My data science portfolio
B”H
How accurately can we predict a home’s sale price using routine listing and property characteristics, and which features most strongly drive residential property values in Ames, Iowa?
This project builds transparent and performant models to predict home sale prices from property characteristics using the widely-used Ames Housing dataset, which contains 1,460 property sales with 79 explanatory variables spanning the 2006-2010 period in Ames, Iowa. The goal is to provide accurate price estimation for real-estate professionals, appraisers, and homeowners while clearly communicating which features most strongly drive price.
The analysis encompasses three major components:
Data Exploration & Preparation - Handling high-missing columns (PoolQC 99.5%, MiscFeature 96%, Alley 93.8%, Fence 80.8%), treating “NA” values that encode true absence vs. missingness, applying log transformation to right-skewed SalePrice target, and preparing mixed-type features (lot/neighborhood descriptors, structure quality/size, basement/garage details, kitchen/bath conditions)
Model Development & Comparison - Training baseline models including Ridge regression, Random Forest, and Gradient Boosting (XGBoost/LightGBM) with 5-fold cross-validation, using median/mode imputation, scaling, and one-hot encoding, with early stopping and hyperparameter tuning to optimize RMSE, MAE, and R² metrics
Interpretation & Feature Analysis - Extracting standardized coefficients from linear models, permutation importance and SHAP values from tree-based models, partial dependence plots, and correlation heatmaps to identify dominant price drivers (OverallQual, GrLivArea, TotalBsmtSF, garage capacity) and communicate stakeholder-friendly explanations
The project demonstrates comprehensive regression modeling practices including proper handling of mixed data types, missing value semantics, transformation strategies for skewed targets, and model-agnostic interpretability techniques. Final results show Gradient Boosting achieving the strongest performance (RMSE ≈ 28.4K, R² ≈ 0.851), with size and overall quality confirmed as dominant price drivers through both correlation structure and tree-based importance analysis.