Comment by organsnyder | Hacker News Reader

Comment by organsnyder | Hacker News Reader

In my experience, Zillow does a terrible job compared to local knowledge. My zip code is weird—our ~$275k house backs up to houses worth $50-70k (the disparity is due to build quality, lot size, and the streets they're on). Not surprisingly, Zillow handles this very poorly, and has consistently valued our house at much lower than it is actually worth (based on city assessments and sales of nearby comps).

Interestingly, Zillow recently bumped up our Zestimate from $190k to $230k. What's puzzling is that they also bumped up the historical Zestimates—there isn't a big jump shown on the graph on their site (though there is a big jump shown in my Personal Capital account, which uses Zestimates as part of their net worth scores). Either they modified their algorithm in a way that affected values retroactively, or they're trying to disguise the fact that their previous Zestimates were way off.

> What's puzzling is that they also bumped up the historical Zestimates

There are two things a "historical estimate" could mean:

1) What did Zillow say the value was on day X?

2) What does Zillow now think the value was on day X?

You and others seem to be thinking of (1), but I think Zillow actually is doing (2). Since the goal is to figure out what the house is worth and how its value has changed, I think (2) is actually the right metric here.

(2), of course, has the advantage of obscuring past mistakes and producing a nice smooth chart.

Unless you think they should never update their algorithm, approach (2) is just making the chart consistent with itself.

The useful part of these charts isn't really the current Zestimate, but rather the trend line of the home values, especially at the neighborhood level.

If you don't recalculate the historical values, then you break the ability to see the trend line.

If you really want to see the results of previous models, then the way to do that is to plot a separate line for each model, but you can see why they wouldn't want to do that from a product perspective.

Every county in the US that publishes historical valuations seem to get by without overwriting history.

Yes, but those are actual valuations based on sales prices, not predicted valuations based on an algorithm. Different thing.

> Either they modified their algorithm in a way that affected values retroactively, or they're trying to disguise the fact that their previous Zestimates were way off.

Probably the latter. Although, it could be that the data points are calculated on the fly or starting from some arbitrary value.

Then again, since it's shady to begin with, disguising their mistakes wouldn't surprise me.

Reply on news.ycombinator.com