SUPERVISED LEARNING · REAL ESTATE ANALYTICS
Housing Price Estimation
Using Machine Learning
A comparative study of Linear Regression, Decision Trees, XGBoost, and Ensemble models for predicting residential property values using real-world housing data.
Best R² Score
99.45%
XGBoost Regressor
↑ Best Model
Lowest RMSE
$6,510
XGBoost prediction error
↓ Lowest Error
Lowest MAE
$1,202
Ensemble Model
↓ Most Consistent
Models Tested
4
LR · DT · XGB · Ensemble
Comparative Study
R² Score Comparison
RMSE vs MAE by Model
MAPE (Mean Absolute Percentage Error) — Lower is Better
Ensemble Model
~0.8%
Decision Tree
~1.2%
XGBoost
~1.5%
Linear Regression
~11.5%
Model Performance Table
| Rank | Model | RMSE ($) | MAE ($) | R² Score | Assessment |
|---|---|---|---|---|---|
| 1st | XGBoost |
6,509.66 | 2,083.93 | 0.9945 | Excellent |
| 2nd | Ensemble Model |
7,007.30 | 1,202.48 | 0.9936 | Excellent |
| 3rd | Decision Tree |
13,426.56 | 1,518.85 | 0.9765 | Good |
| 4th | Linear Regression |
30,580.89 | 15,701.19 | 0.8781 | Limited |
Algorithm Details
XGBoost
Extreme Gradient Boosting
Sequential ensemble method that trains models iteratively, each correcting the errors of the previous one. Features built-in L1/L2 regularization, parallel processing, and automatic missing value handling.
R²: 0.9945
RMSE: $6,510
Handles non-linearity
Anti-overfit
Ensemble Model
Stacked LR + DT + XGBoost
Stacked ensemble combining three base models (Linear Regression, Decision Tree, XGBoost) with a Random Forest meta-learner. Uses 5-fold cross-validation to generate out-of-fold predictions.
R²: 0.9936
MAE: $1,202
Best generalization
cv=5
Decision Tree
CART Regressor
Hierarchical splits on feature values using MSE as the splitting criterion. Interpretable and computationally efficient. Risk of overfitting mitigated by depth constraints and cross-validation.
R²: 0.9765
RMSE: $13,427
Interpretable
Fast inference
Linear Regression
Multiple OLS Regression
Fits a hyperplane to minimize MSE via gradient descent. Assumes linearity, independence, homoscedasticity, and normality of residuals. Struggles with the non-linear interactions in housing data.
R²: 0.8781
RMSE: $30,581
Baseline model
Low complexity
Literature Comparison
| Study | Best Model | R² Score | Key Notes |
|---|---|---|---|
| This Project | XGBoost | 0.9945 | Lowest RMSE, highest accuracy across all models |
| This Project | Ensemble Model | 0.9936 | Lowest MAE, strong overall generalization |
| Abdul-Rahman et al. | XGBoost | 0.9120 | Boosting models for Kuala Lumpur market |
| Shalini et al. | Random Forest | 0.9656 | Focused on preprocessing and data-driven pipeline |
| Z. Li | Decision Tree | 0.8016 | Comparative study of regression models |
House Price Estimator
Enter Property Details
Estimation Result
$0
Estimated current market value
5-YEAR PROJECTION
—
MODEL USED
—
HOUSE AGE
—
REMODEL AGE
—
EST. PRICE RANGE
—
CONFIDENCE
—
Fill in the property details and click Estimate Price to get a prediction.
Model Error Bounds (Expected ±)
XGBoost
±$6,510
Ensemble
±$7,007
Decision Tree
±$13,427
Lin. Regression
±$30,581
Feature Importance
Top Features by Predictive Impact
Values represent Pearson correlation coefficient with SalePrice. Negative values indicate inverse relationship.
Feature Correlation Radar
Engineered Features
🏚️
HouseAge
Derived from YearBuilt. Older homes tend to require more maintenance, affecting price negatively. Correlation with SalePrice: −0.41
2025 − YearBuilt
🔨
RemodelAge
Years since last renovation. Recently remodeled homes command higher prices. Correlation with SalePrice: −0.36
2025 − YearRemodAdd
📐
TotalBsmtArea
Combined BsmtFinSF2 + TotalBsmtSF for total basement coverage. High correlation with TotalBsmtSF (0.94). Correlation with SalePrice: +0.52
BsmtFinSF2 + TotalBsmtSF
📊
LogSalePrice
Log-transformed SalePrice to reduce right skewness. Makes distribution more normal, improving linear model performance. Correlation: +0.59
np.log(SalePrice)
🗺️
LotAreaPerRoom
Adjusts lot size by overall condition. A large lot in poor condition has lower value than a smaller lot in excellent condition.
LotArea / (OverallCond + 1)
⚖️
Class Imbalance
MSZoning (0.78), BldgType (0.83), LotConfig (0.73) show mild imbalance. OverallCond and temporal features are well-balanced. No resampling required.
Mild: MSZoning, BldgType
6-Stage Development Pipeline
01Business Understanding
→
02Data Understanding
→
03Data Preparation
→
04Modelling
→
05Evaluation
→
06Deployment
Data Preprocessing Steps
01
Missing Value Imputation
Mode for categoricals (MSZoning, Exterior1st); Median for numerics (BsmtFinSF2, TotalBsmtSF)
02
Duplicate Removal
Checked and removed duplicate rows to prevent data leakage
03
Normalization
StandardScaler (mean=0, std=1), MinMaxScaler, and RobustScaler applied to numerical features
04
Label Encoding
Categorical variables converted to numeric for ML compatibility
05
Train/Test Split
80% training / 20% testing with 5-fold cross-validation for stacking
Evaluation Metrics
RMSE — Root Mean Squared Error
Penalizes large errors more heavily. Good for detecting outlier predictions. Lower = better.
MAE — Mean Absolute Error
Average absolute deviation. More actionable in real-world pricing decisions. Lower = better.
R² — Coefficient of Determination
Variance explained by the model. XGBoost explained 99.45% of price variability. Higher = better.
MAPE — Mean Absolute Percentage Error
Error as a percentage of actual value. Undefined when actual = 0. Best for relative comparisons.
Challenges Encountered
• Data Preprocessing — missing values, inconsistent formats, and outliers required careful handling
• Feature Selection — identifying high-impact features and managing multicollinearity
• Computational Resources — XGBoost and ensemble training caused memory issues on limited hardware
• Overfitting — some models trained well but failed to generalize to test data
• Frontend Integration — connecting ML backend with interactive web dashboard
Future Work
• Data Enrichment — proximity to schools, transport, hospitals; mortgage rates; demand trends
• Advanced Features — total livable area, renovation history, crime rates, air quality
• Neural Networks — Deep learning for complex non-linear relationships
• Web App Deployment — Flask or FastAPI backend for real-time predictions
• Model Monitoring — drift detection and retraining pipeline for market adaptation