An example of the Dark Bullion Effect (GLD)

The GLD inventory the other week had a small uptick of inventory - just 96 gold bars - not much in the scheme of things but enough to do an analysis of 'where did those bars come from?'.

I consider the dark bullion effect significant because it invalidates more than one general thesis out there (about the metals). For anyone new to the discussion, I am defining dark bullion as 'A bar previously known and registered, re-appearing on a list after an absence.' I stand resolute on this topic simply because the signal (of the dark bullion effect) is unmistakeably clear and provides a different view of metal inventories - I have been pursuing this ever since I first detected the effect in 2011, and I want to talk about it more, with the hope that with enough discussion we might be able to squeeze some secrets out of the data.

Every good test starts with a theory or premise. For the recent 96 'added' bars I originally expected a high percentage (at least 80%) of 'previously known and registered, re-appearing on the list' bars, given the large amount of outflows we've seen in early 2013. At first I rushed some analysis and concluded to my colleagues that only 24% of the 'new' bars were previously seen. But during that same week I did more measured studies using some of my new tables and while I'm embarrassed that my first quotation was incorrect, I am glad to find the new figure is 66%. Getting different results from the same test always sucks, so finally these figures are checked and cross-checked. The curious critic who wishes to debunk my work is invited to do so - the spread-sheet contains references to the original PDF source documents which are downloadable. Figures shown are for all Dark Bullion GLD Additions detected so far this year (i.e. 701 dark bars in total, returned - 45% average for 2013).

Cutting a long story short *, across the last 3 years GLD show a steady 30% 'average rate of return', meaning that on average 30% of the bars we see being added to the inventory, are bars that we've seen before in the data. Later we hope to look at the other 70% of bars and figure out whether these are old or new, but this will take a fair amount of work to sift through the different production sequences from each refiner.

In my last article I used high-school-level-math (statistical probability) to estimate the total stock of gold in the HSBC vault, but now that I've processed the three year's worth I see that the sample sizes available are too small and (as such) a small variation in assumptions about delivery can skew the estimate wildly. Still the back-of-the-envelope calculations are useful and the percentages involved are surprisingly consistent.

In recent results, we've identified gold bars which regularly swim on and off the register:
GLD Bar movements - the '1' is not special, but the '2 times' is, because it means
these bars moved off the bar list once, came back and vanished again.
Measurement is from over 3 years of data (440+ source documents).
What you're looking at here is that there are some bars (quite a few) which were removed from the bar list more than once, a handful of bars which have been removed from the inventory at least three times in total, and even some bars (121) which have a count of 4. So yes, that actually means that for some bars, they were removed from GLD, then added back, removed again, added back later, removed once more then reappeared on the list again and is gone again. This could be expressed as percentages against GLD's bar holdings, if you took 100,000 bars as your 'base', then just under 8% of GLD bars have been removed from the inventory at least twice, ~ 2% have been removed three times. Fun! With a big enough sample size and enough time, randomness will explain these kind of movements, but this curious effect requires more study. With SLV, the same pattern applies and is greatly amplified; you'll have to wait for the next article for that.

That's all I have for today. In one way the study is a failure again because the data creates more questions than it answers. This won't be a satisfactory state of research for casual critics, but I'm happy to be at the stage of asking more advanced questions.

Warren James

Oh, and just so it's clear — the reason for writing this is there is proof of SOME bars that don't appear to have left the vault when they disappeared from the GLD bar list the first time. Although we're talking about a small percentage, we've at least moved on to discussing PERCENTAGES and not absolutes. i.e. these particular bars did not go to China (argument type otherwise known as Falsifiability. I do acknowledge that the GLD drain is an unusual event but I would argue it can't be described in a simplistic fashion).

* In regards to data integrity, my new tables revealed for the first time, gold bars with changed signatures. Aided by a data method kindly provided from, I managed to clean up about 1,000 bar records (which had previously been showing up as same-bar-different-signatures). Now my GLD figures are now 0.5% 'more accurate' and somewhere in the piece, the day spent cleaning up the core data will make a difference. The adjustments were incredibly slight - some extra digits on a Russian bar serial number, that sort of thing. At time of writing I'm in the process of using the same mechanic to clean up the silver data, which we know have a lot of changed signatures from when the bars got moved to the Brinks vault - current estimate is this produces ~5% limit of reading error (on SLV data).


costata said...

Hi Warren,

Great work (as usual). I keep coming back to this:

"Cutting a long story short *, across the last 3 years GLD show a steady 30% 'average rate of return', meaning that on average 30% of the bars we see being added to the inventory, are bars that we've seen before in the data."

I imagine it's a tonne of work ;) but I'm wondering if the vintage of the returning bars could allow us to extrapolate from this 30% rate of return. Assuming that very few of these bars are ancient.

Say, all bars are from vintages no older than 3 years. Estimated total fabrication of LBMA GD bars in all vintages x 30% then "dark pool" = XXX m/t.

If there is a taper in the age of the bars that could be factored in to smooth the data i.e. 60% of returns are less than 1 year old, 30% are 2 years old and so on. (I'm assuming an underlying trend of accumulation = "lost" bars.)

Even if it gave us a range of roughly not-less-than to not-more-than it could be indicative of the size of the dark pool.


Bron Suchecki said...

I would not be surprised to find a bunch of really old bars that never leave, those are probably the ones up the back of the vault that the vault staff couldn't be bothered moving :)

The problem with the ETFs is that their holdings have been far too stable, just growing and growing. We haven't had enough volatility with lots of periods of withdrawals and periods of accumulation. That is what we need to really build up some data on the size of the stocks in London.

Warren James said...

Thanks Costata, Bron, for the feedback.

Yep, our new techniques should allow us to identify the bars which never move (I'm calling this 'Sticky Bullion' and it will feature in 'SLV Database 6'). Physical location within the vault may indeed be indicated by the bars which go back and forth regularly (i.e. bars close @ hand), but we still need more data to build up those vault profiles.

The 'method of selection' for bars being removed and added is one of the standing questions that remains unanswered for now, but with enough pattern analysis we might have some answers.

I too, am guessing there's some kind of pattern in regards to the appearance of new and old stock and I'm (still) hoping there is some relationship with price. But yes, it's a metric tonne of work to accurately flag each of the production runs.

re: volatility - Bron yes you're absolutely correct that bar list volatility creates better data points for us. In this manner the current 'drain' on GLD actually works to our advantage ... especially once bars get added back in GLD again. In that case on average we could expect at least 30-45% of the added bars to be bars we've seen before.

In preparing the data this weekend I have suddenly realized the significance of the ETF Securities inventory - many bars of which share the same vault as GLD - once that data is folded in it should have the effect of making the slope of the dark bullion graph even smoother. This will be reflected in my next article.

Bron Suchecki said...

Certainly you should combine data from any ETF which shares the same custodian and vault, although I would argue for combining all data for metal stored in London in one data set.

costata said...


"..I would argue for combining all data for metal stored in London in one data set."

I presume you mean treat it as a single vault. If that is correct the suggestion makes a lot of sense to me.