Zillow's implosion

[Mejappz]

Sometimes, after I finish an appraisal in higher-end markets, I will look at Zillow to see what they valued the home. Last week, I completed one where the neighborhood price ranged from $500,000 to $10,000,00.

I appraised the property at $1,680,000; I looked at Zillow, and the Zestimate had it at $2,386,000, Lol. Zillow and all AVMs can't keep up with this market.

I have been looking at other higher-price neighborhoods on Zillow. It seems like they are simply taking the median sale price from the market area or neighborhood and applying it to all the properties regardless of property characteristics, location, condition, quality, etc. It has become an absolute joke.




Just a bit outside.

 

[Essexfenwick]

Real estate just frustrates the heck out of the technology crowd. It refuses to be people free.

 

[Alsie35]

Real estate just frustrates the heck out of the technology crowd. It refuses to be people free.

Let's hope so. Otherwise we will eventually be replaced. (Spoiler: Don't bet on it.)

 

[Terrel L. Shields]

They are idiots showing wild gyrations in a market that has simply and steadily gone up and not a day has the market gone down since 2014 here but this is a local house where the end value was made to match its recent sale price and they now say it is worth exactly $100 less than when it sold.

 

[Sandra Koutsopoulos]

I had an owner today who insisted the property had previously been measured wrong which cost her money. Turns out she was relying on Zillow which showed her LOT size smaller by 100sf than expected. I'm not sure she was convinced that Zillow had nothing to do with diddly when I told her it was not a reliable info source.

 

[residentialguy]

 

[RCA]

Sometimes, after I finish an appraisal in higher-end markets, I will look at Zillow to see what they valued the home. Last week, I completed one where the neighborhood price ranged from $500,000 to $10,000,00.

I appraised the property at $1,680,000; I looked at Zillow, and the Zestimate had it at $2,386,000, Lol. Zillow and all AVMs can't keep up with this market.

I have been looking at other higher-price neighborhoods on Zillow. It seems like they are simply taking the median sale price from the market area or neighborhood and applying it to all the properties regardless of property characteristics, location, condition, quality, etc. It has become an absolute joke.


View attachment 59504

Just a bit outside.

I sense that Zillow if finally beginning to realize that it's not that simple. They can give you 80% accuracy, but that other 20% is a killer. Even if it is 90% accuracy, the point is you cannot ever know for sure that it isn't way off. Zillow (or any AVM) does not know when it is accurate or way off. It is just a question of probability. And that can cause major problems.

I not long ago did a home, where my regression model had a high R2, but the value for the subject was clearly higher than reality ( it last sold in 2017 ). Looking at the data, I couldn't figure it out. I called the Listing Agent - and she said the property was hard to sell because a lot of traffic noise was funneling up to the house which was at the top of a slope facing Hwy 101. I did the inspection on an early Saturday morning, and am a little hard of hearing, so didn't notice unusually loud traffic.

In fact, if you create a column for distance from the subject, earth/MARS will add a function that shows that the house is at the lowest point of the distance (to subject) vs sale price graph. In other words, there is something specifically bad about the location. In that area, Hwy 101 runs North-South, so we also get a longitude function that also is at its lowest price point near Hwy 101. But the subject, aside from being relatively close to Hwy 101, is also in a rather unique position with regard to getting a lot of noise "funneled" up the side of the hill (hills on each side of the highway act as an amplifier).

I typically spend a lot of time rerunning R/earth over and over, tweaking the data which of course is usually missing this or that. Garage = 0 does not necessarily mean there is no Garage. There may indeed be a garage for two or three cars. But there may be a carport and extra parking spaces. So, you need three columns: "Garage Spaces", "Carport Spaces", "Parking Spaces". --- And you need to check all the rows with a value of 0 - because 0 doesn't always mean 0. It seems nearly every property has special problems, either with the real situation, or with the data. It may be an issue with geographic competence (and if nothing else, MARS will make you geographically competent very quickly), - or it may be a problem with poor data. You can often fix data problems - but it may require some time to get the data corrected - because if you are using MARS or any type of regression - you have to correct for possibly very many properties.

For the AVMs like Zillow, there is no one formula that is going to work in all cases. It really requires some human knowledge and understanding to create a reasonable value model in all circumstances. Although, a regression tool like MARS (and I specifically use the earth package from R), can be a critical aid in building such a valuation model.

 

[ZZGAMAZZ]

I sense that Zillow if finally beginning to realize that it's not that simple. They can give you 80% accuracy, but that other 20% is a killer. Even if it is 90% accuracy, the point is you cannot ever know for sure that it isn't way off. Zillow (or any AVM) does not know when it is accurate or way off. It is just a question of probability. And that can cause major problems.

I not long ago did a home, where my regression model had a high R2, but the value for the subject was clearly higher than reality ( it last sold in 2017 ). Looking at the data, I couldn't figure it out. I called the Listing Agent - and she said the property was hard to sell because a lot of traffic noise was funneling up to the house which was at the top of a slope facing Hwy 101. I did the inspection on an early Saturday morning, and am a little hard of hearing, so didn't notice unusually loud traffic.

In fact, if you create a column for distance from the subject, earth/MARS will add a function that shows that the house is at the lowest point of the distance (to subject) vs sale price graph. In other words, there is something specifically bad about the location. In that area, Hwy 101 runs North-South, so we also get a longitude function that also is at its lowest price point near Hwy 101. But the subject, aside from being relatively close to Hwy 101, is also in a rather unique position with regard to getting a lot of noise "funneled" up the side of the hill (hills on each side of the highway act as an amplifier).

I typically spend a lot of time rerunning R/earth over and over, tweaking the data which of course is usually missing this or that. Garage = 0 does not necessarily mean there is no Garage. There may indeed be a garage for two or three cars. But there may be a carport and extra parking spaces. So, you need three columns: "Garage Spaces", "Carport Spaces", "Parking Spaces". --- And you need to check all the rows with a value of 0 - because 0 doesn't always mean 0. It seems nearly every property has special problems, either with the real situation, or with the data. It may be an issue with geographic competence (and if nothing else, MARS will make you geographically competent very quickly), - or it may be a problem with poor data. You can often fix data problems - but it may require some time to get the data corrected - because if you are using MARS or any type of regression - you have to correct for possibly very many properties.

For the AVMs like Zillow, there is no one formula that is going to work in all cases. It really requires some human knowledge and understanding to create a reasonable value model in all circumstances. Although, a regression tool like MARS (and I specifically use the earth package from R), can be a critical aid in building such a valuation model.

Why don't companies like Zillow hire an appraiser to conduct brief site inspections to address "Condition." The minimal fees to hire appraisers presumably would be valuable because of the enhanced reliability of the AVM results.

 

[RCA]

Why don't companies like Zillow hire an appraiser to conduct brief site inspections to address "Condition." The minimal fees to hire appraisers presumably would be valuable because of the enhanced reliability of the AVM results.

Hmmm. In my opinion that would certainly be worth something. However, based on my experience, you need an appraiser who can "talk to" the regression and debug it when the models it produces don't make sense. Oftentimes, you wind up patching or appending to the data to make the regression tool generate a reasonable model. If the regression software is greek to the appraiser, he can't fix things. And the regression software certainly doesn't know anything about appraisal.

An advanced appraiser, or valuation engineer, at least in this day and age with almost no specialists in the field except a handful, has to be a jack of many trades to get things done - because only he can tie everything together to make sense of the whole.

It's actually like that in a number of fields where different specializations have to be mastered and "interfaced." My son-in-law, who has a Ph.D. in Geostatistics from Stanford is always saying he works at the interface between X, Y, Z, .... I think, if I recall correctly, it's the interface between Geology, Petroleum Engineering, Statistics, Green Energy, AI ....

I have to work with:
1. Statistics (principally MARS aka R/earth).
2. Data Collection, Management and Analysis - or simply Data Mining.
3. Inspection which includes photography (standard plus drone), mapping, measurement, floorplans, GIS, ....
4. Report writing

Don't forget all the gadgets you have to master and refresh yourself on. These are not simple, and you can eventually forget what the various icons and settings do or how to calibrate them. I've just given up on my old Canon 5D Mark II because the focus is simply not good enough anymore - and there is almost no one around to fix a camera this old. Canon won't take them any longer. So, I'm moving to the mirrorless R6 - which will still take my old Canon lenses (the 16-35mm f2.8L, 100-400mm IS, 28-70mm f2.8L, ....). In fact with the EOS-R circular ring adapter, you can use the old EF lenses the same way as the new R lenses - which is pretty cool. I can still keep the 5D as a backup. It is just that I am not hat happy with its performance any more. There is also the "simple" X4 which I bought a couple of months ago. I still have to drill myself on getting things done with it quickly. Then there is the Phantom IV drone, which is a fairly complex device.

--- But the truth is most of my time is spent with MARS, the R scripts and cleaning up or appending the data until I get something I am happy with. The latest thing I did was split up the swimming pool description into 4 fields which I input into the regression:

A. PoolYN? 0=No, 1 = yes

#1: AboveGround? 0= No, yes = 1
#2: InGround? 0=No, 5= Yes
#3: Gunite? 0= no, 1= yes
#4: Heated? 0=no, 1= yes.

I did the above using an Excel Macro like: =IF(IFERROR(FIND("Gunite",AQ2)>0,0)>0,1,0) where AQ is a column with values like:

"Pool - Yes, Pool - In Ground, Pool - Gunite, Lap Only, Pool - Cover, Other, Pool - Sweep, Pool/Spa Combo, Heated - Solar, Spa/Hot Tub"

B. PoolScore: Is the sum of the #1-#4, which can be: 0,1,2,5,6,7

I enter the variables A and B into the regression. ... And that was the final step that completed the model to my satisfaction.