This Is the End


Markets, Risk and Human Interaction

August 28, 2009

Not with a Bang but a Whimper – The Risk from High Frequency and Algorithmic Trading

Skynet begins to learn, at a geometric rate.
It becomes self-aware at 2:14 a.m. eastern time, August 29.
In a panic, they try to pull the plug. -- Terminator 2
There is a general view that one way or another the end result of all the high frequency and algorithmic trading will be a blowup. But I don’t think the risk is as big as many are making it out to be.
First, let me point out the difference between high frequency trading and algorithmic trading. Both execute using computers, and since computers work really fast, both can be accused of whatever sins are embodied in millisecond trading.
High frequency trading is a type of proprietary trading. The trader (or his computer) sees a profit opportunity and trades accordingly. This profit opportunity might occur because the high frequency trader observes signals in the way the market is trading that makes him think the price is moving up temporarily because someone needs to buy. He supplies the other side of that person’s demand, and once the demand is satiated the market price will most likely revert, and the high frequency trader will make a profit. And in doing so, he will be providing a service to the market – he will be a liquidity provider, and by getting into the market faster he will provide that liquidity for a lower price. Put another way, the better he is at his business, the less the price will be moved by the person who is requiring liquidity, and thus the lower transaction costs will be. Another way the high frequency trader will make money is by getting into the market before others do when there is information that is moving the price. Which explains the arms race in getting news feeds and executing based on the news a few milliseconds faster than others.
Algorithmic trading uses computer algorithms to facilitate trade execution. For example, some investor has decided to buy ten thousand shares of a particular stock. Once that decision is made, the question remains of how to execute the trade. One way to do it is to put a ten thousand lot buy order into the market. Another way is to have somebody sit on the phone and call the order in a hundred shares at a time in ten minute intervals until it is all done. Or, another way is to program a computer to do it. The computer can be programmed to do it any way a person can be told to do it. It can parcel the order out in fixed intervals, it can parcel more of it out during periods of high volume, it can throw orders out at random times and in random quantities. The point is that all this computer program is doing is facilitating a buy or sell order that has already been determined, and doing it based on a trading algorithm of the investor’s choosing. It is cheaper and more exacting than having someone do it on the phone, but really is not much different. If we are going to pose a horror story based on the huge volumes of computerized trading, we should not count the substantial portion of that volume that is occurring due to this algorithmic trading. Because one way or another, these trades are going to be done, and it is simply cheaper to have the computer do it.
So, having made the distinction, why am I not as worried as many others about a computer based cataclysm as a result of this sort of trading?
To answer this, let’s look at the example that often comes up as an object lesson for what can go wrong, the 1987 crash. On the face of it, the 1987 crash seems to be a reasonable historical case study, because the crash was, in a sense, the result of computers gone mad. It was the computers behind portfolio insurance that dictated more and more selling of futures to hedge equity portfolios as the market dropped, and that selling of futures in turn added to the drop. I gave a blow-by-blow description of this in Chapter 2 of my book – I had a front row seat because I was one of the people doing portfolio insurance, throwing my computers into the fray for such blue-chip clients as Ford, Chrysler and Gillette.
I wrote about portfolio insurance to illustrate one of the demons that we have created that contributes to the crisis-prone nature of our markets, namely, tight coupling. Tight coupling is a term I borrowed from engineering. A tightly coupled process is one which, once it gets going, cannot be easily stopped. If there is a problem – if the train starts going off the tracks, so to speak – no one can pull an emergency brake to pause the process until a committee convenes to figure out what to do. A computerized system like portfolio insurance is tightly coupled. The liquidity crisis cycle that comes about from the forced liquidation of highly leverage investors is a tightly coupled process. So are processes in other areas, like a space shuttle launch, a nuclear power plant moving toward criticality, or even something as prosaic as the process of baking bread.
But just because something happens in milliseconds doesn’t mean it is tightly coupled. And, for that matter, just because something is tightly coupled doesn’t mean it is prone for disaster. You need something more to cause such a cataclysm than computers trading quickly. You need a lot of them all doing the same sort of thing, and doing it in a way that feeds back on itself. And doing it in a way that is not brought to a stop when it seems to be going awry. That is why portfolio insurance was a problem. While computers were involved, which of course makes for a better story, it could have happened even if people had been making the calculations on an abacus and phoning the orders in. Because the problem was that many people were all pursuing the same strategy, and that strategy was one that reinforced a drop in the market – that is, it was a strategy that sold into a market drop, leading to a further market drop. I don’t see this essential element in either high frequency trading or algorithmic trading.
For algorithmic trading, the issue is simple. Some investors have determined to buy and others have determined to sell. They have reached these determinations however they have done it in the past; it doesn’t have anything to do with the advent of millisecond trading. Now they happen to decide to execute these trades over time as a function of the bid-offer spread, the volume of trading, the level of prices – that is, based on the same sorts of things that they would have without computers. And they can monitor what is going on with their trades over the course of the day. So this is not a tightly coupled process. A call to their broker, and the trading stops. Just like if the broker had someone doing it on the phone.
For high frequency trading, the issues are not as simple. It is possible to construct a scenario for high frequency trading where a strategy is widely shared which has a reinforcing feedback and which pushes forward without anyone intervening. But I can construct such scenarios even without the need for millisecond trading. Not just construct such scenarios – I have seen them, and so have you, in any number of bubbles and crashes.
But speaking specifically about the risk due to high frequency trading, the risk of a cataclysm is constrained by the lack of feedback and lack of tight coupling. High frequency trading is a matter of individual trades – or possibly baskets of trades. It is not a process, much less a tightly coupled process. And while no doubt similar strategies are employed by many of the traders, if they are liquidity or information oriented, they are not going to be subject to reinforcing feedback. The sort of trading strategy that could be a problem is a trend following strategy with participants piling on in ever-increasing size. And, again, you don’t need a computer to have that occur.
I think the popularity of high frequency trading will end not with a bang but a whimper. As the field gets increasingly crowded, market impact will rise and opportunities will diminish. Day by day more and more of the high frequency traders will see small negative numbers rather than small positive numbers in their P&L columns. The nice thing about high frequency trading is that it doesn’t take long to know if a strategy is working or not. The trader gets many draws of his strategy every day. And there is no cost to closing a strategy down – often every position is set to zero at the end of the day.


  1. In my business, we turn off our high frequency trading programs for all sorts of trivial reasons, say because our database didn't process a corporate action on a single stock correctly. Or because we are taking a week or a month's vacation. Certainly if a strategy stopped behaving as modeled and expected, we would turn it off. HFT is more like a bunch of electric razors looking for stubble than Skynet.

  2. Glad to see you shedding some light on this topic.

    There are other factors adding stability to the system:
    Most HFT is market neutral so it isn't pushing the market one way or another.
    Most HFT doesn't use (or have) tons of capital or leverage. So even if it accidentally went directional to the extent of its capital limits it wouldn't have a systemic effect.
    Most HFT is based on mean reversion which counteracts volatility
    HFT adds liquidity which lowers volatility
    HFT algorithm writers are have incentive to be very careful because if their algorithms mess up they get taken to the cleaners by other traders (including HFT).

  3. Rick,

    A few thoughts, no particular order:

    Given that you're part of what I'd refer to the "quant elite" would you disagree that much of your cohort has similar education, similar training, and similar outlook/approach? Can you honestly say that quantitative trading approaches across the big boy funds are really much different? Are D.E. Shaw's factor models much different than Renaissance's? Are the various tweaks to GARCH (or whatever) model really that different fundamentally, or do they produce fundamentally different results?

    From my vantage point several (dozen) IQ points below you, it seems like you discount the herd mentality which I think has become too prevalent across the quantosphere over the past say 5-10 years. Remember last year, say, November, when so many funds with quant businesses all had the same/similar books, same/similar leverage, same/similar buy/sell signals?

    Just because the math is harder, the education more prestigious, and the frequency higher does not mean that quant and HFT are infallable; quite the contrary, I'd argue that some of the "smartest guys in the room" have, and will continue to fall victim to the same semi-willful ignorance/arrogance/hope that any joe schmo retail investor inevitably succumbs to.

    It was a please to see/meet you @ the FT Alphaville event many months back, a pleasure to get your take on things, and hope to see/hear from you soon.



  4. I agree with the point of possible group-think brought up by Anal_yst, I mention this in my second post on the quant blow-up in August, 2007. See where I say there is group-think, and ascribe it to the use of the scientific method and similar training:

    "The reason I think this is for the same reason the quant funds ended up with very similar strategies in the first place: they use the scientific method.

    The key to the success of the scientific method is the reproducibility of results. If you do the same experiment I do, you will reach the same conclusion. More than that, if you and I start with the same hypothesis and apply the same data, you will discover what I discover. That is why we see so many races for the prize in the pure sciences. Even with the spectacular discovery by Watson and Crick, there were others nipping at their heels.

    With the quant funds, we have well-trained professionals applying the scientific method to capture anomalies in the market. Most are trained at the same handful of institutions, they have read the same academic literature, and they are applying the same statistical tests, using the same analytical tools, to the same sets of data. So it is no surprise they come up with similar models."

  5. Thanks for the reply, Rick!

    I'm curious though since I don't seea/hear it discussed too often (admittedly I may not be looking in the right places) but I think the "group-think" issue is one that many of the critics of "hft" latch-onto.
    The type of questions/concerns I see aren't implicitely invalid prima facie; e.g. Big banks (regulators, etc)relied on VaR which we know isimperfect and furthmore mis-used/abused; what's to say tens of billions of $'s of AUM aren't similarly relying upon imperfect "risk management" approaches, or predicated upon let's call them "conditional" assumptions?
    Personally I'm not too concerned with hft, but I think many critics doubt the ability of the stewards to unplug the machine, so to speak, when everyone's models/algos/etc all start doing the same, self-defeating thing.

    Btw I checked some of your older posts but apparently I didn't go back far enough, whoops?!




Note: Only a member of this blog may post a comment.