📈 economics

Macroeconomic Predictability and Machine Learning in REIT Returns: Out-of-Sample Evidence and Property-Sector Heterogeneity

This study finds that while nonlinear machine learning models offer economically meaningful out-of-sample forecasting gains for medium-horizon U.S. REIT returns—particularly in the health-care sector and driven by macroeconomic factors—they do not achieve statistically significant aggregate superiority over linear benchmarks after multiple-testing adjustments, with predictive signals concentrated in common time-series components rather than cross-sectional selection.

Original authors: Mahsa Khoshnoud

Published 2026-10-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mahsa Khoshnoud

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of investing, some assets behave like stocks, while others act more like bonds or physical buildings. Real Estate Investment Trusts, known as REITs, sit right in the middle of this divide. They are companies that own income-producing properties like shopping malls, apartment complexes, and hospitals, and they trade on the stock market just like any other company. Because their value depends on both the health of the stock market and the physical reality of the economy—such as interest rates, job growth, and inflation—predicting their future performance is a complex puzzle. For decades, investors and economists have tried to solve this puzzle using simple, straight-line math, assuming that if the economy goes up, these stocks go up in a predictable way. But the economy is rarely a straight line; it twists and turns, and the relationship between a recession, a spike in interest rates, and a property's value can change depending on the context. This is where modern computer science enters the picture, offering tools that can look for complicated, hidden patterns in data that traditional math might miss.

A researcher set out to test whether these advanced computer tools, known as machine learning, could actually predict the future returns of REITs better than the old, standard methods. They gathered a massive amount of historical data, tracking 383 different real estate companies over more than two decades, from the year 2000 through 2023. This period covered some of the most dramatic economic shifts in modern history, including the global financial crisis, a long period of very low interest rates, the sudden shock of the pandemic, and the rapid inflation that followed. The researcher fed the computer models a specific set of clues: current interest rates, unemployment numbers, inflation data, how fast the economy is growing, and how the broader stock market was performing. They then asked the computer to guess what would happen to REIT returns over the next few months. To ensure the test was fair, they trained the computer on past data and then asked it to predict the future without ever letting it peek at the answers beforehand.

The results revealed a story that is more nuanced than a simple victory for the new technology. When the researcher asked the computer to predict what would happen in the very near future—one, two, or three months ahead—the machine learning models performed no better than a simple guess based on the average past performance. In fact, for these short timeframes, the computer models were slightly worse than the basic benchmarks. However, the story changed when the researcher looked at a medium-term horizon, specifically asking the computer to predict the cumulative returns over a six-month period. Here, the advanced models began to show their strength. The most sophisticated algorithms, which work by building many small decision trees to find complex patterns, were able to reduce their prediction errors significantly compared to the traditional linear methods. They captured about 9.3 percent of the variation in the six-month returns, a meaningful improvement over the standard approach, which actually performed slightly worse than a simple historical average.

Despite this improvement in accuracy, the researcher was careful not to declare a total victory. When they applied strict statistical tests to see if the computer's success was a genuine breakthrough or just a lucky fluke, the results were less definitive. The improvements, while economically noticeable, were not statistically strong enough to prove with absolute certainty that the machine learning models were superior to the older methods across the board. The study suggests that the computer models are not finding a magic formula that works for every single property or every single month. Instead, they are picking up on a specific type of signal: the broad, shared movement of the entire real estate market driven by the overall economy. The models are good at timing the general market, sensing when the tide of interest rates and employment is likely to turn, but they are not good at picking individual winners and losers among the different companies.

The researcher also discovered that the usefulness of these tools depends heavily on the type of property being analyzed. The machine learning models worked best for health-care REITs, where the predictions were not only accurate but also statistically significant even after rigorous testing. Retail properties also showed promise, while residential and diversified real estate companies were better predicted by the simpler, traditional methods. This finding highlights that the economy affects different types of buildings in different ways, and a one-size-fits-all computer model cannot capture every nuance. Furthermore, the study showed that the computer models were not just memorizing the past; they were learning to weigh different economic factors, such as how changes in unemployment or interest rates interact with each other, to form a more complete picture of the future.

Ultimately, this research offers a clear, grounded perspective on the role of artificial intelligence in finance. It demonstrates that while machine learning can uncover valuable patterns in how the economy influences real estate, it is not a crystal ball. The technology shines when looking at medium-term trends and specific sectors, but it does not replace the need for careful statistical scrutiny or the understanding that the economy is a shifting landscape. The study concludes that the value of these advanced tools is real but conditional; they are powerful instruments for understanding the general direction of the market, provided they are used with the right expectations and applied to the right kinds of assets.

Technical Summary: Macroeconomic Predictability and Machine Learning in REIT Returns

Problem Statement
Publicly traded Real Estate Investment Trusts (REITs) occupy a hybrid position in financial markets, exposed to both equity risks and property-market fundamentals driven by macroeconomic conditions (interest rates, inflation, labor markets, and real activity). While linear predictive regressions have traditionally modeled these relationships, they impose fixed functional forms that may fail to capture state-dependent nonlinearities. For instance, the impact of interest rates on REIT returns may differ depending on whether they are accompanied by strong real activity or persistent inflation. Although recent literature suggests machine learning (ML) can extract nonlinear predictive signals, it remains unclear whether these gains are statistically robust under strict out-of-sample evaluation, whether they generalize across forecast horizons and property sectors, and whether they represent aggregate timing signals or cross-sectional security selection. This study addresses these gaps by evaluating whether nonlinear ML models improve REIT return forecasts beyond linear benchmarks and historical means, specifically under rigorous chronological constraints.

Methodology
The study utilizes a dataset of 47,075 firm-month observations covering 383 U.S. REIT securities from January 2000 to December 2023. The predictor set is deliberately parsimonious, comprising macroeconomic variables (federal funds rate, real 10-year rate, median CPI, real GDP per capita, unemployment), market variables (CRSP value-weighted returns), changes in macro variables, three lags of REIT returns, and firm size.

The forecasting design employs a strictly chronological split to prevent look-ahead bias:

Estimation Sample: 2000–2015.
Validation Sample: 2016–2019 (used exclusively for hyperparameter tuning).
Test Sample: 2020–2023 (evaluated once with frozen hyperparameters).

The study compares three tree-based ML models—Gradient Boosting, Random Forest, and XGBoost—against two benchmarks: a pooled Ordinary Least Squares (OLS) regression and a historical-mean forecast. Performance is evaluated using out-of-sample $R^2$ ( $R^2_{OOS}$ ), RMSE, and MAE across direct (1–3 month) and cumulative (3–6 month) horizons.

Crucially, the study employs formal forecast-comparison inference. It uses Diebold–Mariano (DM) statistics with HAC adjustments for overlapping windows and the Harvey–Leybourne–Newbold small-sample correction. To address multiple testing, the Holm procedure is applied across the family of forecast comparisons. The study also utilizes SHAP (SHapley Additive exPlanations) values to decompose feature importance and analyzes the distinction between time-series (aggregate) predictability and cross-sectional (security selection) predictability.

Key Results

Horizon Dependence: Forecastability is strongly horizon-dependent. At direct short horizons (1–3 months), all models, including ML, produce negative $R^2_{OOS}$ relative to the historical mean. However, at medium-horizon cumulative returns, ML models generate economically meaningful gains. At the six-month horizon, Gradient Boosting and XGBoost achieve $R^2_{OOS}$ values of approximately 9.3%, whereas pooled OLS remains slightly negative (−0.6%).
Statistical Significance: Despite the economic magnitude of the gains, formal forecast-comparison tests do not establish statistically significant aggregate superiority of ML models over linear or historical-mean benchmarks after multiple-testing adjustment. The smallest unadjusted p-value for XGBoost against the historical mean is 0.063, which becomes 0.756 after Holm adjustment. This suggests that while ML reduces forecast error magnitude, the evidence is not sufficient to claim statistical dominance in the aggregate sample.
Nature of Predictability: SHAP analysis indicates that predictive importance is driven primarily by real activity, labor-market conditions, policy rates, and long-term real interest rates, rather than simple return persistence or firm size. The forecasting signal is concentrated in the common time-series component of REIT returns. The correlation between the cross-sectional average of realized six-month returns and the XGBoost forecast is 0.530, but the average cross-sectional Spearman correlation between predicted and realized individual REIT returns is near zero. Thus, the models capture aggregate timing information rather than cross-sectional stock-selection ability.
Sector Heterogeneity: Predictive gains vary significantly across property sectors. Health-care REITs exhibit the strongest evidence of nonlinear predictability, with XGBoost achieving an $R^2_{OOS}$ of 11.5% that remains statistically significant relative to the historical mean after sector-level adjustment. Retail REITs show positive point estimates but fail to reject equal predictive accuracy in formal loss tests. Conversely, diversified and residential REITs favor the linear OLS benchmark.
Robustness: A pre-2020 pseudo-out-of-sample exercise (2014–2016 holdout) shows that while ML models consistently outperform pooled OLS in relative error terms, they do not reproduce positive $R^2_{OOS}$ relative to the historical mean, indicating that the absolute advantage is regime-dependent.

Significance and Claims
The paper contributes to the literature by shifting the focus from whether nonlinear models can generate positive out-of-sample fit to whether those gains survive strict statistical scrutiny. The author claims that the usefulness of machine learning for REIT return forecasting is conditional rather than universal.

Distinction Between Metrics and Inference: The study highlights that sizeable improvements in conventional performance metrics ( $R^2$ , RMSE) do not necessarily imply statistically established predictive dominance once multiple-testing adjustments and small-sample corrections are applied.
Component Identification: The evidence suggests that ML models in this context predict a smooth, macro-sensitive common return component rather than enabling cross-sectional security selection.
Boundary Conditions: Predictive gains are shown to be contingent on the forecast horizon (emerging at medium cumulative horizons), the economic regime (stronger in the 2020–2023 period), and the property sector (strongest in health care).

The author concludes that while machine learning can reveal useful predictive structure in listed real estate, the strength of this conclusion depends critically on the specific horizon, benchmark, return component, economic regime, and inferential criteria used. The findings do not support a claim of universal machine-learning superiority but rather a nuanced view where nonlinear methods offer selective, regime-dependent value.

Technical Summary: Macroeconomic Predictability and Machine Learning in REIT Returns

More like this