Overview
This graduate capstone develops a residential prediction system alongside a client-facing model evaluation effort. The team maintains a shared data foundation and evaluation contract while developing independently owned modeling approaches.
My work combines project coordination, data preparation, model implementation, and reproducible evaluation. This public case study describes the engineering approach without exposing client identity, property records, private outputs, or client infrastructure.
Problem
Property-sale history supports several different questions, and those questions require different targets. A recorded sale does not reveal every listing attempt, withdrawal, or homeowner intention. It also does not directly measure exact days on market.
The current workflow therefore separates sale-price regression from sale likelihood within 1, 3, 6, or 12 months. These tracks share cleaned history but use distinct modeling tables and outcomes.
What I Built
- Contributed to a reproducible foundation for property and transaction history, with feature availability and temporal leakage treated as evaluation constraints.
- Implemented CatBoost model packages for sale price and four binary sale-likelihood horizons, preserving native categorical handling and numeric missingness.
- Implemented separate cascade-forest regression and classification packages for comparison.
- Supported a shared model-development workflow with consistent dataset versions, partitions, submission contracts, and centralized scoring.
- Separated smoke artifacts from evaluation-eligible models and required frozen model metadata before benchmark prediction.
Results and Current State
The repository contains distinct modeling packages and a common evaluation structure, replacing an undifferentiated price-and-time-to-sale framing with explicit prediction tasks. Development experiments and model comparisons remain separate from final-test and client-benchmark claims.
The capstone is in progress. This page does not claim a selected production model, a final client outcome, or validated deployment performance. Private model metrics and row-level outputs are intentionally excluded.
What I Learned
The target definition is part of the product contract. Separating sale likelihood from listing duration, and development evidence from final evaluation, makes the work easier to compare and harder to overstate.
Limitations and Next Steps
Recorded sales cannot recover unobserved listing attempts or seller intent. Model comparisons depend on the approved data release and temporal split, and final evaluation must remain separate from tuning.
The next milestones are governed model comparison, error analysis, and integration of an approved model. Any public result needs a client-safe summary and evidence appropriate to the claim.