Executive Summary

The explosive rise of generative artificial intelligence (AI) has opened a new frontier for real-estate investors: using large language models to decide where to deploy capital. This white paper tests the proposition head-on. We build a fully autonomous GPT-4o agent that ingests Zillow Home Value Index (ZHVI) price histories, demographic trends, macro rates, and Google Trends search intensity for 433 U.S. metropolitan areas. The agent writes its own Python code, selects the best-performing forecast model in each window, and forms 5-, 10-, and 20-city housing portfolios.

We find that GPT-selected portfolios earn Sharpe ratios up to 0.53 points higher (p < 0.01) than S&P Case-Shiller composites, traditional size-ranked baskets, and 1,000 random portfolios over twelve rolling out-of-sample tests (2009-2022). The edge is largest during migration shocks—such as the COVID-19 pandemic—when conventional diversification rules failed to anticipate regional booms.

Transparency matters: when city names and dates are hidden, the Sharpe advantage shrinks by 15–25 percent and geographic concentration rises. The clear takeaway is that label-rich data plus generative AI equals real alpha. Therefore, industry players should integrate well-labeled data streams with AI-driven model pipelines, adopt governance that balances human judgment with algorithmic insights, and target 15–20-city sleeves to maximize the diversification-versus-concentration trade-off.

This study investigates the data analysis capabilities of GPT-4o in real estate portfolio selection by integrating predictive modeling, model evaluation, and investment decision-making into a fully autonomous AI-driven framework. Unlike the earliest large language models (LLMs) that primarily process textual data or recent LLMs such as OpenAI's o1 and DeepSeek's R1, which are designed for complex reasoning, GPT-4o actively executes code and conducts quantitative analysis using the Code Interpreter tool. Leveraging a dataset of Zillow home price data and several predictive factors, the AI-generated portfolios consistently outperform various benchmarks in our out-of-sample backtest. Further, we find that data obfuscation -- removing city names, states, and dates -- reduces geographic diversification and produces lower Sharpe ratios than the unobfuscated portfolios. Overall, our findings highlight the potential of generative AI in advancing data-driven portfolio management.