[thoughts] On Predicting when Bullion will Return to GLD (FAILED PREDICTION) PLUS a new permanent spreadsheet resource

A number of months ago I predicted the return of some specific GLD bars, and I'm sorry to say they haven't yet returned and statistically there is now less than a 5% chance that they will. So that is a FAIL; my method for detection of unusual GLD withdrawals did not create an accurate prediction. I'm disappointed and it's back to researching that area. I will of course update this article if I ever see them back on the ledger, but because it's outside the three to four month range expected, I'm admitting defeat.
For what it's worth, I still think there is scope for the bars to return - a large increase in inventory will undoubtedly result in a number of 'previously lost' bars to be rediscovered, but in retrospect I should have picked an easier target, for example it would have been more accurate to predict bars which were 'most likely' to return given their history - the record is now 5 times for some GLD bars (meaning they have disappeared and reappeared back on the GLD ledger five times). I hope there is some novelty value in a blogger stating publicly 'I was wrong about my prediction', you won't often see it in the precious metals space.

The larger question still remains regarding the significance of certain bars swimming on and off the ledger (my term: "dark bullion"). Being 100% measurable, there is no debate the effect exists but the reasons behind why it occurs and what significance it holds is in question. The tally of bars ever seen on the GLD ledger stands at 154,182 gold bars (currently visible 64,263), which gives you an idea of the amounts of gold in the London system. I am now of the opinion that the returning dark bullion we see is coincidental, i.e. the bars which 'disappear and re-appear' regularly - do so because in such a large sample you will get some bars which will show up again. And in such a model, a prediction like mine SHOULD fail. We'll keep running tests and see where this ends up BUT so far it's leading to proof of being no link between specific inventory and price (not a conclusion I wanted).

Science is built on failures so this can still be counted as a progression ;p I have been doing more work on the data which I'll share later but I am also using this post to launch a new resource public which will be a permanent feature on the Bullion Bars Database page. Basically it's a public spreadsheet (updated occasionally) which contains a dump of all the bar records from our database, updated in the cloud. This holds every signature we've ever seen for Gold (over 250,000), along with their last known location. Any suggestions on how to make the spreadsheet better, are most welcome. The goal of publishing this is to make the data more accessible to independent researchers who might want to answer easy questions like where is bar ZJ6752 or how many Russian or Chinese bars are known. You'll need Excel 2007 or higher (or equivalent) to view the file.
(20 mb)
I have plans to update it periodically (maybe monthly) but the frequency of updating will be correlated directly with public demand and interest. Provided AS-IS with no warranty implied, yada yada.


Bron Suchecki said...

The other use of your spreadsheet is for those lucky enough to be able to afford a (or more) 400oz bar - they can check to see if their bar has been in an ETF, or, if they have sold, if their bar subsequently appears in an ETF.

Warren James said...

@Bron, yes that's right - the spreadsheet lowers the difficulty of doing a basic check.

I'm hoping this will help anyone doing bar-spotting out there, e.g. comparing serial numbers and weights against pictures of gold bars.

btw, the quantity of silver bar signatures is much higher > over 1 million. We could also do a spreadsheet for that too, but it would depend on how much interest there is.

Marks said...

What is the "Unidentified Refiners Bucket"? Some of these bars have serial numbers, yet one still cannot identify the refiner?

Also, what is Undisclosed location? You have the City but what causes you not to know exactly where the bar is stored in the City?

Warren James said...

Hi Marks,

Yes that's right. Some of the refiner brands we find in the lists are ambiguous or historically obscure > a good example is 'NMJA' which might be an acronym for New Mexico Jewelers Association, but we've not been able to find any correlating evidence to confirm the name (and no indication they actually produced any bars). No idea what CMX is. While I was writing this I realised that 'NA-G' stands for 'Norddeutsche Affinerie AG' which is a long shot but confirmed it from the pattern of serial bar numbers - so you can see the amount of detail which has to go into each unknown. The 'REFINERS' tab is the full database dump (Gold, Silver, Platinum) ... in retrospect I should have only included Gold refiners in this list to make it smaller.

For the vault locations titled 'Undisclosed London', 'Undisclosed New York', 'Undisclosed Zurich' ... yes you are correct. Mostly it is BullionVault records - the city is known and it's either ViaMat or Brinks but we don't know exactly which one because there are multiple Brinks London vaults. Likewise for Switzerland I don't know whether the ViaMat airport facility at Kloten is the only vault they have - it's my understanding that the facility is new so therefore there are at least two. I could probably amalgamate some of these into 'Brinks London' and 'ViaMat Zurich' as a general category. It was always the intent to do cleaning up but over the years BullionVault has made it more difficult to get their data and changed their URL's three times so the effort in cleaning the BV data is more trouble than it's worth.

Marks said...

It would be useful if we could calculate the average fineness of a 400 oz bar. Had to correct your data to make all fineness a 4 digit number and then divide by 243,128 bars. This yields average fineness of 99.8575% if my math/adjustments are correct.

About 138,000 of the 243,000 bars are 99.99% fineness = 57%.

I do not see a column for current location? That is, is the bar still in GLD/IAU/PHY or not...?

Is there some way your data can verify the following quote from Rickards/Alex Stanczyk trip to Swiss Refiner:

"We met with the managing director of the largest refinery in Switzerland and spend about two hours talking to him, we learned some very interesting things.
At this Swiss refinery there have been several times this year on which they were unable to source gold. They’re bringing in good delivery bars, scrap and dore from the mines, basically all they can get their hands on. This gentleman has been in the business for 37 years, he was there during the last bull market in the late seventies. I asked him when was the last time this has happened, that he was unable to source gold, he said never. And I clarified it, I asked: in the last 37 years you’ve worked in the gold industry this has never happened? He said: this has never happened.
He said sometimes when they get gold in, it’s coming from the back corners of the vaults. He knew this because these were good delivery bars marked in the sixties. This is a huge supply squeeze and its worse than anything that has happened in the last four decades."

Would assume the gold from the 1960's at the back of the vault would not be refined up to 99.99%...? Can your data somehow verify whether these Rickards quotes above are accurate? Was relative gold fineness via GLD withdrawals overly low fineness during 2013? Is there some way to verify the Rickards quotes from your data...?

Probably too much to ask from the data... If my math correct, your database totals 96 Million ounces which totals 2,982 tonnes of gold. That is only just above 1 year gold production.

Marks said...

GLD started off with 100 tonnes of gold inventory in November 2004:

Your first database purchase for GLD only shows total of about 24,400 ounces dated November 2004 in GLD = under 1 tonne.

Your second database purchase for GLD only shows total of about 498,800 ounces dated February 2005 in GLD = 16 tonnes.

How can GLD gold inventory be 100+ tonnes as of Feb 2005 yet your database only shows total of 16 tonnes.

FWIW, this Nov 2004 and Feb 2005 GLD gold inventory should be at the back of the vault... (also Mar 2005 purchases which are modest). This is the old GLD inventory, any special inferences you can make of this old inventory, is it more likely to still be held as GLD inventory than recent purchases?

Thanks for any comments.

Bron Suchecki said...

The reason Warren's data jumps like that is that is because he only has bar lists for those specific dates - he doesn't have bar lists for each day in those early years.

I kept track of the various ETF's daily ounce balances from the first one in Australia in March 2003 until Dec 2006 but not bar lists. My records show the early GLD daily balances as:

18 Nov 04 260,000
19 Nov 04 1,859,994
22 Nov 04 2,799,953
23 Nov 04 2,799,953
24 Nov 04 3,100,000
25 Nov 04 3,100,000
26 Nov 04 3,230,000
29 Nov 04 3,330,000
30 Nov 04 3,330,000

Warren James said...

@Marks, lots of questions there, in answer:

1. The figures from the Zeal chart seem correct according to our 2004 and 2005 sources but the 'DateFirstSeen'+'DateLastSeen' combination hides all the activity in the middle. i.e. the bars first seen in 2004 were still present and contributed to the 2005 total, so any query or filtering has to include the earlier dates and be cumulative. But the spreadsheet only shows total signatures seen, so those totals from the spreadsheet should be more than what we see in the official tally. But apart from that, if you do a filter where 'DateFirstSeen' = '2005-02-25 00:00:00' AND 'CustodianCode' = 'GLD' then you should see 12471 bar records, which is the first 2005 document being added. Take 400oz bar weighs ~ 14 kilos, that gives 174 tonnes for the 2005 document in the right ball park (and add the 2004 weight as well), just make sure autofilter isn't filtering out any additional items.

2. Current location for bars is tricky, we can only ever display the 'last known location', since there might be a document we haven't yet processed or there might be a wait until the next document issue (e.g. Sprott's documents). So if the last known location is recent (most of the GLD records are updated daily/weekly) then that is pretty much their current location. A lot of the GLD signatures in the spreadsheet include all the 'lost bullion'.

3. You're right, some of the bar signatures are not clean, for example the three-digit fineness records. We've actually preserved the original signature as-is however recently we've discovered some more patterns which confirm that in most cases the missing digit is a zero from the end. The spreadsheet actually shows all known signatures, so some entries are double-ups ... but in the database we have the ability to track the different signatures as the same bar. Therefore you could shave about 5% of these totals, but yes even with all these bars represented it is still just a small sample of the investment-grade gold available on the London market.

4. Re: 0.9999 fineness ... You're right, the data can't show us information about bars leaving the ETF for being re-melted, the best we can do is track that it is currently missing, and flag it as lost bullion or dark bullion if we ever see it come back. There were some significant specific withdrawals during 2013 and data is tantalizing ... seem to have it's own intriguing story to tell but I've never been able to decipher or slice the data to make a cohesive narrative out of this. What I can tell you is that the composition/ratio of 0.9999 bars in GLD hasn't actually changed that much over the years, and that not all recently refined bullion is necessarily 0.9999. It's my opinion that other purities are not necessarily indicative of age ... it's more about market conditions; sometimes there is demand for lower purities, sometimes not.

Hope that helps. Some of what you're interested in might be best served by some additional spreadsheets, so please let me know what you might be interested in seeing and I can try and publish the relevant data set. The more questions we ask of the data, the more we'll get from it. p.s. I confirm your average for all bars in the sample is 0.9985

Marks said...

Thanks, Warren, your comments appreciated.