## <span class='transparent70'> AI and Real Estate</span> ## <span class='subtitle transparent70'>—Data, Appplications and Challenges— </span> <br> ## <span class='subtitle transparent70'>Thies Lindenthal</span> ## <span class='subtitle transparent70'>Grosvenor Professor of Real Estate Finance<br>Department of Land Economy / University of Cambridge</span>

## The big visions 1. AI changes the way we work—and that shifts the demand for real estate. 2. AI makes buildings better. 3. AI reduces frictions in real estate markets and investments. Real Estate becomes a more mainstream investment. 4. We witness an AI infrastructure bubble. 5. No, humans will not be replaced by robots: real intelligence, creativity, curiosity and emotions will prevail.

## <br><br><br><br><br><br><br> <mark></mark> <span class="subtitle"><br><br/><mark>*"If you have visions, you should see a doctor."* <br>Helmut Schmidt</mark></span>

## <mark>So much research!</mark> <span class="subtitle"><br><br/><mark>85K+ papers on Google Scholar on "ML and Real Estate". A success story?<br>Let me share 3 views from an academic perspective. </mark></span>

## <mark>1. AI gets us closer to what matters.</mark> <span class="subtitle"><br/><br/><br/><br/><mark>People and buildings make real estate so interesting. <br>AI helps us grasp both better.</mark></span>

## Rich data are in reach...<span class="subtitle"><br/>So much to learn from looking at buildings, spaces, people... </span> <iframe data-src="https://www.google.com/maps/embed?pb=!4v1615400218848!6m8!1m7!1sfrGO3071bDjlMFWKCuPROg!2m2!1d52.20218228800385!2d0.116240823654906!3f214.30357533449632!4f12.499881249341925!5f0.7820865974627469" data-preload width="900" height="500"></iframe> Note:

<h2>... powerful algorithms too!<span class="subtitle"><br />Computer vision + ML classification, all available in easy to use libraries.<br /></span></h2> <img src="imgs/research_feature_vectors.png" height=400> * Examples: <a href='https://www.mit.edu/~naik/assets/pdf/naik_etal_aer_16.pdf'>Naik et al. (2016)</a>, <a href='https://link.springer.com/article/10.1007/s11146-021-09845-1'>Lindenthal & Johnson (2021)</a>

## Global scale, low cost <span class="subtitle darkonlight"><br/>The world's surface, at high resolution, over time (and it's open source!)<br/></span> * <a href="https://github.com/ucam-eo/tessera">Tessera (2025)</a> offers pre-processed feature vectors based on satellite imagery (incl. radar), at 10 m² resolution, going back ~10 years <img src="imgs/tessera.svg" height="400" />

## <span class="transparent70">A closer look at us humans</span><span class=" transparent70 subtitle"><br/>What is it that people pay attention to when looking at houses?</span> * <span class="transparent70">Project with Carolin Höltken & Wayne Wan</span>

## Black boxes are problematic<span class="subtitle"><br/>Are our systems unbiased and lawful? ML systems need testing! </span> * Example: System testing for ML classifiers (<a href="https://doi.org/10.1111/1540-6229.12416">Wan & Lindenthal, 2022)</a>. What do systems see? <div class='figure'><img src="imgs/house_cars.jpg" height=400></div>

## More low hanging fruits<span class='subtitle'><br>High-dimensional data where functional form is not evident very suitable for ML</span> * Example: REIT returns are partially predictable with ML (<a href="">Leow & Lindenthal, REE 2025</a>). Traditional regressions fail. <div class='figure'> <img src='imgs/fig-2a.png' height=400> </div>

## <mark>The costs continue to fall</mark><span class="subtitle"><br><mark>Example: <a href='https://github.com/Mozilla-Ocho/llamafile'>Locally hosted LLM model</a></mark></span>

## The best hunting grounds<span class="subtitle"><br/>Where can ML offer value over traditional econometrics?</span> <div class='figure'> <img src="imgs/complexity-figure.png" height="400"> <p class='source'>Source: <a href="">Geltner et. al (2026)</a></p> </div> * Buildings, cities and people don't fit into spreadsheets well. The real world is a high-dimensional network.

## <mark>2. ML is great for thin data!</mark><span class="subtitle"><br/><mark>Making sense of (re)development in Phoenix, AZ</mark></span> <br><br><br><br> Notes: New research with John Clapp and Joe Nichols. Three interrelated research questions 1. <span class="transparent70">How to estimate land values in urban areas?</span> 2. <span class="transparent70">Will improved land value estimates enable better automatic valuations<br>and risk assessments?</span> 3. <span class="transparent70">Where will new supply in cities emerge?</span>

## Property is a bundled good <span class="subtitle"><br/>Property Value = DCF of rents from structure at location + Option Value * If structure has been developed to highest and best use - Option value = 0 - Land Value = Property value - replacement cost of structure * If structure is not at HBU - Land value can be encumbered: current structure impairs option value * <span class="transparent70">For background, see <a href="https://link.springer.com/article/10.1007/s11146-021-09834-4">Clapp & Lindenthal (JHE, 2022)</a> and <a href="https://doi.org/10.1016/j.jhe.2022.101872">Clapp, Cohen & Lindenthal (JREFE, 2021)</a>.</span> <img src="imgs/land-structure.svg"> <br><br><br><br>

## Intense dynamics<span class="subtitle"><br/>Maricopa County is ideal for studying house and land values throughout the cycle!</span> <div class="figure"> <p class="title">House Price Indices, Maricopa County: 2007–2018</p> <img src="imgs/hpi-dynamics.png" height=450> <p class="source">Source: <a href="https://doi.org/10.1016/j.jhe.2022.101872">Clapp & Lindenthal (2022)</a></p> </div>

## Land values are more volatile<span class="subtitle"><br/>Land values fall/rise faster than property values</span> <div class="figure"> <p class="title">Land-value ratios, based on residual approach (Don't do this at home)</p> <img src="imgs/neg-residual-land-values.svg" height=450> <p class="source">Source: <a href="https://doi.org/10.1016/j.jhe.2022.101872">Clapp & Lindenthal (2022)</a></p> </div>

## Data needs<span class="subtitle"><br/>Estimate land values, highest and best use configuration, property values</span> * Maricopa County assessor's office data, 2007–2021 - Complete, reliable, various angles offered. <div class="figure"> <img src="imgs/data.png" height=450> </div>

## > 13 million obs.—still sparse!<span class="subtitle"><br/>Spatial data offer structure: half-mile grid, 12K points</span> <div class="figure"> <p class="title">New Supply of Real Estate, Maricopa County, 2007–2021</p> <img src="imgs/map-new-construction.png" height=450> </div>

## Trad. spatial econometrics?<span class="subtitle"><br/>No. We exploit spatial dependencies with Artificial Neural Networks</span> * Three models independently trained 1. Vacant Land Value = $f$(Land-related hedonics, time, distance to grid points) - Trained on <mark>sales of vacant, developable land</mark> 2. Highest best use (m²) = $f$(Est. vacant land value, land hedonics, time, distance to grid points) - Trained on <mark>all newly built homes</mark> (n=160K) 3. Property values = $f$(Est. vacant land value, HBU gap, land and structure hedonics, time, distance to grid points) - Trained on <mark>all sales</mark> (n=350K) * ANN's offer functional flexibility and interactions

## Simple Models<span class="subtitle"><br/>Not very deep, moderately wide neural network designs perform well</span> * Basic hedonic features + matrix of <mark>inverted distances to half-mile grid</mark> * Runs on a modern laptop (sufficient RAM helpful) <div class="figure"> <img src="imgs/ann-layout.png" height=300> </div>

## Better value surfaces<span class="subtitle"><br/>Machine Learning is great for <del>BIG</del> <ins>thin</ins> data</span></span> <div class="figure"> <img src="imgs/intspace.png" height=500> <p class="title">Highest and best use? Interior floor space of new residential units.</p> </div>

## Land Values<span class="subtitle transparent70"><br/>$/sqft</span> <p> </p> <p> </p> <p> </p> <p> </p> <p> </p> <p> </p> <p> </p> <p> </p> <p> </p> <p> </p> <p> </p> <p> </p> <p> </p> <p> </p>

## Actual use vs. HBU<span class="subtitle"><br/> If you could start again, on which parcels would you build more?</span> <div class="figure"> <img src="imgs/gaprel.png" height=500> <p class="title">Observed vs. optimal interior floor size</p> </div>

## Even thinner: Missing data<span></span><span class="subtitle"><br>% Missing in NCREIF sales data is significant and not random. Not always in lockstep.<br/></span> <div class="figure"><p class="title">Lease income and vacancy information</p><img src='imgs/fig8.png' height=430/> <p class=source>Source: <a href="https://www.lindenthal.eu/missing-data/">Leow & Lindenthal</a></p></div>

## Ongoing research project<span class="subtitle"><br/>["Machine Learning and Missing Data in Real Estate"](https://www.lindenthal.eu/missing-data/) (with Kahshin Leow)</span> * Real estate data plagued by missing data: What do we do, drop observation or variable? - Some ML algorithmns can deal with sparse data and will not necessarily drop obs or variables with NaN values. * Price prediction accuracy can increase by incorporating observations with missing predictors in the case of commercial real estate. * Example: For NCREIF sales data, missing data does not occurr at random, dropping observations or variables will bias estimates

## Missing values in NCREIF's Sales<span></span><span class="subtitle"><br>Missing data more frequent with smaller buildings, lower asset values and returns...<br/></span> <div class="figure"><img src='imgs/tab1.png' height=430/> <p class=source>Source: <a href="https://www.lindenthal.eu/missing-data/">Leow & Lindenthal</a></p></div>

## 3. It's graphs, not tables<span class="subtitle"><br />Don't throw away all this detail when pressing data into spreadsheets.</span> * Example: Trace investment performance back to the asset and the value created for the tenant. * Multiple layers of information... how to best aggregate them? Or Not? <div class="figure"> <img src="imgs/information_flow.png" height="450"> </div>

## Capturing the connections<span class="subtitle"><br/>Investor/Assets/Tenants in Graph Attention Networks?</span> <div class='figure'> <img src="imgs/hierarchical-graph.png" height="400"> </div>

## One last view<span class="subtitle"><br/>Mixed skills win: Data science is too important to be left to the data scientists!</span> <div class='figure'> <img src="imgs/skills2.png" height="500"> <p class='source'>Source: <a href="">Geltner et. al (2026)</a></p> </div>

## The robots have arrived<span class="subtitle"><br/>But they make us more productive, not obsolete.</span> <div class='figure'> <img src="imgs/skills1.png" height="500"> <p class='source'>Source: <a href="">Geltner et. al (2026)</a></p> </div>