Section 1: The Problem
More than 2.9 million eviction filings were recorded in the United States in 2023, and filings in many cities now exceed pre‑pandemic levels, pushing families toward homelessness, job loss, and long‑term health problems. Eviction is strongly linked to material hardship, depression, adverse birth outcomes, and long‑term housing instability, especially for low‑income renters and people of color. Losing housing also damages credit, making it harder to secure future rentals and deepening poverty and inequality.
The economic stakes are large. Emergency shelter, foster care placements linked to family homelessness, and downstream healthcare costs mean that a single family eviction can ultimately cost the public sector thousands of dollars, far more than short‑term rental assistance. HUD’s 2024 homelessness assessment notes that 2024 median rent was 18 percent higher than in 2020 and that a severe shortage of affordable units is pushing more renters to the brink. Yet most jurisdictions still lack systems that can identify those at highest risk early enough to intervene.
Traditional approaches depend on tenants self‑identifying in crisis, legal aid waiting at the courthouse, or broad‑brush area targeting that sends outreach workers to entire neighborhoods instead of the specific blocks or buildings where evictions are about to spike. These methods are reactive and inefficient: by the time a family shows up in court, options are narrower, arrears are larger, and the chance of avoiding displacement is much lower.
Section 2: What Research Shows
Researchers have used rich court, demographic, and economic data to build models that forecast where evictions will happen and which renters are most likely to be filed against. One deep multi‑view neural network, MARTIAN, forecasts the number of tenants at risk of formal eviction in each census tract of Dallas County one to three months into the future using court filings, labor statistics, and census data. Averaged across horizons of 1–3 months, MARTIAN achieved a root mean square error (RMSE) of 4.443 filings per tract, compared with 4.744 for the best traditional baseline (Random Forest), about a 5 percent error reduction and higher rank‑ordering of hotspots (Spearman 0.685).
Other work shows that model‑generated risk scores can meaningfully improve who gets help. A 2024 study, “Beyond Eviction Prediction,” used random forests, XGBoost, and neural networks to predict which properties would receive court eviction filings by combining historical filings, neighborhood indicators, and ownership characteristics. When the risk scores were used to simulate targeted outreach, a fixed‑size team of caseworkers could reach 8.5 percent more eviction‑prone properties than a neighborhood‑based strategy and 28 percent more than a strategy focused only on buildings with recent evictions.
At the individual level, logistic regression and random forest models trained on national homelessness and housing data can concentrate risk: in one HUD‑VASH cohort, the top 10 percent of veterans by predicted risk accounted for roughly 28–32 percent of actual negative housing exits, compared with 10 percent under random targeting. A 2025 scoping review of homelessness prediction models in high‑income countries found growing use of machine learning and good discrimination in many models, but noted that only two had undergone external validation, underscoring how far practice lags behind technical development.
Section 3: What the Real World Shows
When prevention support actually reaches households, outcomes change. HUD’s 2024 Final Report on the Eviction Protection Grant Program, covering more than 50,000 tenant households who received legal aid, found that nearly 36 percent had their tenancy preserved, 28 percent negotiated a settlement, and another 28 percent achieved other protective outcomes such as prevented filings, sealed records, or additional time to move; fewer than 2 percent experienced default eviction judgments. Among households receiving even brief legal services, 35.1 percent kept their tenancy, 7 percent prevented a filing altogether, and only about 13 percent had an eviction judgment or displacement.
These are not model‑only simulations; they are real cases where timely legal and financial assistance altered the trajectory of families’ lives. A Homelessness Prevention Unit (HPU) launched in 2021, using a predictive model to identify people at highest risk of homelessness and offer intensive services, reported that enrollment during the pilot period was associated with a 71 percent decrease in subsequent use of homelessness services among participants, ahead of a formal randomized trial. Early analyses of similar predictive programs suggest that focusing limited prevention dollars on those with the very highest predicted risk can avert more shelter entries per dollar than first‑come‑first‑served approaches.
A 2025 scoping review concluded that homelessness and eviction risk models have real potential to improve targeting and that several systems now use risk scores to allocate prevention slots, but the review also highlighted major gaps in calibration, external validation, and reporting of equity impacts. Together, these studies suggest that the combination of good prediction and proven interventions—like legal representation and short‑term rental aid—can materially reduce evictions and shelter entries when deployed in the field.
Section 4: The Implementation Gap
Despite strong retrospective performance and promising pilot outcomes, very few cities actually use these tools day to day. The MARTIAN model has been developed with the Child Poverty Action Lab in Dallas and evaluated extensively, but as of its publication the authors noted that “the usability of MARTIAN is under review by domain experts,” meaning it was not yet embedded in operational decision‑making. A 2024 housing data paper on making eviction data actionable describes how local advocates spend significant effort just cleaning and interpreting basic filing data, with little capacity left to integrate advanced risk scores.
Several concrete barriers show up repeatedly. First, data fragmentation and latency: eviction filings live in court systems, income and employment data in separate agencies, and many courts still publish only PDF dockets, making real‑time feeds into ML tools difficult. Second, workflow fit: legal aid and housing nonprofits are understaffed and already triaging walk‑ins; asking them to reorient around algorithm‑generated lists without more staff or flexible funding can feel unrealistic.
Third, trust and fairness concerns slow institutional buy‑in. Advocates worry that predictive tools trained on biased historical data will concentrate surveillance or enforcement in already over‑policed neighborhoods or miss undocumented tenants and informal rental arrangements. The 2025 scoping review found that very few homelessness prediction models report equity metrics by race or ethnicity, which makes it harder for agencies to assess disparate impact. Finally, incentives are misaligned: prevention budgets are often small, short‑term, and siloed, while the savings from avoided shelter, hospitalizations, or school disruptions accrue to different systems than the one paying for rental assistance.
If we zoom out, the field looks like this:
- Many academic models built, few externally validated.
- Even fewer connected to concrete interventions.
- Only a tiny fraction embedded into routine government or nonprofit workflows.
This “last mile” gap explains why, even as research papers show 5–10 percent gains in predictive accuracy and real programs show 30–70 percent improvements in housing stability for those served, most renters at risk of eviction never appear on any risk dashboard at all.
Implementation gap illustration
- In research, roughly all models are “developed,” but only a small minority are externally validated and a smaller subset deployed in real systems, as summarized by recent reviews.
- Across jurisdictions, expert commentary from housing labs suggests that only a small single‑digit percentage of cities use any machine‑learning‑based eviction or homelessness risk tool; most rely on traditional, reactive methods.
(The chart below encodes that pattern as an illustrative 100/10/2 and 5/95 split based on these qualitative estimates.)
Section 5: Where It Actually Works
Where data‑driven eviction prevention has gained traction, a few common ingredients show up. Dallas’s partnership between Penn State researchers, CPAL, and Texas Housers demonstrates how sustained collaboration can align a model like MARTIAN with the questions local advocates actually face: which tracts are heating up, and how should limited rental assistance be distributed over the next quarter. In these settings, risk forecasts are presented as simple heat maps and ranked lists, not black‑box scores, and they feed into existing processes like outreach campaigns rather than replacing human judgment.
HUD’s Eviction Protection Grant Program is another example of success without heavy tech: by funding legal aid and requiring systematic outcome tracking, HUD effectively created a nationwide “implementation lab” for what happens when tenants actually have representation. Emerging prevention units that combine these proven supports with predictive targeting—like the HPU described by the California Policy Lab—are beginning to show that pairing algorithms with flexible money and seasoned caseworkers can measurably cut shelter use.
Section 6: The Opportunity
Eviction and housing‑loss prediction is a domain where the models are already good enough, the interventions clearly work for those who receive them, and the main challenge is building trustworthy, well‑funded systems that act early instead of reacting late.
What would actually move the needle?
- Build shared data pipelines from courts, housing authorities, and social services so that risk models and dashboards update weekly instead of yearly.
- Pair every predictive tool with guaranteed prevention capacity—legal aid slots and rapid rental assistance—so risk scores translate into real offers, not just lists.
- Require external validation, calibration, and equity reporting (by race, income, neighborhood) for any publicly funded eviction or homelessness risk model.
- Invest in co‑design with tenant organizations and legal aid so scores are understandable, contestable, and embedded in familiar workflows.
- Align incentives by letting local agencies that invest in prevention share documented savings from reduced shelter use, foster care entries, and emergency healthcare.
Beyond Eviction Prediction: Leveraging Local Spatiotemporal Public Data to Inform Housing Policy, ACM Conference on Fairness, Accountability, and Transparency (FAccT), 2024.
U.S. Department of Housing and Urban Development. Evaluation of the Eviction Protection Grant Program, Final Report Preview, 2024.
California Policy Lab. Early Outcomes from the Homelessness Prevention Unit, 2024.
Tabar, M. et al. Forecasting the Number of Tenants At‑Risk of Formal Eviction, IJCAI, 2022.
HUD. 2024 Annual Homelessness Assessment Report (AHAR) to Congress, released January 2025.
Homelessness prediction models in high‑income countries: a scoping review, 2025.
Actuarial Prediction versus Clinical Prediction of Exits from a Housing Program (HUD‑VASH), 2022.
Data‑Driven Models for Eviction Prevention, Data‑Smart City Solutions, Harvard Kennedy School, 2025.
Making Public Eviction Data Actionable for Housing Justice, ACM Digital Library, 2024.
Rental assistance, housing insecurity, & well‑being: Spillover effects amid rising rents, Social Science & Medicine, 2025.
Aethel. 
Leave a comment