>And the list goes on... Is it that difficult to predict?
They do likely have a ML algorithm in the background. They even offer an Excel sheet where, for any county, you can see what % of sold homes are within, say, 20% or 10% of the Zestimate (it can be quite bad). And 20% is a huge error. Even 10% off is huge for most people.
What is hard to capture:
1. Condition of the house
2. Design of the house
3. Color of the house
4. Neighbors being annoying
5. Alignment of incentives for both party's real estate agents. Both profit more from a high price.
>You think estimating the value of every home in the country within 10% is bad?
It may be good from an "interesting problem" standpoint. But if your house's Zestimate at $300K, 10% means it's real value is somewhere between $273K and $334K.
Does anyone have a source of data for the variance of the home prices in the U.S.? Getting within 10% sounds spectacular, considering that I think you could get close to 10% price differences between different combinations of buyers and sellers on the exact same property. If someone were to come up with an algorithm for predicting the price of a steak dinner at a restaurant, using spotty 5 year old data, I think everyone would be impressed at getting it correct within $3 of $30.
>If someone were to come up with an algorithm for predicting the price of a steak dinner at a restaurant, using spotty 5 year old data, I think everyone would be impressed at getting it correct within $3 of $30.
As I said, as an academic exercise it's great. But we're not discussing whether Zillow deserves goodie points. The question is: Is this Zestimate useful for anyone?
From a buyer's or seller's POV, 10% is too high an error. Sure, if your house is only $100K, $10K off is not much. If your house is $300K, you're not likely going to get far by underbidding with $270K, and the seller will be deemed too optimistic for asking $330K (although with bidding wars, it occasionally does happen).
For either the buyer or the seller, being told the price could be between $270-330K is virtually useless in deciding what to buy/sell for. It's just too much variance.
With high numbers, absolute amounts matter more than relative percentage.
They do likely have a ML algorithm in the background. They even offer an Excel sheet where, for any county, you can see what % of sold homes are within, say, 20% or 10% of the Zestimate (it can be quite bad). And 20% is a huge error. Even 10% off is huge for most people.
What is hard to capture:
1. Condition of the house 2. Design of the house 3. Color of the house 4. Neighbors being annoying 5. Alignment of incentives for both party's real estate agents. Both profit more from a high price.
And so on.