There has been a lot of media chatter over the last week regarding the possibility of the London gold price having been rigged, with Bloomberg leading the pack and with a lawsuit now having been filed. Shades of Auric Goldfinger haunt the trading desks. The way that this has been presented is that a set of researchers have found, using a tool that previously found manipulation in LIBOR, evidence of a fix in the gold fix. There may or may not be a market cartel rigging the gold market but my reading of what we have so far suggests that the evidence simply does not permit this determination to be made. Extraordinary claims require extraordinary evidence. What we have is …nothing much. Let me explain
The issue arose last week when Bloomberg published a piece that stated that evidence existed of a decade long price manipulation. But, the first problem we have is that they, and only they, have seen the paper. I understand that they are refusing, on the request of the authors, to release the paper until a draft is complete. This is shoddy academic practice of the first water. Publication by press release is not how good science is done. How good science is typically done is for a paper to be circulated in draft form to people knowledgeable of the area. They critique the paper and the authors incorporate these issues. Then a paper is perhaps presented to a conference, or placed on an open access repository of research papers such as SSRN.com. More comments, more critique, more iterations of the paper. Then it is sent to a journal or an edited book and yet more reviews are made and incorporated. If you have something really really exciting then when the paper is accepted you maybe release a press notice. Occasionally a paper that has gone through many rounds of review and rewrite will be press released or attract media attention when it hits a repository or a conference. It seems that the paper here is not even a complete draft. If the authors didn’t give it to Bloomberg then they have been remiss in noting loudly and constantly that they are not able to stand over the findings; if they did they are equally remiss in short-circuiting the usual process which is designed to catch errors and to spare all our blushes.
The paper that is alluded to as having “helped uncover the rigging of the London interbank offered rate” appears to be at the heart of this. In so far as we can tell, and remember we haven’t seen the paper, the methodology used is a variant of it. That paper is found in its 2008 draft form here and was published in Journal of Banking and Finance (of which I am an associate editor, but I had no involvement in the paper) in 2012. It takes about thirty seconds to read the abstracts. The 2008 paper abstract concludes “while there are some apparent anomalies within the individual quotes, the evidence found is inconsistent with an effective manipulation of the level of the Libor” and the 2012 final version states “We find some anomalous individual quotes, but the evidence is inconsistent with a material manipulation of the US dollar 1-month Libor rate.”. So, far from uncovering the LIBOR scandal they claimed there was none. There was. This doesn’t seem a very powerful test, with a Type 2 error returning a false negative . A more powerful and simpler test, but one that is not possible to be applied here as we do not have the quotes of the fix participants, was that of Snider and Youle, which DID suggest that there was a problem.
To properly analyze the fix in the manner which they analyzed the LIBOR fix they would need to have details on the positions and quotes taken in the fix by the fix members. This they do not have, and I am not aware that this is even collected. Thus they are (it seems) looking at the events around the period of the PM fix. Herein lies a problem. It is not clear what window they use, that is to say the frequency of the data they are analyzing. This is really important in the context of a market which may have upwards of 50% of its trading now coming from algorithmic trading positions. A tick-by-tick or better yet quote-by-quote analysis would be needed to see what was happening. A proper analysis would need to examine the timestamp of when the fix is released and the trades/bids immediately thereafter. It would need to see if the fix banks were able in fact to take consistent superior profits. To see how machines and people interact on information releases, which is what the fix is in essence, see here
The collusion test they use is in essence the following: where there is collusion there should be no real difference in average returns (as opposed to a non collusive market) but there should be lower variance. An immediate problem arises in that lower cross sectional variance is a hallmark not only of possible collusion but of herding. And we know that gold market participants herd (see McAleer , and a number of papers by Pierdzioch . Herding is a feature, at some stage, of nearly all markets. A test that can conflate herding with collusion might be one that would give us pause to think before we leap to a conclusion of one or the other being present.
Even leaving aside this, a further issue is how the authors calculate the intraday variance of trades. From their LIBOR paper they seem to use the coefficient of variation. This is very simplistic and when one is dealing with intraday variance it is really not at all clear that this is what one should do. Particularly when dealing with higher frequency data, and bear in mind we do not know what the data frequency they are using is in fact, there are microstructural issues to deal with in the estimation of volatility. Personally, I prefer to use range based estimators such as the Parkinson or Garman-Klass estimators of variance at intraday frequencies. Theres a very highly cited paper by Alizadhe (available in a working paper form here) on the superiority of these and other range based estimators. Some limited work on ultra high frequency gold suggests that volatility measurement needs to be taken very seriously
Even if we were to agree that the volatility measure used was adequate, if we know one thing it is that volatility is not constant. Engle even got a Nobel prize for a whole family of measures (ARCH and its many offshoots) to estimate and counter this. Looking over a 15 year period as they seem to do it is abundantly clear that volatility changes. Thus any statistical test that looks at the likelihood of a particular outcome as being within normal ranges (as the paper seems to do we are told) absolutely has to take into account the shifting nature of how likely likely is at any given time over any given window. Does it do this? We don’t know. Do the tests take account of the non normality of the data? Are they non parametric perhaps? We don’t know
A further issue is that we do not know what was going on on the days that this anomalous behavior was detected. Gold acts as a safe haven against extreme stock and bond movements (see here for the original paper and here for an extension to oil and currencies) A full investigation of how other markets were evolving on those days and why, whether it be in reaction to news or macro releases or whatever would be most useful. The PM fix is at 10am New York. Is the market in London reacting to New York market issues? We know that the locus of price determination shifts from London to New York and back again. See here. Is that what is happening? Models of the gold price routinely incorporate a range of real and financial variables, which themselves change and which have announcement dates. These need to be taken account of. We simply do not know if the detected changes were explicable by other , non collusive, market reaction.
The paper, we are told, because we don’t know what is in it, finds a break in the behavior of the fix in 2004. This again needs to be looked at. First 2004 represents the first real takeoff of gold from the doldrums of having been stuck in the $200-$400 for the previous 20 years. As markets accelerate the behavior of people changes. Second, we have in mid 2003 the first gold ETF. The market changed fundamentally in 2004-5 and merely finding a break in its behavior there doesn’t indicate anything other than that the market changed.
Bottom line – without seeing the paper its hard to tell what’s going on in it (duh…). However, I would be astonished if there were not a fairly reasonable set of explanations at least potentially available for anomalous (if they be so) changes around the fix in certain days. Extraordinary claims need extraordinary proof. We don’t have any, yet.