Monday, March 25, 2013

Celent Report on Electronic Trading in US Corporate Bonds

Here are some statistics from the recent Celent report on US Corporate bond trading.

  • As of January 2013, holdings of corporate bond inventories at the 21 dealers that trade with the Federal Reserve have declined by 74% to $56.4 billion since the 2007 peak. 
  • Only 30 bonds a day have more than five trades on either side in institutional size with names changing all the time due to new issues, according to MarketAxess. 
  • According to ITB, since 2009 odd lots/super-odd lots notional (US$100K$1M) have seen an increase in daily notional volume of 33%. 
  • There has been only a modest increase in ADV (up 6%) for trade size US$1M$25M since 2009. 
  • Average trade size on RFQ platforms (i.e., MarketAxess) has been declining 56% per year.
  • Average trade size on MarketAxess high grade is now ~$400K. 
  • In the US, corporate bonds volumes compiled with TRACE data are only slightly up year on year at just under $12 billion of average daily volume, and it has been rather stable for the past four years 

I've integrated some platform specific statistics from Celent into the listing of electronic venues

Monday, January 21, 2013

Univariate statistics penalize multivariate filters

A while ago I wrote about simple accuracy statistics that, when applied to commercial pricing services, appear to be more than sufficient to delineate the good from the not so great. But I also mentioned that univariate statistics are somewhat simplistic, especially for assessing multivariate de-noising of market data. That comment was based on my numerical experiments, though it was somewhat underdeveloped.

This writeup (thanks to Nick West and George Papanicolaou) is a clean exposition of the point and is based on an independent numerical study. Here is the abstract:
In this report a simple piecewise constant curve based model is used to analyze the ability of various assimilation methods to be calibrated to observed market yields. To simplify the analysis, we make the following assumptions: (i) the yield of a bond (the observed quantity) is simply the integral of the hazard function through the maturity of the bond; (ii) the levels of the hazard function follow independent random walks; (iii) different bonds on the curve (with di erent maturities) trade with different frequencies, some of them continuously and others in bursts; (iv) observations that occur near (in time) to other observations have correlated observation errors. Assumptions (i) and (ii) allows usto easily apply the the Kalman Filter to obtain the optimal estimates of price; (iii) and (iv) are derived from our observations and experience with bond data. It is shown that a multivariate Kalman Filter's estimate of the current yield is both more accurate and precise when compared to the true, unknown yield. Furthermore, it is demonstrated that when the "next trade" is chosen as a target, serial correlation in observation errors cause the last trade to be chosen as the most accurate predictor of next trade; however, it is much less accurate at estimating the true yield; this bias is most notable is actively traded bonds.

To be clear, the report is not advocating reliance on a Kalman Filter per se. Rather, it is using a simplified model where the true prices are known to illustrate that simple univariate tests give misleading results (unfairly penalizing multivariate filters).

Thursday, January 10, 2013

Listing of New and Old Corporate Bond Trading Venues

Venue Launched Players and structure Liquidity, stats and so forth
New York Stock Exchange 1792 Specialist Bond Liquidity Providers (BLPs) Sample liquidity report
BondDesk 1995 Closed (BondDesk Institutional) and open (BondDesk ATS) Current offerings. Sample transparency report
TradeWeb 1997 Sell side to buy side RFQ model
Knight Bond Point 1999 Closed EOB Began as ValueBond
Market Axess 2000 Sell side to buy side RFQ model
Bonds.Com 2011All-all
BondVision 2011 Sell side to buy side
Aladdin 2012 Blackrock clients.
Vega-Chi 2012Select ATS High Yield only (stats)
GSessions 2012 Goldman clients On hold?
UBS PIN-FI 2012 UBS clients. "Price improvement network"
CitiCross 2013 Citi clients (hybrid electronic/voice) Story
JP Morgan 2013 JP clients
Bond Pool 2013 Morgan Stanley clients Story
DelphX 2013 All to all Larry Fondren (of InterVest)
If you'd like your bond trading venue to flash on this table send $20 in a brown paper bag to the author.

Friday, December 14, 2012

Mirror, mirror on the wall. How shall we define "fair value", if at all?

Today I managed to sneak away from the office for some quiet contemplation on the matter of pricing accuracy. I don't think anyone noticed I was gone.

Does anyone care about the accuracy of their vendor prices?

Pricing accuracy in the bond markets ought to be a hot topic. After all, how can one cross trades effectively, or monitor customer markups (something we know isn't really done) or optimize one's portfolio, or perform any number of front, middle or back office activities with confidence? Many industry participants bemoan the poor quality of third party bond pricing and perhaps unremarkably, no vendors of bond or CDS prices dare to quantify their accuracy. The question of accuracy is never broached except through vague references to confidence.

On the vendor side it is sometimes argued that customers are insensitive to pricing quality and the service is therefore sticky. But that argument presumes there will be no material change in market structure or competitive forces - an argument that was admittedly correct in the past. Major buy side firms are gearing up to better quantify their relative transaction costs (it provides an excellent marketing opportunity if nothing else). Others wish to use their inventory to generate alpha. And many are looking for lower cost means of making markets or supervising the same.

So the relevant question in a couple of years might not be "is accuracy of end of day pricing important?" but "why would I buy an additional service for end of day pricing only that is less accurate than the real-time services I have recently taken on? Who would want to generate day one P/L issues for themselves, for example?

Thus in the interest of fighting accuracy apathy, either real or perceived, we consider here two simple targets that vendors might be asked to aim at when pricing the "fair market" value of a bond. They are imaginatively called the "fair value target" and the "fairer value target". The first is easier to explain. The second slightly more logical. Both are flawed, but let's not make the perfect the enemy of the good.

A Fair Value Target

The Fair Value Target at time is a "size", "money" and "time" weighted average of the subsequent interdealer trades, where loosely speaking, "money" = \(\int\) "size". Specifically, if we fix some moment \(t\) at which a price is supplied by a vendor and consider the \(J\) subsequent interdealer trades (say \(J=25\)) one might compute $$ FVT(t;J) = \frac{ \sum_{j=1}^{J} p_j s_j e^{-(t_j-t)} e^{-M^-_j} } { \sum_{j=1}^{J} s_j e^{-(t_j-t)} e^{-M^-_j} } $$ where \(p_j\), \(s_j\) and \(t_j\) are the price, size and time of subsequent interdealer trades with time measured in business days. This is just an exponentially decaying weighted average where imminent trades are weighted more heavily than distant ones. But there is also an additional "money" decay term I have thrown in - at the very least to provoke discussion amongst the \(2\frac{1}{2}\) readers of this blog. Here $$ M^{-}_j = \frac{1}{c}\sum_{k=1}^{j-1} s_k$$ is the cumulative trading volume up to but not including the trade in question, and we set \(c=$1,000,000\), say, so that "money" is measured in millions. Some motivation for this additional term comes from the notion of a risk limit or rather, the notion that a sufficient volume of trading is sufficient to establish a market price (and therefore render subsequent trades irrelevant). For example it seems unreasonable to argue that $20M of thursday's trades should be included in the assessment of accuracy of Tuesday night's end of day price if there was, say, $10M of trading on Wednesday.

A Fairer Value Target?

At time of writing the Fair Value Target has been road tested with valuation specialists at a couple of bulge bracket banks, and with traders looking to auto-quoting bonds. It has proven reasonably popular, though that may reflect more on the gaping void in this space than anything else. The fair value target is open to several critiques and I decided to mention one here before anyone else noticed: the slightly unnatural use of \(M^{-}_j\). Indeed the fair value target, as written above, is not invariant to splitting of future trades. We can easily fix this, however, by integrating in money instead of time. Thus we might write $$ FVT'(t;J) = \frac { \int_{m=0}^{M^{+}_J} p(m) e^{-m}e^{-(t(m)-t)} dm } { \int_{m=0}^{M^{+}_J} e^{-m}e^{-(t(m)-t)} dm } $$ where \(M^{+}_J := M^{-}_{J+1}\) is the total amount of money under the bridge up to and including the \(J\)'th trade, \(p(m)\) is the price when \(m\) dollars of trading has occurred, and $t(m)-t$ is the time we have progressed when \(m\) dollars of trading has occurred. The reader may verify that this amounts to the following changes in the fair value formula when expressed as a sum over future trades: $$ FVT'(t;J) = \frac{ \sum_{j=1}^{J} p_j e^{-(t_j-t)} \left( e^{-M^-_j} - e^{-M^{+}_j} \right)} { \sum_{j=1}^{J} e^{-(t_j-t)} \left( e^{-M^-_j} - e^{-M^{+}_j} \right) } $$ where again, \(M^{+}_j\) is shorthand for \(M^{-}_{j+1}\), the money that has flowed under the bridge by the time the \(j\)'th trade is "over".

Which target is best?

Now I personally prefer the aesthetics of the "fairer" over the "fair", but admittedly the former is slightly harder to explain. As an aside I certainly make no claim that either is optimal for assessing the de-noising of multivariate data with serial correlation, for example, such as appears to characterize the corporate bond market. But if we restrict ourselves to univariate statistics and specifically the "fair" and the "fairer" as defined above, then as a pragmatic matter I have not yet pushed for the "fairer" over the "fair". I'm just not sure it makes any difference to the results so why limit the audience? Simplicity is a virtue.

For those who may be interested here is a histogram of the ratio of the two fair value proxies when computed for a reasonably large collection of trade time series. The ratio is reasonably close to unity, though I am not suggesting we dismiss the difference on this basis. My temporary opinion is based more on the fact that overall results tend to be robust to much bigger changes in the target than the ones we contemplate.

So who wins?

This is, as yet, no competition in the real-time bond pricing space. We can say, however, that Benchmark's realtime "Magenta Line" sub-sampled at the end of the day is roughly \(25\) percent more accurate than any other end of day data your author has been able to get his hands on, including that supplied by vendors and also internal marks from banks. Perhaps the next step would be for audit firms to perform their own, independent analysis.

Thursday, June 28, 2012

Should rule G-43 call for 3rd party reference prices?

The Bond Buyer covered the new MSRB rule G-43 intended to protect retail muni investors from predatory broker's brokers' practices. Broker's brokers are one of several intermediaries in the muni bond industry characterized, in part, by the so called bid-wanted auctions. Rule G-43, in part, is intended to police this activity and not dis-similarly to the corporate market involves a somewhat arbitrary notion of a roughly reasonable price.

That is because in sourcing interest for bonds broker's brokers will sometimes contact individuals who bid prior to the conclusion of the auction, even without informing the seller of this contact. To limit the nefarious possibilities arising from this sort of communication, rule G-43 lays down criteria under which this communication may take place. Incidentally it also makes other demands, including reasonable dissemination of the bid-wanted lists:
Unless otherwise directed by the seller, a broker’s broker must make a reasonable effort to disseminate a bid-wanted widely (including, but not limited to, the underwriter of the issue and prior known bidders on the issue) to obtain exposure to multiple dealers with possible interest in the block of securities, although no fixed number of bids is required.
Here, however, is where the notion of a roughly reasonable bid enters:
If the high bid received in a bid-wanted is above or below the predetermined parameters of the broker’s broker and the broker’s broker believes that the bid may have been submitted in error, the broker’s broker may contact the bidder prior to the deadline for bids to determine whether its bid was submitted in error, without having to obtain the consent of the seller.  If the high bid is within the predetermined parameters but the broker’s broker believes that the bid may have been submitted in error, the broker’s broker must receive the oral or written permission of the seller before it may contact the bidder to determine whether its bid was submitted in error.
If the high bid received in a bid-wanted is below the predetermined parameters of the broker’s broker, the broker’s broker must disclose that fact to the seller, in which case the broker’s broker may still effect the trade, if the seller acknowledges such disclosure either orally or in writing.
Thus the rules related to communication with bidders come down to an internal determination of whether bids were above or below a given threshold. It would seem that an independent, third party estimate may be the preferable criteria.

Tuesday, June 26, 2012

A Few Statistics from the GAO Report on Municipal Bond Market

The U.S. Government Accountability Office issued a congressional report on the municipal bond market a while back. A few notes:

  • One percent of muni's trade once a day or more
  • There are 46,000 issuers and between 1-1.5m securities in play
  • FINRA oversees 98% of 1,800 MSRB-registered broker-dealers
  • FINRA investigated 5,764 times resulting in 51 occasions where G-30 violations were pursued, of these 37 resulted in cautionary action, 11 resulted in a compliance conference. (The Office of Compliance, Inspections and Examinations - part of the SEC - oversees FINRA's oversight of MSRB rules, albeit infrequently)
  • Average markup for $10,000-$20,000 trade is 1.8 points
  • Average markup for $50,000-$100,000 trade is 0.9 points
  • Average markup for $250,000+ trades is on the order of 10-30 bps
  • Markdowns are smaller

The GAO report also mentions ongoing studies by the SEC and MSRB, also due this year and distinct from the regular statistical summaries from the MSRB (such as this one, for example).

Friday, June 15, 2012

On crossing at the Magenta Line versus paying the bid-offer

More bond market participants are considering guided crossing networks as an alternative to paying the bulge bracket bid-offer spread. The idea is very simple, at least in theory. An independent third party real-time price is computed in real-time, and anonymous parties can cross there. Of course they could also modify their desired price or allow it to float with the reference price (or reference spread). In 99.9% of cases this sort of crossing appears to be a no-brainer, since the typical bid offer in U.S. corporates is quite wide. It is visually obvious that buy side participants will save a lot - arguably on the order of several billions per year.

Because the development costs are very high, the only real-time reference price for U.S. corporates is currently provided by Benchmark Solutions. It is known as the Magenta Line. Crossing at the Magenta Line is a real, dare I say imminent possibility. For years the available reference prices have been inaccurate and produced no more than a few times a day. Needless to say new technology brings new commercial opportunities.  

This short note preempts a possible objection: the quality of the reference price in fast markets. Market participants can see the sort of display shown below and can easily make their own judgement, at least for liquid bonds (and they may exercise prudence by waiting for the market to settle down, of course). But what about illiquid bonds, that must be priced off "the" curve (or really two curves - credit and basis, plus idiosyncratic estimates)? Fortunately the Magenta Line is really a Magenta Curve, so the performance can be assessed by looking at relatively liquid bonds.

The other day we saw a big move in NAV that occurred in the space of twenty minutes, thus providing an opportunity to test whether the Magenta Line is suitable for crossing in fast markets. In fact the Magenta Line performs extremely well - a couple of years of research helps. Below you see the result of a recent code cut and the curve response. The even mildly technical reader will correctly infer that non-gaussian price dynamics are being modeled (together with numerous micro-structure devices that are less obvious in this particular example).

There are two things worthy of note here. First, this is actually an intermediate result in the sense that there is additional post-filtering applied after the fact - which in practice would improve the performance further (particularly in the period 3pm-4pm where I think there is a little room for improvement).

But at the risk of repetition, I am showing the response for one particular bond on the curve which is liquid. The point is that the entire term structure of credit and basis (as well as interest rates, of course) will also be moving. That how customers crossing at the Magenta Line on a far less liquid bond can still receive a fair price, no matter when they choose to trade.