Applied predictive modeling
Pricing Predictions
An archived public-data pricing study comparing regularized regression and random forests, with evaluation issues documented for correction.
data → model → 2023 holdout
Illustrative project overview
The problem
The paper evaluates whether home attributes can support more disciplined price estimates for manufactured homes using a nationwide public-use survey dataset.
How accurately can home characteristics such as size, bedrooms, foundation type, elevation method, and securing method explain 2023 manufactured-home prices?
Manufactured-home pricing depends on physical attributes and construction choices; the study frames the task as a reproducible supervised prediction problem for buyer comparison.
How it works
- 01
HUD / Census survey records
- 02
Attribute preparation
- 03
Earlier-year training data
- 04
Model comparison
- 05
Evaluation review
From input to output
- 01Use HUD Manufactured Housing Survey public-use data spanning 2021 through 2023.
- 02Address missing values according to the report’s documented preprocessing choices and inspect non-linear size effects.
- 03Compare Lasso, Ridge, and Random Forest models with ten-fold cross-validation for model selection.
- 04Train on 2021-2022 data and evaluate on 2023 as an explicit out-of-sample holdout.
Data sources
- HUD Manufactured Housing Survey Public Use File
- U.S. Census Bureau administered survey
- 2021-2023 manufactured-home records
Technical approach
- OLS exploratory analysis
- Cubic size terms and interactions
- Lasso and Ridge regularization
- Ten-fold cross-validation
- 200-tree Random Forest
What you can inspect
- Research paper and embedded R output
- Feature and model specification
- Historical model comparison
- Written interpretation of practical use
- The HTML deliverable documents the R preprocessing and model-comparison workflow.
- Numeric performance claims are withheld because missing target prices were imputed and the published figures do not fully match the visible output.
The study demonstrates a reproducible analytical artifact, but its current evaluation does not establish reliable out-of-sample pricing accuracy.
Where this is useful
- Evaluate only observed target values and fit preprocessing on training data.
- Reconcile the metric definition and saved model output before publishing revised performance.
Checks & engineering decisions
Verification evidence
- The current target-imputation and metric-definition issues require a corrected evaluation on observed target values.
- Published model figures and saved HTML output are inconsistent; no numeric performance conclusion is presented.
Key decisions
- Keep a later calendar year outside training, with train-only preprocessing.
- Compare regularized linear and non-linear models on the same observed targets.
- Document the R² definition and omitted local-cost context before interpreting results.
Role & status
Academic research - data preparation, predictive-model comparison, validation design, and written interpretation.
Research archive · Evaluation under review Source review: 2026-10-14.
Limitations
- Imputed target prices cannot be treated as observed ground truth in evaluation.
- The R² denominator and inconsistent historical figures require correction.
- Unobserved upgrades, local installation costs, and manufacturer differences limit the feature set.
Explore the evidence
Archived research paper and embedded R output. The public-use data is not copied into the portfolio, and unverified performance figures have been removed.