🏠 California Housing Price Prediction

Capstone Project – Purwadhika JCDS 3004-009 (Module 3: Machine Learning)

Author: Bonifasius Sinurat

---

Overview

This project builds a machine learning model to predict the median house price in California using the 1990 census dataset. Beyond accuracy, the project translates predictions into business actions: ROI analysis, negotiation range, and key price drivers for decision makers.

Stakeholders: Chief Marketing Officer (CMO) & Sales Director Goal: Reliable pricing support for faster decisions and better margins.

---

Repository Structure

• California House/ – dataset
• catboost_info/ – CatBoost training logs
• png/ – images used in docs/README
  • actual_vs_predicted.png
  • feature_importances.png
  • pipeline.png
• app.py – Streamlit app
• CA_housing_price_regressor.sav – trained model (pickle)
• California Housing Price Prediction.pdf – slides
• capstone_m3_machine_learning.ipynb – main notebook
• requirements.txt – dependencies (or requirements)
• runtime.txt – Python version for Streamlit (e.g., 3.11)

---

Dataset

• Source: Pace & Barry (1997), California Housing (1990 Census)
• Rows: 14,448 districts
• Target: Median House Values
• Features: 9 numerical + 1 categorical (oceanProx)
• Notes:
  • Censoring at $500,001 for high-end prices
  • medIncome is recorded in tens of thousands USD (e.g., 3.5 → $35,000)

---

Methodology

Preprocessing & Feature Engineering

• Missing values: KNNImputer (137 NaN)
• Outliers: Winsorizer (Gaussian, fold 2.5)
• Scaling: RobustScaler
• Encoding: Binary / Ordinal / One-Hot
• Target transform: log ↔ exp (via TransformedTargetRegressor)
• New features:
  • roomsPerHouseholds, bedroomsPerRoom, popPerHouseholds
  • isManyRooms (rooms > 95th percentile)
  • housingAgeBin (young/middle/old)

Models Evaluated

Linear, Ridge, Lasso, KNN, Decision Tree, Random Forest, XGBoost, CatBoost

Best Model – CatBoost

• MAE: $27,533
• MAPE: 14.97% (meets ≤ 15% target)
• RMSLE: 0.2113

Top Drivers (Feature Importance)

1. Ocean Proximity (~25.1%)
2. Median Income (~21.6%)
3. Latitude (~14%)
4. Longitude (~14%)

---

Visuals

Pipeline

Feature Importance

Actual vs Predicted

---

Business Insights

• Negotiation Anchor (around a $260,000 prediction)

  • MAE-based range: $222,467 – $277,533
  • MAPE-based range: $212,575 – $287,425

• ROI Check (example scenario)

  • Total investment: $236,000
  • Selling fee: 6%
  • ROI target: ≥ 10% (profitability threshold)
  • → No Deal Below $276,170

• Actionable ROI Framework

  ROI Range	Decision	Description
  < 8%	❌ Reject	Below minimum margin
  8–10%	⚠️ Review	Requires manual validation
  ≥ 10%	✅ Proceed	Meets business target
  ≥ 14%	💰 Prioritize	High-margin opportunity
  ---	---	---

• Financial Impact

  • Each $1 prediction error affects ≈ $0.06 in commission.
  • Reducing MAE by $3,223 (CatBoost vs RandomForest) → ≈ $193K annual revenue uplift (based on 1,000 transactions).

Business Takeaway

With a consistent ROI-based decision rule, the model enables property investors and sales teams to act with confidence. Every $1,000 improvement in prediction accuracy protects roughly $60 in commission, turning data accuracy into profit efficiency.

---

How to Run (Local)

1. Install dependencies:

   pip install -r requirements.txt

2. Run the notebook:

   Open capstone_m3_machine_learning.ipynb in Jupyter/VS Code and run all cells.

Limitations & Next Steps

• Error grows on luxury/high-end prices due to censoring and limited samples
• Add geospatial signals (distance to coast/city center/POIs)
• Expand high-end segment coverage
• Consider quantile regression / conformal prediction for price intervals
• Regular retraining & validation; use SHAP for explainability

---

Deployment

This repo includes runtime.txt (e.g., python-3.11) for Streamlit Cloud.

Live App: California Housing Price Predictor

---

License

Educational use for Purwadhika Module 3 – Machine Learning capstone project.

For other uses, please provide proper attribution.