Forecast CRE Submarket Demand with Custom AI Models

Most investment firms rely on platforms like CoStar and REIS for market data. These tools are excellent repositories of historical comps and property data. They can tell you the cap rate for every multifamily sale in a submarket for the last five years, but they are built to report on the past, not predict the future.

Consider an analyst evaluating two submarkets for a new acquisition. They pull reports from CoStar and see Submarket A has 1.5% higher rent growth over the last 12 months. What the report does not show is that a major employer just signed a 500-person office lease in Submarket B, and local permit filings for competing new construction have dropped 40%. These are powerful leading indicators of future demand, but they live in separate, unstructured data sources.

The structural problem is that these platforms are closed databases, not dynamic analytical engines. You cannot feed your own proprietary data or alternative data feeds into their models. An analyst might try to bridge this gap with a BI tool like Tableau, but this still requires dozens of hours of manual data exporting and cleaning. Tableau can visualize trends, but it cannot run a regression analysis across 25 disparate variables to find predictive patterns.

The result is investment committees making multi-million dollar decisions based on lagging indicators. The firm that sees the office lease and permit data first has a significant edge. Without an automated way to ingest and analyze these signals, firms are constantly looking in the rearview mirror.

The engagement would begin with a discovery phase to map the data signals most relevant to your investment thesis and target asset class. We would audit public data sources like city permit portals and Bureau of Labor Statistics employment data, alongside paid APIs for foot traffic (Placer.ai) or company growth signals (LinkUp). The outcome is a data strategy document outlining which 5-10 signals have the strongest predictive potential.

The technical approach involves building a series of Python data pipelines on AWS Lambda. These serverless functions would run on a schedule to pull data from each source, normalize it, and store it in a central Supabase database. The core forecasting model, likely a time-series or gradient-boosted model using XGBoost, would be wrapped in a FastAPI service. This architecture keeps hosting costs low (under $50/month) and ensures the system can handle new data sources as they are identified.

The delivered system provides an API endpoint your team can query with a submarket name to get a 12-month demand forecast with confidence intervals. For non-technical stakeholders, we would build a simple web interface on Vercel to run forecasts and see which factors are driving the prediction. You receive the full source code, a runbook for retraining the model, and all associated documentation.