Problem statement

Eviction risk modeling is often undermined by data leakage, ad-hoc feature pipelines, and inconsistent train/test splits that do not reflect how the model would be used in production. The goal of this project is to build a reproducible framework that (1) enforces strict temporal separation, (2) supports rapidly iterating on feature sets and models, and (3) produces comparable evaluations across experiments.

Solution approach

I built a modular, interface-driven pipeline with:

• Configurable temporal cohorts (train and multiple future test windows) to enforce leakage-free evaluation.
• A feature-generator architecture (pluggable “generators” for property-, owner-, and case-level features) with consistent naming, documentation, and validation.
• A modeling layer supporting baseline models and more expressive learners (e.g., gradient-boosted trees), with cross-validation and tracked hyperparameters.
• Evaluation utilities for PR-AUC / ROC-AUC, calibration checks, and cohort-level breakdowns.

Results

• A single framework that can run repeated experiments across time windows and feature bundles without re-writing data prep logic.
• Clearer diagnosis of performance drops after refactors by isolating differences in feature definitions, cohort construction, and training protocol.
• A maintainable foundation for adding new feature families (e.g., recency/volatility trends, network/owner aggregation, and policy-relevant derived signals).

Main conclusions

A rigorous eviction-risk workflow requires the pipeline to be as “auditable” as the model: when temporal splits and features are first-class objects, it becomes far easier to (a) prevent leakage, (b) compare experiments honestly, and (c) ship improvements without regressions.