As real estate investors navigate AI and integrate it into their work, academics are shining a light on specific areas where AI outperforms status quo decisions, especially investment decisions, and market selection for a real estate portfolio is one. The authors of a new working paper gave GPT-4o several historical public data sets and asked it to optimize risk-adjusted returns through market selection for a national portfolio of home purchases. The AI consistently beat investment benchmarks, including Case-Shiller Home Price Indices.

Applying the model to out-of-sample data sets, authors tested three-year hold periods between 2009 and 2022 and multiple portfolio sizes. They revealed higher Sharpe ratios relative to benchmarks in nearly all cases. The average three-year returns across all testing windows ranged from 15% to 21%, with larger portfolios exhibiting particularly strong Sharpe ratios. Furthermore, the model also made some contrarian market picks that paid off well.

These findings speak to how to optimally diversify a portfolio, echoing insights from broader investment research: that AI today has the power to help you see deeper into complex investment decisions and be a better investor. Diversification is a challenging but proven way to improve returns, and the authors address that fundamental tension of focus vs. diversification, saying about equity markets that from 1926 to 2018, “merely 4% of listed companies generated the entire net gain” for the stock market. On the other hand, they remind us sub-optimally diversified equity strategies, often marketed as “active” strategies offered by financial services firms, “tend to underperform market averages.”

For real estate, diversification by MSA has several benefits. One is the avoidance of concentrated risk, especially risk of natural disasters like fires and hurricanes and local economic shocks. Many other benefits are strategic and upside-related, like spreading bets among markets with high population and job growth. Our recent article on Work From Home (the Haystack’s most widely read issue, btw), has something to say about that.

Appropriate diversification is hard, so how do you really win this thing? The point of the study is that AI can help. But there are some cautionary notes in the study as well. Those combined with the headline results all provide a well-rounded perspective on this moment in AI’s growth. These caveats are crucial for a balanced understanding of AI's role.

To get to those yellow flags, let’s first address what the model “knows.” Inputs included things you would expect - variables about historical home pricing (source: Zillow), population change (U.S. Census), employment change (Bureau of Labor Statistics) and mortgage rates (Freddie Mac) - and one less-anticipated variable: Google trends data. The idea of using search data came from an unrelated, earlier study’s finding that cities with high rates of real estate-related searching on Google outperformed low-intensity cities by up to 8.5% over two years.

These yellow flags arise when the model knows more than you tell it. A general risk when backtesting predictive AI models is the “look-ahead” bias, which occurs when the AI’s “predictions” are subtly informed by actual historical outcomes, such as future housing price movements in any given MSA. In other words, the model can cheat. Not cool! This potential flaw overstates the predictive power of the model.

The authors tested for this by running the base models again but without place names or dates, replacing those specifics with random strings of digits as placeholders. From this revised “obfuscated” data set, some of the model’s portfolio construction changed and resulted in diminished Sharpe ratios. So, it’s possible the look-ahead bias was real, although in many cases the obfuscated data set still produced index-beating portfolios. Importantly, in the more recently tested time periods, the obfuscated models and the normal models constructed nearly identical, high-performing portfolios, suggesting the look-ahead bias was much less of a factor.

A final note: Today’s issue is part of the Haystack’s occasional series on AI in real estate. We’ll continue to write about it as we find compelling academic research. Reach out to us you know of research that merits broader attention within this community. Special thanks to Prof. Ryan Chacon at DU for the head's up on this week's research.

Who: Kore Investments

What: A real estate sponsor actively focused on… drumroll please… value-add office acquisitions.

The Sparkle: Not all distressed office is created equal, KORE will tell you. Many office properties are stranded assets with declining occupancy and essentially no tenant prospects. KORE seeks out another type of opportunity: They are expert in identifying high-quality office assets that are or can be stabilized and that have long-term tenant appeal but distress in the cap stack, buying them at 12%+ cap rates. They bring value-add capital and active management (old school!) to optimize medium-term IRR-driven holds.

You know the U.S. economy produces goods and services. In terms of dollar value, did you know the ratio is 1:4? We really don’t make stuff anymore. Also, finance and real estate drive 25% of all the services. We saw this statistic and thought of last week’s issue and again felt like getting more stability in real estate valuations would be a smart economic priority given the size and importance of our business.

Link: https://usafacts.org/articles/what-does-the-us-produce/

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.