Predicting Housing Complaints Before They Occur

Using large-scale predictive modeling to identify where future complaints are most likely—despite sparse and uneven historical data.

Independent applied ML project using public complaint records from the New York City Department of Housing Preservation and Development (HPD). This was not a paid client engagement through Kogan AI Consulting.

Results at a glance

Results are from a retrospective evaluation of previously generated scores against 2026-Q2 outcomes. Lift compares the top score decile with the overall complaint rate; it is not a reduction in complaints or a business ROI figure.

The challenge

A citywide forecast cannot tell an operations team which building or apartment is most likely to need attention. This project asked a harder question: for each building × apartment × complaint category, what is the probability of a complaint in the next quarter?

Most combinations have few relevant prior complaints, many have none, and different complaint categories behave differently over time. A useful model needed to distinguish genuinely higher-risk records from millions of low-risk ones, while producing probabilities that could be evaluated against later outcomes.

What I built

I designed a quarterly forecasting pipeline using LightGBM, historical complaint activity, building attributes and time-based features. It combined short- and longer-term history, within-category signals, and patterns from other complaint types.

A key design choice was to score both combinations with established complaint history and cold-start combinations without it. I used forward-looking validation, risk-ranking metrics, and probability calibration to understand performance across these different populations—not simply report a favorable global accuracy score.

What the results showed

A later evaluation compared approximately 7 million existing scores with actual complaints recorded in 2026-Q2. The top 10% of scored records had a complaint rate 6.3 times the overall average, demonstrating meaningful risk concentration.

The evaluation also revealed where the model needed caution: some rare cold-start categories were less reliable, and Heat/Hot Water required attention to seasonal calibration. Those findings informed recommended category-level guardrails rather than a claim that every score was equally actionable.

This was evidence of predictive usefulness—not proof of adoption by an inspection team or downstream operational savings.

Why this matters

This project demonstrates how I approach difficult predictive problems: define the decision-relevant target, work with imperfect and sparse data at scale, test performance on future outcomes, and identify where a model is dependable enough to guide decisions—and where it isn’t.

Explore the technical approach · View services · Discuss a similar project