Vaastu WTP

Do buyers pay a premium for Vaastu-compliant homes? How much?

Key Findings

Summary Across Data Sources

Source	N	Specification	Premium	95% CI	Sig.
CampusX (Gurgaon)	3,173	+ sector FE	+2.0%	[-0.5%, 4.6%]	n.s.
Magicbricks (multi-city)	30,981	+ city FE	+3.8%	[2.9%, 4.8%]	***
Housing.com (multi-city)	1,596	+ city FE	+5.6%	[-3.9%, 16.0%]	n.s.

Aggregate Pooled Analysis

Specification	N	Premium	95% CI	Sig.
Pooled (raw)	59,895	+6.2%	[4.4%, 8.0%]	***
+ bhk	59,895	+0.5%	[-0.8%, 1.7%]	n.s.
+ city FE	59,895	+5.6%	[4.4%, 6.9%]	***
+ source FE	59,895	+15.4%	[13.9%, 16.9%]	***
+ city + source FE	59,895	+11.5%	[10.2%, 12.9%]	***

Specification Robustness by Source

CampusX (Gurgaon only):

Specification	N	Premium	Sig.
Raw	3,836	+45.0%	***
+ bhk	3,836	+37.9%	***
+ structural	3,836	+6.1%	***
+ sector FE	3,173	+2.0%	n.s.

Magicbricks (multi-city):

Specification	N	Premium	Sig.
Raw	32,940	+32.4%	***
+ bhk	32,940	+12.6%	***
+ structural	30,981	+7.5%	***
+ city FE	30,981	+3.8%	***

Magicbricks Apartments Only (with developer FE):

Specification	N	Premium	Sig.
+ structural	26,149	+6.4%	***
+ city FE	26,149	+2.9%	***
+ developer FE	26,149	+3.8%	***
+ city + developer FE	26,149	+3.0%	***

Developer FE is only meaningful for apartments/flats (90% of MagicBricks data) since houses/villas typically don't have developers listed. Within-developer comparisons still show a significant +3% Vaastu premium.

Housing.com (multi-city):

Specification	N	Premium	Sig.
Raw	1,599	+27.4%	***
+ bhk	1,599	+12.3%	**
+ city FE	1,596	+5.6%	n.s.

Key finding: Raw correlations show 27-45% premiums, but these attenuate substantially with controls and location FE. With city/sector FE, source-specific premiums range from 2-6%. Magicbricks shows a significant +3.8% premium with city FE, and even with the most stringent within-developer comparisons (apartments only), the premium remains +3.0% (***). Pooled analysis with city + source FE yields +11.5% (***p<0.01, N=59,895).

Data Sources

Magicbricks Multi-City Data (data/raw/magicbricks/<city>/)

• Source: magicbricks.com listings (apartments, houses, villas)
• Collected via: scripts/magicbricks/ 6-step pipeline (Playwright-based scraper)
• Cities: Delhi-NCR, Pune, Navi Mumbai, Bangalore, Jaipur, Lucknow, Patna, Chandigarh, Rajkot
• Sample: 49,845 listings (30,981 after quality filtering for analysis)
• Composition: 90% apartments, 10% houses/villas
• Preserves: Raw HTML for reproducibility

99acres CampusX Data (data/raw/99acres_campusx/)

• Source: 99acres.com listings
• Obtained via: Public GitHub repository campusx-official/dsmp-capstone-project
• Collection method: Unknown (no documentation in source repo)
• Files: flats_cleaned.csv, house_cleaned.csv, gurgaon_properties_cleaned_v2.csv
• Sample: ~6,943 listings (Gurugram only)

99acres Kaggle Data (data/raw/99acres_kaggle/)

• Source: 99acres.com listings
• Obtained via: Kaggle dataset arvanshul/real-estate-dataset-99acres
• Cities: Gurgaon, Mumbai, Hyderabad, Kolkata
• Sample: ~38,000 listings (~21,512 after filtering)
• Limitation: Excluded from vaastu analysis. Features/amenities stored as numeric codes without decoder; vaastu detectable only in free-text descriptions (~5.5% rate vs ~50% in comparable 99acres data from CampusX)

99acres Multi-City Data (data/raw/99acres/<city>/)

• Source: 99acres.com listings (houses + flats)
• Collected via: scripts/99acres/ 4-step pipeline (Playwright-based scraper)
• Cities: Bangalore, Chennai, Hyderabad, Pune, Mumbai, Delhi, Noida, Gurgaon, etc.
• Preserves: Raw HTML + text for reproducibility

Housing.com Multi-City Data (data/raw/housingcom/<city>/)

• Source: Housing.com independent-house listings
• Collected via: scripts/housingcom/ 4-step pipeline (Playwright-based scraper)
• Cities: Bangalore, Chennai, Hyderabad, Pune, Mumbai, Noida, Gurgaon, etc.
• Preserves: Raw HTML + text for reproducibility

Methodology

• Hedonic regression: log(price) ~ vaastu + controls + sector FE
• Controls: area, bedrooms, bathrooms, balconies, floor, age, facing, furnishing
• Robustness: Nearest-neighbor matching within sector/property-type

Data Cleaning

Price Filters

Raw listing data contains significant noise requiring filtering:

1. Price range filter: 0.1 - 100 crore

   • Removes ~14,000 rows with price < 0.1 crore (likely rentals mislabeled as sales or data entry errors)
   • Removes ultra-luxury outliers > 100 crore

2. Price per sqft filter: 1,000 - 100,000 INR/sqft

   • Removes ~500 rows with implausible price/area ratios
   • Catches data entry errors where price or area is wrong

3. BHK filter: 1 - 10 bedrooms

Outlier Examples

Some egregious outliers removed by these filters:

Price (Cr)	BHK	Area (sqft)	Price/sqft	Issue
357.0	2	850	42 lakh/sqft	Price entry error (should be ~0.357 Cr?)
0.00105	2	700	15/sqft	Rental mislabeled as sale
216.0	4	2,455	8.8 lakh/sqft	Extreme outlier

These filters reduce MagicBricks from 49,845 to ~31,000 analysis-ready rows.

Repository Layout

scripts/
├── 02_analyze.py                # Main analysis script (99acres/CampusX)
├── 02_download_kaggle.py        # Download Kaggle dataset
├── 03_unify_99acres.py          # Unify 99acres data sources
├── 04_analyze_magicbricks.py    # MagicBricks hedonic regressions
├── 04_analyze_housingcom.py     # Housing.com hedonic regressions
├── 05_analyze_by_source.py      # Cross-source comparison
├── 05_validate_kaggle.py        # Kaggle data quality checks
├── export_dataverse.py          # Export data for Dataverse
│
├── 99acres/                     # 99acres 4-step pipeline
│   ├── 01_collect_search.py
│   ├── 02_extract_urls.py
│   ├── 03_collect_detail.py
│   └── 04_parse.py
│
├── housingcom/                  # Housing.com 4-step pipeline
│   ├── 01_collect_search.py
│   ├── 02_extract_urls.py
│   ├── 03_collect_detail.py
│   └── 04_parse.py
│
├── magicbricks/                 # MagicBricks 6-step pipeline
│   ├── 01_collect_search.py
│   ├── 02_extract_urls.py
│   ├── 03_collect_projects.py
│   ├── 04_extract_listing_urls.py
│   ├── 05_collect_listings.py
│   ├── 06_parse.py
│   └── run_pipeline.py
│
└── utils/                       # Shared utilities
    ├── parsing.py
    ├── scraping.py
    ├── feature_extraction.py
    └── analysis.py

data/
├── raw/
│   ├── magicbricks/<city>/      # MagicBricks scraped data
│   ├── 99acres/<city>/          # 99acres scraped data
│   ├── 99acres_campusx/         # CampusX Gurgaon data
│   ├── 99acres_kaggle/          # Kaggle dataset
│   └── housingcom/<city>/       # Housing.com scraped data
├── v1/                          # Consolidated export (parquet)
├── config/                      # City URL configurations
└── derived/                     # Analysis outputs

ms/                              # LaTeX manuscript
tabs/                            # Generated tables
figs/                            # Generated figures
tests/                           # Unit tests

Quick Start

Prerequisites

• Python 3.10+
• uv package manager

Installation

uv sync
uv run playwright install chromium

Data Collection

MagicBricks (6-step pipeline, or use run_pipeline.py):

uv run python scripts/magicbricks/run_pipeline.py --city delhi

99acres:

uv run python scripts/99acres/01_collect_search.py --city bangalore
uv run python scripts/99acres/02_extract_urls.py --city bangalore
uv run python scripts/99acres/03_collect_detail.py --city bangalore
uv run python scripts/99acres/04_parse.py --city bangalore

Housing.com:

uv run python scripts/housingcom/01_collect_search.py --city mumbai
uv run python scripts/housingcom/02_extract_urls.py --city mumbai
uv run python scripts/housingcom/03_collect_detail.py --city mumbai
uv run python scripts/housingcom/04_parse.py --city mumbai

Analysis

uv run python scripts/04_analyze_magicbricks.py
uv run python scripts/04_analyze_housingcom.py
uv run python scripts/05_analyze_by_source.py

Caveats

• List prices, not transaction prices
• Vaastu = text mention, not structural certification
• Within-sector support for houses is thin
• Kaggle data excluded from Vaastu analysis: features stored as numeric codes without decoder; Vaastu detectable only in free-text descriptions (~5% vs ~50% true rate)