How The Pursuit Of "Light Speed" Broke Equity Markets, Or Why The NBBO Is No Longer Relevant
After clearly demonstrating that last year's flash crash was essentially a byproduct of massive quote stuffing-induced churn surge, which cascaded into a full blown liquidity collapse, coupled with the bulk of HFT "liquidity" providers simply turning their machines off and subsequently reverberating by ETF "amplifiers" to the nth degree, last week Nanex released what is probably one of the most critical white papers on why in its pursuit of ever higher speed (or, at least speed that is physically capped by the theory of relativity at 300,000km/sec) to gain a frontrunning advantage over everyone else, courtesy of a massively two tiered market in which there is the collocated Ph.D. braintrust focused on millisecond trading (day trading is so 20th century), and then everyone else, the core premise of the "fair and efficient market" - the National Best Bid Or Offer, which is at the heart of Regulation NMS which in turn sets forth the guiding principles behind modern "capital markets" is now an anachronism and is being overrun on a daily basis as the fastest to the market gets to set just what their own NBBO is, in the process literally destroying the premise of market fairness and efficiency, as those who have the newest and shiniest toys are guaranteed to win, while everyone else fights for a fraction of the scraps. Said otherwise, steroids are forbidden in sports but when it comes to capital markets, they are very much accepted.
The reason for this bold claim is that the Security Information Processor, the aggregator of NBBO data from the Consolidated Quote System is no longer used and instead exchanges, of which there are hundreds courtesy of fragmentation, make up rules literally on the fly. The implications of this are dramatic and imply that in order to restore market efficiency as envisioned by Reg NMS, the way modern market topology is evaluated will need to be gutted and overhauled completely, in the process eliminating the bulk of parasitic HFT, which in turn will reduce market volume to an even worse trickle, leading to more Wall Street layoffs, higher equity market volatility, a greater facility for market manipulation by the FRBNY, and an increased exodus by daytraders to other exchange and OTC-based markets such as commodities, rates and FX. In other words, the daily flash crashes we have so grown to love in stocks are coming to a 10 Year near you.
Full text from Nanex:
The NBBO lies at the heart of Reg NMS and is the key concept that assures investors get the best price when buying or selling stocks. However, due to the industry trend that emphasizes speed at all costs, the NBBO, in practical terms, no longer exists. There is no audit trail that can show definitively whether an investor received the best price on their trade. The regulators do not appear to understand the root of the problem because they continue to promote new regulation, when a simple and effective solution exists: enforce Reg NMS. Likewise, getting rid of quotes that only serve to manipulate others is as easy as enforcing Section 9 of the Securities Exchange Act of 1934.
If any new regulations are needed after enforcing existing ones, we think a minimum quote life of 50 ms makes the most sense. Just the discussion of implementing such a rule would expose how deeply flawed the system is today and would be sure to raise a lot of eyebrows. But to be clear, we do not advocate new regulation.
The Speed Of Light
As the lifetime of a quote approaches zero, the arguments for and against a minimum-quote-life rule become more interesting. Let’s suppose, for example, that the day has arrived where a few of the top HFT systems are able to send and cancel quotes in 1 nanosecond (ns). For reference, light travels 30 cm (1 foot) in 1 ns. At this rate, we could see 1 billion quotes per second per stock (qpss). A thousand active stocks trading at this rate would generate 1 trillion quotes per second (qps).
OK, that was an extreme example, but with all the hyperbole from some HFT marketing groups, we couldn't resist. So let’s slow things down by a factor of 1,000 and imagine a world where HFT systems can send and cancel a quote in 1 microsecond (us). At this speed we could expect 1 million qpss and a thousand stocks would generate 1 billion qps.
Still extreme? Let's make things 1,000 times slower again and imagine a world where HFT systems can send and cancel a quote in 1 millisecond (ms). At this speed, we could expect 1,000 qpss and a thousand stocks would generate 1 million qps. At this rate, many quotes will expire before leaving exchange networks. Today qpss rates of 2,000 occur frequently, with peaks in the 5,000 - 30,000 range. These numbers are growing rapidly.
"No one uses the SIP for the NBBO anymore"
Perhaps the definition of a quote is what needs to be modernized, or maybe we need a new name for what used to be called a quote. Not long ago, when a trader (or auto-trading software) received a quote marked auto-execute, there was a reasonable expectation of hitting that quote — the only real exception being that another trader might beat them to it. Today, under the same circumstances, the same auto-execute quote would almost certainly have expired during the same time period.
The change in semantics becomes most important when we look at the definition of the NBBO. Per Reg NMS, the NBBO for a stock is defined to mean the best bid/offer sent by a market center to the Security Information Processor known as the SIP (CQS, UQDF). But no one uses the SIP for that anymore, we are often told, which would seem to violate the letter and spirit of Reg NMS.
So what do they use?
Each exchange computes the NBBO internally from their direct connections to other exchanges. As the speed of trading increases, the likelihood of two exchanges having the same NBBO decreases. Most of this is because of the pesky speed-of-light limitation.
So how does a trader know whether a trade was routed properly to the exchange with the best price?
He doesn't. It is impossible.
You see, each exchange’s view of the other exchange prices only exists in memory on that exchange's machines. It is not recorded. There is no audit trail. Sure, each exchange provides book-level data, but that only includes prices for that one exchange — not the prices that existed on the other exchanges at the time of each order.
"It is impossible to verify that a trade received the best price"
If we are going to allow machines to trade faster and faster, then at the very minimum, there must be audit trail data that includes each exchange’s view of top-of-book quotes for every linked exchange trading that stock. In other words, what now only exists in an exchange routing computer’s RAM, needs to be captured and made available. The exchange routing simulation video below may help you visualize this. In the video, each box represents one exchange and the price information it has from the other exchanges -- that is the information which needs to be recorded to assure trade through price protection. Essentially, each interlinked exchange would end up recording its view of every exchange's top-of-the-book prices -- exactly what the SIP (the box at the bottom) does now.
So why not use just use the SIP? Auditing, maintaining, and improving one system is much better than 14 (one for each exchange). Well, because, as some more informed people (including the SEC), have told us: The SIP is not as fast because it has to aggregate information from the other exchanges.
Wait a minute.
How can an exchange insure every order receives the best prices, if it doesn't aggregate information from all the other exchanges? Take a close look at the simulation video and note that every exchange must essentially replicate the SIP's aggregation function. There is no way around this. Which means the claim that aggregation causes the SIP to run slower is absurd. Quote rate overload is the primary cause of latency in the SIP and direct feeds.
"The claim that aggregation causes the SIP to run slower is absurd"
There is more.
A computer that aggregates information from other computers uses an aggregation policy that details how it selects between messages coming in at the same time from multiple computers (things like time-slice period, scheduling quantum, size of input queues, overflow behavior, etc.). Let's just say it's very complicated and has many dependencies. The aggregation policy is important because it affects the 3rd criterion in NBBO selection: price, size, time. The SIP's aggregation policy is something that is fairly easy to verify. How easy is it to obtain and verify the 14 aggregation policies used by the 14 other exchanges?
It gets worse.
We have discovered a few anomalies when analyzing aggregation characteristics of CQS, which is the SIP for stocks listed on NYSE, AMEX, and NYSE Arca (home of many ETFs). One disturbing anomaly is that the SIP appears to be ignoring the timestamp in the quotes it receives from an exchange, using instead the actual time the SIP receives the quote, even if the timestamp in the quote is over 5 minutes late. That is, we found an example of the SIP treating 5 minute old data as real-time, affecting the NBBO in thousands of stocks. We will be publishing the results of our analysis shortly. The main point, is that this type of detection is impossible to carry out for individual exchange aggregation behavior; and if it was, it would take 14 times more work.
"We found an example of the SIP treating 5 minute old data as real-time, affecting the NBBO in thousands of stocks"
We understand how difficult it can be to grasp these problems. Understanding complex networked systems where the speed-of-light is the dominant source of latency is hard. When the fate of your nation's financial infrastructure is at stake, we think regulators should have a solid understanding of these issues.
Finally, we'd just like to point out to those who say no one uses CQS, that last year, 2.5 million subscribers spent over $450 million in exchange subscription fees to receive and process CQS. We think they deserve better.
The video shows how orders flow between linked exchanges trading one stock. Orders are represented by triangles, and begin from the edge of the display and feed into one exchange. That exchange then updates and checks its internal table to determine whether to route the order to another exchange with a better price, execute it locally, or simply update the top of the book for that exchange. Only orders that affect the top of the exchange book are included to reduce clutter. The order is then transmitted to every other exchange making a market in the stock (via premium direct exchange feeds) as well as CQS which is shown at the bottom. When other exchanges receive the order, they update their internal tables in order to keep track of the other exchanges' top-of-the-book quotes.
The price data within each box therefore represents how that exchange views the prices on other exchanges trading that stock. Unlike this simulation which assumes a perfect world: due to the variations in distance and connection quality, system load, and other real-world imperfections, exchanges won't always have identical NBBO information. As update rates increase, the percentage of time that all exchanges have identical NBBO information will rapidly drop to zero.
Download the exchange routing simulation application. Simply unzip to an empty directory, and drag-and-drop one of the text files over the executable.