Oops. Source: Bloomberg
Oops. Source: Bloomberg

One criticism of high-frequency trading that you hear a lot is that it destabilizes markets, because all the computers trading with each other don't have the common sense famously possessed by stock exchange floor traders, and so they sometimes do dumb stuff like sell stock for a penny or buy it for $99,999.99, and then later when you yell at them they just point to their algorithms and shrug.

A modest version of that happened in a bunch of stocks yesterday. AOL, for instance, pictured above, just trucked along, then WHOOM, then went right back to trucking along. MoneyBeat has some intraday charts -- of AOL, Marathon Petroleum, Lorillard, Canadian Natural Resources, Nabors Industries and, awkwardly Nasdaq OMX -- and is it weird that I want to get large prints and hang them on my wall? They're kind of beautiful, although I guess "my kindergartner could have drawn that" would be a valid criticism. They have a distinctly scribbled look.

What happened? Here is Bloomberg News on Caterpillar, which had a brief but exciting 1-percent-ish jump at 3:49 p.m.:

More than $11 million worth of shares changed hands at the elevated levels, according to data compiled by Bloomberg. The shares were moved by orders tied to the close of trading that were incorrectly entered.

All of the weird trades occurred at about the same time, just over 10 minutes before the close, and that explanation makes sense. There you are at 3:49 p.m., entering your market-on-close order to buy and sell a bunch of shares in the closing auction. You enter the number of shares you want, you go to the drop-down box on your computer system, you pick "Market on Close," your mouse slips, you actually click on "Market," and whoosh, off your order goes to be filled in the market.

The closing auction tends to have lots of people buying and selling, and procedures to make sure the price reflects supply and demand, so you can be reasonably confident of getting a reasonable price.1 As long as you click on the right thing! If you click the wrong thing, you get the price you can get in the open market at 3:49 p.m., which is far less forgiving.

But not entirely unforgiving. Actually, the New York Stock Exchange fixed this mistake for some trades, but not most of them. AOL, which started 3:49:41 p.m. yesterday trading at $36.85, dipped briefly as low as $32.77, and finished that second trading at $36.09, will break trades at or below $33.17. I count 276 trades for a total of 78,567 shares -- about $2.8 million worth -- at or below $33.17.2 The exchanges won't break any trades in the rest of the stocks. Why is that?

Well, there is an actual answer and it is fascinatingly kooky. NYSE has rules for breaking "clearly erroneous" trades, and those rules are standardized with the Securities and Exchange Commission and all the other exchanges, and they don't, like, strictly speaking, make sense, but they say some words and I guess the words mean "you can break trades that are down more than 10 percent from the real price." You gotta squint to see that, but I think they say that.3 If AOL's "real price" was $36.85 -- where it started its unpleasant second -- then trades at or below $33.17, 10 percent below $36.85, were breakable. And were broken.

The rest of the stocks moved less than 10 percent: Marathon and Lorillard were down more than 6 percent, Nasdaq OMX was down almost 5 percent, Canadian Natural Resources was up almost 4 percent, and Nabors was up a bit over 9 percent. So no broken trades.

Now it's perfectly reasonable to dislike the practice of breaking trades. Breaking erroneous trades rescues the dopes who made them, at the expense of people on the other side who didn't mess up, and who probably relied on their confirmed trades as they went off to do other trades. It's no fun to think that you bought at $33.10 and sold at $36 only to find out that you never bought at $33.10, because the guy who sold to you at $33.10 is a dope and wants his stock back. As pseudonymous financial blogger Kid Dynamite says:

Since The Flash Crash in May 2010, regulators have instituted circuit breakers in our markets to prevent the kind of price action that is often associated with “erroneous trades.” ... If you’re going to cancel trades, figure out ahead of time what constitutes a trade that can be canceled, and prevent the trade from happening in the first place.4 I kinda thought this was what the circuit breakers did.

And it is! NYSE, again in coordination with every other exchange, has "limit up/limit down" rules designed to prevent exactly these trades from happening. So the NYSE rules say that stocks can't trade more than 10 percent away from their reference price (the average price over the previous five minutes): Those trades just aren't allowed.5

Except! In the first 15 and last 25 minutes of trading, the limit up/limit down bands are doubled, to 20 percent, because those tend to be more volatile and higher-volume times and you want to restrict trading less. The "clearly erroneous trade" rules, on the other hand, don't adjust for the last 25 minutes. If someone had fat-fingered AOL at 3:30, trades below $33.17 would never have happened, so they'd never have to be broken. But they fat-fingered it at 3:49:41, so the stock could go down as far as $29.48 before the circuit breakers kicked in.6 But any trades below $33.17 could be broken after the fact.

That is just self-evidently dumb: If you're going to allow some trades and not others, and you have the technology to prevent the bad trades before they happen, and you use that technology to prevent some bad trades before they happen, it seems very silly to just let other bad trades happen, and then break them up after the fact. (Or, put another way: Trades not prevented by the technology are not actually bad, and should not be broken after the fact.)

One possible lesson here -- beyond "hey the exchanges should change the rules to fix this silly problem" -- is that they should change the rules carefully. The circuit-breaker rules instituted after the 2012 flash crash did a lot to reduce erroneous trading, but they were not particularly well coordinated with the erroneous-trade rules, leading to weird situations where NYSE allows trades to happen and then cancels them after the fact. As we currently live in a golden age of rushing to fix perceived market-structure problems, this might be worth keeping in mind.

Of course, one of the big perceived market-structure problems is precisely what happened here: When there's a big order, liquidity evaporates. AOL traded $103.8 million worth of stock yesterday; the fat-finger order that pushed its price down 11 percent in half a second seems to have been for well under 10 percent of that amount.7 I guess that sort of confirms some fears that liquidity can evaporate very fast in the face of a big order.

But on the other hand, it came back just as fast. And any moderately competent working of this order -- anything other than just dumping a giant market order all at once because you hit the wrong button -- would presumably have had much less market impact. So I'm not sure one should read these mini-flash-crashes as proof that high-frequency trading destabilizes markets, or that the liquidity that it offers is particularly fragile. It is, perhaps, proof that high-frequency trading is ruthlessly efficient at taking advantage of dumb trading. But the right response to that might be to cut back on dumb trading rather than on the ruthlessly efficient kind.

1 For context, on Monday, Caterpillar traded 146,195 shares at the closing trade, per Bloomberg AQR, or around $15 million worth (source: Bloomberg CAT <equity> AQR). AOL traded 41,063.

2 Source is Bloomberg page AOL <equity> QR from 15:49:40 to 15:49:42 on May 13, at prices from $0 to $33.17, dumped into Excel. I hope it's right! Also note that I count 275 of those 276 trades as occurring on NYSE, with one on EDGX. During that same period, EDGX, Nasdaq InterMarket and NYSE Arca were all printing trades at "normal" (above $36) prices, even as NYSE was printing in the $33s and $32s. (Bloomberg News reports, "about $10 million of shares changed hands" in the time period, but that includes the non-NYSE, non-messed-up trades.) This would seem to raise Regulation NMS questions that I am not smart enough to answer.

3 The trick is that this would seem to be a "Multi-Stock Event," since more than five (but fewer than 20) stocks are involved. Multi-stock events have a 10 percent threshold, as opposed to the lower 3-5 percent threshold for many single-stock errors. (The Nasdaq rules make it reasonably clear that this is because there's more need for finality when a bunch of stocks are messed up, so they make it harder to break trades.) The "Reference Price" is "generally" the consolidated last sale though I don't know how that works if there are hundreds of erroneous trades inside of a second; apparently they just used the price at the start of the second.

4 I have said something similar: "Maybe just don’t match orders that will be broken up as clearly erroneous?"

5 Basically if there's no bid within 10 percent below the reference price, there are no trades. The market gets 15 seconds for a bid to come in; if it doesn't, a five-minute pause is called, which concludes with an auction to establish the market price. (See page 2 of that rule overview.) The relevant bands are in the table on page 3 ("Phase II," for all national market system stocks): 10 percent for stocks with a price greater than $3, 20 percent for $0.75 to $3 prices, etc.

6 That number is fake/illustrative; I did not calculate the 5-minute average, just assumed it was around $36.85. Actually Bloomberg AOL <equity> AQR from 15:44 to 15:49 gets a volume-weighted average price of $36.8407, so that assumption is pretty reasonable.

7 It's hard to tell exactly, but see footnote 2: Some $10 million of stock traded during the glitch. Presumably a lot of that was not from the fat-finger order.

To contact the writer of this article: Matt Levine at mlevine51@bloomberg.net.

To contact the editor responsible for this article: Tobin Harshaw at tharshaw@bloomberg.net.