Predicting Eviction Risk in Philadelphia

(JMC)^2 Team

Jed Chew, Jun Luu, Mark Deng, Mohammad AlAbbas

Final Challenge: Predicting Evictions

Iron Chef × Shark Tank × Kaggle

• Secret ingredient: Philadelphia eviction data
  
• Goal: Build a predictive model that city officials could actually use
  
• Our focus: Eviction risk in Philadelphia

Roadmap

1. Eviction in Philadelphia

2. Policy Context
   

3. Data & Feature Construction
   

4. Modeling Strategy
   

5. Results: How good is the model?
   

6. Limitations
   

7. Recommendations & Implementation

Eviction and Policy Context

The Eviction Crisis in Philadelphia

Source: LSC Evictions Law Database (Last updated: 11/01/2025

Policy Context

Source: City of Philadelphia Website

• Eviction filings are down approximately 40%.

• Landlords and tenants reach agreements in approximately 70% of the cases where they participate in mediation.

Data Sources & Integration

Data Source

• Philadelphia Eviction Data Source: Eviction Lab Philadelphia tracking data, 2020-2025, monthly

• Socioeconomic Data (ACS data): 2023 5-Year American Community Survey (ACS) from the United States Census Bureau

• Philadelphia Crime Data: Philadelphia crime data for 2020-2025 from Open Data Philly

• 311 Calls Data (Eviction-relevant): Philadelphia 311 Service and Information Requests from 2020-2015

• Tax Data: Philadelphia real estate tax balances data by Census Tract from Open Data Philly

Data Cleaning & Filtering

• 311 Data Filtering:
  • Clean 311 data and create a small set of “eviction-relevant” 311 indicators (e.g., Sanitation / neighborhood disorder; Housing condition / disrepair; Safety hazards).
• Crime Data Filtering:
  • Classify crime data into violent and non-violent crimes based on violent type.
• ACS Data Filtering:
  • Filter the tract level ACS data and computes six key socioeconomic variables: poverty rate, unemployment rate, renter share, vacancy rate, rent burden (30%), and nonwhite population share.

Exploratory Data Analysis

Bivariate Maps of Eviction and Crime Types

• North & West Philadelphia show both high evictions and high crime.
• Low evictions + low crime in South, Northwest & Northeast.
• The High-Low and Low-High tracts are areas that evictions can hardly be explained by violent.

Spatial Dependence & LISA Map

The result from global Moran’s I test indicated that there is a significant spatial autocorrelation exist in evictions data

Feature Engineering

Flag for Vulnerabe Tract

Spatial and temporal features

• Spatial features:
  • Distance to Hotspots (High–High) and Coldspots (Low-Low) from LISA map
  • Spatially lagged temporal lags (Spatio-Temporal Lag)
• Temporal features:
  • Temporal Lags for 1 month, 3 months, 6 months, and 12 months

Modeling Strategy

Model Specification

Component	Description
Models	Poisson and Negative Binomial models (handle over-dispersion)
Target Variable	Evictions (monthly eviction counts at the tract level)
Candidate Predictors	Violent and non-violent crime incidents; 311 service calls; Poverty rate, unemployment rate, renter share, vacancy rate, rent burden (30%), nonwhite population share, vulnerability flag; Distance to the nearest hot spot and cold spot; median gross rent, median home value; Number of tax-delinquent properties, average tax balance, total tax penalties.
Temporal Lags	1-, 3-, 6-, and 12-month lags of the target variable
Spatio-Temporal Lags	Spatio-temporal lags (spatially lagged temporal lags, W × past values) capturing spatial dependence

Training and validation

• Train–test split:
  • We use a forward-chaining temporal split to respect the time ordering of the data and avoid information leakage.
  • All observations from 2020 to 2024 are used for model training and all observations from 2025 are held out for testing.
• Spatial Cross-validation:
  • We use spatial cross-validation to ensure that model performance generalizes across geographic areas.
  • Each census tract is randomly assigned to one of K = 10 spatial folds and all months for a tract are placed in the same fold to avoid leakage across time.
• Baseline comparison:
  • Baseline models: Poisson and NegBin models using only contemporaneous predictors.
  • Add temporal lags: Add 1-, 3-, 6-, and 12-month lags.
  • Add spatio-temporal lags

Model Performance

How well does the model predict?

Model	MAE	RMSE
Poisson (baseline)	1.943	2.827
NegBin (baseline)	1.961	2.856
Poisson (+ temporal lags)	1.752	2.522
NegBin (+ temporal lags)	1.822	2.961
Poisson (+ temporal + spatio-temporal lags)	1.756	2.527
NegBin (+ temporal + spatio-temporal lags)	1.817	2.892

Performance across the city

Where the Model Struggles?

Limitations

Key Limitations

• Pandemic artifacts: Model trained on 2020-2025 (moratoriums, emergency programs). Patterns may not reflect baseline behavior.
• Spatial mismatch: We predict at census tract level, but RTC program operates by zip code.
• Missing detailed-data: We have filing counts, not record-level data on tenants, landlords, or outcomes.
• Historical bias risk: Model learns from past patterns—which reflect decades of discrimination. Could automate inequality if not careful.

Recommendations & Implementation

How the City Can Use This

Monthly Predictions Department of Planning and Development generates tract-level eviction risk forecasts. Which areas will see elevated filings in the next 30-90 days?

Concentrate Resources Target outreach, mediation capacity, and rental assistance to high-risk tracts. Use risk maps to design where to pilot new rental assistance or mediation programs. Amplify where need is highest—don’t abandon low-risk areas.

Bottom line: Eviction risk is predictable, concentrated, and actionable with oversight.

Q & A

Thank you.