We’re getting better at accurately estimating property values. Researchers at the forefront of this field developed a pricing model that consistently predicts a property’s price within 10% of its market value, and in an even tighter range if that property has sold more than once in the data set. You may think that’s a wide margin for error, but it outperforms what most practitioners would consider best-in-class valuations today: in a few seconds their model provides an estimate of value that is more accurate than the average appraisal. And with a little work the researchers believe these results could be improved “quite considerably.”¹

Their approach encompassed a set of 2400 apartment, office, industrial and retail properties in Phoenix and is notable for its effectiveness. But to get there, they had to combine “two cultures of statistical modeling.”

The first is a traditional econometric approach. This is conceptually simple and familiar–using historical data to measure how different factors can impact an outcome. It usually starts with intuition, like “property location and property values are related,” and uses regressions to test for a relationship. You can also test how several variables - location, building quality, building size, building age - might impact an outcome, like price, by feeding the model historical information.

The econometric model will tell you for every change of X amount in an input, you should expect a change of Y in the outcome. With that linear relationship, you can forecast and build more thoughtfully diversified portfolios.

But these models are imperfect because relationships between inputs and outcomes, especially in real estate, are often non-linear. As property size grows, for example, the price paid for each additional square foot drops. Short-term interest rates rising from 2% to 3% will have a different impact than going from 5% to 6%. In addition, econometric models rely on history, making them unreliable predictors in markets where structural shifts have broken from historical patterns — urban office being the obvious current example.

The second “culture” they draw from is algorithmic, or machine learning models. These models require much more data than econometric models and test every factor to find its relationships to the others, with no relationships pre-assumed. Instead of starting with an intuition, you ask it a question like “what are property values related to?” and it looks for relationships in the data.

It will answer you with factors that correlate to what you want to know, like econometric models, but also find where those relationships might break down, like “age only matters for Class B apartments” or “short drive-times matter more for Midwestern retail than in the Southeast.” Algorithmic models capture non-linear relationships and can handle a massive number of inputs and thus typically get you better prediction power than traditional regressions.

But the logic an algorithmic model uses to predict an outcome can be very complex and hard to explain. In addition, they can “memorize noise” and overfit their predictive model such that it performs great with data it trained on but flops with “out-of-sample” data. This is especially problematic in real estate, which has highly heterogeneous assets that don’t often trade.

The recent research combines the two approaches. Their machine learning model by itself predicts prices with an average error rate of 13.9%. Similarly, their economic model’s average error is 14.7%. But combined in an iterative approach, the average error falls to 10.9% and 9% for properties that sold more than once. Critically, these average error rates are for out-of-sample data, properties the model had never seen before.

The study focused on ways to combine these two modeling approaches; it did not focus on “optimizing the fit per se.” Hence their thought that improved results “should easily be achievable in the future.” We would welcome that progress!

Interestingly, the average error rate for properties that sold more than once does not drop much even if two critical inputs, NOI and WalkScore, are omitted from the models. The insight here is that having multiple property sales gives a sophisticated model Cliffs Notes to estimating current valuation.

In all, we are at a moment when valuation tools are becoming more disciplined, more data-driven, and significantly more accurate. Although these models aren’t available “off the shelf” yet, they’re coming, and the road map is clear.

By combining structure (econometrics) with flexibility (machine learning), prediction accuracy improves to levels we have not seen before. For investors, that matters less as a headline and more as infrastructure worth building into our day-to-day practice. We’d use this to evaluate owned assets as well as potential acquisitions and developments. Over time, this should benefit the entire industry, as even small improvements in valuation precision compound into better capital allocation decisions.

Who: Lunar Energy

What: A low-cost, home-swapping alternative to Airbnb: all members must host their own home to travel.

The Sparkle: Honestly we’re not sure what the sparkle is here but they just raised more than $100M in a Series D.

The $100M is interesting, but the more relevant detail is what Lunar is quietly revealing about where residential infrastructure is heading. California homeowners enrolled in Lunar's Virtual Power Plant program can sell excess stored energy back to the grid when demand peaks — effectively turning a home battery into a revenue-generating asset. The system also tracks incoming storms and pre-charges automatically, a feature that speaks directly to how grid reliability has degraded in high-risk markets like coastal California and fire-prone inland metros. For owners and developers in those markets, this isn't a consumer gadget story. It's a signal that resilience infrastructure is becoming a legitimate amenity — and eventually, a line item that underwrites will need to have a view on.

.

If you’re at your desk and want to feel like you’re someplace great…

Temple Bar, Dublin. Cobblestones, festoon lighting, and the ambient hum of a city that has been doing this for centuries.

It's a webcam, yes. There's a brief ad. Watch it anyway.

Link: https://www.earthcam.com/world/ireland/dublin/?cam=templebar

Editor’s Note: The Real Estate Haystack believes in sharing valuable information. If you enjoyed this week's newsletter, subscribe for regular delivery and forward it to a friend or colleague who might find it useful. It's a quick and easy way to spread the word.