Modelling Spatial Heterogeneity in Singapore's HDB Resale Market: A comparative Analysis of OLS and Euclidean Distance-Based GWR

This repository contains the full analytical workflow, data-processing scripts, and modelling code for this project. This project investigates spatial variations in housing price determinants across Singapore's public housing (HDB) market using a combination of global and local spatial regression techniques.

📌 Project Overview

Housing prices in Singapore exhibit pronounced spatial variation driven by accessibility, amenities, built-environment characteristics, and neighbourhood differences. Traditional global regression models (e.g., OLS) assume spatial stationarity, which often fails in a highly differentiated urban landscape like Singapore.

This project develops a comparative framework:

• Global OLS hedonic modelling
• Local Euclidean Distance-based Geographically Weighted Regression (ED-based GWR)(adaptive kernel, nearest-neighbour bandwidths)

📂 Repository Structure

  │code/
  ├── 00DataPreprocessing.ipynb
  ├── 01DataPreparation.ipynb
  ├── 02Prepare Analysis-ready HDB features.ipynb
  ├── 03Exploratory Spatial Analysis OLS.ipynb
  ├── 03Exploratory Spatial Analysis OLS & Moran's I.ipynb
  ├── 04Local Modeling (GWR).ipynb
  ├── SpatialJoin_datapreprocess_OLS_GWR.ipynb
  │
  ├data/                   # raw & cleaned datasets
  ├── figures/             # generated maps, diagnostics, visuals
  ├── results/             # OLS & GWR outputs, CSVs, summaries
  │
  └── README.md

📊 Key Research Questions

1. How do structual, locational, and accessibility attributes influence HDB resale prices?
2. Do these relationships vary across space, and to what extent?
3. Can ED-based GWR improve explanatory power and reduce spatial autocorrelation compared to OLS model?
4. How do local coefficients and local R^2 reveal subzone-level housing market heterogeneity?

🧰 Methodology Summary

1. Data Processing & Integration Datasets include:

• HDB resale transactions (2023)
• Subzone boundaries (Master Plan 2019, Singapore)
• Accessibility metrics (Raw data processed):
• • Distance to nearest MRT station
  • Distance to nearest healthcare facility (hospital, older house)
  • Distance to nearest recreaional facility (sports centre, recreation spot)
  • Bus stop counts within 400m each HDB flat
• Housing attributes (HDB type, floor area, building age, storey)

Processing tasks:

• Spatial join (subzone boundaries, each HDB transaction)
• Euclidean distance calculations
• Variable engineering (log resale price)
• Construction of analysis-ready dataset

Relevant notebooks: 00DataPreprocessing.ipynb, 01DataPreparation.ipynb, 02Prepare Analysis-ready HDB features.ipynb.

2. Global Spatial Analysis (OLS + Moran's I) Performed tasks:

• Global OLS model

• Diagnostics:

• • R^2, adjusted R^2, AICc
  • VIF multicollinearity checks
  • Residual analysis

• Global Moran's I on OLS residuals (Detects spatial autocorrelation indicating OLS deficiency)

Notebook: 03Exploratory Spatial Analysis OLS.ipynb, 03Exploratory Spatial Analysis OLS & Moran's I.ipynb

3. Local Spatial Modelling (ED-Based GWR Model) Key features:

• Adaptive bisquare kernel
• Optimal bandwidth selection using AICc
• Subzone-level coefficient estimates
• Maps of local R^2, t-statistics, and parameter surfaces

Notebook: 04Local Modeling (GWR).ipynb

📈 Main Findings (Summary)

• ED-GWR dramatically improves model fit (R^2 increases from 0.762 to 0.978).
• Spatial autocorrelation in OLS residuals is largely eliminated under GWR.
• Housing price drivers vary substantially across Singapore, particularly in central vs. outer subzones.
• Local R^2 reveals strong heterogeneity, identifying areas where accessibility or structual attributes dominate price formation.

🗂 Reproducibility Guide

Requirements Recommended environment:

• Python 3.10+
• GeoPandas
• libpysal, esda
• mgwr
• shapely, pyproj
• contextily
• pandas, numpy, matplotlib

Install dependencies:

pip install -r requirements.txt

How to Run

1. Clone the repository (Use SSH):

git clone git@github.com:yanjie99/HedonicPriceModel.git
cd HedonicPriceModel/code

2. Open JupyterLab or Jupyter Notebook.
3. Run notebooks in numerical folder:

00 → 01 → 02 → 03 → 04 → SpatialJoin_datapreprocess_OLS_GWR (subzone-level)