Yesterday, we looked at Quote Stuffing and other HFT oddities. Today, I want to direct your attention to an indepth WSJ article about the flash crash:

“Philip Vasan, who heads the Credit Suisse prime-brokerage unit catering to hedge funds, began hearing from fund managers who were ratcheting back on trading because, they told him, stocks were behaving strangely. The funds were acting like “a dog that growls before an earthquake,” Mr. Vasan told several clients.”

When we moved to cash, it was because (in part) of how squirrely the markets were.  That is an apt description of strangely behaving markets: A dog that growls before it bites.

Back to the Journal:

“When the quake hit on the afternoon of May 6, the Dow Jones Industrial Average suffered its biggest, fastest decline ever, and hundreds of stocks momentarily lost nearly all their value. So many things went wrong, so quickly, that regulators haven’t yet pieced together precisely what happened.

A close examination of the market’s rapid-fire unraveling reveals some new details about what unfolded: Stock-price data from the New York Stock Exchange’s electronic-trading arm, Arca, were so slow that at least three other exchanges simply cut it off from trading. Pricing information became so erratic that at one point shares of Apple Inc. traded at nearly $100,000 apiece. And computer-driven trading models used by many big investors, apparently responding to the same market signals, rushed for the exits at the same time.

Three months later, many market veterans have arrived at a disquieting conclusion: A flash crash could happen again because today’s computer-driven stock market is much more fragile than many believed. Many investors, still gun-shy, have been pulling money out of stocks.”

It is your weekend home work assignment . . .

Legacy of the ‘Flash Crash’: Enduring Worries of Repeat
WSJ, AUGUST 6, 2010

Category: Trading

Please use the comments to demonstrate your own ignorance, unfamiliarity with empirical data and lack of respect for scientific knowledge. Be sure to create straw men and argue against things I have neither said nor implied. If you could repeat previously discredited memes or steer the conversation into irrelevant, off topic discussions, it would be appreciated. Lastly, kindly forgo all civility in your discourse . . . you are, after all, anonymous.

18 Responses to “Origins of the Flash Crash”

  1. NoKidding says:

    And still half cash as of last week.

    I was excited to see what effect the crappy (priced in?) employment report and big revision would have and got… nothing.

    1) Too many people know what the data will be before it becomes public. e.g. Goldman beating the concensus.

    2) Too much volume has become technical-only, with high volumes and low margins unplayable at retail. e.g flash crash.

    3) Too much market forecasting must be counterbalanced by government intervention forecasting. e.g. get out because of low PCE vs. get in because it means QE2 is closer.

    The advantage of being a minute player has always been that the battleships do not threaten or even acknowledge your presence, so you could live off the small gains they throw over the gunwales. Computers are destroying that advantage.

    The disadvantage has always been that the big boys hear the news first and can prepare accordingly. No changes there.

    Its driving me back to dividend yield as the only viable metric.

  2. wally says:

    I remember a competition – way back in the early days of PCs – to write a winning program for the Prisoner’s Dilemma game. There were some interesting articles (I think by Hofstader) about the competition.

    It would be very interesting to set up a similar competition for market trading programs. Maybe get somebody to pony up a cash prize. You could assign equal beginning portfolios to all entrants… but also inject some random-walk ‘personal investors’ into the action from time to time. The random walkers would show up to buy or sell for various personal reasons, either at market or with limits. Probably an hour or so would be sufficient time to run the show.

  3. Take a name says:

    Don’t need a weekend. What happened was an answer to a question that many had been asking for over a year. What will the HFT’s algos do if there is a sell off/crash? Will they continue to provide liquidity on the way down or will they unplug? We got our answer

  4. Petey Wheatstraw says:

    The NYSE: a finely crafted, high tech pachinko machine.

  5. rob says:

    That’s the drawback to using things that work faster than humans… unintended consequences with no way to monitor it real time. As evidence, it’s been HOW LONG and they still don’t know what happened. IMO the market is being priced by algorithms that work by pricing boundaries and technicals, and not fundamentals. Ironically the market is being purchased by the same algorithms using the same principles and algorithms. The only liquidity they are providing is to the same players, they are just scaring the s&^% out of the retail investors. So it just becomes a privileged rocking boat with cash sloshing from port to starboard, May 6 was when the gunwales touched the waterline. Please ignore the small iceberg ahead.

  6. NoKidding says:


    That competition was won by a group of entrants who gamed the system by encoding an identity recognition sequence in their first set of decisions. It allowed their programs to cooperate with each other, but attack outsiders.

  7. BR,

    w/this: “It is your weekend home work assignment . . .”

    “Only, if there’s a Test on Monday..”

    for the fingerlings that may be in these Waters, that type of thing may be a good way to incent them..

    they could, for instance, earn Credits toward a copy of “Bailout Nation” … in conjunction with their “Teachers”, receive ‘Extra Credit’…discounts on TBP T-Shirts, etc., ad infi..
    Have I mentioned that you need a 501(c)(3) ? (:
    but, past that, when you’re out hooking Osteichthyes, see if you can shoot (with your Nikon) some of these

  8. “Its driving me back to dividend yield as the only viable metric.”


    as used to be heard: “Dividend Checks don’t Lie.”

    one of the motivations for this Book Title..

  9. Joe Retail says:

    NiKidding: No kidding!

    Joe Retail is starting to seriously look at preferreds, on the assumption that he might have some chance of knowing what he’s buying.

  10. obsvr-1 says:

    Terminator 3 … Skynet in fact does not have a core but instead exists as software in cyberspace running on computers all over the world, making it effectively impossible to shut down. It begins a series of nuclear attacks on various cities, commencing Judgment Day.


    Mitigate the impact of the HFT machines with the circuit breakers (as implemented) – use the market trading systems to monitor and regulate the HFT systems, effectively implementing a negative feedback mechanism into the system. There will likely be modifications to adjust the circuit breaker algorithms over time as we continue to learn what the HFT responses to the the new game.

  11. Chief Tomahawk says:

    “A flash crash could happen again because today’s computer-driven stock market is much more fragile than many believed.”

    Seems to make the case for holding a call on the VIX at all times as a precaution for as long as mad scientist/alchemist Bernanke is at the controls of The Fed.

  12. jamie_2002 says:

    The computer simulation was an iterated prisoner’s dilemma– programs made moves over a number of turns (200) and the total payoff was determined.

    The winner (in two separate competitions!) was “tit-for-tat”. It had two rules:
    1) Turn one: cooperate
    2) All following turns: use the other’s move from the last turn

    Tit-for-tat is
    nice (it doesn’t try to sneak a defection)
    retaliatory (you defect, it defects)
    forgiving (you defect once, it defects once)

    It is able to minimize losses against the not-nice, and maximize gains with the nice.

    “Tit-for-tat” is only effective if there are repeated turns with an identified opponent. It doesn’t identify the opponent ( that is assumed by the structure of iterated PD). If you don’t know it is the same opponent, the game is a one-off prisoner’s dilemma, and tit-for-tat is not effective.

  13. tab says:

    I’m not much of a financial guy, more of a computer guy, but here is a scenario. If I wrote an AI program and let it learn while interact with the market through buying and selling, it could learn a range of seemingly bizarre behaviors. For example it could learn that the response time to selling small quantities at low prices is a good predictor of future prices. And what if there is another AI that has learned that small quantities of stock being sold low is a good predictor of a fall in price? What happens is a sort of feedback loop operating at extremely high speeds. Learning programs can only learn about situations they have been in before – they have no way to say “this is crazy” and stop themselves.

    There is also another possibility. If am using an AI, and a competitor is using another, then I might try to use an AI to disrupt the activity of the other – either intentionally or not.

  14. Gator81 says:

    A question for all here, and if it’s a dumb one, you tell me and I’ll pour myself a cup of BR’s special “STFU”.

    Is the following scenario possible? Plausible?

    A proprietary trading desk at a large firm (which is also a market-making brokerage with very deep pockets) uses a HFT system. They also use a projection model which statistically evaluates the ocean of price data available at such a firm, correlating this data with the seemingly limitless economic and financial data also available to them, streaming in real time. The projection model is thus able to say, for any stock in its database, based on some recent set of prices for the stock, what it “should” be worth right now. This projected value is then compared to the universe of bids and offers in the HFT system, where every bid over the projected price (plus a minimal margin) is sold, and every offer under the projected price (minus the minimal margin) is bought. Over and over again, in small fractions of a second.

    On a day when the volume is sufficiently low, with a virtually limitless supply of stock to sell and money to spend, could such as system not only make the market but also set the price?

    If they could do that with one stock, is there anything more than buying another 29 computers to keep them from doing that with the DJIA?

    As I type, the DJIA is at 10,540, MoL. It did a swan dive at the open (while the computers were tuning up the models?) and then settled at, you guessed it, 10540, where it has been for the last 3 ½ hours. Dave Fry (ETF Digest) calls that a “ramp and camp”, or perhaps in this case a “de-ramp and camp”.

  15. traderdc says:

    I found the following to be the best discussion about the Flash Crash.

  16. NoKidding says:

    Gator, yes but it takes really deep pockets.

    If the master-bot is holding the market up, then what is the exit strategy – can’t get out without crashing?
    If the master-bot is holding the market down, then what is the entry strategy- can’t get in without ramping?

    More likely the master-bot would want a controlled pseudo-random oscillation and alternately scalp 50% of all buyers and sellers a few cents per trade. Or scalp 100% of all buyers and sellers if it could both see orders in advance and frontrun them by a few trades.

    That strategy requires no entry/exit strategy unless volitility is high. Ideally prices would stay flat on average. Evidence of such a system would be price stability in the face of changing value. The all-knowing big pockets could be reduced to a large number of relatively dumb medium sized pockets.

  17. VennData says:

    As dumb as day traders who says “…a dog that growls before an earthquake…” is that dogs can not predict earthquakes…

    …anymore than a day trader can predict the stock market…

    …when they do, let’s hear it BEFORE “the crash,” OK?

  18. Bruman says:

    The flash crash reminded me of something my father said he had read in some science journal: that epileptic fits happen when the brain circuitry becomes synchronized (some corroboration here: ). Apparently the brain functions well when most neurons are doing their own thing and are not too synchronized with whatever else is going on – in market speak: when they are basically uncorrelated.

    It’s always made me worried about computers managing enormous portfolios. If they become synchronized, phase locked, or something like that, correlations go to one and everything goes haywire.

    I don’t have enough data to confirm that’s what happened in the Flash Crash, but I think it’s an enormous danger, and I think the suggestions that insist that orders be valid for a minimum of half a second or so are pretty sensible. Or a fee of a tenth of a cent per order, so that flooding becomes expensive.

    I would like to see more conclusive demonstration of the value-add of HFT (other than to their owners) before we make our financial system more brittle by handing it over to the computers.