The Philadelphia Eviction Early Warning System

A Predictive Model for Proactive Resource Allocation

Angel Rutherford
Ixchel Ramirez
Tess Vu

2025-12-08

Research Question


“Where should renter’s assistance programs be targeted in Philadelphia?”


The Goal: Move from reactive crisis management to proactive triage.

Philadelphia Context

331,415 Philadelphia households are renters.

Policy Gap: The Emergency Rental Assistance Program (ERAP) just ended in September 30, 2025, removing a critical safety net.

  • Trend: Filings are rebounding to pre-pandemic levels but are affecting neighborhoods unevenly.

Data Sources

  1. Eviction Lab: Monthly Filing Counts
  2. OpenDataPhilly:
    • Tax Delinquency: Proxy for landlord financial distress.
  3. Census ACS (2023): Poverty Rate, Rent Burden, Single-Parent Households.

Data Cleaning

  • Sealed Tracts (Privacy): We identified “Sealed Tracts” containing mass eviction events (up to 694 filings/month).
  • Tax Data: Cleaned to focus on true delinquency with active penalties rather than minor overdue bills.
  • Outlier Management: We capped training data at the 99.75th percentile (20 filings) to stabilize the model, but tested on raw data to prove real-world robustness.

Exploratory Analysis

Eviction data is zero-inflated and overdispersed.

  • Conclusion: OLS and Poisson models are invalid, use Negative Binomial Regression.

  • Zeros: 37% of tract-months have zero filings.

  • Dispersion: Variance > Mean.

Methodology: Spatio-Temporal Modeling

Negative Binomial Model that predicts filings using:

  1. Momentum: Filings from month before.
  2. Contagion: Average filings of neighboring tracts.
  3. Structure: Neighborhoods, Poverty, Tax Delinquency, and Demographics.
  4. Policy: Moratorium active/inactive status.

Key Drivers

Model Performance

Accuracy: Within < 2 filings per tract.

Stability: RMSE change is -5, so it’s better on unseen data.

Risk Triage

  • Tracts flagged as “Critical Risk” saw 5x more actual evictions than those flagged as “Low Risk.”
  • Model works as a Triage Tool for resource allocation.

Structural Inequity

Black-majority tracts face higher risk.

Model Accurately Predicts Disparity

Model reproduces structural bias to identify communities most at risk.

Limitations

  1. Mass Events
  2. Stale Data
  3. Sealed Tracts
  4. Variability

Recommendations for Implementation

Triage Dashboard

  1. Run Monthly: Input new filing data on the 1st of the month.
  2. Generate List: Output Top 50 “Critical Risk” Tracts.
  3. Targeted Intervention:
    • Canvassing / Legal Aid: Deploy teams, allocate resources.
  4. Safeguard: Tool strictly for adding resources, never for automated decision-making.