Back to AI projects
NYU Stern Team Project · Nov 2025

Rental Prediction

Rental Investment Strategy with a Data-Driven Solution

Stack Python · scikit-learn · XGBoost · category_encoders
Data source 100K U.S. apartment listings (UCI) + IRS AGI by zip code
Team Anuraag Jain · Jacqueline Liu · Mechelle Tong · Rachel Lu
The story

Traditional apartment underwriting doesn't scale: listings are incomplete, pricing is inconsistent, and no analyst can manually weigh regional, city, and neighborhood effects across a 100,000-listing market. We built a machine-learning model to predict fair rental prices from a property's features, amenities, and neighborhood income data, with Random Forest performing best overall and separate models built per state to capture regional pricing differences.

Why it matters

A screening tool for three different desks

The result: a screening tool that flags listings priced significantly above or below prediction — a faster way to narrow a large market down to the opportunities worth a closer look, not a direct purchase recommendation.

That distinction matters because each team uses the output differently:

In every case, the model's job is to narrow the search — a mispriced rent signal doesn't mean the property is available to buy at a discount, so it's a starting point for due diligence, not a purchase trigger.

Data quality lift
8% zip coverage 100%
Missing zip codes imputed by mapping each listing's latitude and longitude
How it works

From Raw Listings to Investment Signal

Four stages · state-specific models · statistically flagged mispricing

Stage 1
Neighborhood quality
Imputed 92% of missing ZIP codes using latitude/longitude. Merged IRS ZIP-level AGI to proxy local economic health
IRS AGI
Stage 2
Engineer features
Amenities, pets-allowed etc extracted from unstructured text. Log-transformed price, square footage, and AGI to reduce skewness. City target-encoded;
category_encoders
Stage 3
Compare models
Ordinary Least Squares (OLS), Lasso, Random Forest, and XGBoost trained models
scikit-learn
Stage 4
Flag mispricing
Gap between actual and predicted prices, benchmarked by state, identifies buy/sell opportunities
XGBoost
Investment signal

Two states, two different mispricing stories

Texas · High-Volume, Supply-Responsive
Mispriced properties
597 total — the highest volume of any state
Pricing gap
Avg. underprice -11.58% · Avg. overprice 16.63%
Pattern
Volume-based: more frequent, smaller price tests
Strategy
Focus on size and construction quality, since value tracks the physical asset
California · High-Barrier, Scarcity-Driven
Mispriced properties
508 total — second-highest volume
Pricing gap
Avg. underprice -16.41% · Avg. overprice 21.31% (highest risk premium)
Pattern
Premium-based: fewer, more severe pricing errors
Strategy
Prioritize acquisitions in high-demand cities, since location is the primary value driver
Machine Learning Regression Modeling Feature Engineering Real Estate Analytics