The study of markets is still in its infancy. We know that, when only three or four people gather together to trade, there is already a recognizable market. At that level, game theory can explain a lot, as participants take account, not only of the fundamentals governing the commodity traded, but also of each other and the views they might take. But even at this level, game theory might explain behaviour, but it cannot predict market outcomes.
As the size of the market increases, so does its complexity. Factors other than fundamentals and gamesmanship come into play – time horizon, liquidity, experience of traders, entry of new traders, reactions to past gains and losses, influence of analysts. Perhaps the most complex existing market is the market for US equities.
Some humans can understand intuitively a great deal about markets. They can scan the complex factors governing a market far more efficiently than any computer. They may not process as much information, but they don’t need to. They can select information with more agility than any computer has yet been taught to do.
The human brain can “appraise things faster than you can blink”, says Arnie Wood, president of Martingale Asset Management, Boston. His example is the instant reaction of most people to the name Lee Harvey Oswald. “Your opinion [of the man who killed John Kennedy] is already formed,” says Wood. A computer might recognize the name but it would take some time, and a huge amount of programming, to make a judgement.
As an example, each year the US government summons various research establishments to a competition on natural-language recognition. The task is to get the researchers’ custom-built computer programmes to scan a body of text – newspaper articles, wire agency reports, news bulletin transcripts – and from the collection of English text, extract the relevant facts on, say, international terrorism or trade embargoes. So far, the computer wizards have been delighted with a 60% success rate.
Computers have been set to work on the US stock market in a quest to make excess returns. The image that springs to mind is that of the mouse climbing an elephant with intent to rape.
But computers have had some success, especially those which find small arbitrages between the price of a stock and its implied option value or its intrinsic value, based on some mechanical calculation. Usually these computers are applying some genetic or learning algorithm which works for a while then loses step after some structural shift in the market. The algorithms make a little money but it seems they do not get nearer to any fundamental truth about the working of markets.
Theory vs practice
Then there are the market theorists who reject almost everything the computers are doing. They have a different agenda, which is to understand more about markets. Putting their theories to work to make money is only a secondary goal, if they think of it at all. Charles Darwin’s On the Origin of Species may have explained the ascent of man, but it didn’t predict what the next evolutionary step would be, says Woody Brock of Strategic Economic Decisions, Menlo Park, California. Einstein’s general theory of relativity, on the other hand, anticipated many subsequent developments, but it didn’t predict the timing or human behaviour.
Likewise, any new general theory of markets may help to explain the phenomenon, without enhancing anyone’s returns.
The theorists don’t like time series. They don’t like finding data to fit hypotheses or vice versa. Mordecai Kurz, professor of economics at Stanford University, is continuing work on his theory of rational beliefs equilibrium (RBE). The theory seeks to explain mathematically the phenomenon that investors with self-consistent rational beliefs about the future nevertheless have views which diverge from each other and so collectively will make many forecasts that don’t come true. The volatility caused by this he calls endogenous uncertainty.
Brock, who is a James Boswell to Kurz’s Dr Johnson, says the theory will take another 10 to 20 years to develop and prove. Investors may fear the market will go away from them if they wait that long.
Somewhere in the middle, between the extreme computer-driven methods and the students of market behaviour, a lot of practical work and testing is being done. David Leinweber, managing director at First Quadrant in Pasadena, having developed a computer tool called MarketMind (now marketed with some success by Investment Technology Group as part of Quantex, which also executes trades), is using genetic algorithms to help optimize investment strategies.
Several algorithms for investment strategies are designed, then chopped up and reassembled in different combinations. Those combinations are tested, the successful ones kept and the less successful thrown out. The new winning combinations are jumbled again into new combinations until some optimum algorithms are found.
The model algorithms are not arbitrary. They are put together as a synthesis of many people’s ideas, says Leinweber. The inputs might come from studying earnings revisions, equity factors or industry models. “We apply a lot of judgement in the models,” says Leinweber. And the selling point of the strategy is “disciplined contrarian investing. You tend to do better when taking contrarian positions”, Leinweber says.
The bottom line is a 3.5% to 4.7% return above the portfolios which do not use genetic algorithms to optimize performance, according to First Quadrant literature, although the examples given are rather selective.
State Street, which for several years has used an aggregate of analysts’ earnings revisions to guide investment decisions, has started to look at signals given by individual analysts. There is less value in an analyst’s earnings revision, says Peter Stonberg, chief investment officer at State Street Global Advisors, if he is simply changing his estimate to catch up with the herd.
An analyst’s revision is more significant if it was higher than the mean and goes up, or was lower than the mean and goes down. For historical data, State Street bought the Sachs database of analysts’ revisions. It found that it first had to clean up “enormous errors” in the database, says Stonberg. Wells Fargo Nikko Investment Advisors – recently bought by Barclays and renamed BZW Barclays Global Investors – is also looking quantitatively at the effect of individual analysts’ revisions. Chief investment officer Blake Grossman says “systematic biases [among 2,500 analysts analyzed] can be very revealing in telling us how to find misvalued stocks”.
Richard Morris of Cursitor Management, London – recently bought by Alliance Capital – offers some warnings to scanners of market data. First: “Choosing the information to monitor is as important as the way we monitor it.” And secondly: “You can never attribute market changes to any specific set of data.”
The late Fischer Black, a partner at Goldman Sachs, was renowned for his suspicion of data-mining. In his final book, Exploring General Equilibrium, he warns that using a model of even 200 variables won’t help with understanding the volatility of financial markets.
Black also notes: “We kid ourselves when we say that one variable helps to ‘explain’ another. This sounds like causation, but really means only correlation.”
Quantitative research, having explored the avenue of pure data-mining and variations of it, is coming back to the primus motor of the market phenomenon – man. The study of man, and his behaviour in markets, both sophisticated and primitive, is the theme of the following two examples.
In the heat of battle
The tank commander’s squadron has come under heavy and accurate rocket fire. He assesses the terrain and sends a pair of his M1 battle tanks on a flanking movement over a grassy knoll. Two miles away, an identical battle commander sums up the situation, hulls his tanks down and radios for support. Unfortunately, the first squadron is wiped out, though it did everything by the book. The enemy commanders, equipped with exactly the same tactical skills, simply occupied a better position.
This isn’t a history of the Gulf War. It’s a video game being played out at Hughes Research Laboratories in Malibu, California. And it’s deadly serious. By 1997 Hughes has to be ready to field its simulated armed forces in full-scale defence department trials called STOW (simulated theatre of war) 97.
The simulations use artificial intelligence (AI) to provide the battle with the confusion and uncertainty that is nevertheless the interaction of expert decisions and military hardware. Each tank commander, each fire platoon leader, reacts and makes decisions according to thousands of rules downloaded from the brain of a real-life professional who is expert at his particular job. Welcome to MKAT (modular knowledge acquisition technology). STOW, if it succeeds, will be a lot cheaper than committing real tanks and troops to field exercises when testing war strategies.
To log the decision-making processes of these military men, researchers at Hughes spent months with a tank commander, and a rifle platoon leader, questioning their responses to battlefield situations and encoding them for the computer. It was a fascinating exercise with occasional cultural stand-offs. For example, shortly into session number one, the tank commander stood up and snapped: “Okay, I’m going to the john for 10 minutes and when I get back I want a clear definition of fuzzy logic.”
There’s less of a culture gap between Hughes researchers and another of their subjects, Christine Downton, chief investment officer of fund manager Pareto Partners, London. Downton has a PhD in monetary economics from the London School of Economics and worked at the Bank of England before becoming an investment manager. One of her passions is economic history.
At the end of 1991 and the beginning of 1992, Hughes Researchers spent several months encoding Downton’s investment decision-making processes. They built a library of around 1,500 rules which Downton applies and then acts on when running a bond portfolio.
Live project
The objective was to create a computer programme that could process far more information than one economist, looking at many more variables and markets, yet which made decisions consistent with that of this single experienced investor with her proven track-record. “Human beings can process only about seven chunks of information at one time,” says Downton.
Hughes and Pareto have a joint venture – Hughes Pareto Research Partners. In September 1994 the project went live with Pareto running fixed-income portfolios for two clients, together totalling more than $100 million.
Of course, it was fortunate for Pareto and its clients that the start of the live experiment came after one of the worst bond sell-offs in history. Downton, for reasons of client confidentiality, won’t give direct results of the live investment returns, but she suggests that the model can be hit by shocks that aren’t value-related: “In February 1994 the extent of the rate increase and the market reaction were hard to predict. If you have 60% to 70% accuracy, it means that 30% of the time you’re wrong, and sometimes you can be badly wrong.”
Back-testing the model over the last 10 years, however, Pareto has seen indications that “there is a potential to outperform the index by 4% to 5%”, says Downton.
The rules are simple judgements based on the relationships between fundamental economic indicators. The model applies them to decide whether bond prices are out of line, taking into account a country’s main economic figures – inflation, GDP, government debt, budget deficit, etc – and the risk of a change. Fuzzy logic is applied to fill the grey area between the broad conclusions drawn by the weighted values of the numbers – for example that Italy’s debt situation is “bad” or “very bad”.
The portfolios are re-balanced every month, strictly according to what the model decides. Downton reviews the model’s output monthly – “checking from a due-diligence perspective” – but, she adds, “we implement exactly what it says, we don’t second-guess it. Usually, if I second-guess it, I’m wrong. And that’s probably because I got out of bed the wrong side.”
The beauty of models, says Downton, is that they aren’t emotional about their decisions. Her philosophy of markets is that the winners are those with consistent long-term views, losers are those who have no view and change their mind on recent data. “Most research into cognitive bias,” says Downton, “suggests that investors tend to overweight recent information.”
But in the Hughes Pareto case, it would be wrong to say that the computer has taken over. Exactly the opposite. The computer is being used to duplicate a very human decision-making process. It is as if Christine Downton had cloned herself many times and set each of her clones to watch a humanly manageable amount of information. The Hughes Pareto global fixed-income model watches 12 major markets and computes the optimum portfolio balance taking at least 12 variables into account in each country. Even with those simple numbers, there are many millions of possible outcomes.
Many clones of herself are better than an investment committee, says Downton: “There’s no one truth but there are plenty of feasible workable hypotheses. Everyone on a committee has his own internal model. But that produces a sort of soup which doesn’t have much flavour.”
Self-consistency
As with the tank commander and the rifle platoon leader, the knowledge-based system MKAT needs an expert brain to download. Not only that, it helps if the mind is as consistent as possible and if the subject knows why he or she reached a certain decision. Downton appears to have these qualities.
“Most experts are self-consistent,” says Charles Dolan, project manager at Hughes Research Laboratories. “But Christine is also pretty knowledgeable about how she reaches a conclusion.”
For Dolan, who has a PhD in artificial intelligence from the University of California, Los Angeles, this project is “more fun than most of the things I’ve done in science”. There is not only the curiosity of devising a tool to outsmart the markets, there is also the clear indication of success or failure. “The rubber really meets the road,” he says.
Dolan is also a convert. His initial view was that learning algorithms, another area of Hughes research, were going to be the right technology for running a bond portfolio. Learning algorithms build models by seeing patterns in the raw data of the market. In the early stages, both learning algorithms and the knowledge-based approach were applied to the Hughes Pareto project. But as Dolan worked with Pareto, he saw that “Christine’s insights kept helping the learning system, giving consistently positive results. So I jumped ship.” This doesn’t mean that learning algorithms have been rejected for modelling other areas of the market. Hughes Pareto is developing other techniques to run a global asset allocation model including equities. “I still believe that eventually the learning approaches will achieve something,” says Dolan.
A knowledge-based system suits the fixed-income portfolio because bonds are governed more by fundamentals than by short-term volatility and trader sentiment. “Market participants can increase the volatility of returns,” says Downton, “but that doesn’t change the reality that, if there’s value in the market, you will make money over time – although the timing isn’t clear.” The currency risk in the bond portfolio is handled completely separately by Pareto’s traders who can intervene 24 hours a day to adjust exposures. “This reduces the volatility of returns,” says Downton, “contributing perhaps 1% to the overall result.” The currency activity is to reduce, not to increase risk. Currency risk management, says Downton, “allows you to be more aggressive in the bonds – you have a safety net”.
Political risk is also ignored by the model, although the volatility caused by political uncertainty is picked up. “The impact of political events is unpredictable,” says Downton, “especially the timing. You’d never have invested in Italy if you counted political risk. You can’t put a value on it. But you can identify that risk is increasing [by the increase in volatility] and reduce your exposure.”
Downton – not a power-dresser, thoroughly down-to-earth – underlines the point that there is no magic or mumbo jumbo in what the fixed-income model does. “Finance is about analyzing information and managing complexity. There has been an information explosion. The capacity to analyze it has increased and costs have plummeted.” The fixed-income model does no more than apply tried and tested concepts, but in a rigorous and comprehensive way. Downton believes there is no evidence that the exploitable inefficiencies in the markets are diminishing: there are always different groups of investors, domestic and global, short-term and long-term.
What about the changing structure of the markets? Doesn’t that mean the model must constantly be adapted? “We designed the software three years ago,” says Downton. “I’m not expecting to modify it significantly. I don’t think the rules of the game change that much.” Dolan says there is scope for fine-tuning, country by country, what the model considers to be “high” or “low”. For example, an inflation figure that looks high in Germany might look low for Italy. But he believes, “with one caveat”, that the fundamental assumptions of the model don’t need to change. “If the inefficiency that Christine believes is there is bid away, then we’d have to improve the model. You can never say you’ve got absolutely everything.”
Downton seldom finds the model doing things she wouldn’t. For example, in the January 1995 sell-off, “I wanted to do what the model did – buy into weakness. I’m a value investor.” But she cautions: “If I had intervened, the whole philosophy of the organization would have opposed it.”
Human laboratory
The auditorium at the Biltmore Hotel in Phoenix, Arizona, is wired for quadraphonic sound. On the big screen is a graph, with the “x” axis representing time and the “y” axis price. Bid prices appear as red blobs, offer prices as blue and actual transactions are logged in yellow.
The market opens. For a while, the traders hesitate, then a few offers appear high up the graph and a handful of bids lower down. Each offer is accompanied by a high electronic blip whose pitch varies with the level of the price. For each bid, there’s a blop at a suitably lower pitch.
What the audience can see, but the traders can’t (although they have the information), is a line sloping down and to the right which represents the theoretical minimum value of the commodity being traded; and another line, higher and farther to the right representing the theoretical maximum.
As the seconds tick by, a few yellow blobs (transactions) appear between the minimum and maximum lines, then more as the market gets into its stride. Traders are getting the measure of each other and developing a feel for where the price should be. It travels tentatively upward as the bids begin to outnumber the offers. Clearly, the market believes the commodity is underpriced. The blops exceed the blips and the blues the reds. The quadraphonic sound conveys a sense of rising frenzy. And sure enough, the frenzy takes the yellow blobs (transactions) higher and higher until they are well beyond the theoretical maximum. Then they hesitate. The audience of sophisticated investors knows there’s going to be a crash as surely as if they were watching a Disney character who had just run over a cliff. The bids dry up, the offers scream into a steep dive and the price drops by 50% in three seconds. The sophisticated investors clap and cheer with delight. They themselves have felt this fear and greed, but for once they are seeing it happen to the other guy.
Charles Plott, professor of economics at California Institute of Technology (Caltech), has been running artificial markets in laboratory conditions for over 20 years. But it is only recently that the investment community has begun to show much interest in his work. Maybe the trigger has been advances in computer graphics which allow him to show vividly and dramatically how closely his artificial markets parallel real markets. They parallel real market dynamics but they show markets stripped of all noise, complexity and extraneous detail. They demonstrate how, in relatively free markets, information is communicated rapidly through the price mechanism. In laboratory conditions, Plott and his colleagues can study, with some hope of accuracy, the effect of introducing a derivative or a new regulation, or of seeding a trader or group of traders with inside information, for example. Regulators and lawyers are naturally interested. In fact, they were quicker than the investment community to spot the potential of artificial markets.
The basis of a laboratory market is as follows. Plott induces Caltech students or other groups of participants to spend an evening trading in his lab situated in the basement of the humanities and social sciences building. The incentive, apart from curiosity, is that they can earn real money, perhaps as much as $200. They sit in cubicles at computer terminals and trade for a series of sessions, typically seven minutes long, according to rules and incentives provided for that particular market. They might be buying or selling a security whose dividend is uncertain but might pay out 200, 400, 600 or zero units (see example below). They might be trading six securities of which only one or two are expected to pay a dividend. There may be an option in the market which puts a floor of 200 on the dividend payout.
However artificial the conditions, it is quickly apparent that, even with as few as three players, a market has been created in which trading skills, psychology and sentiment play as big a role as fundamentals.
“The people who caused that bubble [beyond the theoretical maximum] and the subsequent crash were not stupid,” says Plott. “They were playing against each other as much as they were watching the fundamentals. Their mistake was to think they could get out before the others.”
To date, the real-world applications of Plott’s markets have been more in the bidding auction and anti-trust area. For example, his simulations of “smart” markets – in which bids for, say, packages of cable TV licences or airport landing slots are made contingent on other bids – have helped anti-trust bodies or governments to find the optimum outcome. “It’s like fitting together a jigsaw whose pieces are constantly changing shape,” says Plott.
For years, Plott has accepted stoically that Wall Street traders think they know more about the real market than he can teach them in a lab. “Fish aren’t interested in hydrodynamics,” he philosophizes. In the 1970s, after contacts with Chicago theorists such as Eugene Fama, it was clear that the market academics were preoccupied with the US stock market, whereas Plott was looking at the dynamics of any market anywhere. In the early 1980s, bond brokers Cantor Fitzgerald tried to apply some of Plott’s findings to real trading, “but it wasn’t very successful”, says Plott.
Most support for Plott’s lab has come from General Motors, Hewlett-Packard, Pacific Bell, IBM and General Electric. The government’s National Science Foundation has funded him for years and has just agreed to put in another $1 million. For Wall Street firms, setting up such a lab would be pin-money. The hardware and software would cost around $100,000 and a series of experiments, including incentives for the traders, would need perhaps $20,000.
In fact, the number of similar labs worldwide is growing. The first one ever was set up by Vernon Smith, professor at the University of Arizona. Now there are labs in Amsterdam, Barcelona, Berlin, Bonn, Bielefeld, Grenoble, Siena, Turin, Stockholm, York and Zurich. The lab in Bielefeld is the creation of Reinhardt Selten, co-winner of the Nobel prize in 1994 for his work on game theory. Game theory is eminently testable in these artificial market conditions.
Applications
At the end of December, Plott was due to run his first artificial market globally, via the World Wide Web, with participants in Sweden, the Netherlands, France, Australia, New Zealand, and “a couple in the States”.
For him, this is a shot across the bows of real-world practitioners. Why not have a real market on the World Wide Web? Up to now, the technology has been too slow for real-time trading, but Plott foresees that computerized custom-built markets could be set up on the web for specific commodities or franchises such as time-share on space stations. A handful of competitors could set up a market for almost no cost. The only problem might be regulation.
For the financial community, many possible applications are offered by artificial markets:
1 By comparing artificial market dynamics with those of real markets, it may be possible to cut out much of the irrelevant complexity and noise – perhaps establishing where the true market sentiment is leading. (Many theorists might say this is impossible, arguing that it is the noise that makes the market.)
2 Learning much more about “cognitive bias” among market players. This refers to the tendency for investors to be influenced by factors other than price information – such as overreaction after suffering a loss, feeding frenzy as the market rises, overweighting of recent data, myopic following of market sentiment and reaction to arbitrary posted prices.
3 Artificial markets might give a better understanding of how derivatives and hedging mechanisms affect overall price discovery and market volatility. (One problem with laboratory markets is that, while students like to win money, they’re reluctant to lose it, and collecting debts from them is difficult. So it has been hard for Plott to create conditions in which traders short the market or otherwise leverage their bets.)
4 It might be a cheaper way to train or classify traders than letting them loose on real markets. Plott has occasionally got professionals to trade his markets and says that “the greed is still there” even if the sums at stake are paltry by professional standards. Plott finds he is easily able to put the traders he employs into one of two categories: wolves and rabbits.
“The rabbits are very myopic, risk-averse but impulsive,” says Plott. “The moment they see something to their advantage – bang – they hit it. So market information goes screaming past them.
“The wolves, on the other hand, understand the rabbits and they’re watching the market. They set traps. The professional market-makers are really good at that.”
Applications for financial lawyers and regulators might be:
1 Evidence from artificial markets to help prove or strongly suggest the presence of insider dealing. Plott’s experiments show that, when information is seeded to a single insider dealer, within seconds, the market seems collectively to acquire that information even if individuals don’t see it. “You are made aware that rationality is not common knowledge,” says Plott.
2 Ways of structuring new markets could be tested in the lab before expensive mistakes are made. This might include markets where parameters such as minimum prices or interest rate levels are set by the authorities. A lab might also be a testing-ground for modifications to existing markets such as the New York Stock Exchange’s specialist market-maker system.
Plott’s audience in Phoenix on November 2 were institutional clients of New York fund manager Scudder Stevens & Clark. Scudder president Ed Villana is an economist with more than a passing interest in Plott’s work. Phil Fortuna, a managing director at Scudder, believes there are “some real applications for Dr Plott’s experiments”. He suggests it would cost next to nothing for a Wall Street firm to supply Plott with tick data from the market, so that real-market bubbles could be overlaid on artificial bubbles and compared in detail. Musing why this hasn’t happened already, Fortuna says: “The trouble is he doesn’t have the tick data and we don’t have the patience.” Another turn-off for the financial community, suggests Fortuna, is the sheer simplicity of Plott’s experiments: “The quants want fractals and neural networks. His stuff comes from a different discipline.” But Fortuna is convinced that Plott’s experiments would merit further study by Wall Street.
Homage to Fischer Black
“Fischer Black is a nihilist!” was one of the more colourful insults Black invited on himself for his distrust of market data as the raw material of quantitative research. “Time itself is a suspicious variable,” Black once told this magazine.
But perhaps his natural opponents will not be so emotional about his final work, published last year, which is an attempt, almost biblical in its simplicity and clarity, to rewrite the basics of utility theory and general equilibrium.
Black, a partner of Goldman Sachs from 1984, and former professor of finance at the University of Chicago’s Graduate School of Business and at the Massachusetts Institute of Technology’s Sloan School of Management, died in August. Together with Myron Scholes, he developed the widely used Black-Scholes formula for option valuation, published in 1973. Here are some passages from his last work, Exploring General Equilibrium, The MIT Press, 1995
Volatility
“Given the volatility of expectations, I’m surprised that asset prices aren’t more volatile than they are. We see asset volatility for the same reasons that we see variation in growth rates across countries and through time, and for the same reasons that we see business cycles. The questions are different, but the answer is the same. Asset prices vary because we make highly specific investments while looking far into the future; as time goes by, we change our beliefs about the future along many dimensions, which makes our investments seem more relevant – or less relevant – and thus changes the values of claims on those investments. Sometimes asset prices change because we become more patient – or less patient – and sometimes they change because we become more tolerant of risk – or less tolerant of risk – but we don’t need those sources of volatility to understand why prices often change for no apparent reason.
“The details matter. We can’t understand the powerful role of varying expectations without looking at outputs that range from peanuts to spaceships, and at inputs that range from truck driving skills to experience in doing coronary bypass operations. We just can’t understand volatility using a model that has two inputs and one output, or even one that has 200 inputs and outputs.
“Prices will continue to change in response to forces that we can’t easily measure, and people will continue to claim that markets have ‘excess volatility’. To me, the volatility in market prices seems normal and consistent with the volatility in national growth rates and in cyclical measures of business activity.”
Econometrics
“I feel that econometric studies of time-series data tell us very little about the economy …
“Econometric studies can tell us about variances and covariances, including auto- covariances. They can tell us about correlations, including partial correlations. But they can’t tell us anything direct about cause or effect, impact or influence, structure or meaning. We kid ourselves when we say that one variable helps to ‘explain’ another. This sounds like causation, but really means only correlation.”
Financial intermediation
“As a country grows, it invests in more and more elaborate forms of financial intermediation. In effect, its markets become more complete. Its residents can diversify better, or hedge the risks they take, or align ownership and control for the investments they make.
“In general, this reduces the price-of-risk part of the cost of capital. It increases the expected payoff for taking more risk, as with added specialization or more roundabout production. Increasing the payoff means the economy takes more risk. Expected growth increases, asset volatility increases, and average unemployment increases.
“Financial intermediation improves the terms on which people take risk, but it actually increases the risk we measure in the economy.”
