Is it really fair to judge their algorithm on "one of the most expensive sales of 2016"? Only a tiny, tiny portion of the half million houses sold in the US each year would be even approaching the 20 million USD mark. His service isn't intended to service the mega rich.
Pretty much this. 5% isn't bad, especially when you're working with something very unique. In that price range, you're buying your subjective taste in addition to the value.
Gotta say, by my tastes, that's one of the nicest places I've ever seen. I really like the layout and the living area flowing into the back patio/yard like that. It's too bad my market is more like 300k homes in Maine :D
I live in the Midwest and our home prices are relatively flat over the last decade for the city I live in. 18 months ago I noticed that my zestimate was inflated and rose to 2x my appraised value of my home (refinanced that year). I looked into it and noticed that Zillow was doubling the sales price of recent homes. This was the main issue. I reported it to Zillow and some in customer service said their engineers would look into it. Only in the last few months have th Zestimates come back to reality.
I bring this up because I was curious as to how this would play out for my city and home prices. Would this increase prices for the area or would realitors ignore the data and go with what they knew would make sense.
I haven't seen a huge increase in prices nor have a done any study of it. But just the idea that a bug could cause such a calamity for a least the size of a city was eye opening.
I have a friend who is a realtor and we talk about this subject occasionally. My understanding is that most realtors will recommend using Zillow to "get a feel" for the market. But Zillow's estimates are only as good as the data that goes into them. Usually, data for an area is publicly available and up-to-date, but sometimes it's very old or just plain wrong.
I think Zillow claims their estimates to be within 10% of actual sale price. To me, that figure seem to be pretty good from a data analysis stand point; but 10% of the value of a house is a lot of money for even a modest home.
Most realtors used closed databases, like MLS, to document sales and compare area housing prices. I believe MLS data is typically manually entered and verified (I'm not 100% sure about this though). I believe there is also a lot of extra metadata added to MLS listings that Zillow can't get from public records. Realtors also tend to look at property tax assessments for their areas in order to gauge market trends (just like Zillow does, except without a hidden algorithm).
It's not surprising that Zillow's AVM [1] isn't accurate for this sale. The model needs local comparable sales data to function properly which it didn't have in this case; it's not everyday that mansions are bought and sold. The next best thing it can do is look at the last recorded sale (or tax assessor's indication of value) and project out by looking at market trends and applying some other proprietary mumbo-jumbo to come up with a price.
I didn't realize people actually cared about zillow zestimates. I'm not trying to be mean, I love Zillow as someone too introverted to go to open houses but still curious what the inside of homes look like around me.
I live in a city (stl) where less than one mile north are houses <30k and less than a mile west re >1m and one block east are 150-200 and one block west are 500k.
Whenever houses sell around me my zestimate goes nuts but we have a smaller house in the middle of 100+ year old 4k square foot homes. They sell for either MUCH higher prices (someone kept it up) or much less (they didn't) causing our estimate as a unrelated newer home to go haywire.
I would never expect an algorithmic prediction of property prices to do well on a very unusual ("one of the most expensive sales of 2016") property.
Zillow's estimates in general may or may not be useful -- and personally, i'd be awfully interested in finding out more about their (presumably proprietary secret) algorithms. But "one of the most expensive sales of 2016" is not a good example to say anything about how the algorithm functions in general.
I don't see how a model is supposed to accurately predict a house's selling price when in many markets people look at those estimated prices and add a percentage.
What do you mean? If there's an easy way to improve on the accuracy of the estimated price, by eg adding a percentage, shouldn't you just bake that simple way into your model in the first place?
So the Redfin and Zillow price of my house I purchased in Oct. 2015 is off about 125K from my purchase price (1.125M) and misses the 400K in remodel. I checked the houses on the same street that sold since and they are off by 200-400K from the ASP. To be fair I live in Los Gatos in Silicon Valley and the housing prices are crazy stupid. Anywhere else in the US the house would be 400-500K.
I wish I could pick up properties that my parents own in western Maine and drop them in seacoast New Hampshire. A town house that might scrape 80k up there would be worth 300-400k down here, assuming it was a barely livable shithole.
Or you could have the situation a couple of blocks from me: 1920s remodel that was awesome in the inside sold for $850k. Crappy 1940s house with one owner for 40 years (and it shows) is Zestimated to be the same price as they have the same stats in square footage and bathrooms. They are listing for a reasonable $575k.
Zestimates are a complete joke. They can complain that they don't have all the facts (but they can infer from the wording of "needs work" and "recently remodeled") but that's like if a frog had wings it wouldn't bump its ass a hopin.
What I don't get is that they have the data on millions of sales. Surely they can come up with something that's not 30% off. At that rate you minus well just guess at a price and be better off.
Because a more accurate estimate of value requires well-formed data. They would need a good way to judge condition and other characteristics that might require a very subtle interpretation of a listing, or might require an in-person visit to a property.
These kinds of judgements are much more complicated than traditional machine learning. My guess is that they are feeding easily-obtained data into a machine learning algorithm, and quality of property, interpretation of a listing, and viewing of pictures, is not something that's easy to feed into a machine learning algorithm.
Readit News
Sounds like a pretty good estimate to me, particularly considering the paucity of back data for this kind of inventory.
Gotta say, by my tastes, that's one of the nicest places I've ever seen. I really like the layout and the living area flowing into the back patio/yard like that. It's too bad my market is more like 300k homes in Maine :D
I bring this up because I was curious as to how this would play out for my city and home prices. Would this increase prices for the area or would realitors ignore the data and go with what they knew would make sense.
I haven't seen a huge increase in prices nor have a done any study of it. But just the idea that a bug could cause such a calamity for a least the size of a city was eye opening.
I think Zillow claims their estimates to be within 10% of actual sale price. To me, that figure seem to be pretty good from a data analysis stand point; but 10% of the value of a house is a lot of money for even a modest home.
Most realtors used closed databases, like MLS, to document sales and compare area housing prices. I believe MLS data is typically manually entered and verified (I'm not 100% sure about this though). I believe there is also a lot of extra metadata added to MLS listings that Zillow can't get from public records. Realtors also tend to look at property tax assessments for their areas in order to gauge market trends (just like Zillow does, except without a hidden algorithm).
[1]: https://en.wikipedia.org/wiki/Automated_valuation_model
I live in a city (stl) where less than one mile north are houses <30k and less than a mile west re >1m and one block east are 150-200 and one block west are 500k.
Whenever houses sell around me my zestimate goes nuts but we have a smaller house in the middle of 100+ year old 4k square foot homes. They sell for either MUCH higher prices (someone kept it up) or much less (they didn't) causing our estimate as a unrelated newer home to go haywire.
Zillow's estimates in general may or may not be useful -- and personally, i'd be awfully interested in finding out more about their (presumably proprietary secret) algorithms. But "one of the most expensive sales of 2016" is not a good example to say anything about how the algorithm functions in general.
Zestimates are a complete joke. They can complain that they don't have all the facts (but they can infer from the wording of "needs work" and "recently remodeled") but that's like if a frog had wings it wouldn't bump its ass a hopin.
What I don't get is that they have the data on millions of sales. Surely they can come up with something that's not 30% off. At that rate you minus well just guess at a price and be better off.
These kinds of judgements are much more complicated than traditional machine learning. My guess is that they are feeding easily-obtained data into a machine learning algorithm, and quality of property, interpretation of a listing, and viewing of pictures, is not something that's easy to feed into a machine learning algorithm.
Dead Comment