We often hear from Haystack readers looking for a forecasting edge. They ask (and we write) about more sophisticated approaches to key assumptions like cap rates and rent growth, or even submarket performance. Others push further into harder questions about forecasting returns across entire portfolios and funds. Recently a reader asked a great question: For the most liquid real estate investments (REITs), how much of their returns is actually forecastable?

That question opened a surprisingly productive rabbit hole. We dug into new machine learning research that quantifies just how predictable REIT returns actually are — and how much edge investors might extract from that insight. Below, we unpack the findings, why they matter more than they seem, and how these models might shape real-world portfolio construction

At first, it felt like an unhittable pitch - we could find no meaningful academic work on the limits of REIT forecasting - but we got lucky. A new study published just months ago analyzed 486 REITs from 1990 to 2021 and found that the best machine learning models can explain 5% of monthly change in REIT returns.¹ That may not sound like much, but it’s actually impressive and very useful. Here’s why:

First, traditional regression models — linear, with fixed factor weights — explain less than 1% of monthly REIT stock price movements. So we had nearly zero predictive power before machine learning came along. Even when ML is used to improve linear models, explanatory power only rises to 2–3%. To go further, you need nonlinear approaches — the kind of models ML excels at building — which deliver meaningfully better results, and which we discuss more in a moment.

Second, even 5% predictability has an impact: REIT investments informed by these models show materially higher risk-adjust returns. For individual REITS, Sharpe Ratios improved 10 to 40 basis points, a meaningful shift in any long-only strategy.

Third, the models are useful for stock selection but they’re even more powerful at the portfolio level. When applied to portfolio construction the researchers saw “a two-to-threefold jump when forecasting” a portfolio’s monthly change. Once again, the most advanced ML models outperformed in terms of average returns and Sharpe ratios.

Fourth, these models learn and get better. The researcher’s best models bombed predictions for 2007-2008 because they could not foresee the systemic shock of the GFC. But they learned. By 2020 it understood those shocks were possible and perhaps even what to look for. The model had fantastic predictability (20%+) for 2020’s Covid correction. The researchers tested and confirmed that the models had, in fact, internalized lessons from the GFC — not just overfitting, but pattern recognition with forward relevance.

So why do non-linear machine learning models outperform? It boils down to what they can see that linear models can not, even if those linear models are ML-informed. Non-linear models see “complex predictor interactions.” In other words, these models don’t just assess how individual factors relate to REIT returns; they analyze how those factors interplay with each other, uncovering patterns invisible to linear approaches.

These research findings are significant for other reasons. One is that most ML-related research in real estate relates to “the prediction of property values,” not stock prices, with the latter being a much more challenging mountain to climb.

The results highlight how fundamentally different real estate equities are from other sectors. This paper compared real estate stocks vs. other industries under the suspicion that the differences in fundamental drivers of return for each would show up in different levels of predictive power. As model sophistication increases, “real estate market predictability starts to vastly outperform stock market predictability.” ML models that could predict 5% of real estate stock’s monthly change could predict only 0.4% of the changes in other industries’ stock prices.

The study also found that large REITs are more predictable than small REITs, and that “specialist homogeneous property types” (who says that?) including retail, residential and hotel “perform well,” but REITs with diversified portfolios were harder to forecast.

Finally, it’s worthwhile to note that the model’s strong explanatory power for one-month changes fell off when you looked longer-term, leading the authors to reflect, “it is clear that nothing good lasts forever.” The model had basically no predictive power at 12 months. However, these models were asked to focus on one-month changes and trained on monthly data. When the researchers retrained the models on quarterly and annual data, its predictive power for those new look-ahead periods was good.

Of course, forecasting remains imperfect, but in the hands of disciplined investors, even a 5% signal — if persistent — compounds powerfully. In markets driven by noise, small signals matter.

Who: Sharing Connexion

What: A non-profit with deep real estate skill sets that helps other non-profits manage/sell real estate donations they otherwise would not know what to do with.

The Sparkle: SC solves a problem you didn’t know existed. When big charitable organizations receive gifts of real estate assets, those assets come in every form, from farmland to Taco Bells, and in every condition. The beneficiary is typically ill-equipped to manage or sell these assets, especially of they come with operating, zoning or environmental complexity. Enter the real estate experts at Sharing Connexion, who underwrite the asset, sometimes invest their own capital into improving it if they know that will add significant value, and then manage a rigorous sale process and earn a percentage of the sale as a fee.

When we drop the expression “let’s not reinvent the wheel” the reference is easy to understand but hard to attribute - we actually don’t know who invented the wheel.

But there’s a new theory backed by archeological evidence that the wheel probably got rolling in a Hungarian copper mine nearly 6,000 years ago. Editorially, if I’m going to work in a mine I’d much rather do it after the wheel was invented.

For a fun, accessible, slightly science-y take on one of humanity’s most important breakthroughs, check out the article below.

Link: https://theconversation.com/how-was-the-wheel-invented-computer-simulations-reveal-the-unlikely-birth-of-a-world-changing-technology-nearly-6-000-years-ago-244038

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.