About this estimator
This is the demonstration application for the paper Predicting Residential Property Prices in the Philippines Using Machine Learning (Section 4.7). It runs the same Random Forest the paper reports, entirely inside your browser. Nothing you type leaves this page.
The data
The model learned from asking prices for houses, condos, lots and commercial properties across the Philippines. The dataset started as scraped listings; duplicates of the same listing were removed. Snapshot date: , as declared by the dataset's author (paper Section 3.2).
The model
A Random Forest is a crowd of decision trees. This one has trees, each at most questions deep. Each tree turns ten facts about a listing into its own price guess: floor area, lot area, bedrooms, bathrooms, latitude, longitude, city, province, region and property type. The estimate you see is the average of all the guesses.
What the numbers mean
- Accuracy. On listings it had never seen, the model's R² was (0 means no better than always guessing the average price; 1 means perfect). The typical estimate was off by of the asking price, and of estimates landed within 50% of it (Section 4.2).
- Likely range. Measured on the same unseen listings, separately for each property type: it is the span that held the actual asking price for 80% of them. It describes past performance. It is not a guarantee for any one property.
- Tree agreement. Shows how closely the trees agree with each other. Trees can agree and still be wrong, so this is not a measure of accuracy.
Five limits that matter
- Asking prices, not sale prices. The data are what sellers advertised. What a property sells for can differ, and an appraisal is a different thing again.
- Unfamiliar cities. When a whole city () was left out of training, R² fell to . That is worse than guessing the average price. The estimator warns you when a city is not in its training data (Section 4.4).
- Lots and commercial properties are much harder. The typical error is for houses and for condos, but for lots and for commercial properties.
- Scraped data contains mistakes. For example, raw listings claim more than 10,000 m², and one claims m². Province and region are missing for of listings and lot area for . The inputs here only accept sensible values.
- The estimate leans high. For the typical listing it is about above the actual asking price, and it was above it for of test listings (Sections 4.2 and 5.2).
Source and licence
Data: Christian (covaga.xyz), Philippines Real Estate Asking Prices, Hugging Face, snapshot , repository on huggingface.co. Licensed under CC BY 4.0.