Putting the above noted large dataset aside, there remains the interesting question as to why aren’t any of the appraisal competitors using the exchange listed unsold information? One possible explanation can be based on scientific findings that humans (and other animals) only notice what has happened and ignore what has not happened, i.e., they notice a sale that has taken place, but ignore the no sales, whereby no transaction has taken place. Obviously, all relevant information needs to be included for more reliable predictions. I will leave other explanations to your imagination.

2. “Human” vs. Machine
If it can be measured, it can be scientifically studied! Thus, not having the right tool to pound a nail can be frustrating. But, it sure beats pounding it with your head. Similarly, we don’t know the true domain name price-generating model, but we can do a lot with what we have and know.
There are a number of plausible explanations as to why consumers seek “human appraisals.”

(a) Humans like to “control” outcomes. Control can be achieved through human involvement in appraisals. Thus, an appraised value controlled by a human is preferred to a potentially superior one generated by an automated system.

(b) Humans overvalue events that they think they control. Thus, they unjustifiably place more value on a “human appraisal.”

(c) Some customers project from a narrow experience. I constantly hear things like: “I have seen some automated real estate appraisals. They are worthless;” and “I have seen few of these domain name appraisals. They are based on a machine that spits random numbers.”

(d) A human appraisal can potentially be based on an appraiser “randomly picking numbers from a hat.” However, it is very likely that’s not the mental image that customers form of a human appraiser. They imagine due diligence. Thus, the mental image of the process and what an appraiser actually does can be out of whack.

(e) Human appraisals can have major disadvantages. For example, in a scientific study, Americans were asked which countries were most similar to each other – Ceylon and Nepal or West Germany and East Germany. Most picked the latter. But when asked which countries were most dissimilar, most Americans also picked the same pair. How can a pair of countries be similar and dissimilar? Does this imply that logical analytics is always superior to gut feeling decisions? Not necessarily! However, when it comes to domain names, due to the massive amount of data, analytical models are superior in detecting comparables.

It should be noted that regression-tree models require human tweaking to determine the optimal tree size. Thus, at least in this sense, they are human based.

Similar aversion to automation has been noted in selecting equity investment funds, despite the strong empirical evidence that automated rule-based stock picking systems have outperformed the “human expert” funds.

3. Statistical Significance
Our regression-tree models are statistically significant. For example, in some of our linear regressions, the relatively naïve model explains over 70 percent of the variations in prices. Regression-tree models are nonlinear and are more robust to outliers.

Moreover, no matter how many times an appraiser “picks numbers from a hat,” it does not make him a significant appraisal expert.

Hence, one can easily design a powerful statistical test to determine whether or not an appraiser is coming up with voodoo reports. Yet, customers are not demanding such proof!

(1) How to find comparables?
a. You seem to suggest that by loosing at sales of domain names that have a common word, such as mortgage or cash, you are able to determine comparables. Does not make sense!

b. Do you use the average of comparables or the median or something else? Again, without statistical models you cannot say anything intelligent about your prediction.

c. Are three letter .com domains comparable? Definitely not! So what do you do?

d. What about various extensions? How do you find comparables?

e. Thus, using statistical techniques is the only way to find comparables. Moreover, using statistical models you can find a more-inclusive set of comparables, thus, a more accurate appraisal.

(2) Human vs. Statistical Models
a. How can you evaluate your human appraisal model’s predictive accuracy?

b. How can you know that it is good? How can you improve on your model? For example, people consider the length of a domain name, which is practically irrelevant. Yet, people keep saying it without any merit. For a second, let’s suppose length is important. How important is it? If you choose a name with, say, 5 less characters. How much more valuable would it be?

c. “You can get an appraisal, but few people place stock in appraisals.” Are you among the few people who place stock in appraisals? Why or why not?

(3) Rich Dataset
a. For example is traffic important? Are links from search engines important in valuation? Yes! But your approach completely ignores it. Moreover, how can you know what the domain name’s traffic and links were at time of sale? Thus, you need a richer dataset to appraise domain names.

b. What about unsold domains? Can they add information? Yes!

There is no doubt in my mind that some domainers may be able to find domain name pricing patterns without an explicit statistical model. However, I am convinced that they should be trading financial instruments or playing professional chess.

I hope this post helps us identify voodoo appraisals.

I look forward to comments and suggestions.