That's what I thought you meant by 50%. Instead of a value of 60, the value was 90. Also, I would suggest it is important whether the output error is biased or random: is it consistently wrong in one direction, or just all over the place.As far as the model being good but the "incorrect" Zestimates being the result of bad data, well, tough. The Zestimates for this county are completely unreliable.
I have no idea whether it is because of bad data or their model because, like every other AVM I have seen, it is a black box AVM. There is no way to know the strengths and weaknesses of it unless they publish their methodology.
The accuracy of an AVM is not difficult to determine.
My only concern for you, is that if you produce something, produce something well-wrtten. Try to appreciate there is a difference between the valuation model and the data. If you refer to one as the other carelessly, you give your potential rebutter too much ammunition.
A friend of mine inherited a house in California a few years ago, and his lawyer mentioned getting an appraisal, so he asked me for advice. I looked up the property on Zillow, and found the software to be amazing. However, I detected a problem with the model. I said to my friend either all the houses in good condition sold first, or this market is going up sharply. The Zillow value is the same one I get with averaging or regression without time adjustments. However, when I put in time adjustments the value goes from 570 tp 630. The next week he got an appraisal that came in at 620, using two of the ten sales I used from Zillow. After that he called a broker to list the house who said 635, without knowing that my friend had done some research.
I seem to recall that there were some tabs on Zillow where one could get ratings of the overall economics of the area with respect to housing. While it looked "nice," in my opinion, it was no substitute for actually at the prices and square-foot prices to see the trend is up, down or flat. In this particular case, I believe the data was sound, but the model was off, because it didn't do time adjustments.
I agree with you the problem is the "black box" component. That is, they are not disclosing a big chunk of the scope of work. This is part of what amazes about the ASB's incomprehensible opinion that these people are not doing appraisals. Of course they are. And because the reporting is sub-standard (not disclosing scope), we can't really assess the problems. For example, given the software's capability, there is no reason that they don't generate the same analysis that you did; along with the ten sales, they could present what the model predicted for those properties. Wouldn't that tell their users something signficant?