Predictive, not causal
TreeSHAP decomposes model predictions. It does not estimate what would happen if a feature were changed.
Scientific documentation
Research design, traceability, and limits
This page describes the evidence already produced by the authoritative notebook and geographic SHAP workbook. The application is a presentation layer, not a second analytical pipeline.
How accurately can 2024 inflation-adjusted property values for owner-occupied one-family houses in five U.S. states be predicted from 2020–2023 ACS PUMS property, household, financing, utility, and State-PUMA characteristics?
Data
ACS PUMS housing microdata
California, Florida, New York, Tennessee, and Texas; 960,182 development records from 2020–2023 and 268,930 temporal-evaluation records from 2024.
The response is `VALP`, adjusted with `ADJHSG` to comparable dollar values. Housing-unit survey weights are used in relevant summaries.
Model
Leakage-safe temporal design
All model comparison and feature reduction occurs inside development data. State–PUMA uses five-fold cross-fitted target encoding. No county column or target-related field enters the predictor matrix.
The selected reduced XGBoost model uses a `log1p` target and 18 documented semantic features.
Interpretation
Survey-weighted TreeSHAP
Exact TreeSHAP values are calculated for all eligible 2024 records, then aggregated with housing-unit weights. Mean absolute SHAP measures contribution strength; mean signed SHAP measures average direction from the model baseline.
SHAP units are log1p property-value units, not direct dollars.
Reproducible workflow
From scientific source to public explorer
- 01Develop
2020–2023 common folds and candidate-model comparison.
- 02Select
Reduced XGBoost chosen by development MAE.
- 03Evaluate
2024 opened once for temporal performance.
- 04Explain
TreeSHAP for 268,930 2024 records.
- 05Aggregate
Survey-weighted state, PUMA, and approximate-county summaries.
Selected inputs
Feature categories
Property
Bedrooms · Other rooms · Lot size · Year built · Structure type · Heating fuel
Survey
Survey year
Household
Household income · Household size · Year moved in · Household type
Financial
First mortgage payment · Condo or HOA fee
Utilities
Electricity cost · Gas cost · Other fuel cost · Water and sewer cost
Geography
State–PUMA
Claim boundaries
What the study can—and cannot—say
Survey-reported response
VALP is a respondent estimate that can be rounded, allocated, and state-specific top-coded; it is not an appraisal.
Geographic precision
State and PUMA are observed. County labels are descriptive assignments from dominant PUMA–county overlap after modeling.
Survey uncertainty
Housing-unit weights support aggregation, while the record bootstrap is a stability check—not an ACS SDR margin of error.
Scope
The results cover five states and short-horizon temporal generalization to 2024, not every U.S. housing market.
Ordinal fields
Year-built, year-moved, and lot-size values are ordered ACS categories, not equally spaced measurements.
Traceability
Every rendered result points back to the completed analysis.
The deterministic exporter parses the geographic workbook programmatically, normalizes notebook-produced CSVs, validates geography and feature consistency, and records SHA-256 source hashes.
Open source manifest