Authors: Lim Zheng Gang, Eugene Matthew Cheong, Pius Yee
Creating a model to help empower our agents to provide customers invaluable insights for making well-informed decisions on pricing of houses.
Our current manual process of evaluating property data for customers is time-consuming and prone to inconsistencies, as it relies on individual agents generating analyses independently.
To address this challenge, we have developed a model for this project, aimed at empowering our agents to deliver consistent property evaluations to customers.
By leveraging this model, our agents can provide invaluable insights and make data-driven decisions, ensuring a standardized approach to property valuation.
HDB resale transaction data - sourced from GA
Supermall location (longitutde and latitude) - data extracted from HDB resale transaction data
| Feature | Type | Dataset | Description |
|---|---|---|---|
| id | int64 | train.csv | Transaction ID for HDB resale |
| resale_price | float64 | train.csv | Resale price of the HDB transaction |
| mid | int64 | train.csv | Middle floor from the range of floor level |
| floor_area_sqft | float64 | train.csv | Floor size of the HDB flat |
| max_floor_lvl | int64 | train.csv | Max floor level for the block of transacted unit |
| Feature | Type | Dataset | Description |
|---|---|---|---|
| mrt_nearest_distance | float64 | train.csv | Distance to the nearest MRT |
| from_centre_distance | float64 | train.csv | Distance from Central area. GA campus set as the central area |
| mh | float64 | sm_location_supermalls.csv | Distance to supermall i.e. the popular malls with very high number of customers |
| Feature | Type | Dataset | Description |
|---|---|---|---|
| mature | int64 | train.csv | Mature HDB estate |
| planning_area_category_Group1 | int64 | train.csv | Includes areas such as Tanglin, Bukit Timah, Outram, Downtown Core, Bishan. |
| planning_area_category_GroupA | int64 | train.csv | Includes areas such as Ang Mo Kio. |
| planning_area_category_GroupCM | int64 | train.csv | Includes areas such as Marine Parade, Changi. |
| planning_area_category_GroupCQS | int64 | train.csv | Includes areas such as Queenstown, Serangoon, Clementi. |
| planning_area_category_GroupJB | int64 | train.csv | Includes areas such as Bukit Merah, Jurong East. |
| planning_area_category_GroupPWC | int64 | train.csv | Includes areas such as Bukit Panjang, Choa Chu Kang, Woodlands. |
| planning_area_category_GroupYH | int64 | train.csv | Includes areas such as Hougang, Yishun. |
| Feature | Type | Dataset | Description |
|---|---|---|---|
| tenure | int64 | train.csv | Transaction year less lease commencement year |
| tenure_buckets_0-10 | int64 | train.csv | Tenure bucket for 0 to 10 years |
| year_category_Group0 | int64 | train.csv | Year categories by the resale price trend. Group 0 is the years with lowest avg resale price |
| year_category_Group1 | int64 | train.csv | Year categories by the resale price trend. Group 1 is the years with 2nd lowest avg resale price |
| year_category_Group2 | int64 | train.csv | <Year categories by the resale price trend. Group 0 is the years with 2nd highest avg resale price> |
| is_pre_war | int64 | train.csv | Flat built before war i.e. 1949 |
| Feature | Type | Dataset | Description |
|---|---|---|---|
| is_premium | int64 | train.csv | Flat model is DBSS or Improved-Maisonette |
| is_terrace | int64 | train.csv | Landed terrace HDB |
| is_superlargeterrace | int64 | train.csv | Terrace with size more than 2250 sqft |
| flat_type_1 ROOM | int64 | train.csv | 1-room flat |
| flat_type_2 ROOM | int64 | train.csv | 2-room flat |
Data cleaning and EDA
Feature engineering and selection
Model selection - Linear regression, Ridge Regression and Lasso Regression
Kaggle submission
Lasso Regression was chosen for the model because it achieved the best performance including R-squared score (0.0903) and RMSE (around $45,000).
It is highly recommended to use our model for HDB resale price prediction.
Through extensive research and experimentation, our model achieves an R-squared score of 90.3%, indicating it can predict up to 90.3% of the variation in HDB resale prices. On average, the predicted price differs from the actual price by around $45,000 (RMSE).