Diamond Price Appraisal with Artificial Intelligence

[Prashant Telang]

Over the last several weeks TransPacific Software is working on various Linear and Neural Network models using Google's TensorFlow AI library to achieve fairly accurate diamond pricing tool. This tool can be used to find the fair price of the diamond in the current B2B market.
Tool is available on: http://52.14.211.195/ai/
Blog post detailing the Machine learning technology used is here This is a pure research project with no commercial value. Please do let us know your feedback

 

[OoohShiny]

Interesting.

I've not clicked the links, but is this basically using Big Data to ascertain minimum / maximum / median values for stones with matching parameters?

 

[Prashant Telang]

[

Interesting.

I've not clicked the links, but is this basically using Big Data to ascertain minimum / maximum / median values for stones with matching parameters?

Diamond pricing is too complex for regular statistical methods. We did try linear regression etc. but error margin was too high especially for big size diamonds like 3ct. above.
However, advanced Neural Network algorithms are able to appraise diamonds with 8-9% of error.
We fed the neural network with more than a million data points (diamond data culled out from P2P diamond exchanges like Rapnet IDEX online in last few months)
The features used are 4C and Crown, Pavilion, table and depth

 

[Karl_K]

The biggest problem is you have no idea what they actually sold for, the list price and the sale price can be very widely diverging numbers.
The bigger the diamond the more likely the numbers are to diverge.

 

[oldminer]

Human intelligence is pretty smart and readily available. Little will be saved with current models of collecting and calculating asking prices and nothing can be forecasted about actual transaction prices as they are nearly never made public. Years ago, a good friend and associate made a long term effort to figure how cut quality affected asking prices and found that the poorer cuts had generally higher asking prices then the nicest and best cut stones. The logical reasoning was that the ability to discount the poorer cuts was thereby increased and the seller could bend further to move those less desired goods along to end users. The best cut goods look far more like a pure commodity and so one had to ask a smart and tight price on those to attract knowledgeable buyers who were participating in a very narrow and highly competitive market with little or far less negotiating room. Asking way too much for fine goods made the seller look ignorant, but asking way too much for poor cut goods made the seller look like a smart dealer who could be manipulated for the final price. Initially this seem counterintuitive, but the reasoning became very clear in a rather short time.

I have seen automated pricing and on-line appraisal tools which do good job at final price estimation for normal items with very nicely cut diamonds of common sizes, but once cut quality is compromised, the advertised pricing found everywhere is off the wall, crazy and erratic. It is all part of the diamond game and this is a deep area where human game playing strategy outdoes AI in the current form it is forced to use to gain market knowledge. Just simply do a Pricescope search for a specific diamond and you will quickly understand approximate asking prices and how tight they are most of the time. I suspect may times they are tighter than 8 or 9 percent variance.

Maybe one day, we will have AI dealers in diamonds and they will share all knowledge, and pricing and actual transaction data evenly to all potential buyers and no one will make a living at it. Its a thought, but not one that I'd hope for. In my mind, our open source of asking prices is found already on Pricescope. Any seller truly wishing to sell any level of cut needs to ask a competitive and interesting price or no one will look twice at the stone. That's human intelligence joined with honest pricing and it makes the best sense out of an often difficult subject.

 

[valeria101]

The next step is asking that price.

[Recalling Martin Rapaport's take on the asking matter.]

 

[denverappraiser]

Much is covered in David’s post above but I”ll point out another one.

Let’s start with a sample diamond search.

1.00-1.03/H/VS1/xxx/GIA/USA

That’s about as mainstream as a diamond can get these days. In the database here there are 352 offers and prices range from $4615 to $8738

I used the same search in one of the very databases you are drawing from. At the moment 68 stones available ranging from $4897 to $7828.

The median stone is $6,362 and the margin is +/- 23%.

Next I did the same thing eliminating GIA from the specs. That gave 105 offers ranging from $2888 to $7828. The median stone is $5358 with a 46% margin.

Your system says it’s ‘worth’ $4900 +/- 8%..

Those are considerable differences.

You ask for Depth/Table/Crown/Pavilion in addition to a cut grade. May I ask what you do with that data? There are other traditional variables, like grading lab, location, and fluorescence, that you don’t seem to consider important. Although I agree you can glean some important cut clues from these, I'm skeptical given the elimination of what I would count as some pretty important variables that lead to an asking price span of nearly a factor of 3.

 

[Prashant Telang]

Thanks a ton for your feedback
We are experts in creating e-commerce websites for Diamond Industry but have limited knowledge in the diamond appraisal. One of the reasons I posted here is to get feedback from the community to improve the system.

Your question what we are doing with Depth/Table/Crown/Pavilion i. With these parameters, we are able to get Holloway Cut score. Currently, the HCA score is accurately computed by the system to determine cut quality of the diamond. But yet we are not displaying it we were not sure if it is worth displaying If it is we will display.

If Grading Lab, Location and fluorescence also matters in a big way we will add it in our AI system for weight age

would be happy to get additional feedback

 

[Karl_K]

"We are experts in creating e-commerce websites for Diamond Industry"
ok!
But you have 0 understanding of the b2b diamond marketplace and diamond value.
This is a forum for consumers to learn about diamonds and seek help from other consumers with some trade experts providing technical support.
I think my time would be better served helping them.
There may be more b2b oriented diamond forums that might better help you.

 

[Prashant Telang]

Yes we have zero understanding of Diamond B2B market and most traditional diamond dealers have zero understanding of complex technologies like AI.We have posted for feedback on other diamond forums and response is very encouraging.

"We are experts in creating e-commerce websites for Diamond Industry"
ok!
But you have 0 understanding of the b2b diamond marketplace and diamond value.
This is a forum for consumers to learn about diamonds and seek help from other consumers with some trade experts providing technical support.
I think my time would be better served helping them.
There may be more b2b oriented diamond forums that might better help you.

 

[Karl_K]

Your biggest problem is going to be and I don't see a way around it is garbage in/garbage out.
You simply wont have access to the data that really matters. Completed sale prices.

 

[Prashant Telang]

Much is covered in David’s post above but I”ll point out another one.

Let’s start with a sample diamond search.

1.00-1.03/H/VS1/xxx/GIA/USA

That’s about as mainstream as a diamond can get these days. In the database here there are 352 offers and prices range from $4615 to $8738

.

@denverappaiser
We detected this same case you have sighted. AI tool is giving pricing of $4650 for cut Good to $7676 for cut excellent, which is fairly matching with listed prices. You have used cut "Good" for all thus such large variation

 

[denverappraiser]

No, I didn't. I used GIA-xxx as one of my search parameters. I should include the options of x-vg-vg in addition to xxx, and that does make a difference, but I was trying to match your list as much as possible. Symmetry and Polish aren’t among your variables.

Yes I think you should use lab, location, and fluorescence as variables.

Outside of this very site, HCA score does not very directly correlate with price. The GIA cut score matters, and people subdivide that is an assortment of ways including HCA, but that’s far from the only or even the most common option.

 

[oldminer]

There are at the very least three levels of "Excellence" in cutting within the range GIA has chosen to use for their round diamond "Excellent overall cut grade". There are those stones which are closely matched to Tolkowsky's formulation of best cut, those just around the outer fringes of that and then quite a reasonably broad level or more of those diamonds which are well cut but graded "Excellent" under the GIA system. That third level of excellent cut has the steepest possible angles and greatest possible depth, the thickest possible girdles, etc, etc. The devil is in the details. Those stones retain the most weight so when you get one of those which weighs 1.02, 1.01 or 1.00ct, you can be assured that if it were better cut within the Excellent grade, it would not weigh 1.00ct, but weigh less than 1ct. That really makes a difference both in how they look, but also in what their real value basis is. We value them as 1ct diamonds, but some adjustment lower is assumed part of their valuation process and likely should be built into an AI system, too. They may have the same GIA grading, but they are distinctly not the same in quite detectable ways. An AI system might be initially easier to create using the far more limited scope of top quality found with AGS000 cut diamonds. Those give a very good clue as to the top end of cut quality pricing without mixing them up with 2nd and 3rd tier stones that the GIA lumps into one grading basket.

Just to be clear, in the overall, GIA EXEXEX diamonds look very nice, but they can be further divided and graded if there were a larger, general will to do so. The more liberal approach GIA took has been welcomed more by dealers and cutters and the reasons are economic, not science based. This is just one of the reasons prices being asked seem to be so broadly spread. Dealers and cutters have very good reasons to want grading done this way and the public is not being badly served, but for those members of the public who want stones at the top end of cut quality, the choices are fewer, but the dealers in these high end products are very driven to provide excellence and to satisfy the most demanding scientific approaches. Their asking prices likely are far tighter bunched in relation to one another.