Philadelphia Eviction Prediction Model

MUSA 5080 - Final Project README

Author

Zhiyuan Zhao & Fan Yang

Published

December 6, 2025

Philadelphia Eviction Prediction Model

An early warning system for predicting eviction filings at the census tract level in Philadelphia, developed for MUSA 5080: Public Policy Analytics.

Authors

  • Zhiyuan Zhao - University of Pennsylvania
  • Fan Yang - University of Pennsylvania

Project Overview

This project develops a predictive model to forecast monthly eviction filings across Philadelphia’s 408 census tracts. The goal is to help the Philadelphia Housing Authority allocate prevention resources more effectively by identifying high-risk neighborhoods before evictions occur.

Policy Context

Philadelphia experiences over 1,000 eviction filings per month, ranking among the highest eviction rates in the nation. Many evictions are preventable with timely intervention—legal aid, rental assistance, and tenant counseling. This model enables proactive resource deployment rather than reactive crisis response.

Key Findings

  1. Evictions are predictable - Past eviction patterns strongly predict future filings
  2. Geographic clustering - Hotspots concentrate in North, West, and Southwest Philadelphia
  3. Racial disparities persist - Majority-Black neighborhoods show elevated eviction rates even after controlling for income
  4. Model performance - Enhanced Poisson model achieves MAE of 1.78 and correlation of 0.44 on test data

Data Sources

Primary Data

Dataset Source Description Link
Eviction Filings Eviction Lab Weekly eviction filing counts by census tract, 2020-2025 Eviction Lab Philadelphia Tracking

Secondary Data

Dataset Source Description Link
Census Demographics U.S. Census Bureau ACS 5-year estimates: poverty rate, median income, race/ethnicity Census API
Renter Population U.S. Census Bureau Tenure data (renter vs. owner occupied) by tract Census API
311 Service Requests OpenDataPhilly Housing-related complaints (no heat, dangerous buildings, infestations) OpenDataPhilly 311
Census Tract Boundaries U.S. Census Bureau TIGER/Line Shapefiles for Philadelphia County Census TIGER

Data Access Notes

  • Eviction Lab data requires download from their tracking page; select Philadelphia and export CSV
  • Census data accessed via tidycensus R package (API key required: get key here)
  • 311 data available as direct download or via API from OpenDataPhilly
  • Shapefiles accessed via tigris R package

Repository Structure

├── data/                              #Data used for this assignment
├── output/                            #visualizations and tables
├── slides/                            #presentation slides package
  ├── eviction_presentation.qmd          # Presentation slides
├── final_Zhiyuan_Fan.qmd              # Main analysis document
├── README.md                          # This file
└── .gitignore

Methodology

Temporal Aggregation

Weekly data aggregated to monthly to address zero-inflation (68% zeros at weekly level → 34% at monthly level).

Feature Engineering

  • Temporal lags: 1-month and 2-month lagged eviction counts
  • Spatial lag: Average evictions in k-nearest neighbors (k=5)
  • Demographics: Poverty rate, renter percentage, median income
  • Distance features: Distance to city center, distance to nearest hotspot
  • Hotspot indicators: Local Moran’s I classification (High-High clusters)
  • Interaction terms: Poverty × racial composition, spatial lag × renter percentage

Models Tested

Model Description McFadden’s R²
Baseline Poisson Temporal lags only 0.14
Full Poisson + Spatial lag, demographics, race 0.23
Full Negative Binomial Same as above, allows overdispersion 0.22
Enhanced Poisson + Hotspots, month effects, interactions 0.24

Validation

  • Train/test split (80/20, temporal split)
  • Spatial autocorrelation check on residuals (Moran’s I = 0.001, p = 0.33)
  • Equity analysis across neighborhood racial compositions

How to Reproduce

Requirements

# R packages
install.packages(c(
  "tidyverse",
  "sf",
  "spdep",
  "tidycensus",
  "lubridate",
  "MASS",
  "knitr",
  "ggplot2"
))

Steps

  1. Clone this repository

  2. Obtain Census API key and store in .Renviron:

    CENSUS_API_KEY=your_key_here

  3. Download Eviction Lab data and place in data/raw/

  4. Download 311 data from OpenDataPhilly and place in data/raw/

  5. Open eviction_prediction_enhanced.qmd in RStudio

  6. Run all chunks to reproduce analysis

Key Outputs

  • Risk Classification Map: Census tracts classified as Very High / High / Moderate / Low risk
  • Monthly Predictions: Forecasted eviction counts by tract
  • Equity Analysis: Model performance comparison across neighborhood demographics

Limitations

  • Filings ≠ Evictions: Model predicts court filings, not completed evictions
  • COVID-era training data: 2020-2025 period heavily influenced by moratoriums
  • Census tract granularity: Cannot identify individual households at risk
  • Equity concerns: Higher prediction errors in majority-Black neighborhoods
  • Policy shocks: Cannot anticipate sudden policy changes

Ethical Considerations

  • Risk of stigmatizing neighborhoods labeled “high risk”
  • Model outputs should inform—not replace—human decision-making
  • Access to predictions should be restricted to prevent misuse
  • Implementation should include additional resources for historically underserved communities

License

This project is for academic purposes (MUSA 5080, University of Pennsylvania).

Acknowledgments

  • Professor’s guidance on spatial modeling techniques
  • Eviction Lab for making eviction data publicly accessible
  • OpenDataPhilly for municipal data infrastructure

Last updated: December 2025