Interview with Dave Cliff

Dave Cliff of the University of Bristol is someone whose work I've been meaning to look at much more closely for a long time. Essentially he's an artificial intelligence expert, but has has devoted some of his work to developing trading algorithms. He suggests that many of these algorithms, even one working on extremely simple rules, consistently outperform human beings, which rather undermines the common economic view that people are highly sophisticated rational agents.

I just noticed tht Moneyscience is beginning a several part interview with Cliff, the first part having just appeared. I'm looking forward to the rest. Some highlights from Part I, beginning with Cliff's early work, mid 1990s, on writing algorithms for trading:

I wrote this piece of software called ZIP, Zero Intelligence Plus. The intention was for it to be as minimal as possible, so it is a ridiculously simple algorithm, almost embarrassingly so. It’s essentially some nested if-then rules, the kind of thing that you might type into an Excel spreadsheet macro. And this set of decisions determines whether the trader should increase or decrease a margin. For each unit it trades, has some notion of the price below which it shouldn’t sell or above which it shouldn’t buy and that is its limit price. However, the price that it actually quotes into the market as a bid or an offer is different from the limit price because obviously, if you’ve been told you can buy something and spend no more than ten quid, you want to start low and you might be bidding just one or two pounds. Then gradually, you’ll approach towards the ten quid point in order to get the deal, so with each quote you’re reducing the margin on the trade.  The key innovation I introduced in my ZIP algorithm was that it learned from its experience. So if it made a mistake, it would recognize that mistake and be better the next time it was in the same situation.

HFTR: When was this exactly?

DC: I did the research in 1996 and HP published the results, and the ZIP program code, in 1997. I then went on to do some other things, like DJ-ing and producing algorithmic dance music (but that’s another story!)

Fast-forward to 2001, when I started to get a bunch of calls because a team at IBM’s Research Labs in the US had just completed the first ever systematic experimental tests of human traders competing against automated, adaptive trading systems. Although IBM had developed their own algorithm called MGD, (Modified Gjerstad Dickhaut), it did the same kind of thing as my ZIP algorithm, using different methods. They had tested out both their MGD and my ZIP against human traders under rigorous experimental conditions and found that both algorithms consistently beat humans, regardless of whether the humans or robots were buyers or sellers. The robots always out-performed the humans.

IBM published their findings at the 2001 IJCAI conference (the International Joint Conference on AI) and although IBM are a pretty conservative company, in the opening paragraphs of this paper they said that this was a result that could have financial implications measured in billions of dollars. I think that implicitly what they were saying was there will always be financial markets and there will always be the institutions (i.e. hedge funds, pension management funds, banks, etc). But the traders that do the business on behalf of those institutions would cease to be human at some point in the future and start to be machines. 

Personally, I think there are two important things here. One is that, yes, trading will probably soon become almost all algorithmic. This may tend to make you think the markets will become more mechanical, their collective behaviour emerging out of the very simple actions of so many crude programs.

But the second thing is what this tells us about people -- that traders and investors and people in general aren't so clever or rational, and most of them have probably been following fairly simple rules all along, rules that machines can easily beat. So there's really no reason to think the markets should become more mechanical as they become more algorithmic. They've probably been quite mechanical all along, and algorithmic too -- it's just that non-rational zero intelligence automatons running the algorithms were called people. 

Algorithms are smarter than people

On the topic of algorithmic trading, I recently posted on some evidence documenting the benefits it brings to markets -- more liquidity, lower spreads and trading costs, etc. On a related topic, Ole Roleberg at Freakynomics has a nice post reviewing some of the evidence that automated decision tools actually make better decisions that real people when confronting many different kinds of problems. As he notes,

There’s a host of studies showing that human judgment is poor at synthesizing and weighting a large number of different types of evidence, and that simple, statistical models can outperform humans on tasks such as predicting recidivism, making clinical judgments (psychiatry and medicine), predicting divorce, predicting future academic success, etc. (for an entrypoint to this literature, see here for a blogpost I found that has some good quotes from J.D. Trout and Michael Bishop).

I guess the point is that algorithmic trading can be good or bad depending on the algorithm – and that the danger it brings is more if the ecology of trading algorithms active in a market is of a kind that could create cascading ripples destabilizing the market: One set of algorithms lowering the price of a set of stocks, triggering another set of algorithms to sell these stocks to avoid loss, triggering another set of… and so on.

This is precisely the point I've made before about the dangers of algorithms -- it's not one algorithm that might blow things up, but potentially explosive webs of feedback running between many.

But I think the superior performance of algorithms at making decisions is itself quite striking and not generally recognized. The article to which Rogeberg links makes the following all-too-plausible remark:

Training of large numbers of experts by universities has probably had the perverse effect of increasing the number of people running around making highly confident but wrong judgements. But the tendency to not notice our errors and to place excessive confidence in our subjective judgements is something that all humans suffer from to varying degrees.

One final interesting read -- again thanks to Rogeberger for pointing this out -- is a profile in The Atlantic of Cliff Asness of the quant hedge fund Applied Quantitative Research. AQR was one of the hedge funds involved in the infamous "quant meltdown" of August 2007 which was driven precisely by a positive feedback loop, in the case one which caused a violent de-leveraging among a number of hedge funds using similar strategies and invested in similar assets. This is one of the few cases in which we have a pretty good quantitative model explaining how these kinds of feedback loops emerge essentially in the same way violent storms (or hurricanes) do in the atmosphere -- through ordinary processes which create the conditions in which explosive events become virtually certain. In the profile, Asness describes the dynamics behind the quant meltdown, which weren't as complex, mysterious or irrational as many people seem to think:

He told the New York Post that he blamed the sudden losses not on AQR's computer models but on "a strategy getting too crowded ... and then suffering when too many try to get out the same door" at the same time.

Algorithmic trading -- the positive side

In researching a forthcoming article, I happened upon this recent empirical study in the Journal of Finance looking at some of the benefits of algorithmic trading. I've written before about natural instabilities inherent to high-frequency trading, and I think we still know very little about the hazards presented by dynamical time-bombs linked to positive feed backs in the ecology of algorithmic traders. Still, it's important not to neglect some of the benefits algorithms and computer trading do bring; this study highlights them quite well.

This paper asks the question: "Overall, does AT (algorithmic trading) have salutary effects on market quality, and should it be encouraged?" The authors claim to give "the first empirical analysis of this question." The ultimate message coming out is that "algorithmic trading improves liquidity and enhances the informativeness of quotes." In what follows I've given a few highlights -- some points being obvious, others less obvious:

From a starting point near zero in the mid-1990’s, AT (algorithmic trading) is thought to be responsible for as much as 73% of trading volume in the U.S in 2009.

That's no longer news, of course. By now, mid-2011, I expect that percentage has risen to closer to 80%.

Generally, when I think of automated trading, I think of two activities: market makers (such as GETCO) and statistical arbitrage high-frequency traders, of which there are many (several hundred) firms. But this article rightly emphasizes that automated trading now runs through the markets at every level:

There are many different algorithms, used by many different types of market participants. Some hedge funds and broker-dealers supply liquidity using algorithms, competing with designated market-makers and other liquidity suppliers. For assets that trade on multiple venues, liquidity demanders often use smart order routers to determine where to send an order (e.g., Foucault and Menkveld (2008)). Statistical arbitrage funds use computers to quickly process large amounts of information contained in the order flow and price moves in various securities, trading at high frequency based on patterns in the data. Last but not least, algorithms are used by institutional investors to trade large quantities of stock gradually over time.

One very important point the authors make is that it is not at all obvious that algorithmic trading should improve market liquidity. Many people seem to think this is obvious, but there are many routes by which algorithms can influence market behaviour, and they work in different directions:

... it is not at all obvious a priori that AT and liquidity should be positively related. If algorithms are cheaper and/or better at supplying liquidity, then AT may result in more competition in liquidity provision, thereby lowering the cost of immediacy. However, the effects could go the other way if algorithms are used mainly to demand liquidity. Limit order submitters grant a trading option to others, and if algorithms make liquidity demanders better able to identify and pick off an in-the-money trading option, then the cost of providing the trading option increases, and spreads must widen to compensate. In fact, AT could actually lead to an unproductive arms race, where liquidity suppliers and liquidity demanders both invest in better algorithms to try to take advantage of the other side, with measured liquidity the unintended victim.

This is the kind of thing most participants in algorithmic trading do not emphasize when raving about the obvious benefits it brings to markets.

However, the most important part of the paper comes in an effort to track the rise of algorithmic trading (over roughly a five year period, 2001-2006) and to compare this to changes in liquidity. This isn't quite as easy as it might seem because algorithmic trading is just trading and not obviously distinct in market records from other trading:

We cannot directly observe whether a particular order is generated by a computer algorithm. For cost and speed reasons, most algorithms do not rely on human intermediaries but instead generate orders that are sent electronically to a trading venue. Thus, we use the rate of electronic message traffic as a proxy for the amount of algorithmic trading taking place.

 The figure below shows this data, recorded for stocks with differing market capitalization (sorted into quintiles, Q1 being the largest fifth). Clearly, the amount of electronic traffic in the trading system has increased by a factor of at least five over a period of five years:

The paper then compares this to data on the effective bid-ask spread for this same set of stocks, again organized by quintile, over the same period. The resulting figure indeed shows a more or less steady decrease in the spread, a measure of improving liquidity:

So, there is a clear correlation. The next question, of course, is whether this correlation reflects a causal process or not. I won't get into details but what perhaps sets this study apart from others (see, for example, any number of reports by the Tabb Group, which monitors high-frequency markets) is an effort to get at this causal link. The authors do this by studying a particular historical event that increased the amount of algorithmic trading in some stocks but not others.The results suggest that there is a causal link.

The conclusion, then, is that algorithmic trading (at least in the time period studied, in which stocks were generally rising) does improve market efficiency in the sense of higher liquidity and better price discovery. But the paper also rightly ends with a further caveat:

While we do control for share price levels and volatility in our empirical work, it remains an open question whether algorithmic trading and algorithmic liquidity supply are equally beneficial in more turbulent or declining markets. Like Nasdaq market makers refusing to answer their phones during the 1987 stock market crash, algorithmic liquidity suppliers may simply turn off their machines when markets spike downward.

This resonates with a general theme across all finance and economics. When markets are behaving "normally", they seem to be more or less efficient and stable. When they go haywire, all the standard theories and accepted truths go out the window. Unfortunately, "haywire" isn't as unusual as many theorists would like it to be.

** UPDATE **

Someone left an interesting comment on this post, which for some reason hasn't shown up below. I had an email from Puzzler183 saying:

"I am an electronic market maker -- a high frequency trader. I ask you: why should I have to catch the falling knife? If I see that it isn't not a profitable time to run my business, why should I be forced to, while no one else is?

You wouldn't force a factory owner to run their plant when they couldn't sell the end product for a profit. Why am I asked to do the same?

During normal times, bid-ask spreads are smaller than ever. This is directly a product of automation improving the efficiency of trading."

This is a good point and I want to clarify that I don't think the solution is to force anyone to take positions they don't want to take. No one should be forced to "catch the falling knife." My point is simply that in talking about market efficiency, we shouldn't ignore the non-normal times. An automobile engine which uses half the fuel of any other when working normally wouldn't be considered efficient if it exploded every few hours. Judgments of the efficiency of the markets ought to include consideration of the non-normal times as well as the normal.

An important issue is to explore if there is a trade-off between efficiency in "normal times" as reflected in low spreads, and episodes of explosive volatility (the mini flash crashes which seem ever more frequent). Avoiding the latter (if we want to) may demand throwing some sand into the gears of the market (with trading speed limits or similar measures).

But I certainly agree with Puzzler183: no one should be forced to take on individual risks against their wishes.