Goldman Sachs Got Lost in Its Own Dark Pool
There comes a time in every financial blogger's life when he feels compelled to write a post about dark pools that uses the phrase "latency arbitrage." So, sorry, but this is that post. The thing is, some of the worrying about high-frequency trading and dark pools is just fear of excessive efficiency, and some of it is just paranoid fantasy. But latency arbitrage is a straightforwardly bad thing, and it is real, and now we need to talk about it.
This is what it is. The market quote for a stock is $10.00 bid / $10.05 offered at time T. Then someone comes in with a bid at $10.02 at time T+1. Then at time T+2, you come in with a market order -- "sell stock at whatever the market price is." But your trading venue -- your exchange or dark pool or internalizer -- has a slow computer, so at time T+3, you sell for $10.00, the old (wrong) bid. And at time T+4, the guy who bought from you at $10.00 sells for $10.02, the bid that you really should have gotten if the computer wasn't slow. He makes an instant 2-cent profit, which he's effectively taken from you.
"Front-running," you might shout, and while it isn't front-running in the legal sense it isn't great either. And it clearly happens; the debate is over how to prevent it, and how economically material it is -- how often it happens.
Here's a data point: at least 395,000 times in eight days at one dark pool.
FINRA found that from July 29, 2011, through August 9, 2011, there were more than 395,000 transactions executed in SIGMA-X where the execution traded through a protected quotation at a price inferior to the National Best Bid and Offer (NBBO).
During the eight-day trading period, Goldman Sachs was unaware that it was trading through a protected quotation in these instances. The trade-throughs were caused by market data latencies at SIGMA-X and were undetected in a timely manner.
Oops! That's from this Finra press release announcing an $800,000 fine against Goldman for "failing to have reasonably designed written policies and procedures in place to prevent trade-throughs of protected quotations" in Sigma X, its dark pool. Finra's "letter of acceptance, waiver and consent" is short and sweet, and, while I have never run a dark pool, Goldman's procedures certainly don't seem reasonable to me:
First, the firm used an exception report, the GSEC REDI Market Data Accuracy Check Report (the "Market Data Accuracy Report"), to compare a sampling of SIGMA-X's market data with an independent data source at the time of an order event. The Market Data Accuracy Report, however, from the date of its inception, November 3, 2008, through August 31, 2011, only reviewed 20 orders per week. SIGMA-X, however, executed millions of transactions weekly. Because of the limited scope and sampling of the Market Data Accuracy Report, it was inadequate for its intended purpose.
If last week's lawsuit against Barclays freaked you out about dark pools, this Finra order ... really shouldn't freak you out too much less? It's smaller and quieter and less full of angry rhetoric about "predators," but it identifies an actual harm to customers. A small one, but still: For eight days in the summer of 2011, customers came to Goldman expecting to trade at the market price, and 395,000 times, they lost out, because Goldman was unable to keep track of its dark pool. The Barclays case is about false advertising of the dark pool's supervision. The Sigma X case is about a failure of that supervision.
The other thing to freak out about, of course, is: How much of this is there? Goldman has updated its systems and presumably no longer has this problem. Perhaps it was a one-week fluke.
On the other hand, it does point to the rickety structure for monitoring dark pools. Dark pool operators are responsible for not hosing their customers, and for having "reasonably designed written policies" to check that they're not doing that. Finra is responsible for checking up on those policies. The Securities and Exchange Commission is responsible for checking up on Finra. Some number of balls were dropped.
One idea would be for the SEC, which has computers and stuff, to actually check in on exchanges and dark pools periodically and see if they're really giving their customers the best market price. Instead of relying on the dark pools to do that. Weirdly, that doesn't seem to be in the cards.
Instead, the problem was caught by Goldman's customers: Customers were reviewing their own execution reports, and "several" of them contacted Goldman on Aug. 8 to complain. Good for them! The current status of the equity markets may not be "rigged," but it's at least "be careful out there." Goldman's customers were. And this case is a good reminder that everyone should be.
I mean. This is a simplified version of one thing that it is.
I mean, it's not the end of the world, either. If you'd just pressed the button on the trade one second later, you'd trade at like time T+1000, and this whole mini-drama of the moving quote wouldn't matter. In fact you'd sort of expect that, if this problem existed, in the long run it would cancel out: Half the time you'd lose a bit trading on a stale quote; the other half you'd win a bit. It only matters if some fast traders can take advantage of it, by only trading on stale quotes when they know that they'll profit. The "markets are rigged" claim is that this is what happens, and that it happens a lot.
Incidentally, I do not entirely understand the numbers here. You can look at ATS data on the Finra website, which shows Sigma X doing 858,724 trades last week. I don't have 2011 data, and it's likely that Sigma X was busier then than now, but "millions of transactions" seems high. And of course, even if you're doing, say, 2 million trades in a week, then:
- 395,119 is a lot of them to mess up!, and
- Even if you're only sampling 20 trades a week, you should catch some errors, no? That's like a 20 percent error rate.
I don't know what to tell you.
- 395,119 is a lot of them to mess up!, and
I periodically mention some biases, so let's do that again. I work for Bloomberg LP, which is in all sorts of electronic trading businesses, including running an electronic trading platform. I used to work at Goldman Sachs, and still own some restricted stock. I left, as it happens, about a month before this problem with Sigma X, and the place immediately fell apart without me.
I'm trying not to talk about the SIP and the direct feeds, but I guess we should. One of the high-frequency trading controversies is the timing difference between the SIP -- the "Securities Information Processor," sort of the official consolidated feed of prices -- and the direct feeds of prices from exchanges. The SIP is slower than the direct feeds, and one of the leading HFT conspiracy theories is that dark pools set their prices off the SIP but let HFTs using the direct feed collocate next to the dark pool. So the HFTs have the faster price -- they know the price is $10.02 when the dark pool thinks its $10.00 -- and can make risk-free money.
HFT defenders say that most dark pools now use the direct feeds, not the SIP, so this is not a real issue -- and that, even on venues that use the SIP, the economic value of the latency arbitrage is tiny to nonexistent.
Wouldn't it be nice to know? The SEC could find out: Just demand trade and quote data from a bunch of dark pools and exchanges for, like, a week. See how many trades happened that were (1) legal (not trading through a protected NBBO quote according to the SIP at the time of the trade) but (2) latency-arbed (trading through what would be a protected NBBO quote according to the direct feeds at the time of the trade). If it's, like, 10 trades, then we all agree to stop worrying about this. If it's a million trades with lots of economic losses for the customers, then maybe we fix the rules.
This footnote hereby urges the SEC to do that experiment.
The New York Stock Exchange complained on the Aug. 9. Finra complained ... I don't know, now?
This column does not necessarily reflect the opinion of Bloomberg View's editorial board or Bloomberg LP, its owners and investors.
To contact the author on this story:
Matthew S Levine at firstname.lastname@example.org
To contact the editor on this story:
Toby Harshaw at email@example.com