What Is a Location Model for Real Estate Data?
A location model ties every listing to a place in a country's official administrative hierarchy, so listings can be counted, compared and mapped by area. In Costa Rica that hierarchy is province, canton and district; in Panama it is province, district and corregimiento. It is what lets you ask "what does a house cost per square meter in Guanacaste?" when the original listing only said "ocean view, five minutes from Tamarindo."
Propiedades.cr is a property search platform for Costa Rica, with a sister site in Panama. It belongs to Ingeniarte Consultores, weKnow's parent company, and was built by weKnow's co-founders. It gathers public listings from dozens of agency and portal websites. Its recommendation engine has its own case study. This one is about the data underneath it: how the listings were tied to real places, and how we published them as the LATAM Real Estate Listings dataset on Hugging Face.
The Challenge: Listings Describe Location in Free Text
A listing's location is whatever the agent typed. That works for a buyer reading one ad, and fails as soon as you want to compare thousands of them.
Informal place names.
Agents write the beach, the neighborhood or the nearest mall: "Playa Hermosa", "near Multiplaza", "Santa Teresa". None of those is a canton or a district.
The same name in many places.
In this dataset alone, 741 Costa Rica listings sit in a district called San Rafael, spread across 17 different cantons. Without the canton, the name tells you almost nothing.
Two countries, two hierarchies.
Costa Rica divides into provinces, cantons and districts. Panama uses provinces, districts and corregimientos, plus indigenous comarcas. One dataset had to hold both without forcing one country's terms on the other.
Values that don't survive a second look.
Prices in colones tagged as dollars, impossible lot sizes, status fields mixing Spanish and English. Published as is, they make every average wrong.
The Solution: One Place Model for Two Countries
Every listing goes through the same path: read the location, place it in the official hierarchy, enrich the place, check the values, then publish.
1. One schema, three administrative levels
Each listing carries admin_level_1, admin_level_2 and admin_level_3: province, canton and district in Costa Rica, and province, district and corregimiento in Panama. Analysts can group by any level in either country with the same query.
2. An AI step, then people
An AI step reads the location each source gives and proposes the province, canton and district. A manual review in the back office corrects what it gets wrong; 1,150 listings in this release carry a human-verified flag.
3. Places enriched with what is nearby
Each province, canton and district has its own page, enriched through Google Places with nearby schools, hospitals, banks and supermarkets, each with map coordinates, next to the listings and the median house price for that place. See the page for Limón canton as an example.
4. Validation before release
The export rejects prices outside realistic ranges, re-reads colones mislabeled as dollars, nulls impossible areas and room counts, keeps condition values in Spanish only, and leaves a field empty when the source doesn't know instead of guessing.
5. Published as open data
The cleaned export is on Hugging Face under the MIT license, with the validation rules documented in the dataset card, so anyone can load it in a few lines of Python.
Example place page: Limón canton on Propiedades.cr
What the Location Model Makes Possible
Once every listing sits in a real place, area comparisons take one line of code. Houses and apartments in the beach provinces of Guanacaste and Puntarenas ask 50 to 60% more per square meter than in San José.
Median asking price per m², houses and apartments for sale
USD, by province. Provinces with 50 or more such listings.
Source: LATAM Real Estate Listings dataset, version 6 (exported November 12, 2026). Asking prices from listings published November 2022 to June 2026, not closing prices. Panama in grey.
Project Outcomes
Every Costa Rican canton covered
Listings land in all 84 of Costa Rica's cantons, across its 7 provinces, and 96% of Costa Rica listings sit on an official province and canton.
District-level precision
94% of listings reach district (or corregimiento) level: 11,293 of 12,030. In Costa Rica alone it is 95%.
Two countries, one query
11,070 Costa Rica listings and 960 Panama listings share the same columns, so one analysis runs on both.
Clean enough to publish
Every column stays inside documented, realistic ranges, so the minimum and maximum a visitor sees on Hugging Face are real listings.
Open to everyone
Free to use under the MIT license, for price models, market research, search and recommendation work.
Use the Dataset Yourself
The dataset loads straight from Hugging Face with the datasets library:
from datasets import load_dataset
ds = load_dataset("weknowinc/latam-real-estate-listings", split="train")
df = ds.to_pandas()
df.groupby("admin_level_1")["price_per_m2_usd"].median()The dataset card lists every column, the validation ranges and the place types for each country.
Why This Is a GIS Case, Not Just a Data Case
Most location data arrives the way these listings did: an address, a landmark, a town name typed by hand. Turning it into something you can group, join and map means resolving it against an official geography, handling places that share a name, and enriching each place with what surrounds it. That is the same work behind utility asset records, field surveys and service territories. weKnow's GIS practice does it for enterprise mapping, and here we did it end to end on our own data.
- ✓ Free-text locations resolved to an official administrative hierarchy
- ✓ Places enriched with points of interest and map coordinates
- ✓ Validation rules documented and applied before every release
- ✓ Data engineering, AI and manual review in one pipeline
- ✓ U.S. time zone overlap for real-time collaboration
FAQs
12,030 property listings from Costa Rica (11,070) and Panama (960), for sale and for rent, with price in USD, price per square meter, built and lot area, bedrooms, bathrooms, amenities, the listing description in Spanish and its location by province, canton and district. It is published on Hugging Face under the MIT license.
Kenny Abarca
CEO & Co-Founder, weKnow Inc. Specializes in web architecture, Drupal and AI-driven development workflows.
More about Kenny →Have Location Data That Won't Line Up?
Our GIS practice resolves messy addresses to real places, enriches them, and builds the pipelines and maps on top, with senior engineers in your time zone.
Talk to our team
