.

Broadly, the single-family housing stock can be grouped into old, small homes and large, new homes. The coexistence of the oldest and newest home types within the same neighborhood signals active transformation and is associated with rapid house price growth.

Share

A key concern is the effect on impacted homeowners, and on the wealth they had invested in their homes. In principle, fire and homeowner insurance should help cover the costs of displacement and reconstruction. The crucial problem, however, is that private insurers had stopped renewing fire insurance policies in the areas that were later hit by the fires. While local homeowners could have switched to the California FAIR Plan, a state-backed insurance option, many may have chosen to remain uninsured. This is especially true for homeowners who had fully paid off their mortgages and were therefore not required by lenders to obtain new insurance. Even those who did obtain a new policy may have underinsured their homes, selecting policy coverages well below market values to reduce premium costs.

Homeowners who are now unable to obtain insurance payouts or financial aid may be forced to sell their damaged homes, because they cannot afford the rebuilding costs and are likely under financial pressure from having to pay for temporary housing while they are displaced. Selling a damaged home in the immediate aftermath of a disaster is likely to result in a low sales price and a large capital loss, even after accounting for avoided rebuilding expenses. This is because few buyers are willing to invest in a severely damaged property, or in a disaster area.

We take an initial look at key patterns in home sales before and after the Eaton Fire in Altadena. To understand how housing market trends in the fire-affected area diverge from those in nearby unaffected areas, we compare listings within the fire perimeter to those in surrounding neighborhoods located within a 2.5-mile buffer zone outside the perimeter.

We find a sharp increase in sales volume within the fire-affected area, with many properties sold as “lots.” Nearly all of these sales involve heavily damaged homes, whether listed as single-family residences or vacant lots. Single-family listings are selling for roughly half the price relative to homes sold in the same area in 2024, and relative to 2025 sales in neighboring, unaffected areas. These post-fire sales prices likely fall below pre-fire bare land values. Lot sales prices also appear to fall below pre-fire land values.

Subscribe now

Volume and Characteristics of Sold Properties

Our analysis combines data from three sources: home and lot sales scraped from Zillow for the period between January 1^st 2024 and June 18^th 2025, parcel-level shapefiles for Los Angeles County, and fire damage information published by CAL FIRE.1

The maps below visualize changes in sales volume in the area struck by the Eaton fire (in yellow), relative to the nearby neighborhoods (in grey), located within a 2.5-mile buffer of the fire boundaries. The map on the left displays the locations of all sales that occurred between January 2024 and June 2024. Red dots represent sales that occurred in the area later affected by the Eaton fire in January 2025. The blue dots represent sales in the surrounding area. The second map presents the same analysis for the period from January 2025 to June 2025, after the fire struck.

It is clear that, in the aftermath of the Eaton Fire, sales volume within the fire perimeter is significantly higher than during the same months in the previous year.

Share

To explore this pattern further, the graphs below display the total number of monthly home sales from January 2024 to June 2025 in both the Eaton Fire area and the surrounding neighborhoods within a 2.5-mile radius of the fire perimeter. They also distinguish between properties sold as “lots” and those sold as single-family homes. While sales volume increases after the fire both within and outside the fire perimeter, the rise is far more dramatic within the fire-stricken zone. In 2024, monthly sales in the fire area ranged between 9 and 20. That number jumps to 35 in March 2025 and to 60 in April.

Moreover, lot sales were virtually nonexistent in both the fire-affected area and the surrounding neighborhoods prior to January 2025. After the fire, they remain highly uncommon in the surrounding areas but increase sharply within the affected zone. They account for approximately 50% of all sales in most months, and nearly all sales in June.

Intuitively, these lots represent properties that were completely destroyed by the fire. However, even the single-family home sales in the Eaton Fire area almost exclusively involve heavily damaged structures. In the table below, we match the 174 properties sold between January and June 2025 with fire damage data provided by CAL FIRE. Nearly all of these properties sustained damage to 50% or more of their structure.

Property Prices

At what prices are the fire-impacted properties selling? The figure below shows the median sales price per square foot for single-family listings, each month from January 2024 to June 2025. In 2024, median prices are similar in both the Eaton Fire area and the surrounding neighborhoods, averaging around $800 per square foot. In 2025, prices in the surrounding areas remain stable, while prices in the fire-affected area drop sharply, and range from $200 to $400 per square foot.

While this decline is consistent with the fact that most of the properties sold in the fire zone are heavily damaged, the magnitude of the drop is particularly striking. A recent research paper estimates that in 2022, approximately 67% of the value of a single-family home in Los Angeles County is attributable to the land it sits on. If we take this share as fixed, it implies that for the median house land alone accounts for a value of $536 per square foot. Thus, the single-family listings in the Eaton fire area are now selling at values below the pre-fire valuation of the bare land.

For our final analysis, we examine the sales prices of properties listed as lots. In this case, a direct comparison between the fire-affected area and surrounding neighborhoods is difficult. This is because lot sales are virtually nonexistent outside the fire zone. Instead, we use an external benchmark: the same research paper mentioned above estimates the value of a quarter-acre parcel of land for a single-family home in Los Angeles County at approximately $918,000. The figure below compares this benchmark to the price per quarter acre implied by the median lot sales prices in the Eaton Fire area. In February, the implied price for a quarter acre is close to $880,000, and it declines and stabilizes between $800,000 and $760,000 in the following months. This represents a discount of roughly 13% relative to the 2022 estimate.

Conclusion

Overall, the evidence suggests a significant increase in sales volume within the fire-affected area, primarily driven by homes that were severely damaged by the fire. These homes are selling at steep discounts, supporting the view that many owners are opting to quickly liquidate their properties before reconstruction efforts have even begun. This may reflect financial pressure among owners who were uninsured or underinsured.

While this trend may create opportunities for large-scale redevelopment, it comes at the cost of significant wealth losses for the impacted households.

Subscribe now

This post addresses a fundamental question: how much capital, combining resources from insurance, state relief, and federal relief, should be set aside for each state to adequately meet anticipated disaster-related losses?

Falling short of the capital required to address losses can have severe consequences for the local population and economy, and lead to slow and problematic recoveries. However, overcommitting resources is also problematic, especially when state and federal funds are involved. It may delay or prevent the implementation of other essential projects.

The central point we would like to stress is that regulators, when determining capital needs, should focus not only on “mean” expected losses, but on the entire distribution of potential losses. Focusing solely on the mean expected loss may leave U.S. states underfunded in many reasonably likely scenarios.

For example, the 2023 FEMA’s National Risk Index (NRI) dataset provides estimates of the expected annual loss for each U.S. county from disasters such as floods, hurricanes, and wildfires. Aggregating expected annual hurricane-related losses across all Florida counties yields a total estimated expected loss of $7.2 billion for buildings (real estate). However, in 2024, two major hurricanes (Helene and Milton) struck the state, causing over $30 billion in property damage. This stark disparity highlights how resources calibrated to cover only the expected loss can fall dramatically short in the face of severe disasters. At the same time, Helene and Milton might have been extreme events, and it is unreasonable to set budget goals based on extremely rare occurrences.

We propose here a middle ground, which is to estimate the minimum amount of capital required to cover losses that are infrequent, but not highly unlikely. For example, those with 1-in-10 (10%) or 1-in-20 (5%) annual probability. This is the Value-at-Risk (VaR) methodology, widely used in finance and insurance.

We find that the “right tail” of the distribution of disaster losses is long, and VaR estimates, capturing the capital needed to cover infrequent, but not improbable, events substantially exceed “mean” expected losses. This is especially true for wildfires, which tend to have low expected annual loss, but comparatively really large VaR.

The Value-at-Risk (VaR) Approach

The Value-at-Risk (VaR) approach provides a quantitative assessment of the minimum amount of capital needed to meet losses from events that occur with probability smaller or equal than p, which is is the Value-at-Risk threshold.

Calculating VaR involves two main steps. First, one must derive the distribution of losses. This can be done either analytically or through simulation. Second, a risk threshold p is selected, commonly set at 10%, 5%, or 1%. As shown in the figure below, the 5% VaR corresponds to the 95th percentile of the loss distribution (the loss amount that is exceeded only by the highest 5% of outcomes).

We construct annual state-level loss distributions by running simulations based on county-level data from the 2023 FEMA’s National Risk Index (NRI). This dataset provides the probability of various disasters and the expected losses to buildings conditional on a disaster occurring.1 A key assumption in our simulation is that disaster occurrences are uncorrelated across counties. To the extent that disasters are instead spatially correlated, our approach will underestimate the upper tail of the loss distribution, and thus the VaR amount. Moreover, conditional on a disaster striking, we do not model the occurrence of extreme losses. Thus, our VaR estimates should be interpreted as conservative. We focus on losses from hurricanes and wildfires, which have been particularly devastating types of natural disasters in the last decades.

The Distribution of Losses in Florida and California

The figures below report the simulated distribution of the dollar amount of building losses for Florida, the state most exposed to hurricanes, and California, the state most exposed to wildfires.

Starting with Florida, the distribution of losses is highly dispersed, with a long “right tail” representing scenarios involving large-scale damage. The red and magenta lines indicate the 10% and 5% Value-at-Risk (VaR) levels, corresponding to losses of $10.2 billion and $11.3 billion, respectively. This implies that Florida faces a 1-in-10 chance of incurring annual losses of $10 billion or more, and a 1-in-20 chance of losses of $11 billion or more. These figures stand in stark contrast to the expected loss (black line) of approximately $7.2 billion.

Calibrating capital resources to just meet the expected loss can leave disaster recovery substantially underfunded with relatively high likelihood.

This point becomes even more striking when we examine the distribution of wildfire-related losses in California. Unlike the hurricane loss distribution in Florida, most simulated scenarios for California involve minimal losses. However, the right tail of the distribution is extremely long, encompassing rare but catastrophic events.

The intuition is that fires occur far more frequently in areas with limited exposed real estate. However, when fires do strike regions with significant property exposure, they can result in substantial losses. Extreme examples are the fires in Pacific Palisades and Altadena.

As a result, the expected loss provides in this context even less guidance than in the case of hurricanes.

The black line marks an expected annual loss of $1.4 billion. In contrast, the 10% Value-at-Risk is $3.8 billion, and the 5% Value-at-Risk soars to $12.2 billion. This implies that California faces a 1-in-20 chance of incurring wildfire losses nearly ten times the expected loss amount in a given year.

Value-at-Risk from Hurricanes and Wildfires across U.S. States

We extend the analysis beyond Florida and California to include all states with substantial exposure to hurricane or wildfire risk. The table below reports the top 20 states ranked by expected loss from hurricanes. For each state, we show the expected loss, along with Value-at-Risk (VaR) estimates at the 10%, 5%, and 1% thresholds.

Florida, Texas, North Carolina, South Carolina, and Louisiana have the highest expected losses as well as the highest VaR figures across all thresholds. Among these states, the 5% VaR exceeds the expected loss by 50% to 200%, with the largest gap observed in Texas. At the 1% VaR level (a more conservative benchmark) Florida would require $13 billion in capital, while Texas would require nearly $9 billion. Even when using the 5% threshold, capital needs exceed $1 billion in each of the top 15 states.

The table below presents the same analysis for wildfire risk. The gap between expected loss and VaR is even more pronounced than in the case of hurricanes. California is the only state with expected annual wildfire loss exceeding $1 billion. In contrast, Florida and Texas (the second and third most exposed states) have expected loss of less than $300 million. However, Florida has a 5% VaR of $3 billion, and Texas $1.2 billion. This again highlights that capital reserves based solely on expected loss would be inadequate for covering even moderately unlikely events.

Similar effects are observed in Arizona and Utah, where the expected loss is below $200 million, but the 5% VaR reaches $600 million and $1.6 billion, respectively. Additionally, some states such as Colorado and Hawaii appear relatively unexposed even under likely tail scenarios, with low 10% and 5% VaR, but face the potential for extreme losses in rare events. Colorado, for instance, has a 1% VaR of $5 billion, while Hawaii’s 1% VaR exceeds $1 Billion.

Conclusion

We believe that Value-at-Risk analysis, and more broadly, examining the right tail of loss distributions, is a valuable tool for agencies involved in planning and allocating resources for natural disasters.

The specific figures in our analysis are subject to some limitations. They rely on statistical approximations and on FEMA estimates published in 2023, which may be revised as new data on climate risk become available.

While the exercise presented here is a simplified example based on readily available data, it illustrates how such methods can offer a more realistic picture of disaster risk exposure and inform more resilient funding strategies.

Subscribe now

See the figure below for an example: the property sold on 4/30/25 for $1.23 million, while Zillow’s estimate that day was $1.324 million.

Buyers, sellers, and agents routinely consult Zestimates when setting list prices or making offers, and research shows that these estimates can influence final sale prices. In other words, market participants generally place trust in Zillow’s algorithmic guidance. But what happens when that trust breaks down? A recent working paper addresses this question.

An exogenous shock to trust

In November 2021, Zillow shut down Zillow Offers, its algorithm-driven “iBuyer” program, which purchased homes with minimal inspection, renovated them, and then resold them. Confronted with an average gross loss of $80,771 per home in Q3 2021, Zillow’s CEO admitted that “the unpredictability in forecasting home prices far exceeds what we anticipated.”

It is reasonable to expect that this news undermined public confidence in Zestimates. The shutdown received widespread coverage in both traditional and social media, and a follow-up survey confirms that it significantly reduced consumers’ trust in Zillow’s valuations.

In standard product-recall settings, diminished trust typically reduces demand, leading to lower prices. On a two-sided platform that provides algorithmic tools, however, a loss of confidence can influence the behavior of both buyers and sellers, even without directly altering overall demand.

In the housing market, increased uncertainty about true market values should lead to greater dispersion in listing prices, resulting in larger absolute gaps—positive or negative—between listing prices and their corresponding Zestimates. Models on the pricing of “one-of-a-kind” goods (such as houses) also predict that sellers would err on the high side when uncertain, thus becoming more likely to set list prices above Zestimates.

In summary, the predictions are that:

1. The absolute gap between listing prices and Zestimates increases after the iBuyer shutdown.

2. Sellers set listings prices above the Zestimate more frequently.

Subscribe now

Testing the predictions

The paper analyzes approximately 28,000 listings in Boston (MA) and Pittsburgh (PA) between June 2019 and May 2022. To identify the causal effect of the shutdown of the iBuyer program—and the associated shock to trust—it compares changes in outcomes in 2021–22 (the treated period) with seasonally matched changes in 2019–20 and 2020–21 (control periods).

The figure below shows estimated monthly changes—relative to the control periods—in the absolute deviation between listing prices and Zestimates, spanning five months before and after the shutdown of the iBuyer program in November 2021 (indicated by the red dashed line). Deviations remain near zero prior to the shutdown but increase noticeably afterward. This pattern supports the hypothesis that diminished trust in Zestimates leads to greater absolute divergence between list prices and Zillow’s valuations.

To give a sense of magnitudes, the absolute difference between listing prices and Zestimates increases by an average of 1.1 percentage points following the iBuyer shutdown. This effect is substantial—it represents a 26.8% increase relative to the average absolute deviation in the pre-shutdown period.

Next, the same approach is used to examine whether home sellers became more likely to set listing prices above or below their Zestimates. Consistent with the paper’s hypothesis and with prior literature, the non-absolute deviation from Zestimates increases by 1.5 percentage points—indicating that sellers, on average, priced their homes higher relative to Zillow’s estimates following the iBuyer shutdown.

What are the consequences for time-on-market and sales prices?

All else equal, higher initial listing prices should slow the arrival of buyer offers, leading to longer time on market, which is generally associated with lower final sale prices.

However, the observed changes in list prices stem from a decline in consumer trust in Zestimates and the resulting increase in pricing uncertainty affecting both buyers and sellers. Because the iBuyer shutdown influenced the expectations of both sides of the market, its impact on time on market and final sale prices is theoretically ambiguous.

The paper finds that the sale-price premium over list prices increased by 0.7 percentage points, while the average time on market declined by nearly 11.2%. Taken together, these results suggest that sellers ultimately benefited from reduced trust in Zestimates and the resulting increase in price uncertainty.

Takeaways

Algorithmic tools are pervasive, and their credibility has real market consequences. The findings in the paper discussed in this post show that when trust in Zillow’s pricing algorithm declined, list-price dispersion increased, sellers raised their asking prices, and—surprisingly—homes sold more quickly and at higher premiums. This suggests that reduced trust in AI can shift market outcomes, even in the absence of changes to underlying supply or demand.

Industrial real estate in Southern California is likely among the commercial real estate segments most exposed to the current set of U.S. tariffs (as of the time of writing). This sector relies heavily on warehouses and logistics operations serving the Ports of Los Angeles and Long Beach, as well as the broader Los Angeles metropolitan area. These are key entry points for Chinese goods into the United States. Although some industries and products are exempt, most Chinese imports are currently subject to tariffs of up to 145%. While these measures may prove to be just temporary, it is reasonable to expect some degree of continued trade friction between China and the U.S., which could dampen import volumes.

Estimating the impact of these shifting conditions on real estate asset values is difficult, particularly in commercial sectors like warehousing, where transactions are infrequent.

However, one of the key players in Southern California’s industrial real estate market is Rexford Industrial Realty Inc. (ticker: REXR), which is a publicly listed Real Estate Investment Trust (REIT). It specialized in holding high-quality assets across the main Southern California markets. Indeed, based on recent financial reporting by Rexford, the entirety of its portfolio of properties is located in Los Angeles County, Orange County, San Diego County, the Inland Empire, and the rest of Southern California; 20.3% of the REIT’s assets are used for transportation and warehousing, 23.5% for manufacturing, and 22.1% for wholesale trade.1

By leveraging publicly available stock price data, information from Rexford’s balance sheet, and a set of reasonable assumptions, we can estimate the decline in the implied value of its assets based on recent fluctuations in its share price.

To start, we plot the evolution of Rexford’s stock price from March 24 to April 18, alongside two benchmark indices. The first is the S&P 500, which reflects the performance of the largest publicly traded U.S. companies. The second is the FTSE NAREIT All REITs Index, which captures the performance of all publicly listed REITs in the United States. To facilitate comparison, we express all series as percentage changes relative to price and index values on April 2, which is the date on which the first set of new sweeping tariffs was announced.

The S&P 500 fell by about 10% in the days following the April 2 announcement. However, it rebounded the following week after most of the harshest duties were suspended. The REIT index followed a more moderate trajectory, with a gradual decline and subsequent recovery.

In contrast, Rexford’s stock experienced a sharp initial drop, a partial rebound, and then another decline, driven by the news that tariffs would remain in place for Chinese imports and by the subsequent escalation of the U.S.–China trade dispute. Since April 10, Rexford’s stock price has been trading at 16-18% below its value on April 2.

This pattern is consistent with Rexford’s high exposure to trade with China. Since the company’s portfolio of operating real estate assets is representative of the industrial sector in Southern California, and in particular of its best and most productive assets, it is reasonable to interpret the decline in Rexford’s valuation as reflecting a broader shift in market expectations about future growth prospects and risk in the entire region’s industrial real estate market.

How, then, can we translate stock price fluctuations into estimates of changes in the underlying value of industrial real estate assets?

The figure below illustrates our approach in an intuitive way, using a simplified depiction of Rexford’s balance sheet. On the left-hand side are the company’s assets, which consist primarily of operating real estate properties, along with other assets such as cash, land, and properties under development. On the right-hand side are the liabilities, comprising both debt and shareholders’ equity. The values of assets and liabilities have to match to satisfy the balance sheet identity.

If we can estimate the value of the company’s equity, debt, and other non-operating assets, we can infer the implied value of its operating real estate assets as:

We can use the formula above to mark-to-market operating real estate assets as stock prices adjust. Indeed, the figure also shows in a stylized way how a drop in the stock price, by reducing the stock market capitalization and thus the equity of the REIT, coincides with a drop in the value of the REIT’s operating assets. In principle, we can then mark-to-market Rexford’s operating real estate as stock prices fluctuate.

For our calculations, we use the simplifying assumption that the values of other assets and liabilities remain relatively stable. While this assumption is admittedly strong, it is acceptable in this context. The value of most other non-real estate assets, such as cash and receivables, are unlikely to be directly affected. Moreover, despite the decline in Rexford’s stock price, the company’s leverage remains manageable, and there is no evidence of a material change in its credit risk since March.

We can then conduct our calculations using data from Rexford’s most recent earnings release, as of April 16. Rexford has approximately 227.4 million fully diluted common shares outstanding. Multiplying this share count by the stock price yields the company’s total market value of common equity. Total debt and liabilities are equal to approximately $4.3 billion.2

Finally, Rexford’s other assets include short-term assets, such as cash, receivables, and other current items, as well as real estate under construction and land held for development. Short-term assets have a book value on the balance sheet of approximately $1.2 billion. Valuing construction in progress and development land is more complex. While these assets are recorded on the balance sheet at $386.7 million based on historical acquisition and construction costs, their market value is likely higher. For example, Green Street Advisors (a real estate intelligence company) valued these assets at $1.3 billion as of the end of 2024. Our calculations below rely on book values for simplicity, though results are similar when using Green Street’s estimates.

The figure below shows fluctuations in the derived market value of Rexford’s operating real estate for each day from March 24 to April 18. We report percentage changes in values with respect to market close on April 2.

The implied value of Rexford’s operating real estate assets declined by 15% following the April 2 tariff announcement, before stabilizing at a level roughly 13% below its pre-announcement value. This pattern suggests that the market has priced in a significant discount on Rexford’s assets in response to heightened tariff-related uncertainty.

Given that Rexford owns some of the highest-quality industrial real estate in Southern California, it is plausible that the discount applied to other, lower-tier industrial assets in the region could be even greater.

While this insight is compelling, it is important to acknowledge its limitations. First, our estimate of operating real estate asset values is derived from market-implied equity valuations. If equity investors are overreacting or underreacting to changes in Rexford’s future prospects, our inference may be biased. As such, the results should be interpreted with caution and viewed as indicative rather than definitive.

Second, this evidence is based on current market expectations. It may be subject to change if there is a sudden breakthrough or shift in policy regarding U.S.–China trade negotiations. Nonetheless, the analysis underscores the extent to which the value of warehouse and logistics assets in Southern California is tied to the trajectory of these negotiations.

Finally, our calculations rely on key assumptions regarding the valuation of non-real estate assets and liabilities on Rexford’s balance sheet. As discussed earlier, real estate under construction and land held for development are difficult to value. Moreover, their market values may also have been affected by the tariff announcements. Similarly, the value of liabilities may have fluctuated in ways we cannot fully account for, as they are not marked to market on a daily basis. However, while these factors may influence the precision of our estimates, they are unlikely to overturn the main conclusion.

First, we examine fire risk exposure estimates that were available prior to the disaster. Specifically, we use data from FEMA’s National Risk Index (NRI), reported at the Census tract level as of March 2023. Following the approach used in a previous post, we convert the annual probabilities of a large fire from the NRI into 10-year probabilities to estimate the likelihood of a major fire over the next decade.1 These estimates represent the publicly available information on fire risk before the Palisades and Eaton fires, which likely shaped both the level of concern among property owners and the countermeasures they took.

Second, we consider the age of the property. Newer structures are generally expected to use better materials and construction technologies, making them more resilient to fire exposure and other risks. In addition, building codes have become increasingly stringent over time. Notably, in 2008, Los Angeles County introduced new regulations for buildings located in Fire Hazard Severity Zones (FHSZs) and Wildland-Urban Interface Fire Areas, requiring fire-resistant construction for new developments and additions. We test whether more recent construction, reflecting advancements in building standards as well as stricter regulations, leads to higher fire resilience.

Overview

Overall, we find that a non-negligible share of the properties affected by the fires were located in areas with very low ex-ante estimated fire risk probabilities.

Interestingly, conditional on being within the fire perimeters, the properties that ex-ante were categorized as low-risk are more likely to have been heavily damaged compared to ex-ante high-risk properties. Although this may appear counterintuitive at first, it may instead reflect that properties in low-risk areas were less prepared to handle a fire event.

We also find that properties within the fire perimeters tend to be relatively old on average, which complicates efforts to isolate the effects of the 2008 fire-resistant building regulations, given that relatively few homes have been constructed in the past 15 years. Nonetheless, we find that homes built within the last 20-30 years proved to be more resilient compared to older buildings.

Ex-Ante Estimated Risk Exposure and Damage

Our analysis focuses on single-family properties, geolocated by CALFIRE as part of its ongoing assessment of property damage,2 and located within the fire perimeters, as defined by shapefiles from the Wildland Fire Interagency Geospatial Services. Using geolocation data, we merge CALFIRE’s damage assessments with parcel-level tax assessment records, allowing us to link each property to its characteristics, including year built. After completing the merge and excluding properties labeled as “inaccessible” by CALFIRE (for which damage information is unavailable) our final sample includes 7,174 single-family homes affected by the Palisades fire and 7,582 affected by the Eaton fire.

First, focusing on NRI fire risk exposures, an important finding is that a non-negligible share of affected properties were located in Census tracts with very low estimated fire probabilities. As shown in the pie chart below, the majority of properties within the fire perimeters were in high-risk areas, defined as Census tracts with 10-year fire probabilities exceeding 10%. However, more than 17% of properties within the fire perimeters were in areas where the estimated 10-year fire probability was below 0.25%.

Of course, this finding does not imply that the probabilities estimated by the NRI are biased. The January fires were extreme events that caused damage deep within urban areas, well beyond what might typically be expected. However, the results do highlight that a substantial share of affected homeowners likely perceived fire risk as a remote possibility, and may therefore have been underprepared for such an event.

Consistent with this observation, the figure below reveals a striking pattern in the extent of damage across areas with different ex-ante fire risk estimates. Properties located in low-risk areas were more likely to suffer severe damage. Specifically, the share of properties experiencing damage that destroyed 50% or more of the structure is 84% in low-risk areas (10-year fire probability < 0.25%), compared to 74% in areas with intermediate risk (0.25%–10%), and 69% in high-risk areas (fire probability > 10%).

Property Age and Damage

We then examine the relationship between property age and fire damage. To begin with, it is important to note that the stock of properties impacted by the fires is relatively old. As shown in the pie chart below, among properties with available year-built data, 77% were constructed before 1980, and only 4% were built after the introduction of stricter fire-resistant building regulations in 2008. While some older homes may have been renovated or expanded in ways that improved their fire resilience, it is reasonable to expect a substantial difference in construction quality and regulatory compliance between homes built 30–40 years ago and those built more recently.

Another important consideration is that the spatial distribution of older and newer properties is not random. As the figures below illustrate, particularly for the Palisades fire, newer construction tends to be clustered in specific areas, which may influence observed patterns of damage and resilience.

Thus, to formally assess the relationship between property age and fire damage, we estimate a simple linear regression model that compares the likelihood of a house experiencing more than 50% structural damage, depending on whether it was built before or after a given cutoff year, and after controlling for property characteristics and location.3

More specifically, to test the effects of property age, we include in the regression a dummy variable equal to one for properties built before a specified cutoff year. The coefficient on this indicator captures the difference in the likelihood of experiencing more than 50% structural damage between properties built before and after the chosen year. If newer homes are more resilient, we would expect this coefficient to be positive and statistically significant.

To control for property characteristics we include dummies for number of bedrooms, bathrooms, and stories, as well as the log of property size (living square feet). To account for micro-location effects, we divide the fire-affected areas into 100-meter grid cells and include fixed effects for each cell. This approach allows us to compare properties of different ages with similar characteristics and within very small geographic areas. This approach helps isolate the specific effect of construction year from other confounding factors.

The figure below presents the estimated coefficients for different choices of the property age dummy, using cutoff years of 1980, 1990, 2000, and 2008, respectively. Along with the point estimates, we also display 95% confidence intervals.

The point estimate for the 2008 cutoff is approximately 3%. This implies that houses built before 2008 are 3% more likely to experience structural damage exceeding 50%, compared to those built afterward. However, this difference is not statistically distinguishable from zero. The fact that the estimate is imprecise should not be taken as evidence that the 2008 building regulations were ineffective. Rather, it reflects the limited number of homes within the fire perimeters that were built after 2008, as discussed above. This reduces the precision of the estimate.

Indeed, we find larger and statistically significant estimates when using 2000 as the cutoff year, which gives us a larger sample of homes built after the cutoff. For homes built before 2000, the likelihood of experiencing structural damage of 50% or more is 5.5 percentage points higher than for those built after 2000. Results are similar when using 1990 as the cutoff year. This difference is not negligible, given that, on average, 73% of properties in the perimeters suffered damage exceeding 50%. Thus, the estimated effect represents a relative reduction of approximately 7.5% (5.5/73). This evidence is consistent with the notion that more recent construction offers greater resilience to fire damage.

When we shift the cutoff year to 1980 or 1970, we continue to find that more recently built properties are less likely to suffer structural damage exceeding 50%. However, the estimated difference declines to 3% for the 1980 cutoff and becomes small and statistically insignificant for 1970. This pattern is sensible: as the cutoff year moves further back, the group of homes built after the cutoff increasingly includes old structures that could not have greatly benefited from improvements in materials and the introduction of stricter building codes.

3

We estimate:

where the dependent variable is a dummy equal to one if affected property i in micro-location l suffered damage exceeding 50% of the structure, the main variable of interest is a dummy equal to one if the property was built after year Y. The specification includes a vector of property characteristics, and micro-location fixed effects.

Our analysis explores the distribution of risk across single-family homes, examines the characteristics and values of risk-exposed properties, and provides estimates of individual homeowners’ wealth exposed to fire risk. We present three key takeaways:

1. A significant number of single-family homes in the county has non-negligible exposure to fire risk. The total number of at-risk homes is close to the total number of single-family homes in Phoenix.

2. These at-risk single-family homes are generally larger and more expensive than the average home in the county. At current market prices, their combined value reaches hundreds of billions of dollars.

3. A rough calculation of home equity suggests that a typical homeowner in a fire-prone area, after 10 years of ownership, has approximately $1 million in equity at risk. Policies aimed at fire mitigation or addressing underinsurance are therefore crucial for protecting homeowners' wealth.

Fire Risk in the County

Our study relies on fire risk exposures from FEMA’s National Risk Index (NRI) Data, reported at the Census tract level as of March 2023. We combine this information with data on residential parcels and with publicly available house price indices from Zillow.

The figure below shows the probability of a large fire occurring over a 10-year period across all Census tracts in Los Angeles County. These 10-year probabilities are based on the annual large fire probabilities made available in the NRI Data.1 While a decade may seem a long time period, it is close to the average tenure of an homeowner in the US.

Share

The map shows that the hilly regions to the west and north of the main metropolitan area have 10-year fire probabilities exceeding 10% (dark red areas). Moreover, vast areas of the county fall in the 5%-10% probability range (red areas), and in the 1%-5% probability range (orange areas).

The figure below shows the proportion of single-family homes in the county located in areas with different levels of fire risk exposure. We identify single-family units and their locations using residential parcel data from December 2022. Note that these calculations, along with the rest of the analysis in this post, include the properties affected by the January 2025 fires.

The vast majority (78.4%) of single-family units in the county have a negligible chance of being exposed to a large fire (below 0.25% over 10 years). However, approximately 21.6% of single-family units in the county, or 309,805 homes, are in areas with non-negligible risk. For context, this is close to the number of single-family homes in Phoenix (approximately equal to 320,000). Among these at-risk homes, 56,382 have probability of exposure to a large fire between 5% and 10%, and 92,752 have probability greater than 10%.

Share

Single-Family Homes in Fire-Risk Areas

When we examine the characteristics of the single-family units in at-risk tracts, we find that the majority are the primary residences of their owners. Among the units for which we observe occupancy status, more than 85% are owner-occupied.

Single-family homes in at-risk areas are on average larger and more expensive than those in areas with negligible risk. The top panel of the figure below shows the average sizes of single-family homes categorized by fire risk. While the average size in areas with negligible risk is 1,660 square feet, the average size in fire-risk areas is 40% larger, and exceeds 2,300 square feet. These size differences can be attributed to the suburban nature of the fire-risk areas.

The second panel of the figure compares home prices across areas with different levels of fire risk. This calculation is an approximation based on estimates of zip code-level median single-family home prices available from Zillow, as of December 2024. We match each property with the corresponding zip code index and calculate averages across all properties in areas with different levels of fire risk. The result is a weighted average of the zip code-level price indices.

The figure shows that in the areas with negligible fire risk, the average price per single-family home is approximately $1.05 Million. Prices are higher in fire-risk areas, peaking at an average of $1.5 Million in areas with a fire probability between 5% and 10%. Average prices are equal to $1.17 Million in areas with a fire probability greater than 10%.

Taking the price estimates at face value and multiplying by the total number of properties in each different fire risk area, our estimates imply that approximately $190 billion of housing value is exposed to a fire probability greater than 5% over the next 10 years, and $108 billion is exposed to fire probability greater than 10%. These staggering magnitudes highlight the potential devastating impact of future fires on the county, particularly if fire risk is underinsured.

Subscribe now

Implications for Homeowners’ Wealth

Our final set of calculations focuses on quantifying the wealth exposed to fire risk per homeowner. While the home price estimates mentioned above provide a good indication, it is important to consider that homes are typically owned with a mortgage, and that California is a non-recourse state. If a home is impacted by fire, and its value falls below the outstanding mortgage balance, a homeowner can walk away from the mortgage.

The amount of individual household wealth exposed to fire risk is then determined by the amount of equity the household owns in the property. Home equity consists of three components: price appreciation since purchase, the initial down payment, and mortgage principal repayment.

We take the perspective of a homeowner who purchased a single-family home in January 2015, with 85% of the purchase price financed by a fixed-rate mortgage, as is common in the US. The homeowner had an initial mortgage rate at origination of 3.67% (the prevailing rate at the time) and refinanced into a lower rate of 2.68% in December 2020.2 The homeowner bought a house of average price within a fire-risk area and experienced the average price increase in that area between January 2015 and December 2024. We believe these calculations provide a reliable proxy for the home equity of a representative homeowner who did not increase her borrowing with a cash-out refinance during the 10-year tenure period from January 2015 to December 2024.

In the top panel of the figure below, we display the average price appreciation for homes in areas with different levels of fire risk across the county. As noted earlier, this is just one component of home equity. We estimate values ranging from approximately $700,000 to $550,000, with the lowest value realized in the areas with the highest levels of fire risk. These large price appreciations reflect the strong growth that the county has experienced over the last decade. Between 2015 and 2024, home prices across the county increased by 93%.3

Then, in the bottom panel of the figure below, we calculate representative homeowners’ total equity in each risk category, including the initial down payment and mortgage principal repayments. The results show values ranging from $1.15 Million to $925,000. Although the lowest value is observed in the area belonging to the highest fire risk category, overall home equity values are comparable across all fire-risk areas, and equal to about $1 Million.

Thus, the typical homeowner in these areas has substantial wealth exposed to wildfire risk. This has two important implications. First, policies aimed at fire mitigation, and at increasing fire insurance coverage are tremendously valuable to individual homeowners in the at-risk areas.

Second, any decline in house prices in fire-exposed areas can substantially deplete homeowners’ wealth. Price declines might be driven by lower housing demand and relocations (households may become less willing to live in at-risk areas after the recent wildfires). However, these effects may take time to materialize. Housing is undersupplied in the entire county, and if the rebuilding effort following the Eaton and Palisades fires progresses slowly, it could force many impacted households to relocate, putting additional strain on housing production in the entire county, and increasing housing demand in certain locations. This may lead to higher prices in the next months and years (see our previous post).

Subscribe now

The fires are still ongoing as we write, so our estimates are preliminary; however, we believe they provide valuable insights into the scale of the impact of this natural disaster.

For our analysis, we use data on residential land parcels and corresponding properties, fire perimeters provided by the Wildland Fire Interagency Geospatial Services (WFIGS) as of January 10, and building permit information from the U.S. Department of Housing and Urban Development (HUD).

The figures below display the fire perimeters in pink, with colored dots representing different types of properties within these areas. As highlighted in news coverage, not all of these properties were destroyed or suffered catastrophic damage from the fires. However, they are all located in neighborhoods directly affected by the fires, meaning residents are likely to face displacement, at least temporarily.

Consistent with the suburban character of both areas, the figures clearly show that single-family homes are the primary property type impacted by both fires. A detailed breakdown, displayed in the figures below, reveals that more than 78% of residential properties within the perimeter of the Palisades Fire are single-family homes. This share is even higher for the Eaton Fire, where nearly 84% of residential properties within the perimeter are single-family homes. Overall, based on our estimates, the perimeters of the two fires combined encompass 15,032 single-family homes and 18,616 residential structures in total, including condos, multi-family buildings, and duplexes.

The estimates above highlight the astonishing scale of this crisis. They also raise critical questions about how the housing markets in Los Angeles County will respond to the supply and demand imbalances created by the large number of potentially displaced households.

Los Angeles County is a massive urban area with a population exceeding 9.6 million. However, significant challenges limit its ability to absorb this shock.

In the medium and long term, the key challenge lies in the ability to adjust housing supply by rebuilding destroyed homes or increasing supply in other areas. However, despite a more than 150% increase in house prices since 2012, the construction of new housing units has been severely limited across the county. This is due to both geographic and regulatory factors, which complicate efforts to produce new housing supply.

The figure below illustrates how this is relevant with respect to the effects of the fires. On the left, we show the combined total number of structures within the perimeters, affected to varying degrees by the fires. As mentioned earlier, these are 15,032 single-family homes and 18,616 residential structures in total. Using data on new construction permits from HUD, we find that in Los Angeles County, on average, between 2012 and 2022, only 5,875 new single-family homes were permitted annually. This accounts for just 39% of the single-family homes impacted by the fires. When considering all residential structures, the total number of permits per year raises to 20,654. However, the number of structures impacted by the Palisades and Eaton Fires is alarmingly close to the total number of structures permitted in the entire county in a year.

When we restrict our analysis to permits issued in the cities near the affected areas (Glendale, Pasadena, Arcadia, Sierra Madre, Monrovia, Santa Monica, Malibu, Beverly Hills, West Hollywood, and Manhattan Beach), the number of housing units destroyed by the fire far exceeds annual new housing construction. Between 2012 and 2022, these cities combined permitted just 449 new single-family homes and 1,684 new residential structures per year.

Of course, part of the reason new housing construction has been so limited in the county is the scarcity of available land. In principle, this is not a constraint in the areas affected by the fire, where homes can be rebuilt on the same land. However, the figures above show that housing production in the county needs to step up substantially in order to meet this new need quickly.

The situation may be even worse if residents impacted by the fires choose to relocate to nearby cities in order to reduce their exposure to future fire risk. If this is the case, given the historically low levels of new housing construction displayed above, it may not be possible to adjust the housing supply in the new destination cities to accommodate the increase in demand. This would then trigger large price increases.

Overall, the fires may result either in substantial house price increases in certain areas of the county or in the relocation of households from the affected areas to locations outside the county.

Even in the short term, a crucial issue is the hot rental market in the county. Many displaced residents will have to temporarily relocate to rental units while basic services are restored to the affected areas and houses are rebuilt. Although the county has lost approximately 4% of its population since 2020,[1] rents for new listings have increased from 2020 to 2023 by around 15%, as shown in a recent study. The Zillow Observed Rent Index shows that the median monthly rent for the county is above $2,800. Either due to the sudden increase in demand or to potential price-gauging, rents could rise even further in the immediate aftermath of the disaster. As a result, even temporary relocation is likely to be extremely burdensome for affected households and may result in increases in housing costs across the county.

[1] Cities located close to the fire perimeters (e.g., Pasadena, Arcadia, and Glendale for the Eaton fire, and Santa Monica for the Palisades fire) have experienced similar decreases in population, ranging from 3% to 4.5%.

In this post, we examine findings from a recent study investigating the use of rental properties as retirement investments and the motivations behind retirees becoming landlords. Drawing on Australian tax filings data, the study documents a large increase in the share of retirees owning rental properties, taking place over the period between 2006 and 2019. Over this period, the share of middle-income retirees earning rental income increased from 12% to 20%. To uncover the motivations for this trend, the study employs analyses based on fiscal and transaction data, complemented by insights from two surveys targeting landlords who purchased properties between 2006 and 2019.

The key explanation that emerges from the analysis is that retirees depend on investment income to support their consumption needs, and favor assets paying regular income streams over realized capital gains. Rental properties are appealing because they offer recurring income payments. This appeal is amplified when bond yields and interest rates are historically low or declining, as during the period covered by the study.

The survey evidence presented in the paper is particularly striking. The first survey was distributed to members of the Australian Landlords Association (ALA), a group representing small, individual landlords who are Australian residents. An open-ended question at the beginning of the survey invited respondents to describe the motivations behind their decision to purchase a rental property.

The figure below summarizes keywords from survey responses using word clouds, where font size and color represent word frequency. The word cloud on the left displays keywords for individuals of retirement age (60 or older), while the word cloud on the right represents keywords for younger individuals. Among individuals of retirement age, “income” and “retirement” are the most frequent terms, appearing three times more frequently than the third most frequent word, “rental”. In contrast, for non-retirees, “income” is overshadowed by several other terms, including “capital” and “growth.”

These patterns are confirmed by responses to close-ended questions, in which participants are asked to assign scores to various motivations for purchasing their rental properties. For retirees, the prospect of earning recurring rental income ranks highest among the competing motivations, including capital gains. Interestingly, the second highest ranked motivation for retirees is the perception that returns on saving accounts are too low.

Subscribe now

The second survey delves deeper into how retirees utilize rental income to support their consumption. It was conducted through Qualtrics and targeted a proprietary panel of landlords in retirement age (60 or older) who purchased rental properties between 2006 and 2019. The figure below presents the distribution of responses to the main survey questions.

The first question asks if rental income is a convenient way to pay for consumption in retirement. Respondents rated their agreement on a scale from 1 to 5, with 65% selecting 4 or 5. The second question asks what is most appealing about rental income. Participants are allowed to choose up to two options. The top responses are “Monthly frequency of payments” and “Safety and reliability of payments.”

The survey also asks about sources of investment income before purchasing the rental property. The most frequently selected option is “Savings accounts”. Lastly, respondents are asked how they funded the down payment for the property. The predominant answer is “Cash from saving accounts,” selected by 48% of participants, while all other options received less than 10% of responses each.

In summary, retirees are drawn to rental property investments primarily for the steady income they provide. Rental income is used to support consumption needs, and its monthly frequency is a convenient fit for budgeting. Before becoming landlords, retirees typically relied on saving accounts as their main source of investment income. They then used the funds in their saving accounts to finance the down payments on their rental properties.

Thus, retirees view rental properties as a form of annuity or bond, generating periodic income payments that help cover their consumption expenses. This is consistent with common perceptions of rental property investments and with narratives about “passive income.”

However, it remains unclear whether this common wisdom holds up from a portfolio or money management perspective. Rental property investments come with risks and frictions that do not affect more liquid assets like annuities and bonds. For one, rental properties are lumpy (meaning that a significant amount of capital is tied up in a single property), and they are also illiquid, as they are not easy to trade.

Moreover, rental income is not a risk-free cashflow. Owning and renting a property is similar to running a business; it requires ongoing monitoring, significant time, and further capital investments. For example, dealing with a problematic tenant could disrupt income collection, and issues like corrosion in the plumbing system might necessitate costly repairs. While experienced and skilled landlords may be able to manage their properties effectively and minimize risks, inexperienced landlords are more likely to encounter losses or even legal disputes.

In conclusion, while small investors are common in rental markets, we still have limited understanding of the motivations driving the decision to become landlords and of the role that rental properties play in household portfolios. This paper sheds light on some of the key factors that drive mom-and-pop investors to turn to real estate.

[1] Badarinza, C., J. Y. Campbell, and T. Ramadorai, 2016, “International comparative household finance’’, Annual Review of Economics, 8, pages 111–144.

[2] Garriga, C., P. Gete, and A. Tsouderou, 2023, “The Economic Effects of Real Estate Investors”, Real Estate Economics, 51(3), pages 655-685.