One of these is not like the other

[musospuso]

I've been cramming and working on my grading chops lately but have been continually frustrated by the near constant contradictions in both photo grades and descriptions from different sources. From auction house photos to PCGS website there's just wild, stray examples in all directions.

Pop quiz: What grade would you grade the attached quarters as? Just broad categories needed i.e. P, VG, EF, no need for numbers.

The answer after a message from our sponsor......
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
OK, did you answer?

So apparently they are all EF (aka XF).

This boggles as clearly one looks more like VF and another closer to AU.

All graded by top-tier TPGs. It's driving me up the wall.

[JTCC]

Just goes to show the true nature of grading companies......

[Daniel]

I see this a lot, but you can’t leave out the reverse it’s a part of the grade as well.

[musospuso]

I see this a lot, but you can’t leave out the reverse it’s a part of the grade as well.

Yes, that is a good point. Particularly for Washington Quarters I tend to rely on the reverse as well.

There was a great quote (from Rod Gillis I think?) from one of the ANA grading videos on YouTube: A reverse will never increase a grade but it will bring it down.

[Daniel]

That was a quote from one person, the reality is much more complicated complicated.

[Earle42]

I agree with JTCC.

... but have been continually frustrated...

There is an easy solution...

Let the data speak for itself to you instead of being frustrated.

Mental Experiment:
Take a blank piece of wood you think would make a good 12 inch ruler. Make marks on the wood where you think the inches should go. Measure ten random sticks to a 16th of an inch and write all of these numbers down.

Have two of your friends make their own rulers and measure the same ten random sticks in the same order you did.

Granted...all three of you probably have a general idea of how long a piece of wood to use to make your rulers. All of you also may have a general idea of where to put the marks to make inches show on the wood you chose.

But I highly doubt any of the three of you would be frustrated your recorded ten measurements were all different!

Yet people can get frustrated when they find what you have discovered about the grading company systems!

The ink on the label is from someone who had a general idea of what has always just been called an MS level coin. And they (measure the stick) go to the nth degree of assigning a grade by saying it's a (for example) MS62, MS63, or MS64. But it is all fantasy done with a homemade ruler that was made with the best guess as to what 12 inches actually is.

Daniel's AI system for grading Morgans was able to see hundreds of (let's say) MS64 labeled coins and make a "fuzzy" approximation of what grading companies have generally labeled as MS64. But reality says get rid of the expectation that if a slab says the coin is MS64, then you can count on it.

And, as you are finding, forget the 16th of an inch accuracy, you can find slabs where they don't get it to an 8th (4th!) of an right.

Click to enlarge. Feel Free to share. Save for reference.

[musospuso]

I agree with JTCC.

... but have been continually frustrated...

There is an easy solution...

Let the data speak for itself to you instead of being frustrated.

Mental Experiment:
Take a blank piece of wood you think would make a good 12 inch ruler. Make marks on the wood where you think the inches should go. Measure ten random sticks to a 16th of an inch and write all of these numbers down.

Have two of your friends make their own rulers and measure the same ten random sticks in the same order you did.

Granted...all three of you probably have a general idea of how long a piece of wood to use to make your rulers. All of you also may have a general idea of where to put the marks to make inches show on the wood you chose.

But I highly doubt any of the three of you would be frustrated your recorded ten measurements were all different!

Yet people can get frustrated when they find what you have discovered about the grading company systems!

The ink on the label is from someone who had a general idea of what has always just been called an MS level coin. And they (measure the stick) go to the nth degree of assigning a grade by saying it's a (for example) MS62, MS63, or MS64. But it is all fantasy done with a homemade ruler that was made with the best guess as to what 12 inches actually is.

Daniel's AI system for grading Morgans was able to see hundreds of (let's say) MS64 labeled coins and make a "fuzzy" approximation of what grading companies have generally labeled as MS64. But reality says get rid of the expectation that if a slab says the coin is MS64, then you can count on it.

And, as you are finding, forget the 16th of an inch accuracy, you can find slabs where they don't get it to an 8th (4th!) of an right.


20190131_PCGS_Slab_Grade_Opt288.jpg

Another great example in your photo.

Still hoping for a Technical Grading-based app.

• Use all existing, documented grading rules and convert to algorithms
• Pass coin through AI+Computer Vision
• Apply algorithms
• ?
• Pofit

[Earle42]

Just get someone who knows how to access the face recognition app on an iPhone and use the 30K data points it takes in an instant (infra-red so now worry about lighting/flash etc.) to make a mathematically percentage based algorithm for assessing damage and wear. Sell it on the App store for for 10.00 and become a millionaire.

Better yet this would remove all of the non-science (nonsense?) and bring coin grading (back into the 20th century where it once was and into) the 21st.

[musospuso]

Just get someone who knows how to access the face recognition app on an iPhone and use the 30K data points it takes in an instant (infra-red so now worry about lighting/flash etc.) to make a mathematically percentage based algorithm for assessing damage and wear. Sell it on the App store for for 10.00 and become a millionaire.

Better yet this would remove all of the non-science (nonsense?) and bring coin grading (back into the 20th century where it once was and into) the 21st.

This will be one of my projects after the new coin flip.

I've gathered a few bids and assessments on what it would cost. I might try kickstarter or something along those lines.

[Earle42]

Great idea. Glad someone has the resources to know how to do this.

I like Daniel's AI system concept also. If you have the means, maybe offering to work with him might also be a possibility.

Something I often hear proposed as a con to a system like this (I have been talking about this for quite some time now) is that eye appeal cannot be told by a computer.

Two points:
1. An AI system likely could get a general idea of what people call eye appealing.
But...
2. I do not see that as a necessity since eye appeal already is not a part of the grading process. I have seen some pretty horrible looking coins graded very highly - meaning eye appeal is already not a part of the system. And things like tarnish (toning) look GREAT to some and UGLY to others.

[musospuso]

Great idea. Glad someone has the resources to know how to do this.

I like Daniel's AI system concept also. If you have the means, maybe offering to work with him might also be a possibility.

Something I often hear proposed as a con to a system like this (I have been talking about this for quite some time now) is that eye appeal cannot be told by a computer.

Two points:
1. An AI system likely could get a general idea of what people call eye appealing.
But...
2. I do not see that as a necessity since eye appeal already is not a part of the grading process. I have seen some pretty horrible looking coins graded very highly - meaning eye appeal is already not a part of the system. And things like tarnish (toning) look GREAT to some and UGLY to others.

Exactly. I'll leave things like "eye appeal" to Market Grading apps, TPGs, and human beings. I just want to get the ultimate Technical Grading tool to take subjectivity out of the equation as much as possible. There can still be debate as to what technically the algorithm between 35 and 40 will be, but once that decision is made (perhaps by open source or other consensus or governance mechanism) the app will apply it consistently and without bias as only a computer program can.

Totally down for collaboration and partnerships when the time comes. Pretty soon I think I'll post some of the information I've gotten from app bids and see what folks think.

[Earle42]

Sounds good - looking forward to it.

And let's face it...the Red Book, for decades before grading companies came along, had a decent standard everyone used and was happy with. Now with computers though, it would be a lot easier to set up an algorithm by having so many measurements and using them compute the actual percentage of wear as the baseline for indexing the grades.

The Sheldon system was based on opinion only and became the used system through marketing. I can see though where modern society in the coin hobby is ripe for a verifiable scientific approach. Well...except all those who put a lot of money into some slabbed coins that are proven to be less than the companies said they were.

Eliminating the Beanie Baby factor (no verifiable, inherent value [in the plastic and ink on the label]) is something I do look forward to.

[Daniel]

We have a system at mintstate.com that can grade Morgan Dollars.

Also, AI can identify the characteristics of a coin that people consider eye appealing or even ugly. People completely underestimate what AI is capable of doing.
However, programming is extremely time consuming and expensive.

[musospuso]

Sounds good - looking forward to it.

And let's face it...the Red Book, for decades before grading companies came along, had a decent standard everyone used and was happy with. Now with computers though, it would be a lot easier to set up an algorithm by having so many measurements and using them compute the actual percentage of wear as the baseline for indexing the grades.

The Sheldon system was based on opinion only and became the used system through marketing. I can see though where modern society in the coin hobby is ripe for a verifiable scientific approach. Well...except all those who put a lot of money into some slabbed coins that are proven to be less than the companies said they were.

Eliminating the Beanie Baby factor (no verifiable, inherent value [in the plastic and ink on the label]) is something I do look forward to.

That's what I've been reading. That Technical Grading was actually the front-runner roughly from the late 70s to early 90s and then Market Grading took over and no doubt has gotten worse as the "Everything Bubble" took off and is still looking for a pin...

Beanie Babys or Tulips, it's all the same jive, man.

[musospuso]

We have a system at mintstate.com that can grade Morgan Dollars.

Also, AI can identify the characteristics of a coin that people consider eye appealing or even ugly. People completely underestimate what AI is capable of doing.
However, programming is extremely time consuming and expensive.

You are correct I think. I do agree that an app can identify the characteristics that you point out ("eye appeal, "ugly"), you just have to define what that is and so far there's two ways that problem has been approached.

If the approach is to algorithmically decide up front what "eye appeal" means to an app (Computer Vision+AI system), that would be the superior method I would think. The community, experts, and all can debate the definition of "eye appeal" but once that definition is roughly decided upon by some governance consensus then that's that, that's the definition until such time that the community iterates and updates the definition. Which does not preclude the possibility of competing algorithms--I'm all for it! Let the best "eye appeal" definition win! Let the free market sort that out. It's up to the app(s) to then apply "the" definition algorithmically--consistently.

The challenge with the other approach, using existing, human-graded examples, with all the inherent flaws we talk about on these forums and at coin shows, as input, begs the question: why would we use that data as a benchmark for grading if there are such problems with it? If we all disagree vehemently when a TPG "mis-grades" our coin for "market" reasons, why would we use those as input? The CS axiom garbage-in, garbage-out still applies.

Numismatics is definitely ripe for tech disruption. Too much red-tape, gatekeeping, Ivory Towers, and Comic-Book-Guys. Bring'em all down!

[Earle42]

with all the inherent flaws we talk about on these forums and at coin shows, as input, begs the question: why would we use that data as a benchmark for grading if there are such problems with it?

In this case I think the GIGO principle is not quite applicable b/c the AI is using a ton of data to find the "average" of what a human mind calls any specific grade. The data it collects from, let's say, 200 MS62 Morgan slabs will definitely show variance (b/c that is how the current system of grading works). But the AI can easily evaluate each slabbed MS62 Morgan using a technical analysis to extrapolate from those 200 MS62s what the middle point of all those MS62s was.

So the more coins it sees, the better it gets at being able to tell what a perfectly impartial human grader with experience would call an MS62.

I guess another way of looking at it is that you are not putting garbage in, you are putting data for the machine to find the perfect midpoint and use that as the standard.

It would be interesting to see how the AI system and an algorithmic system would agree or disagree.

[musospuso]

with all the inherent flaws we talk about on these forums and at coin shows, as input, begs the question: why would we use that data as a benchmark for grading if there are such problems with it?

In this case I think the GIGO principle is not quite applicable b/c the AI is using a ton of data to find the "average" of what a human mind calls any specific grade. The data it collects from, let's say, 200 MS62 Morgan slabs will definitely show variance (b/c that is how the current system of grading works). But the AI can easily evaluate each slabbed MS62 Morgan using a technical analysis to extrapolate from those 200 MS62s what the middle point of all those MS62s was.

So the more coins it sees, the better it gets at being able to tell what a perfectly impartial human grader with experience would call an MS62.

I guess another way of looking at it is that you are not putting garbage in, you are putting data for the machine to find the perfect midpoint and use that as the standard.

It would be interesting to see how the AI system and an algorithmic system would agree or disagree.

Absolutely, if one wants a Market Grading app, then this approach is perfectly logical and acceptable--it provides Market Grading because that's what its data set is based off of and if that's your goal/desire, then mission accomplished.