The reasons for a bad AVM vs a bad appraisal imo would be different. The AVM and the appraiser both have access to the same data. The appraiser qualifies the data, the AVM can not do that on a personal each one analyzed basis, but by a rote programming basis. And the rote part is the flaw...maybe in future years AI will replicate human experience and judgement but we are not there yet .For example, the fannie CU comps satisfy their rote within a mile within X % sf, but can range from good to mediocre to terrible comps due to various reasons.
If an appraisal is bad, other than a one off event, it is either the appraiser was not competent for the assignment or the appraiser set out to mislead/ inflate value - either way especially the second the clients should not be using that appraiser -
The other flaw with an AVM is the AVM concludes X $ value, is that a range or a point value with a confidence score around it ? Either way, who on the client or user end gets to choose the value from that AVM or whether to rely on it, order another one? idk but seems a murky area..
Well, I think there is some misunderstanding here.
I usually don't say much. Appraisers on this forum like to show how much statistics they know. But what they know is usually what is taught in basic statistics classes --- parametric statistics, normal curves, standard regression. In data mining nonparametric statistical tools like MARS, Random Forests, CART, LASSO or many other types are typically used. And they are used with:
Bootstrapping: A way of splitting your data set up into partitions, using maybe 80% for analysis to create a model and then the other 20% for testing. They draw data from the test partition with replacement to simulate a more general population and then test on the remaining. They may do this thousands of times, continually creating new partitions or test data sets from the original, then either selecting the best model, or combining them through an
Ensemble: Merging a number of differently created models to create a model that is more robust on predicting unknown or new data.
Bagging: Another term for splitting up your data set into partitions ... Go to YouTube for lots of videos.
The result of these techniques are data models that are robust even when created on rather small data sets.
About the only thing that matters on non-parameterized statistics is the R2 or the percentage of variance in the data that is accounted for by the model. Admittedly there are other specific measures created for specific techniques like logistic regression, but again these are not your typical statistics parameters. Anyway, the R2 or GCV-R2 is a "normalized" measure that makes it possible to compare models run against different data sets with different complexity models. For example we we make two runs against the same data set and one model uses 10 basis functions with N knots while another model gets by with 8 basis functions and N or fewer knots, and they both have the same standard R2, we would prefer the simpler model and that is where the GCV-R2 comes it, it will tive you a lower value for the more complex model. While MARS will also display things like MAPE or sum of errors - these are impossible to accurately compare between different sets of data or even different strategies in bagging and bootstrapping.
The GCV-R2 is an R2 that is adjusted for the effective number of parameters or the complexity of the model. In other words it penalizes for complexity that would for example indicate overfitting of the model to the data. It is usually the lowest in value of all R2 variants.
My guess is that Zillow in fact uses Salford Systems MARS or something very close. Furthermore, because they are so large they have a big advantage: Users, i.e. the public, will call them up or send emails if they see problems in the data. So, they get additional verification that appraisers don't get. I am also pretty sure they have people that constantly go over the data coming in to find and remove the most flawed property data.
Zillow provides free real estate information. Search homes for sale, home prices, home values, recently sold homes, mortgage rates, apartment rentals, and more.
www.zillow.com
Of course appraisers have a big advantage (now anyway) in that they are more familiar with the areas they work in than probably many of the people working at the companies that write software for and manage the AVMs.
And, the AVMs like Zillow are simply to big to be well managed, IMO.
Hard to say how things are going to play out in the end.
community for the discussion of real estate appraisal. Register a free account to be able to post and unlock additional forums and features.
Maybe they aren't really that interested after all.
.