Yesterday I and others pointed out that, while UBS was not alone in getting screwed by Nasdaq failures on Facebook, it was alone in losing 10x as much as other, more competent market makers like Knight Capital, and ha ha ha. This apparently had a jinxing effect:
Knight Capital Group Inc., one of the largest trading firms, told brokerages to send their orders elsewhere and was probing a software problem, according to people involved in the matter. U.S. exchanges said they were examining potentially erroneous trading in more than 100 securities that saw big price swings or unusually high volume. Knight saw a fifth of its own market value wiped out. …
The system error and reports of irregular trading stoked suspicions that trades had been accidentally duplicated via computer algorithms, rather than the problem being contained to one server, as has happened in the past, traders said.
Knight is down ~21%, vs. ~4% yesterday for UBS and its costly Facebook fail, a useful reminder that focusing on perfecting your market-making business may make you less likely to fuck it up, but when you do fuck it up it goes far worse for you. That’s maybe some sort of a metaphor for high-frequency electronic market-making generally, which it will not surprise you to learn is coming in for some flak today.* Algorithmic high frequency trading makes it more likely that your small trade will be executed quickly and cheaply, but it also makes it more likely that larger orders will go horribly awry as prices move away from them.
Which is why this coincidence (?) pointed out by the Journal is kind of tantalizing:
The episode struck on the same day that NYSE Euronext introduced a new program designed to produce more competitive prices for retail investors. The so-called retail liquidity program enables market-makers to offer improvements on stock prices in fractions of a cent—a new function for stock exchanges. A spokesman for NYSE Euronext said Wednesday that the system was functioning properly.
More documentation on that here; basically it allows market-makers to submit pre-programmed undisplayed offers to improve price for retail liquidity takers in $0.001 increments. This would seem to improve pricing for retail investors – the name “Retail Price Improvement Orders” is a dead giveaway – but, trading costs being zero-sum, that’s gotta come from somewhere.** Traditional theory is that, the more that market makers can distinguish uninformed retail traders from big and/or informed institutions, the more they can price discriminate, offering well-priced liquidity to retail and jacking up costs for big/informed orders. The extra $0.001 per share that retail investors save comes out of an extra $0.05 (or whatever!) per share that mutual funds lose, with Knight and its friends pocketing the spare $0.049. Zero sum.
Obviously if a new system comes into effect and you are an algorithmic trading firm, you would want to build new algorithms to take advantage of it; those algorithms would presumably make you more aggressive in some places – to try to get that retail flow – and less aggressive in others – to back away from less desirable trades. So you might expect more aggressive orders in some circumstances and widening bid-asks in others. And, of course, when you implement new algorithms that haven’t been tested in the market before, sometimes they go wrong in embarrassing ways – so your more aggressive orders, and others’ backing away, might land you in a bunch of erroneous trades. Like Knight’s.
But that’s all baseless speculation. Here’s another theory for what happened (also here):
Traders said they believed a large order for a number of stocks was executed in five minutes rather than a longer period of up to five days. The far shorter transaction period pushed prices sharply higher or lower as the scale of the orders overwhelmed the natural liquidity of the market, which is limited for smaller capitalised stocks.
Which, if true, has nothing to do with any new systems and structures: it’s fat-finger-y rather than fat-algo-y. And it jives with the fact that the problems are only with 148 names rather than all of the NYSE. So I guess I’ll buy it. Computers are extremely efficient at propagating screw-ups widely and quickly, but they do their best work with human assistance.
Volatility Rocks U.S. Stock Market, Halts Some Trading [WSJ]
NYSE reviews trading after algo glitch [FT]
* Like: “‘Today is a day that proves that you just can’t rely on machines,’ said Joseph Saluzzi, co-head of equity trading at Themis Trading,” or “‘This is another example of a market structure that has failed. Another example how they have ruined the market structure to appease the machines,’ said Saluzzi, author of Broken Markets,” or the straightforward “Kill the quants before they kill us!”
** We talked about a not-unrelated proposal before, to improve the social functioning of HFT by reducing the minimum price increment. At the time I thought it sounded like a purely good thing, and all the knowledgeable commenters were like “you are an idiot,” so, duly noted. It is not a purely good thing.