About the model
Valora learned from thousands of real house listings across Greater Jakarta. This page explains, without the jargon, how the model was built, how it arrives at a price, and how far you can trust it.
3,352
listings it learned from
13,007
nearby places mapped
1,000
small decision trees
685
houses held back for testing
Follow ten short steps: how the model was built, then how it values a house.
Step 1 of 10
A model learns from examples. We used 3,553 house listings from Rumah123 across nine cities in Greater Jakarta, each with its price, size, rooms and location.
Think of a property agent who has seen thousands of houses. Ask them about a house and they don't use a formula. They start from what a typical house costs, then adjust: "bigger than usual, add some; far from the centre, take a bit off." Valora works the same way, except every adjustment is learned from data.
Click the answers to follow a path. A simplified illustration, not real model numbers.
Floor area above 150 m²?
Within 15 km of SCBD?
Electricity 2,200 VA or more?
A single tree is crude: it only knows a few questions.
The real model adds up about a thousand trees, each one correcting what the previous ones got wrong. Many small, careful adjustments add up to one estimate.
A real estimate from the model: a 100 m² house on 120 m² of land near BSD City, 3 bedrooms, 2 bathrooms, 2,200 VA.
Rp 1.84 B
Each factor moves the price by a percentage. Small factors are grouped under "all other factors".
Average size of each factor's effect on the price, across all houses in the data.
These results come from the 685 houses the model never saw while learning.
Out of 100 houses, 63 were priced within 20% of their real asking price.
14.8%
Half of the estimates were closer than this to the real asking price.
20.2%
A few big misses pull the average up. Lewis (1982) calls 20–50% reasonable and 10–20% good; Valora sits right at the border.
62.8%
Share of estimates that landed within 20% of the real price.
0.84
How much of the difference in price between houses the model can account for (1 = all).
Average error per region on the test houses. Shorter is better.
No one can know the exact price of a house before it sells. That's why Valora shows two ranges: the most likely range, where the price should fall half of the time, and the wide range, which should catch it 8 times out of 10. Each region gets its own ranges, because prices in some regions are harder to predict.
The ranges for the BSD example above:
83 of 100
test houses had their real price inside the wide (80%) range
per region
ranges are calibrated separately for each region
Asking prices, not sale prices
The data are advertised prices, which are often negotiated down. Treat an estimate as a well-informed starting point.
Not a formal appraisal
Valora can't see the condition, view or legal status of a house. It supports, but doesn't replace, an on-site valuation by a licensed appraiser.
Locations are approximate
Listing coordinates are at area level, not exact addresses.
Thin data in some places
Jakarta and very expensive houses have fewer examples, so estimates there carry more uncertainty.
A snapshot in time
The listings have no reliable date and are no longer online, so the model doesn't follow the latest price movements.
Patterns, not causes
The factors show what the model associates with higher or lower prices, not what causes them.
In the data, luxury listings often leave electricity blank. The model learned that "blank" goes with "expensive" and raised estimates by about 10% for no real reason. Making these fields required removes that bias.
That simple rule is exactly what we compared against. It missed by about 33% on average; Valora misses by about 20%, because it also weighs land size, rooms, electricity and what's around the house.
Studies on house prices repeatedly rank it among the most accurate methods for this kind of table data. It copes with missing information, captures "it depends" effects (a big house matters more in some areas than others) and runs fast enough to answer instantly.
No. The model is fixed after training. Your valuations are saved only to your account so you can revisit and compare them, and your feedback helps us judge where the model needs work.
The best way to judge the model is to test it on a house whose price you already know.
Data: Rumah123 listings via Kaggle · Places © OpenStreetMap contributors (ODbL) · Accuracy scale: Lewis (1982). Estimates come from a statistical model trained on listing asking prices (not transaction prices). Use them as a starting point, not as a substitute for a formal appraisal. Main factors show associations the model learned, not cause and effect.
These are associations, not causes. "Distance to the airport" mostly tells the model whether a house is in the west or the east of Jabodetabek; it doesn't mean the airport itself raises prices. Electricity and bathrooms act as signs of a larger, more upmarket house.
Together, location explains about 28% of the model's decisions.
Distance to 8 landmarks
Distance to the nearest…
| Place | Points mapped | Typical distance |
|---|---|---|
| School | 9,912 | 0.4 km |
| Hospital | 655 | 1.2 km |
| Market | 451 | 1.4 km |
| Mall | 368 | 1.8 km |
| Toll gate | 406 | 1.9 km |
| Train station | 139 | 4.4 km |
| University | 580 | – |
| Supermarket | 496 | – |
Jakarta has the fewest listings (651 of 3,352) and the widest spread of prices, so estimates there are less certain.