Dr. Stace Sirmans

The surprise

During the pandemic, remote work pulled housing demand away from expensive coastal job centers toward cheaper, easier-to-build places. The natural guess was that generative AI, which loosens the tie between high-skill work and location even further, would push that dispersion along.

The data went the other way. After ChatGPT’s late-2022 release, the metros most exposed to AI didn’t keep falling behind; they pulled ahead. The premium shows up after 2023 (not during the remote-work years), in both Zillow and FHFA price indices, and in rents, and it is strongest exactly where new housing is hardest to build.

Why: you can’t build there

Why would AI raise home prices instead of just wages? Because where housing supply can’t expand, extra demand capitalizes into land and prices rather than into new construction. The premium is a local demand shock landing on scarce housing, and it lives almost entirely in supply-constrained metros.

The evidence stacks up

Why REITs need their own factors

REITs are not small industrial firms. Their payout requirement limits the retained-earnings channel behind standard profitability and investment factors, and their balance sheets, lease structures, and property-market exposure differ from the firms used to build broad equity factors. Models calibrated on the general stock market can leave REIT-specific return patterns unexplained, which is why a sector-specific benchmark is worth testing.

Not just the stock factors

A natural worry is that REIT factors simply mirror their stock-market twins. They do not. The REIT versions of momentum, quality, and reversal have higher Sharpe ratios than analogous stock factors and earn statistically significant alpha after controlling for those stock versions. That points to return patterns specific to listed real estate, not just imported equity anomalies.

The evidence stacks up

Why REITs are a clean test

REITs make a sharp laboratory for pricing AI. They monetize it through operations: setting rents, converting leads, screening residents, underwriting deals, and managing collections, not through products or patents. The 90% payout rule shuts off the internal research budgets that drive AI value at technology firms, so AI matters only if it changes the routines behind recurring cash flow. That lets us ask one question cleanly: can markets tell real operating AI from AI talk?

What the market pays for

We sort the AI language in filings and earnings calls into internal operating use, generic AI commentary, or tenant demand. Only internal operating use moves prices. Disclosing fully deployed internal AI raises the three-day earnings-call return by about a percentage point; generic AI language, the kind any firm can write, draws no significant response.

Talent is what makes it credible

The reaction is not automatic. It appears only for REITs that had already hired AI-skilled workers before the call, the costly input that should accompany a real capability claim. And it is specific to that input: replace prior AI hiring with general IT and analytics staffing and the effect disappears.

The evidence stacks up

The biases are real, and large

Across the board, a single framing change, anchor, or social cue moves the model’s answer by amounts that would change a real decision. Reframing risk in loss versus gain terms is the strongest: it shifts a 1-to-10 rating by 1.62 points on average, enough to move a call from below the midpoint to above it. Social and narrative cues, such as authority, herding, representativeness, and availability, are nearly as large.

Smarter is not de-biased

More capable models are not simply cleaner thinkers. Higher intelligence reduces some distortions, like framing and representativeness, but it increases others, including sunk-cost thinking, loss aversion, and disposition-style asymmetries. Capability reshapes the bias profile rather than removing it.

The evidence stacks up

Insurance early, encumbrance later

Extending a lease does two opposite things. It insures the owner against near-term rollover and vacancy, which is valuable and pushes the cap rate down. It also encumbers the owner with decades of exposure to obsolescence, climate adaptation, regulation, and shifting demand, which is a liability. The first lease years are mostly insurance; the far-out years are mostly encumbrance. So the price of duration should bend, not run in a straight line.

What bends the curve

The same trade-off predicts who gets a steeper or flatter curve. Anything that lowers rollover risk, like a reliable tenant or a tight market, makes the insurance more valuable and steepens the compression. Anything that raises long-horizon risk, like climate exposure or an aging building, makes the encumbrance bite sooner and flattens it.

The evidence stacks up

Same firms, two crowds

Corporate bonds and stocks are claims on the same companies, but bonds are held mostly by ratings-mandated institutions, insurers and pension funds, that reprice on common credit conditions. Their flows imprint on the systematic, factor-level part of bond returns, not on any single bond. Equity investors weigh a broader, more firm-specific information set, so the credit signal reaches stock prices with a lag.

Information, not a risk premium

The spillover behaves like slowly diffusing information, not a fixed risk premium. It is far larger when markets are stressed or pessimistic, and it even reverses when sentiment runs high. The effect is also stronger where the bond and equity investor bases overlap less, the opposite of what a shared-clientele risk story would predict.

The evidence stacks up

Two cautionary tales

Two episodes frame the field. In 2021, Zillow shut down its iBuyer arm after $881 million in write-downs, even though its price model was strong by industry standards, with a median error near 5%. A good prediction is not the same as a good market outcome. In 2025, the Department of Justice and ten states settled with RealPage, whose rental-pricing software was estimated to add about $25 a month across roughly 4.2 million units by helping landlords coordinate. The same efficiency that helps can also redistribute surplus away from households.

An equity-efficiency map

We separate two questions the literature often blurs: does AI improve the task, and does deploying it change who wins and who loses? Mapping the evidence across residential domains and AI mechanisms shows the research clusters in automated valuation and mortgage credit, while rental pricing, leasing automation, tenant screening, property management, and agentic workflows stay lightly studied, exactly where deployment is moving fastest.

The evidence stacks up

Four frictions, three steps

Commercial real estate is defined by four frictions that liquid markets lack: opacity (value hides in leases, memoranda, and zoning), search costs (there is no central exchange), illiquidity (stale prices and smoothed appraisals), and heterogeneity (every asset is unique). We test AI against each in three steps: does it reduce the friction, and does that reduction actually show up in outcomes like tighter spreads, faster sales, or smaller forecast errors? The literature clears the first hurdle and stalls at the second.

The bimodal wall

Adoption is sharply uneven. AI splits tasks by type, not difficulty: information-processing work (pulling listings, abstracting leases, answering client questions) shows heavy use, while tasks that need physical presence, interpersonal contact, or legal judgment sit at zero. And it engages the highest-importance tasks, not the most routine: in 12 of 15 occupations, the tasks AI already does are more important than the ones it leaves alone. Those important, still-human tasks define the agentic frontier.

The evidence stacks up

Access is the wrong question

The evidence on AI at work looks contradictory: large gains in some studies, nothing in others, losses in a few. That is what you get when research measures access or adoption, a binary “did they have AI,” instead of the priced quantity of inference actually applied to a task. Generative AI is different from earlier workplace tools: every inference is metered, priced, and logged against a worker and a task, so for the first time we can measure intensity, not just availability.

Four regimes hidden in one number

A binary “uses AI or not” pools four very different situations: a worker who uses AI well, one who rightly abstains, one forced to overuse it, and one held back from using it. Access and adoption estimates are a mixture of these, so their sign tells you little about whether AI is valuable. The fix is to measure the priced inference at the chosen margin.

The evidence stacks up

Why belief moves the price

A property’s cap rate is roughly its discount rate minus its expected income growth. An investor who believes climate change will worsen sees both a higher discount rate (riskier cash flows and costlier capital) and slower income growth (rising insurance, repair, and compliance costs). Both push the cap rate up and the price down, so the same hazard is discounted more wherever investors take it seriously.

The more present climate change feels, the harder it bites

Within high-belief areas, the discount grows as climate change becomes more salient. Where residents expect more personal harm from it, and where they discuss it often or see heavy media coverage, cap rates respond even more sharply to the same underlying hazard.

The evidence stacks up

Belief works through demand

Buyers who take climate change seriously discount homes in risky areas and avoid moving there, while skeptics are more willing to move in and prop up demand. Builders, anticipating weaker demand, pull back too. So where belief is high, the same disaster risk pushes price growth down harder, even in places not facing an imminent physical threat.

Fear of harm bites harder

General belief is not the whole story. The drag is sharpest where residents expect climate change to cause real harm, and sharper still where they fear harm to themselves. Concern about personal vulnerability moves housing markets more than abstract agreement that climate change is happening.

The evidence stacks up

Where the citations went

Over five years, finance’s most-cited work clustered around sustainability above all. ESG and climate finance account for 35 of the top 100 papers, more than any other theme, with the rest spread across banking and intermediation, FinTech and crypto, the pandemic, corporate governance, retail investors, machine learning, and econometric methods. The top 20 papers alone draw just over half of all citations to the top 100.

Volume is not impact

Counts and influence are not the same thing. Econometric methods is the smallest category, just four papers, yet those four earn nearly 9% of all citations, and the average methods paper gains about 217 citations a year, more than ESG or AI. The single most-cited finance paper of the period is a critique of difference-in-differences. Quietly, the field has been auditing its own tools.

The evidence stacks up

Insurance is a cost that gets capitalized

Like property taxes, a higher recurring insurance bill lowers what buyers will pay. A one-point rise in the premium-to-income ratio is associated with about 0.25 percentage points less annual appreciation; moving across the interquartile range (1.5 points) implies about 0.37 points less, the same order as the effect of property taxes. The link holds both before and after the pandemic.

And the bite is biggest where risk is lowest

The pattern runs opposite to a simple physical-risk story. The effect is stronger in low-FEMA-risk counties than high-risk ones, and stronger where insurer competition is limited. Where insurance stress is new, a premium increase carries more news: it tightens mortgage qualification and updates expectations about future cost and insurability, even before any loss occurs.

The evidence stacks up

Cap rates are tied to the capital markets

A cap rate is the price of a property’s income: a lower cap rate means investors pay more per dollar of rent. The paper shows these prices are not set in isolation. From 2005 to 2019, the net-lease cap rate rose and fell with corporate bond spreads and the S&P 500 earnings yield, moved with retail sales growth, and fell when investor sentiment ran high. Broad macroeconomic indicators, by contrast, did little.

The evidence stacks up

Disagreement predicts weaker returns, not higher ones

One school of thought holds that disagreement is a priced risk: investors who buy into a contested stock should be paid extra for the chance of being wrong. The opposing view, going back to Miller (1977), is that when the most optimistic buyers set the price and pessimists cannot easily short, prices drift too high and future returns fall. The REIT setting lets us referee this debate using analysts’ valuation estimates rather than next-quarter earnings.

Sorting REITs each month into three equal groups by NAV dispersion, the low-dispersion group earns 1.27% per month (t = 2.26) while the high-dispersion group earns just 0.67% per month (t = 0.94). The low-minus-high spread of 0.61% per month (t = 2.24) is statistically significant and survives the Fama-French six-factor model as an alpha of 0.65% per month (t = 2.24). Disagreement on value lines up with overpricing, not with a risk premium.

Where the disagreement comes from

Why do analysts disagree more about some REITs than others? We trace it to the difficulty of valuing the underlying property portfolio. NAV dispersion rises with the number of properties a REIT holds, the number of metropolitan markets those properties span, and how physically far-flung the portfolio is. Owning buildings outside the United States carries the largest effect of all. Concentrating in smaller or non-gateway markets, by contrast, does not move dispersion, and neither does spreading across property types.

The evidence stacks up

Splitting the premium into fundamentals and mood

The NAV premium is the percentage price-to-NAV ratio: how much the stock market pays for a REIT relative to the appraised value of its properties. The authors regress that premium on a large set of firm characteristics. Size, market share, and asset growth push the premium up; profitability, leverage, and interest coverage push it down. Together these characteristics explain about 56% of the variation, which means more than 40% is left unexplained.

That unexplained residual is the key object. It is transient and fast-moving, the signature of investor sentiment rather than slow-moving fundamentals. When the authors use each piece to predict the next month’s return, the characteristics-driven part is a weak positive predictor (not significant), while the sentiment-driven residual is strongly negatively related to returns. In other words, the well-known negative link between NAV premiums and returns is coming from mood, not fundamentals.

The evidence stacks up

Momentum is strongest where ratings matter most

A CDS momentum strategy that buys recent winners and sells recent losers earns 0.59% per month on average. The effect is not uniform: it is small for high-grade firms, where ratings rarely move, and large for junk-grade firms and for the most actively quoted contracts, where the CDS market has the sharpest informational edge over the rating agencies.

The signal jumps the fence into bonds and stocks

The information in past CDS returns does not stay in the CDS market. Sorting a firm’s bonds or stocks on the firm’s own past CDS performance, while holding past bond or stock performance fixed, produces significant long-short returns in those markets too. Past CDS returns better anticipate upcoming rating changes than past bond or stock returns do.

The evidence stacks up

Coverage tracks value; disagreement tracks risk

Two analyst signals point in opposite directions. The number of analysts estimating a REIT’s NAV behaves like a quality marker: it rises with firm value and falls with volatility. NAV coverage carries a positive, significant coefficient on Tobin’s q of +0.96 (t = 6.58) in the baseline, holding at +0.76 (t = 4.31) with size, leverage, return on assets, volatility and property-type fixed effects in the model, while Ln(Volatility) enters at −0.45 (t = −3.62). The spread of those same estimates is the opposite signal. NAV dispersion is highest for leveraged, volatile REITs: leverage enters its regression at +0.88 (t = 4.90), return volatility at +0.45 (t = 4.86), and the idiosyncratic share of risk at +0.47 (t = 2.98). When analysts cannot agree on what the buildings are worth, the firm is the riskier, more levered kind.

The evidence stacks up

Effort raises the price, but not the commission rate

Using virtual tours to stand in for observable effort, the 2SLS system shows a clear and consistent pattern. Homes marketed with a tour list for about 3% more (coefficient 0.030, t = 3.21) and sell for about 2.5% more (coefficient 0.025, t = 4.28). They also stay on the market about 17.6 days longer (t = 5.55). The one thing effort does not move is the price of brokerage itself: the contracted commission rate is statistically unaffected by the presence of a tour (coefficient 0.008, t = 0.45). The competitive brokerage market does not let agents bill the cost of a tour back to the seller through a higher rate.

Why a value-creating tool stalls

Because the rate will not move, the agent recovers the cost of a tour only through the larger dollar commission that a higher sale price generates. The authors work the arithmetic at larger firms, where the effect is strongest. A 4.30% selling-price gain on the median sale price of $233,303 is about $10,032 of extra commissionable price. At a 3% rate that is roughly $300 to the firm, and a top-producing agent keeping 75% nets about $225, less than the roughly $250 cost of the average tour. A tool that demonstrably helps the seller can still be unprofitable for the agent who must pay for it, which is the paper’s limit-to-innovation: a transaction-level agency conflict that can keep a beneficial innovation from becoming standard.

The evidence stacks up

Five forces, one map

The central claim is organizational: REIT returns are not driven by a single factor but by five recurring forces, and the literature is most useful when read as a whole. Valuation studies tie returns to fundamentals like funds from operations and cash flow news. Volatility studies show equity REITs swing harder than stocks or bonds. Asset-growth studies link returns to how managers expand, finance, and diversify, and document intra-industry REIT momentum of about 1.33% per month after 1990. Leverage studies trace how debt amplifies both returns and crisis losses. Sentiment studies show that non-fundamental investor mood moves REIT prices, returns, and the timing of offerings. Read together, fundamentals set the level and sentiment plus leverage set the swings.

What the evidence says in each theme

Pictures price into cap rates

Adding the 54 image features to a model that already absorbs the asset format, the market, the lease and ownership structure, and the time of sale lifts the explained variance in log cap rates from 0.567 to 0.595. That gain is highly jointly significant and closes about 6.5% of the residual variance the controls leave behind. The contribution is variation within property type, not a vision model simply learning that a pharmacy looks different from a warehouse.

The useful layer is the one you can name

The paper separates interpretable image content from latent content and asks which carries the signal. Named CLIP scores plus a few engineered summaries account for almost all of it. The 20 principal components of DINOv2 embeddings, the part of image AI that resists explanation, add only 0.005 more once the nameable cues are in the model. That latent increment is statistically significant but economically minor.

The evidence stacks up