DATA SCIENCE CASE STUDY

Predicting property age without year leakage.

A deterministic random forest classifies pre-1980 homes from structural, condition, architectural, and transaction attributes—while explicitly excluding the build year.

Read the analysis

FIXED 20% HOLDOUT · 4,583 ROWS

Strong performance, measured from several angles.

92.36%Accuracy
92.13%Balanced accuracy
0.9384F1
0.9768ROC-AUC

METHOD

The feature boundary is part of the test suite.

The model never receives yrbuilt, before1980, or the parcel identifier. A stratified split, fixed random seed, pinned dependencies, and generated metrics make the result repeatable.

Confusion matrix showing 1,568 correct newer-home classifications, 2,665 correct older-home classifications, and 350 total errors.
Bar chart of the twelve most important random-forest features.

EXPLAINABILITY

Architectural and structural signals lead.

One-story architecture, living area, story count, bathroom count, garage type, transaction values, and quality indicators carry substantial model signal. These are model importances—not causal claims.

LIMITATIONS

A portfolio result, not a production appraisal.

The data represents one assessment context. External validation, drift monitoring, and clearer source licensing would be required before operational use. The raw CSV remains upstream and is downloaded only when the analysis runs.