Abstract
The US Real Estate Investment Trust (REIT) market is a critical component of global investment portfolios, yet its volatility makes accurate return prediction challenging. This paper evaluates the efficacy of the Group Method of Data Handling (GMDH), a self-organizing neural network, in forecasting REIT returns. By comparing GMDH against the traditional GARCH (Generalized Autoregressive Conditional Heteroscedasticity) model, the study demonstrates that GMDH's ability to automatically select variables and optimize its architecture leads to superior predictive accuracy. Using data from the US REIT index between 2016 and 2021, the results show that GMDH consistently achieves lower Mean Squared Error (MSE) than GARCH, even with small training datasets.
Introduction
Real Estate Investment Trusts (REITs) provide investors with exposure to large-scale, income-producing real estate without the need for direct property ownership. However, the REIT market is characterized by high volatility and is influenced by a complex array of non-linear economic factors.
Traditional financial modeling has long relied on parametric approaches like the GARCH model to manage this volatility. While robust, GARCH relies on pre-defined mathematical assumptions that may fail to capture the intricate, "messy" patterns of the modern financial landscape. This study explores the application of the Group Method of Data Handling (GMDH) as a data-driven, non-parametric alternative for the US REIT market.
Methodology: Self-Organizing Neural Networks
The core of this research is the Group Method of Data Handling (GMDH). Unlike standard Artificial Neural Networks (ANNs) that utilize a fixed architecture designed by human experts, GMDH is a self-organizing system.
As shown in Figure 1, traditional ANNs require the pre-specification of layers and neurons. In contrast, GMDH builds its structure iteratively. It adds layers one by one, automatically selecting the most relevant input variables and discarding those that contribute to noise. This process continues until the model reaches an optimal level of complexity.
The researchers tested three specific kernel functions to connect neurons within the GMDH framework: Linear, Linear Covariance, and Quadratic. These kernels determine how the model processes the relationships between variables at each layer.
Data and Experimental Setup
The study utilized the US REIT index, encompassing 135 stocks across various sectors including residential, retail, and healthcare. The dataset spanned from October 6, 2016, to July 30, 2021. To train the model, 26 technical indicators were used, categorized into volatility (e.g., Bollinger Bands), trend (e.g., MACD), momentum (e.g., RSI), and return rates.
A key focus of the study was the model's robustness across different training sample sizes, ranging from 30 to 600 daily observations. Performance was benchmarked against a standard GARCH model using Mean Squared Error (MSE) as the primary evaluation metric.
Key Results and Discussion
The quantitative analysis revealed a significant performance gap between the machine learning approach and traditional methods. In nearly every tested scenario, GMDH achieved a lower MSE than GARCH.
For instance, in a test involving 100 training samples and a 30-day prediction window, the GMDH model with a Linear kernel achieved an MSE of 0.72, while the GARCH model lagged significantly with an MSE of 3.79. The Linear Covariance kernel also showed high stability for longer-term (60-day) predictions, yielding an MSE of 0.82 compared to GARCH’s 1.93.
Interestingly, the study found that "more complex" does not always mean "better." The Quadratic kernel, while mathematically more sophisticated, proved difficult to train and was highly sensitive to outliers, sometimes leading to invalid results. This highlights the importance of kernel selection in financial applications.
Conclusion
The Group Method of Data Handling offers a powerful, automated alternative to traditional parametric models for predicting REIT returns. Its ability to perform well with small datasets and its self-organizing nature make it particularly suited for the dynamic US real estate market. While GMDH remains a "black-box" model—offering limited interpretability regarding the specific drivers of its predictions—its superior accuracy suggests it is a valuable tool for modern financial analysts.
Summary updated